The Sunday Signal

The 15-Minute Lawyer

David Richards MBE — Sun, 26 Jul 2026 06:57:00 GMT

BOTTOM LINE UPFRONT: High-priced hourly professional work is over, and this week produced the evidence at both market ends. At the top, a City investment banker showed me an email where a law firm ran up nearly £250,000 of billable time, 12.5% of the growth capital, on a stock market listing raising £2,000,000, one of the reasons for the decline in London listings. At the bottom, a Sheffield legal tech firm reduced the time for a remortgage from weeks to 15 minutes per file. Professional fees are leakage, capital that investors subscribe to and companies never receive. With a fixed cost of a London stock market admission of nearly £900,000, they have driven the UK AIM stock market from almost 1,700 to barely 600. This week, Centrica announced 1,300 jobs would go alongside a £600m digital and AI customer service overhaul, the clearest sign yet of automation arriving on UK shores. The billable hour always survived because there were no alternatives. The alternative now exists, and what stands between it and the market is the professions’ defence of their own clock.

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Subscribe now

Nobody raising £2m can afford £250,000 of lawyers

The email was from an investment banker who specialises in taking small companies public. I have anonymised the names of all companies, law firms, and people, and intentionally kept the details of the transaction vague. The numbers are as follows.

The transaction was a reverse takeover (RTO), the type of transaction the London markets were intended to facilitate, raising around £2m in growth capital. The law firm’s engagement letter set out a fee of £160,000 based on an agreed scope, and then the clock started. In the words of the lawyer, “we ended up with circa £250k of time” by the end of the transaction. The firm billed £170,000 and wrote off the rest, around £80,000.

A quarter of a million pounds worth of legal time to raise £2m equates to 12.5% of the funds being raised. Before the company has hired a single engineer, or shipped a single product, one adviser has consumed an eighth of the funds, and this is before the brokers, accountants and nominated adviser bill the company. No rational founder pays that toll twice. Most of them now refuse to pay it once, which is why they walk into a venture fund’s office, where legal expenses are capped, controlled and relatively small.

It’s worth looking at the explanation for the cost overrun because that’s the system confessing. When justifying the overtime, the renegotiation of relationship agreement documents and out-of-scope items were mentioned. To put that in normal terms, two groups of lawyers were, for weeks, fighting over the wording of standard boilerplate documents. Each hour they spent fighting was billable hours for both groups. The billing model didn’t just tolerate the friction, it rewarded it. And think about what the £80,000 write-off actually tells you. Nobody writes off £80,000 of genuine value. The write-off is the firm conceding the hours were never worth the rate.

AI can easily perform all of those tasks. To renegotiate a relationship agreement, a draft can quickly be reviewed against the current market and deviations flagged. AI can do in seconds what would cost a company thirty hours of billable time. To verify a prospectus means comparing a number of source documents, a task AI can parse quickly and accurately better than several first-year associates. Once the labour collapses, so too does the cost, and the fixed-fee quote will become what it is in the construction industry: a real number the contractor must stand behind.

What is holding it up? The banker’s covering note indicated exactly that. Nominated Advisers carry the regulatory risk of a listing, and they are terrified. In his words, there “will be a gap in trust” that needs to be filled, and “a legal face to hold hands” will be needed. A Nomad does not buy legal work. A Nomad buys someone to blame, backed by a large firm’s professional indemnity policy, for the regulator when the time comes. Until the FCA creates a safe harbour for AI-verification and fixed-fee legal work, the gatekeepers will continue to demand the expensive human ritual, and the small-cap market will continue to be strangled by it. London does not have a valuation issue. London has a toll booth issue.

Subscribe now

Take the IPO bill apart line by line. Almost none of it survives

Take apart a toll booth, and you’ll find that the £250,000 on the clock is divvied up into four categories of labour. Each of these categories must answer the same question: what does a human do that a machine cannot?

Let’s start with verification. It is the largest portion of most small-cap listings and is the most misunderstood. A team of junior lawyers would build a verification matrix that lists hundreds of statements and claims, each of which is cross-referenced to a lease, a contract, a set of accounts, and board meeting minutes, among others. A verification matrix eats up associate time and comes at a massive cost to the client.

An AI data room achieves the same task by ingesting source documents, linking each assertion in a draft document to evidence, and identifying gaps in hours instead of weeks. The purpose of verification, which is to safeguard the directors and the investors from a false statement, remains intact. What is rendered obsolete is the manual matrix. The work of the lawyer shifts from creating a matrix to auditing the exceptions, which is the singular task that requires judgement.

Legal due diligence proceeds similarly. Reviewing due diligence reports at a small target company means scanning through all relevant contracts for change of control provisions, termination rights, and consents. Then comes title, litigation, employment, and IP. Automated contract review and clause extraction is the most advanced application of legal AI. This particular technology has existed in larger M&A data rooms for years. The judgement call on whether a clause, which legal AI flagged, is material to the transaction still remains a human task. The two hundred hours spent reading does not.

Once the tedious task of data review is complete, the actual negotiation can begin. The email from the banker was blaming the negotiation for the time overrun. The renegotiation of a relationship agreement requires the comparison of draft agreements to existing market agreements and the justification of the variance. An AI system that has been trained in redlining agreements against market standards can complete this comparison in a matter of seconds. In contrast, traditional methods require over thirty hours of negotiation. There is no reason that the redundant tasks of document assembly should require billing at a fee earner’s hourly rate.

For humans, the remaining areas are real but small: advice on structuring, negotiation strategies, and regulatory judgement calls, as well as conversations with directors to interpret the meaning of a responsibility statement. Priced to reflect that true value, clients would get a fixed fee, at a fraction of the current price. Clients wouldn’t have to bear the risks that are common in hourly billing, as the risks would be eliminated throughout the process.

Why this decides whether London has a small-cap market at all

The toll booth tolls are not simply an irritation. They are a capital leakage for the market. These are funds subscribed by investors that are never received by the companies, since they are collected at the toll booth before any funds begin to operate. One long-time AIM specialist has calculated that the cost to access the AIM is around 3.43% of the total funds raised, and then a ticket of £897,000 is added. Using our banker’s £2m target, the estimated cost would be around £966,000. This is why the deal was executed via a reverse takeover instead of an initial public offering. The fixed cost, charged to all companies, would have priced our company out of doing a deal.

The market data shows what the toll booth has done. AIM held 1,694 companies at the end of 2007. It holds 612 today. The past two decades have brought more than 2,100 delistings, while new admissions ran at just 11 a year in both 2023 and 2024, fewer than in the depths of the financial crisis, raising £119m in 2024 against £6.6bn at the 2006/7 peak. There were two IPOs in London in the first quarter of this year. Colin Wright, chairman of UHY Hacker Young, points to the “excessive red tape” as a reason companies prefer to remain private. Delistings still outpace listings, though the proceeds have improved since spring thanks to listing reforms and stamp duty exemptions. The recovery is only at the top of the market. The small end remains shut.

Perform the calculations in reverse. The exchange’s fees are so small that they are a rounding error, so the reforms that reduce them are welcome and pretty much meaningless. The cost of entry is the advisers. The cost of the advisers is professional hours. The next story will show that cutting the professional hours required to complete the task by 80% to 90%, which is the reduction that has already been achieved in conveyancing, will make the near-£900,000 fixed cost fall to around £200,000. A £2m fund raise would then mean the entry costs would be around 13% of the fund rather than 50%, and a category of company, which has been unable to access the public markets for the past fifteen years, would then have a route to enter the public markets. The leak is not a natural occurrence. It is a model for conducting business. Business models can be changed. The question is whether the FCA and the exchange will allow it to be changed. Nobody in the official conversation is willing to pull the one reform lever that is big enough to matter.

Subscribe now

The 15-minute conveyancer has already arrived

If the corporate Bar thinks this is all theoretical, they should look at what is happening two hours up the East Coast Main Line. To be completely up front, LEXcelerate, the company in question, is part of the Yorkshire AI Labs portfolio where I’m the Managing Partner. Judge the numbers on their merits.

A typical UK remortgage takes six to eight weeks to complete. In client surveys, more than 50% of borrowers said the process takes too long. No one involved in the process knows what takes up all the time. LEXcelerate can complete the process in two weeks. However, that is not the disruptive factor; the disruptive factor is the labour involved. Total human fee-earner time per matter: 15 minutes.

That is possible because the firm views conveyancing as a manufacturing process, not a craft. Where a traditional firm would move a file through a poorly managed dozen stages, LEXcelerate would break the entire process down into more than 160 micro-stages. Most of these micro-stages are automated via API integrations with the Land Registry, lenders, and ID checks. Automated workflows drive the micro-stages. Human judgement is reserved for six mandatory gates. The system enforces the gates: a file cannot progress until the risk checks at each of the gates are completed. The lawyer’s role shifts to that of a verifier.

The founders matter here. This isn’t a pair of outsiders speculating on law’s mechanisms. Paul Firth, the Chief Executive, was DLA Piper’s managing partner for the UK. He also led Irwin Mitchell’s national real estate practice. He spent a career at the summit of the billable-hour model and now sees it as the end of the road. His co-founder, Mark Hewitt, has spent thirty years in the field of automation and built a legal software company that was sold to Verisk. When the people who profited most from the old model start building its replacement, the argument about whether it works is already over, and the only remaining question is how fast it can be done.

Next, let’s follow the economics, because they are harsh. In order to grow, a traditional firm has to hire more associates. This means costs will rise as revenue increases. A platform that caps an employee’s time at 15 minutes per case file builds like a software company, which is why LEXcelerate predicts that within the next three years, profit margins will be more than 70%. If a case file only requires small amounts of human oversight, the salaried employee who is doing junior-level case work is no longer a viable option. The lawyers on the platform operate like locums and register for a shift when they wish to perform a task to help clear the verification gates. The conveyancing market is worth about £2.6bn per year. By cutting the amount of human time to do the work by 90%, the market will need 90% fewer junior-level conveyancers. No lobbying from a professional work body is going to change this.

A defence for the profession has already been employed: remortgages are commoditised, corporate work is tailored, an IPO cannot run on rails. Solicitors made the same sort of defence in 1985, when licensed conveyancers breached the monopoly on property work, and conveyancing fees never recovered. Once the market understands that risk can be systemically managed instead of manually, and that a fixed price is a standard instead of a concession, the £250,000 clock cited in the first story has no excuse remaining to run.

Subscribe now

AI has landed on our shores, and the call centre goes first

The wave rolling through the law does not stop at the law. For over a year, this newsletter has followed it in American earnings calls and restructuring memos from Silicon Valley. This week, for the first time, the wave reached the UK, and it did so via the most ordinary company possible.

Centrica, owner of British Gas and a FTSE 100 firm, announced alongside its half-year results that it will cut 1,300 jobs over two years. 500 of the jobs are contact centre positions in Glasgow, Edinburgh, Cardiff, Leicester, Stockport and Leeds, and the remaining 800 are offshore, outsourced, and back-office operational support. In total, Centrica will cut 14% of its customer operations workforce and 3% of its operating costs. Some of these roles will be cut through attrition, while the remaining will be eliminated through redundancy.

Centrica is trying to convince the public that it’s not the machines. CEO of Centrica, Chris O’Shea, told the press that “AI isn’t driving these particular job reductions,” and instead described the change in customer behaviour: more than 90% of customers now use digital channels first, calls have reduced by 20% year on year, and inbound contact has fallen by a third since 2023. The GMB union doesn’t agree. GMB’s national secretary, Charlotte Brumpton-Childs, said staff are “being replaced by artificial intelligence”.

Both things can be true, and the results illustrate how. Centrica’s adjusted EBITDA dropped from £900m to £737m, a decline of 18%, because of asset divestments, outages, and milder weather, yet the interim dividend increased by 9%, retail profits even increased to £346m from £338m, and even with British Gas’s customer base falling to 7.45 million, the company is investing £600m to digitise and incorporate AI into the customer operations. Read the numbers this way. Earnings under pressure, dividends protected, technology funded, and payroll paying for it all. Whether the chatbot took the job or the job became unnecessary is a distinction that is important for the press office and unimportant for everyone in Stockport.

Centrica’s case won’t be unique, but rather a template, and it’s coming at the worst possible time. The S&P Global UK Services PMI reported 48.8 in June, indicating a second month of contraction and the largest decline since January 2023. Services account for around 80% of the UK economy. Banks, insurers, telecoms and retailers provide customer service and back-office functions at large scale, and all of those firms watched Centrica cut 14% of them this week while increasing its dividend. During the same week, Uber announced that it would cut 10% of its global customer service function and, for the first time, named AI as the reason. These are not City jobs, but they are essential to the economy of the towns and regions in the UK that were promised service jobs would replace the roles lost in manufacturing.

Centrica points, fairly, to the other side of its ledger: it keeps hiring engineers and plans 500 apprentices this year. But a 45-year-old call handler in Leicester does not retrain as a heat-pump engineer between redundancy notice and final pay cheque. The jobs being created and the jobs being destroyed belong to different people in different places with different lives. If the rest of the UK service economy runs the Centrica play, and the incentives say it will, the bill lands on the state, one Universal Credit claim at a time. The Institute for Public Policy Research has warned that millions of UK jobs sit exposed to this exact automation. This week, that ceased being a projection, and instead has become the announcement of results.

Subscribe now

The Signal Tech & AI Layoff Tracker

Week 29 wrap, 18 to 24 July, built directly on last week’s verified baseline and synthesised from Layoffs.fyi, LayoffHedge, SkillSyncer, state WARN registries and SEC disclosures. For informational purposes only. Here is where the data stands:

This week was all about explicitness. For the past two years, companies have obfuscated the AI-induced restructuring behind euphemistic references to efficiency improvements and customer behavioural shifts. This week, three of them chose to say the quiet part out loud in regulatory filings and internal memos.

Monday.com, ~620. The work management platform announced workforce reductions impacting 20% of their global employees, approximately 620, in a Form 6-K filed with the SEC on 22 July. Staff were told by co-founders Roy Mann and Eran Zinman that the company was reorganising, centring around an “AI Work Platform,” which entailed a shift from helping customers manage work towards doing the work via AI agents. They also announced plans to flatten management tiers by organising staff into smaller, fully autonomous teams. They argue that the changes should not be seen as cost-cutting or direct AI replacement of work, and the numbers half support them: while restructuring costs the company between $45m and $55m, they maintain that the company is on track for 19% to 20% revenue growth. However, the signal sent by the move was far from ambiguous. A SaaS company that was built in the platform era has determined that its own organisational structure is the obstacle to competing in the AI era, and the company has made the difficult decision to reduce staff by 620.

Uber, headcount undisclosed. Uber laid off 10% of its global Community Operations team. For the first time, the company said artificial intelligence contributed to the layoffs. The team is responsible for supporting Uber’s riders, drivers, and couriers. Vice President Megha Yethadka, who leads Community Operations, told her team it had become “too complex and siloed” for Uber to optimally implement AI. This is Uber’s second round of layoffs in two months, and since the company has not disclosed the total headcount, the layoffs are reported but have not been fully incorporated into the baseline.

Amazon, headcount undisclosed. Reported layoffs in Amazon’s Artificial General Intelligence unit, the team developing the Nova models, continue to affect roles in model customisation and post-training. Since no numbers have been disclosed, the layoffs will not be included in the baseline. However, the context is important. These layoffs add to the approximately 30,000 layoffs affecting Amazon’s corporate positions that were announced since October of last year. These layoffs also illustrate that even the teams building the technology are not sheltered from the corporate restructuring.

And on this side of the Atlantic, the largest UK cross-sector event of the week is Centrica, reported in full above, with 1,300 layoffs, logged by LayoffHedge. For UK readers, this is the domestic counterpart to the international reports this tracker has been compiling. Read it as the arrival on our shores of everything in the numbers above.

Week 30 Watch List: Customer Operations. This week, Centrica and Uber eliminated the same function on different continents. Now, every business with a contact centre is analysing the impact of these events as they prepare their Q3 earnings. Also, expect the same type of middle-management reduction you see with Monday.com, as SaaS firms under valuation pressure realise that AI is eliminating organisational layers, not just headcount.

Final Thought 🚀

In these three stories, a single machine begins to form. At the top of the professions, a law firm runs a £250,000 clock on a £2m fundraise and writes off £80,000 as if it’s a giveaway. In Sheffield, a man who used to run one of the largest law firms in the world shows that the same type of work can be done in 15 minutes at a fixed price. And, at Centrica, the machine meets the workforce, giving us 1,300 letters, a protected dividend, and £600m for the tech.

The professions have been saying to themselves for the last two years that the wave would break on someone else. Lawyers said it would be the call centres. The call centres have been taken, and this week the wave reached the timesheet, which is the profession. The billable hour was a way to measure scarcity, and the scarcity is finished.

There will be no obituary when it finally dies. There will be a client who asks for a fixed price, a Nomad who finally accepts an algorithm’s verification, a founder who lists in London because the toll booth is gone. Until that day, every hour on the clock is a choice, and the ones who choose it no longer get to say it’s a cost of doing business. It is the business, and that is precisely the problem.

📩 Read the full issue free at thesundaysignal.ai. New issue every Sunday. AI disruption, UK innovation, and the future of work, with no hype and no hedging.

Until next Sunday,

David

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Subscribe now

What Would the Victorians Do?

David Richards MBE — Sun, 19 Jul 2026 05:30:12 GMT

THE SUNDAY SIGNAL · Issue #63 · Week 29 · Sunday 19 July 2026

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube.

Subscribe now

Ten days ago I climbed down into the Megatron, the vaulted Victorian culvert beneath Sheffield station, and watched a river that engineers buried a century and a half ago still running through the dark. On Thursday a Beijing lab founded in 2023 shipped a 2.8-trillion-parameter AI model that matches the American frontier on several of the benchmarks that matter, and promised to hand out the weights within a fortnight. One is a monument to how fast Britain once built. The other is a measure of how fast somebody else is building now.

BOTTOM LINE UPFRONT: Britain wants Victorian results without Victorian methods, and at the moment we get neither. Run the thought experiment properly and the conclusion cuts both ways. A resurrected syndicate of railway barons would make Britain the compute capital of Europe within three years, and they would poison rivers, flatten villages and ruin small investors to do it. Then look east. China has already adopted most of that playbook, swapping the speculative chaos of the Railway Mania for the discipline of the five-year plan, and this week’s launch of Kimi K3 shows what the swap buys. The lesson of both centuries is the same. Speed has a price, and somebody always pays it. The only question worth asking is whether they got a vote.

Subscribe now

They would not reform the planning system. They would dismantle it

Start with a thought experiment. Resurrect a syndicate of Victorian railway barons, together with the parliamentarians who served them. Hand them the blueprints for a one-gigawatt AI data centre. Tell them Britain has a compute bottleneck and that our rivals do not. Then stand well back, because what follows would be terrifying, spectacular and brutally fast.

They would look at our fifteen-year waits for a grid connection, our planning inquiries and our environmental impact assessments, and they would not reform them. They would dismantle them.

Today a hyperscaler that wants to build a data centre in Britain faces years of council battles, objections about visual impact and noise complaints about cooling fans. The Victorians would not fight that system. They would go over its head, exactly as they did in the 1840s: a private bill in Parliament. Once passed, the AI Infrastructure Act would grant the Great British Compute Company the power of compulsory purchase. If the most efficient site for a 500-megawatt campus were flat land in the Home Counties, they would draw a straight line on the map, buy the villages in the way at a price Parliament had set, and level them. No judicial review, no appeal. The Midland Railway did precisely this to Sheffield in the 1860s, which is why a river runs beneath Platform 5. More on that later.

The single biggest obstacle to AI in Britain is that the National Grid cannot supply the electricity. A Victorian engineer would find the idea of waiting a decade for a wire from a regulated monopoly absurd, so the syndicate would build the power station inside the data centre fence. If nobody would licence a small modular reactor quickly enough, a private gas-fired station would go up beside the server halls, burning whatever fuel was cheapest and most reliable. They would sign no offshore wind agreements and make no Net Zero pledges. The exhaust stacks would blanket the surrounding country in smog, and they would call the smog progress, because to them it was.

Then there is the water. An AI campus generates astonishing heat and drinks millions of litres to cool its chips. Modern operators negotiate with water companies for years to avoid draining the local supply. The Victorians solved water problems by force. The men who entombed the Sheaf in brick would appropriate a major waterway, perhaps the Trent or the Ouse, divert it through the cooling systems and return it to its bed at 40 degrees. Everything downstream would die. The server racks would hold their optimal temperature. They would consider that a fair exchange, and they would not commission a study on it.

Subscribe now

A burst bubble still leaves the infrastructure standing

The financing would mirror the Railway Mania of the 1840s, when Parliament passed 272 railway acts in a single year at the top of the bubble. Rather than depend on three American hyperscalers, the Victorians would float hundreds of speculative AI infrastructure companies on the London Stock Exchange, promising impossible returns on “compute yields”. The bubble would inflate and it would burst, ruining thousands of middle-class investors and bankrupting half the companies on the list. And when the dust settled, the concrete shells, the fibre and the power stations would still be standing, bought out of receivership for pennies in the pound. Britain would emerge scandalously overbuilt, which is another way of saying ready. It is exactly how a financial catastrophe left us with the best railway network on earth.

Subscribe now

The Workshop of the World would reopen, whatever the cost

Give the syndicate a second morning and they would spot the deeper weakness. Modern Britain depends on Taiwan for semiconductors, on China for steel and batteries, and on globalised supply chains for copper. To an industrialist who built an empire on domestic coal and iron, that is not a talking point. It is a survival problem. So the programme would widen.

Semiconductor fabs are the cotton mills of this century: the most complex factories on earth, needing billions in capital, enormous power and millions of gallons of ultra-pure water. The Victorians flooded the Thirlmere valley to water Manchester; their heirs would dam valleys in Wales and Scotland to wash silicon wafers, and they would run the chemical waste into the nearest estuary rather than regulate it. Cornwall’s lithium would come out of the ground through deep shafts and thousands of workers rather than through five-year environmental surveys, and any picturesque village sitting on a vein of critical minerals would be moved, on the Victorian principle that what lies beneath the ground matters more than whatever happens to be sitting on top of it.

They would treat the closure of primary steelmaking at Port Talbot as national self-sabotage and reverse it, expanding the blast furnaces rather than replacing them, and reopening the coking coal seams of Yorkshire and Cumbria to fire them. And to staff it all they would resurrect the company town: Saltaire with fibre broadband. The corporation would build the factory, then build the workers’ houses, own the shops and run the schools, in exchange for a workforce that is regimented, non-unionised and entirely dependent on the employer for its livelihood. They had a word for that arrangement: paternalism. We would have another: feudalism with a pension scheme.

Would it work? Almost certainly. Within three years Britain would have the largest AI infrastructure in Europe and be closing on the world. We would fab our own chips, mine our own lithium and forge our own steel, insulated from every supply-chain shock the Pacific can produce.

And the price would be a return ticket to the nineteenth century: dead rivers, demolished villages, ruined investors, abandoned climate targets and a workforce with no vote on any of it. The speed we envy in the Victorians was never a technique that we have somehow mislaid. It was a willingness to send the bill to people who could not refuse it: the poor, the unborn and the landscape. Deindustrialisation, seen from this angle, was not only an economic retreat. It was a choice to buy clean air and clean rivers at the cost of sovereign output. The thought experiment does not tell us to copy the Victorians. It tells us the price we are currently paying for the right not to.

That would be a comfortable place to stop, if the syndicate were safely dead. The trouble is that a version of it is alive, well and shipping product…

Share The Sunday Signal

Meanwhile in Beijing, the syndicate is alive and shipping product

On Thursday, Moonshot AI released Kimi K3. The lab was founded in Beijing in 2023 by Yang Zhilin, formerly of Google Brain and Meta, and the numbers deserve a moment. It is a 2.8-trillion-parameter mixture-of-experts model that activates just 16 of its 896 experts per token, with a full million-token context window and native vision. Moonshot claim their new attention design decodes up to 6.3 times faster at long context, and they trained it quantisation-aware from the start, so the weights land at roughly 1.5 terabytes rather than something far worse. API pricing is $3 per million tokens in and $15 out, about a third of Claude Fable 5’s rate. And Moonshot promise the full open weights on 27 July.

Now the reality check, because vendor launches deserve one. On the broadest independent composite, Artificial Analysis’s Intelligence Index, K3 scores about 57: fourth in the world, behind Claude Fable 5 on roughly 60 and GPT-5.6 Sol on 59, and level with or ahead of Claude Opus 4.8. On Terminal-Bench 2.1 it posts 88.3, ahead of Fable and just behind Sol. It opened at number one in human preference rankings for front-end coding. It trails the top closed models on pure reasoning, on several software engineering suites and on vision, some of its best scores come from Moonshot’s own harnesses, and early reports describe it as verbose and only moderately fast. Treat the exact rankings as provisional until independent runs accumulate.

None of that blunts the headline. Open-weight models used to trail the closed frontier by six to twelve months. K3 is the first model at this scale with a firm open-weights commitment, and it lands within a point or three of the leaders on axes that matter commercially: long-context coding, agentic work, cost per task. From 27 July, any government, enterprise or lab prepared to run a serious cluster can self-host near-frontier capability, fine-tune it, distil it or audit it. And here is the detail that connects it to everything above: Moonshot built this under American export controls, without reliable access to the best silicon. Denied the finest tools, they got there on extreme sparsity, architectural cleverness and brute force: more chips of lesser quality, more electricity, more space. Investors have noticed; the company raised roughly $2 billion at a $20 billion valuation in May. A three-year-old lab, working under embargo, is now within touching distance of the frontier.

Subscribe now

The real heirs of the Victorians work in Shenzhen, not Silicon Valley

Which brings me to the argument I actually want to make. If you want to see the ghost of the Victorian industrialist at work today, do not look to London, and do not look to Silicon Valley either. Look to Shenzhen, Hangzhou and Beijing. The parallel between nineteenth-century Britain and twenty-first-century China is one of the most robust in economic history: both achieved world-altering industrialisation by choosing scale and speed over local objection, democratic consultation and environmental caution. K3 is the perfect lens for it.

Take compulsory purchase. Just as the Midland Railway used an Act of Parliament to bury a river and build a station on top of it, the Chinese state uses centralised authority to abolish objection altogether. Under the “Eastern Data, Western Computing” programme, thousands of miles of fibre now connect the coastal tech hubs to state-built data centre clusters in Guizhou and Inner Mongolia, where land and energy are cheap. When a high-voltage line needs to cross a province to feed an AI cluster, the state mandates it, relocates whoever is in the way and builds it in months.

Take the mania. In 1846 Parliament passed 272 railway acts. If that number seems to be following you around this issue, it is rather the point. Most of those companies failed; the tracks remained. China’s version is called the Hundred Models War: billions of RMB poured into dozens of competing AI labs on heavily subsidised compute, in the full expectation that 90% of them will fail or merge. The Victorians ruined private shareholders. Beijing absorbs the losses into the state. Both treat the waste as an acceptable tax on the road to supremacy. Moonshot is simply what a survivor looks like.

And take the bottleneck. Faced with a collapsed tunnel, a Victorian engineer went through it rather than around it. Faced with export controls, Chinese engineers networked vast clusters of less efficient domestic chips and accepted the energy penalty, throwing cheap coal and subsidised solar at the data centres until the compute went through. It is the modern equivalent of burning more coal to force a heavy train up the gradient.

The execution is identical: rapid, aggressive, unsentimental. The engine is not. The Victorians ran on fragmented, private, speculative capital. The Midland buried the Sheaf to beat a rival and enrich its shareholders. China runs on state direction: Moonshot are a private company, but they float on subsidised compute, a command-driven grid and a national mandate for AI sovereignty at any cost. China has kept the Victorian tolerance for landscape-altering speed and replaced the chaos of privateering with the discipline of the plan. Whether that combination can outlast its own demographic arithmetic, and whether Western export controls can slow it, are questions for a future issue. What it has already done is expose our own position: we have given up the Victorian methods and not replaced them with anything that builds.

Subscribe now

The River They Buried

All of this started underground. My Yorkshire Post column this week, reproduced in full.

Last week I climbed down into the Megatron. It sits beneath Sheffield station, a Victorian chamber so large that urban explorers compare it to a cathedral, though no cathedral I know of has a river running through the nave. The River Sheaf, which gave the city its name, has been flowing through this brickwork in the dark for a century and a half.

The Victorians put it there. In the 1860s the Midland Railway needed a direct line into Sheffield to replace an awkward branch connection from Rotherham Masborough, and the obvious route ran through the Sheaf Valley. A river was in the way. So they entombed it in brick culverts and built their station on top, and when the trains started running in 1870 nobody standing on the platforms had any idea what was rushing under their feet. Most passengers still don’t, though a grated lightwell built into Platform 5 now lets the inquisitive look down at the water.

What struck me underground is that none of it is crude. The great parabolic arch, three bricks thick, was designed to withstand the weight of trams crossing Exchange Street as well as to swallow floodwater. This was the foundation on which modern Sheffield was built, and it went from plan to platform in a decade.

Before anyone holds that decade up as a rebuke to modern Britain, though, the ledgers deserve a look. Brunel told investors the Great Western Railway would cost £2.5 million and spent more than £6 million. Parliament passed 272 railway acts in 1846 alone, at the top of a mania that ruined thousands of small investors, and the navvies who dug the Woodhead Tunnel died, reportedly, at a higher rate than soldiers at Waterloo. Victorian speed was subsidised by people who never got a vote on it.

The river got no vote either. Bury a watercourse and you abolish its floodplain, which is fine until the weather has other ideas. In December 1991 the Sheaf backed up and flooded the station, halting services on a main artery of the regional economy; the tree branch it dumped on Platform 5 was kept and given a commemorative plaque. Keeping the 750-metre culvert clear now falls to a £3 million Environment Agency screen at Granville Square, while over at Castlegate the council is demolishing a 1917 culvert altogether, daylighting the Sheaf to make a park that can hold floodwater. The rule the Victorians never wrote down: skip the environment’s invoice and you settle it a century later, with interest.

And yet. Anyone who concludes that caution is free has not read the other column of the ledger. The Victorians built first and posted the bill to their great-grandchildren; we send the bill before the building starts. A passing loop takes a decade of consultation. HS2 shed its limbs one assessment at a time, and the grid connections our AI data centres queue for can take longer to approve than the entire station took to build, river and all. Every reservoir we fail to dig lands on the people of 2126 just as surely as the culverts landed on us, except that they will get nothing in exchange.

Bill Drummond of The KLF painted a question on the culvert wall down there: “Is the river really beautiful or is it just the gradient of the land?” I came up blinking into the daylight with a different question, which is whether Britain can still build anything important and leave something worth inheriting at the same time. We managed the first half 150 years ago and have spent a fortune on the consequences. At the moment we manage neither. When you read this on Friday morning, trains will be pulling out of Sheffield above a hidden river, proof that a serious country has to hold two thoughts at once: build the thing, and remember who settles the account.

Guided tours are available at The Sheaf and Porters Rivers Trust

The Signal Tech & AI Layoff Tracker

The corporate theme of Q3 has solidified into a clear compute-over-compensation trade. Profitable mega-cap organisations across tech and telecom are aggressively flattening operational, customer support and traditional engineering layers. The goal is no longer margin defence against cooling revenues but an active capital transfusion: freeing up liquidity to fund AI infrastructure, custom silicon development and data centre cluster deployments.

Verizon, 7,800. The largest single restructuring event of the summer, in pursuit of the targeted $5 billion operating-expense reduction. Verizon eliminated 4,800 domestic customer care, operations and administrative support roles, with internal documentation explicitly outlining the permanent transition of routine customer support to conversational LLM agent frameworks. Concurrently, the company announced the divestiture of 274 corporate-owned retail storefronts to third-party franchise operators, shedding a further 3,000 corporate retail field positions from its direct payroll. Chief executive Dan Schulman has said AI will replace “a large percentage” of customer service positions.

Thomson Reuters, ~500. The week’s clearest structural blueprint for technical workers. Thomson Reuters are cutting roughly 500 legacy engineering and tech operations roles, 1.8% of the global workforce, while planning to hire 250 specialised “AI-native” engineers over the next 24 months. Read that arithmetic the way a legacy developer will: generative developer tools compressing headcount, with identical or superior output demanded from smaller, leaner teams. Expect mid-market software firms to copy it before the Q3 earnings calls.

Mid-market and infrastructure trims, 457. GoKwik, an e-commerce infrastructure platform, executed a sharp 20% workforce reduction, 120 positions, citing a deliberate shift toward AI-automated merchant workflows and support structures. Charter Communications formally logged a Missouri WARN notice eliminating 107 network engineers and IT managers at its Town and Country operations hub, outsourcing those layers to Capgemini. And Rogers Sports & Media dismissed 230 employees and shut six radio stations across Canada as programmatic, AI-driven ad sales hollow out legacy regional advertising.

Watch list for Week 29: enterprise software QA and release engineering tiers, where automated, LLM-driven code-testing suites are being treated as drop-in replacements for human QA benches; and B2B customer success operations, where telecom has proved that customer care can be reliably offloaded to fine-tuned agentic models, putting heavy account-management overhead in the frame before late-Q3 earnings cycles.

Read My Weekly Column

Final Thought 🚀

Hold this issue’s three stories together. A thought experiment says Britain could have Victorian speed tomorrow if it would accept Victorian costs. A lab in Beijing shows a country actually paying those costs, with a five-year plan where the share certificates used to be. And the tracker shows how the West is financing its own build-out: not with navvies this time, but with payroll.

Every era funds its infrastructure with the thing it is most willing to spend. The Victorians spent the environment and the poor. China spends whatever the plan requires. We are spending jobs, quietly, one restructuring memo at a time. And here is the part that should sting: we are paying without building. The layoffs run at roughly a thousand a day while the data centres queue fifteen years for a grid connection. That is the worst position on the board: Victorian costs, modern delays, nobody’s benefits.

The Victorians built the thing and never asked who settled the account. We audit the account so carefully that we never build the thing. Somewhere between the Megatron and the planning inquiry sits a country that can do both: build fast, and leave the people of 2126 something they will be glad to inherit rather than a bill they never voted for. Finding it is not a nostalgia project. On this week’s evidence, our competitors have already chosen their century. We should choose ours.

Until next Sunday,

David

THE SIGNAL PLAYBOOK · No. 1 · Turn Claude Into Your Travel Desk

David Richards MBE — Sun, 12 Jul 2026 05:31:17 GMT

The Signal Playbook is an occasional companion to The Sunday Signal. The Sunday Signal interprets change. The Playbook turns it into an operating procedure.

Subscribe now

Bottom Line Upfront

Most travel prompts fail for a simple reason: they ask for ideas when the traveller needs a decision. Ask Claude for the best route or a great hotel and you’ll get a polished average. A serious travel brief does something different. It defines the job, supplies the relevant facts, requires current sources, states what the model must not invent, and dictates the form of the answer. The fix isn’t better manners. It’s better structure. This playbook gives you twenty such workflows, organised the way a trip actually unfolds: booking, briefing, the ground, the mission, and the moment it all goes wrong. At the end, I’ll show you how to install them permanently so you never paste them again.

Structure beats politeness

Long chat sessions suffer from drift. You start with a sharp brief and twenty messages later Claude has slumped back into tour-guide mode. The cure is block architecture: XML tags that help Claude separate the role, the inputs, the constraints and the output. Every workflow in this guide uses the same six blocks, in the same order: , , , , , . Learn the architecture once and you can build your own.

Two of these blocks matter more than the rest. The constraints tell Claude what it must exclude, flag or decline to invent, which is where most of the value hides. And a defined output format nudges Claude to produce clean, standalone documents (itineraries, comparison tables, evidence packs) you can save straight to your phone.

Each workflow carries a small metadata strip so you can see at a glance what it needs, what it produces and how careful to be with the result.

Subscribe now

The standard of proof

Here’s the part most prompt guides skip. A detailed prompt makes Claude sound authoritative. It cannot manufacture access to a live fare, a current border rule or a policy clause. Specific is not the same as reliable, and a travel desk that invents timetables is worse than no travel desk at all.

So use Claude as a research and planning layer, not the final authority. Confirm anything that can stop the trip, cost significant money or affect health and safety: schedules, fares, availability, entry rules, road rules, tax, medication, insurance and passenger rights. If Claude can’t verify a claim from a current source, it must say so.

The protocol below enforces that. Install it once in your Project instructions (more on that at the end) and it governs every workflow in this guide. Using a workflow standalone? Paste the protocol above it.


- Use web search for any claim involving current schedules, fares, prices,
  availability, laws, regulations, safety, medicine, events or company
  data.
- Cite every material current claim and state the date it was checked.
- Prefer official government, regulator, airport, airline, railway, venue,
  hospital, insurer and company sources.
- Distinguish clearly between:
  1. VERIFIED FACT
  2. REASONED INFERENCE
  3. UNVERIFIED OR UNAVAILABLE
- Never invent a fare, timetable, policy clause, room feature, phone
  number, processing time, legal entitlement or medical rule.
- Return fewer options rather than fill a requested quota with weak
  evidence.
- Where information conflicts, show the conflict instead of silently
  choosing.
- For legal, medical and tax issues, identify what must be confirmed with
  an official source or qualified professional before the traveller acts.
- End every response with a "Before You Act" checklist containing the
  items that still require direct confirmation.

Good prompting isn’t only telling the model what to do. It’s telling it what counts as evidence and when it must stop.

Book like an analyst, not a romantic

1. The Flight Friction Matrix

Looks past the fare aggregators at what actually shapes a long-haul day: the aircraft, the cabin altitude, and the misery of a bad terminal transfer.

Needs: web search, exact dates · Output: comparison table and verdict · Risk: operational


You are a corporate aviation logistics consultant managing flight routing
for demanding executives.



Route: from [Origin] to [Destination]
Departure window: [Date/Time range]
Cabin and budget: [Class / Maximum fare]
Baggage: [Carry-on only / Checked bags]
Loyalty programmes: [Airlines and status, if any]
Maximum connections: [Number]
Primary objective: [Minimise jet lag / Maximise uninterrupted sleep /
Bulletproof connection margins]



Identify the routings that best serve the primary objective, judged on the
friction booking engines hide rather than marketing copy.



Apply the research and verification protocol. Check live schedules and
fares by web search and timestamp every fare. Aircraft assignments change:
label them scheduled, not guaranteed. Include delay statistics only with a
citable source; otherwise omit them.



- Return up to 3 routings. Do not include a weaker option to fill the
  quota.
- Express prices and connection margins as ranges where the evidence does
  not support precision.
- Compare: connection friction (terminal realities at each hub), hard
  product (scheduled aircraft and its cabin altitude, humidity and noise),
  and sleep/work value (elapsed time vs dark-cabin windows).



A clean Markdown comparison table, then a standalone "Smarter Pick"
verdict with the operational rationale.

Subscribe now

2. The Rail Strategist

European multi-city rail, planned by timetable rather than brochure. The pass-versus-tickets question answered with arithmetic, not vibes.

Needs: web search, exact dates · Output: routing plan with full cost arithmetic · Risk: operational


You are a European rail routing specialist who plans by timetable, not by
brochure.



Itinerary: [Cities in order]
Dates: [Date range]
Travellers: [Number, ages if relevant]
Luggage: [What you are carrying]
Priority: [Cost / Speed / Scenery / Flexibility]



Produce the best rail routing for this itinerary and settle the pass
versus point-to-point question with complete arithmetic.



Apply the research and verification protocol. Price fares from official
operator sites and state when each was checked. General estimates are not
acceptable where official prices are published.



- THE PASS MATHS: show the complete calculation for this exact itinerary:
  pass cost vs point-to-point tickets, including compulsory seat
  reservation fees, booking fees and delivery charges on both sides.
- RESERVATION TRAPS: which legs require compulsory reservations, which
  sell out, and how far ahead to book each one.
- STATION INTEL: for every transfer, a realistic minimum connection time
  as a range, the platform layout risk, and where stairs or luggage will
  hurt.
- THE SCENIC OVERRIDE: any leg where the slower train is objectively the
  better decision. Recommend which side to sit on, flagged as a general
  guide: carriage orientation can change.



A leg-by-leg routing plan, the pass arithmetic as a table, then a single
recommendation.

3. The Anti-Tourist Hotel Scout

Applies a short-seller’s bull and bear case to hospitality marketing. Every hotel gets both.

Needs: web search, exact dates · Output: bull and bear cases per property · Risk: operational


You are a hospitality industry short-seller and a cynical local culture
editor.



Location: [City]
Exact dates: [Check-in and check-out]
Occupancy: [Adults, children, rooms]
Nightly budget: [Amount in local currency]
Essential facilities: [e.g. lift, air conditioning, step-free access,
workspace]
Aesthetic: [e.g. Brutalist, high-design boutique, classic luxury]



Find properties in neighbourhoods with real character that beat the
tourist centre on my dates, and give me both sides of the case for each.



Apply the research and verification protocol. Confirm each property is
open and bookable for my dates. Label room-tier, floor and view advice as
VERIFIED (from the hotel's own published information or credible recent
reviews) or INFERRED, and say which.



- Exclude major tourist epicentres, cruise-ship bottlenecks and
  over-indexed influencer hubs.
- Focus on walkable neighbourhoods with distinct architectural integrity
  and a high density of independent local businesses.
- Return up to 3 properties. Fewer is fine if the evidence is thin.



For each property:
- NEIGHBOURHOOD INTEL: the social fabric of the area and why it beats the
  tourist centre.
- THE BULL CASE: the objective design merits, service quality or privacy
  perks.
- THE BEAR CASE: the cynical downsides (micro-rooms, street-noise
  corridors, weak water pressure, poor transit access).
- THE ROOM HACK: the room tier, view orientation or floor range to
  request, each labelled verified or inferred.

Subscribe now

4. The Insurance Interrogator

Reads your policy the way a claims department will: for the exclusions. Best used inside a Project with the policy PDF uploaded.

Needs: policy document · Optional upload: full policy PDF · Output: coverage-gap assessment and evidence checklist · Risk: high-stakes


You are a claims-side insurance analyst who reads policies for the
exclusions, not the promises.



Policy: [Uploaded to project knowledge, or pasted]
Destination and duration: [Where, how long]
Activities: [e.g. skiing, scooter hire, working with expensive equipment]
Valuables carried: [Items and approximate values]
Pre-existing conditions declared: [Yes/No/Details]



Find the gaps between what I think I've bought and what the policy
actually says, before an incident rather than after.



Apply the research and verification protocol. Every finding must quote the
policy verbatim with the section number and page. If a clause is ambiguous,
show the wording and say it is ambiguous. Do not paraphrase exclusions.



- THE GAP ANALYSIS: every exclusion and limit relevant to my trip and
  activities, with special attention to clauses that void cover entirely:
  alcohol clauses, unattended baggage, undeclared conditions.
- THE SINGLE-ITEM CEILING: check my valuables against single-item caps and
  state plainly what is effectively uninsured.
- THE CLAIM-EVIDENCE CHECKLIST: the exact evidence pack to collect at the
  moment of any incident: police report deadlines, airline property
  irregularity reference, receipts, photographs, medical notes.
- This is a coverage-gap assessment, not legal advice. Present findings as
  questions to put to the insurer where the wording is unclear, and
  recommend professional advice for anything material.



A gap table (clause, quotation, section/page, what it means for this
trip), the claim-evidence checklist, then the open questions for the
insurer and the one upgrade most worth investigating.

5. The Border Bureaucrat

Visas, ETAs, transit rules and the paperwork border officers actually check. Built on official sources, because entry rules move.

Needs: web search, official sources · Output: entry requirements dossier · Risk: high-stakes


You are an immigration paralegal who works only from official government
sources.



Passport: [Nationality, plus any second nationality]
Destination: [Country]
Transit points: [Airports/countries en route]
Purpose: [Tourism / Business meetings / Remote work / Other]
Duration: [Length of stay]



Establish exactly what I need to enter and transit legally, and flag
anywhere my planned activity does not fit the visa category I assumed.



Apply the research and verification protocol. Entry requirements come from
official government and immigration sources only, each cited with the date
checked. Flag any requirement that has changed within the past year. Name
the official source I must verify against before booking.



- ENTRY MECHANICS: the visa, ETA or waiver I need, the fee, the official
  application site, and published processing times expressed as ranges.
  Include transit requirements for every layover, including whether a
  terminal change forces me through immigration.
- THE PAPER TRAIL: what officers may ask to see: proof of onward travel,
  accommodation, funds. Distinguish requirements that are routinely
  enforced from those that rarely are, and label that distinction as
  inference.
- THE STATUS CHECK: I will answer all border questions truthfully and
  completely. If any planned activity (business meetings, remote work,
  freelance work) may require a different immigration status than the one
  I'm applying for, say so plainly and identify the correct route. Do not
  advise me how to characterise my plans.



A requirements dossier ordered by deadline: what to obtain now, what to
carry, what to verify on the official source before booking.

Subscribe now

Arrive briefed, not hopeful

6. The Culture Briefing Officer

One novel, one film, one map of the conversations that need care. Arrive knowing what the locals are actually arguing about.

Needs: web search, recent local sources · Output: dated cultural dossier · Risk: operational


You are a cultural attaché and global socio-political journalist.



Destination: [Country/City, plus specific regions you'll visit]
Trip dates: [Dates]
Context: [Tourism / Business / Meeting specific people]



Brief me on the contemporary cultural identity, historical turning points
and current anxieties of this place, without treating the country as
uniform. I reject surface-level trivia.



Apply the research and verification protocol. Ground the zeitgeist section
in recent local media, not travel guides, and date the briefing. Flag
where attitudes differ by region or generation.



- THE SYLLABUS: one regional novel, one definitive film and one piece of
  audio or journalism that distils the local mindset.
- SUBJECTS REQUIRING CONTEXT AND CARE: the historical and political fault
  lines where a visitor needs background before raising them, and how
  views differ across regions and generations.
- THE ZEITGEIST: the domestic news story, economic worry or cultural
  debate currently dominating local conversation, from local sources, with
  dates.
- THE RESPECT VEHICLES: 2-3 regional idioms or customs that signal
  sophisticated respect for the specific region I'm visiting.



A briefing dossier headed with the date it was compiled.

7. The Counter-Scam Actuary

The defensive briefing: which ATMs to trust, how to exit the airport, and the scams currently being reported at your destination.

Needs: web search, official advisories · Output: defensive briefing · Risk: operational


You are an international security consultant specialising in travel-sector
fraud prevention.



Destination: [Country/City]
Arrival airport and time: [Airport, local time]
Profile: [Travelling alone / couple / family; language ability]



A defensive blueprint for my finances, data and physical safety, built on
what is currently being reported rather than folklore.



Apply the research and verification protocol. Source scam patterns from
recent official material where available: police, airport authorities,
banks, government travel advisories. Date every claim. If a widely
repeated scam cannot be verified recently, label it as unconfirmed.



- THE BANKING PROTOCOL: the local ATM landscape. Which banks run secure,
  low-fee networks; which commercial ATM networks carry predatory fees.
  The on-screen choices that matter, including declining dynamic currency
  conversion.
- THE TERMINAL EXIT: the mechanics of leaving my arrival airport safely:
  the authorised taxi queue, the rideshare platforms that verifiably
  operate there, and the fair price to the city centre as a range.
- LOCAL STRATAGEMS: up to 3 scams currently reported as targeting
  travellers here. For each, the warning signs and the safest action that
  exits or de-escalates the situation. Do not promise that any phrase ends
  a con instantly.



A briefing ordered by when I'll need it: before landing, at the airport,
in the city.

Subscribe now

8. The Medical Contingency Planner

Which of your medications are restricted where you’re going, how the local pharmacy system works, and where serious care is. A research brief for you and your clinician, not medical advice.

Needs: web search, official sources · Output: contingency dossier for professional confirmation · Risk: high-stakes


You are a travel health researcher who prepares briefing documents for
travellers to review with their own doctor and pharmacist.



Destination: [Country/City] for [Duration]
Medications: [List, with generic names and dosages if known]
Conditions and allergies: [List]
Accommodation area: [Neighbourhood, if known]



Prepare a contingency dossier I can act on and verify: medication
legality, pharmacy access, emergency escalation and prevention.



Apply the research and verification protocol. Medication import rules come
from official sources (the destination's health ministry, embassy or
customs authority), cited and dated. Never advise changing, replacing,
stopping or substituting any medication: flag the question for my doctor
or pharmacist instead.



- THE LEGALITY AUDIT: check each medication against the destination's
  published controlled-substance rules. Flag anything banned or
  restricted, the documentation required (prescription, doctor's letter,
  import permit), and the official source to verify before flying.
- THE PHARMACY MAP: how pharmacies operate locally (prescription rules,
  hours), and the likely local generic names for my medications, labelled
  clearly as requiring confirmation by a pharmacist.
- THE ESCALATION LADDER: the emergency number, how ambulance services are
  organised and charged, and hospitals near my accommodation noted for
  emergency care, with the evidence for that reputation cited.
- PREVENTION: required and recommended vaccinations per official travel
  health sources, and destination-specific risks prioritised by evidence,
  severity and relevance to me.



A dossier in two parts: what is verified with sources, and a "confirm with
a professional" list for my GP, pharmacist and travel clinic.

Subscribe now

9. The Packing Architect

A one-bag packing list built for the actual microclimate, the local dress code and your airline’s actual baggage rules.

Needs: web search · Output: capsule packing list · Risk: operational


You are a minimalist textile expert, wardrobe consultant and international
cultural etiquette specialist.



Trip: [Number of days] in [Destination] in [Month]
Activities: [e.g. heavy walking, fine dining, regional rail]
Airline and cabin bag: [Carrier, so baggage dimensions and weight limits
can be checked]
Laundry access: [Yes/No/Unknown]



A strict one-bag carry-on list that works for the real conditions, not the
seasonal average.



Apply the research and verification protocol. Check my airline's published
cabin baggage dimensions and weight. Use climate data for the location and
month, and flag that a forecast only becomes reliable close to departure.



- MICROCLIMATE ANALYSIS: the weather realities for this location in this
  month: diurnal temperature swings, humidity spikes, rain probability.
- CULTURAL DRESS CODE: the local visual norms. Flag items that mark me as
  a pickpocket target or breach religious and social etiquette.
- THE CAPSULE MATRIX: a compact set of interchangeable pieces (roughly
  five tops and five lower garments or layers) that combine into at least
  10 distinct outfits. Show the combinations.
- FOOTWEAR PROTOCOL: footwear for the actual local surfaces (wet marble,
  cobblestones, steep terrain) without sacrificing style.



The packing list grouped by function, the outfit matrix, and the total
item count against my airline's limits.

Subscribe now

10. The Jet Lag Strategist

Sports science translated into an hour-by-hour light, sleep and caffeine plan. Jet lag can be managed; this is the protocol for managing it.

Needs: flight times · Output: hour-by-hour adjustment schedule · Risk: high-stakes


You are a sports scientist who designs circadian adjustment protocols for
travelling athletes, and you are conservative by default.



Origin and destination cities: [Cities]
Departure and arrival: [Dates and local times]
Baseline sleep window: [e.g. 11pm to 7am]
Health context: [Any conditions or regular medications, or "none"]



An hour-by-hour schedule from 24 hours before departure to 48 hours after
arrival that reduces jet lag as far as the evidence supports.



Apply the research and verification protocol. Label each recommendation by
evidence quality: well-established (light timing, caffeine timing) vs
promising but less certain (meal timing strategies). This is not medical
advice: if I've listed any condition or medication, flag which parts of
the protocol need checking with a doctor first, and default to the
conservative option throughout.



- LIGHT MANAGEMENT: windows to seek bright light and windows to avoid it,
  the highest-evidence lever available.
- SLEEP AND NAP PROTOCOLS: strategic sleep windows on board vs wakefulness
  targets, with realistic fallbacks if sleep fails.
- MEAL TIMING: sensible adjustment of eating times toward the destination
  clock. Do not prescribe aggressive fasting.
- CAFFEINE: strategic intake and cut-off times.
- No supplement or medication recommendations. Mention melatonin only to
  note it is regulated differently by country and is a question for a
  pharmacist or doctor.



A chronological table: time, action, why, evidence level.

Subscribe now

Own the ground

11. The Day-Batcher

Builds a day around geography rather than a wish list, so you stop donating hours to transit.

Needs: web search, opening hours · Output: chronological itinerary with contingencies · Risk: operational


You are an elite urban concierge specialising in time-optimised,
high-yield itineraries.



Target city: [City]
Base: [Accommodation or starting point]
Specific day: [Day and date]
Core focus: [e.g. Art and architecture, contemporary history, retail]
Pace and mobility: [Leisurely / Efficient / High-yield; any step-free or
accessibility needs]



A day where the geography does the work: minimal transit, maximum yield,
and a fallback at every step.



Apply the research and verification protocol. Verify opening hours and
closure days for my specific date, and flag anything that requires advance
reservation with the booking lead time.



- GEOGRAPHIC CLUSTERING: group all targets within a tight micro-radius.
  Transit between consecutive morning or afternoon stops should be on foot
  or a short ride, treated as an estimate that varies with traffic.
- THE ANCHOR ATTACK: put the biggest, busiest cultural asset first thing
  in the morning to beat the tour buses.
- FRICTION PADDING: allow for realistic walking speeds, queues and
  check-in times.



A chronological morning-to-night itinerary as a clean, scannable Markdown
list. For every step, a "Plan B" block covering weather, oversold venues
and fatigue.

12. The Street-Side Historian

Turns the pavement outside a monument into a sharp history lesson, delivered at walking pace. No romance, no textbook dates.

Needs: nothing beyond a location · Output: 180-word briefings on demand · Risk: operational


You are an elite historian specialising in regional architecture,
socio-economic flashpoints and political history. You are my private,
real-time walking tour guide for [Location/Trail].



I will name buildings, plaques and monuments as I reach them.



Deep, academic-grade history delivered as sharp storytelling I can read on
a phone while walking. Prioritise gritty realities, human flaws, economic
motives and lesser-known episodes over textbook dates.



Apply the research and verification protocol. Label contested
interpretations, legends and disputed attributions as exactly that. A
good story flagged as legend beats a legend passed off as fact. Note your
main source type (scholarship, contemporary accounts, local tradition) in
one line.



- BREVITY: under 180 words per stop.
- STRICT PACING: never jump ahead or summarise the whole trail. Only
  analyse the exact site I specify.
- NO HALO EFFECT: do not romanticise historical figures. Present their
  real geopolitical and financial motives.



1. Name of the site
2. The raw history (180 words maximum, with disputed points labelled)
3. One-line source note
4. The hook: exactly two hyper-specific questions that let me steer the
   depth (e.g. "Do you want the financing scandal behind this building,
   or the plot that failed on these steps?")

Subscribe now

13. The Cynical Gastronomer

Filters out the laminated menus and the TikTok queues using signals you can actually observe. Up to four venues, each with a reason to exist.

Needs: web search, dates · Output: classified recommendations · Risk: operational


You are an uncompromising culinary critic and investigative food
journalist who despises food trends and curated social media hype.



Location: [City/Neighbourhood]
Date and day of week: [When I'm eating]
Budget: [Per head, per meal]
Dietary needs: [Any]



Exceptional, ingredient-driven food and drink, selected on observable
quality signals rather than fame.



Apply the research and verification protocol. Confirm each venue is
currently open and check closure days against my dates. Judge venues on
observable criteria: menu breadth and language, what locals order, chef
provenance, queue composition. Do not claim knowledge of paid placement
you cannot verify.



REJECT: venues whose fame rests on views rather than cooking, menus
translated into five languages with photographs, and anything built on
sensory gimmicks.
Return up to 4 venues across these categories, and fewer if the evidence
is weak:
1. THE INSTITUTION: a generational venue still executing flawlessly.
2. THE INSIDER'S PROGRESSIVE: a modern, ingredient-obsessed kitchen
   quietly favoured by local chefs.
3. THE HOLE-IN-THE-WALL MASTERPIECE: an unpretentious specialist
   mastering a single dish.
4. THE LIQUID ARCHITECT: a bar or pub devoted to serious beverage craft,
   history or atmosphere, with no club dynamics whatsoever.



For each venue: the exact dish or pour that justifies the visit, how to
beat the booking system or the queue, and the red flag that signals
quality has slipped.

14. The Event Sniper

What’s actually on during your exact dates, and how to buy real tickets without feeding the touts.

Needs: web search, exact dates · Output: event shortlist with official ticket routes · Risk: operational


You are a city desk arts and events editor with a season ticket to
everything.



City: [City]
Exact dates: [Dates]
Tastes: [e.g. football, opera, live music, food festivals, exhibitions]
Budget: [Per ticket]



The events genuinely worth attending during my exact dates, each with a
legitimate route to a ticket.



Apply the research and verification protocol. Search against my exact
dates. Link only official sellers or resale platforms with buyer
protection, and state the date each price was checked. Refuse to recommend
unofficial sellers without buyer protection.



- THE REAL CALENDAR: fixtures, performances, festivals and closing
  exhibitions on my dates. Exclude tourist-trap dinner shows and anything
  staged nightly for visitors.
- THE TICKET ROUTE: for each event, the official seller, the realistic
  face price as a range, when tickets are released or returned, and the
  legitimate resale route if sold out. Name the tout patterns to avoid.
- THE SEATING INTEL: where to sit or stand at each venue, and the ticket
  tiers that are poor value, labelled verified or inferred.
- THE LOCAL RITUAL: the pre- or post-event custom (the pub, the square,
  the late bar) that turns a ticket into an evening.



A shortlist ordered by date, with ticket routes and deadlines flagged.

Subscribe now

15. The Hire-Car Defence Brief

Written from the other side of the rental desk. The upsells, the deposit traps and the tourist-fining road rules, judged against your actual booking.

Needs: web search, official sources · Optional upload: booking confirmation and rental terms · Output: desk-to-return defence brief · Risk: high-stakes


You are a former rental car branch manager turned consumer advocate. You
know every technique used at the desk because you trained staff to use
them.



Hire: [Country/City], [Company if known], [Duration]
Licence: [Issuing country]
My cover: [Card benefits or separate excess policy, if any; upload or
paste the rental terms if available]
Route: [Cities/regions I'll drive through]



Prepare me for the desk, the road and the return, based on my actual
contract position rather than blanket rules.



Apply the research and verification protocol. Road rules come from
official sources for each country on my route, cited and dated. Base
insurance guidance on my actual rental terms and existing cover: if I
haven't provided them, list what to check rather than a blanket answer.



- THE DESK GAUNTLET: the upsells and pressure tactics I will face
  (insurance upgrades, fuel schemes, vehicle upgrades). For each, the
  questions to ask and how the answer depends on my existing cover. Do
  not issue a blanket instruction to decline insurance.
- THE DEPOSIT TRAP: the expected deposit block as a range, card
  requirements, and the step-by-step photographic walkaround protocol
  that wins damage disputes.
- THE ROAD RULES THAT FINE TOURISTS: for my route: ZTL zones, vignettes,
  toll transponders, average-speed cameras, parking colour codes,
  drink-drive limits. Official sources only.
- THE RETURN: the fuel policy reality, the out-of-hours return risk, and
  how long to keep my evidence afterwards.



A brief in the order I'll need it: before the desk, at the desk, on the
road, at the return.

16. The Haggle & VAT Reclaimer

Negotiation craft for the market, warning signs for the goods, and the airport VAT refund procedure nobody explains properly. Two separate briefs, because bargaining is culture and customs is law.

Needs: web search, official customs sources · Output: negotiation brief plus customs and VAT brief · Risk: operational


You are a veteran market trader and, separately, a customs and tax
researcher. You keep the two jobs distinct.



Shopping in: [Country/City]
Target goods: [e.g. rugs, leather, ceramics, electronics]
Returning home to: [Home country]



Help me negotiate like someone who has stood on both sides of the stall,
and get my goods home legally with the tax I'm owed back.



Apply the research and verification protocol. Customs allowances,
prohibited goods and VAT refund rules come from official government
sources, cited and dated. Do not state standard market markups as fact
without evidence; present negotiation norms as cultural guidance.



Part one, the market:
- THE NEGOTIATION CULTURE: how bargaining works here, where it is
  expected, and where it is an insult. The walking-away script.
- THE WARNING SIGNS: physical checks for my target goods (weight,
  stitching, smell, markings) presented as red flags that justify walking
  away, not as proof of authenticity.

Part two, the law:
- THE CUSTOMS LINE: my duty-free allowance going home, what must be
  declared, and the goods (ivory, antiquities, certain foods) that will
  be seized regardless of receipts.
- THE VAT RECLAIM: whether I qualify, the minimum spend, the paperwork to
  demand at purchase, and the airport procedure: which desk, before or
  after check-in, and how the refund is actually paid.



Two clearly separated briefs: "At the stall" and "At the border", each
with its own checklist.

Subscribe now

Travel with a mission

17. The Family Logistics Officer

Plans around the least predictable members of the party. The itinerary rebuilt around nap windows, pram-hostile terrain and boredom contingencies.

Needs: web search · Output: family-paced daily plans with bail-outs · Risk: operational


You are a military logistics planner retrained as a family travel
specialist. You plan around the least predictable members of the party.



Destination: [City/Region]
Base: [Accommodation location]
Party: [Ages of children; any mobility, sensory or dietary needs in the
party]
Transport: [Walking, pram, public transport, car]
Duration: [Days]
Hard limits: [e.g. nap 1pm to 3pm, early dinners, maximum continuous
walking time]



Days that survive contact with real children: built on their rhythms,
with an exit at every stage.



Apply the research and verification protocol. Verify opening hours,
lift/step-free access claims and family facilities from venue sources
where possible; label the rest as inferred.



- RHYTHM FIRST: build each day around the children's sleep and meal
  windows, not the sights. Schedule the flagship attraction in their
  strongest window.
- TERRAIN AUDIT: flag pram-hostile ground (stairs, gravel, stations
  without lifts) on our actual routes, with the workaround for each.
- THE BOREDOM CONTINGENCY: for every stop over 45 minutes, one specific
  child-facing angle: the gruesome story, the thing to count, the machine
  to watch.
- THE BAIL-OUT: for each day, the pre-identified point where the plan can
  be cut short without wasting it, plus the nearest playground, pharmacy
  and reliable family facilities.



A day-by-day plan with timings, terrain notes and the bail-out point
marked in each day.

Subscribe now

18. The Business Trip Compressor

The 36-hour trip built around one thing: the meeting. Hotel by geography, briefing by dossier, etiquette by rulebook.

Needs: web search, company sources · Output: logistics plan plus counterpart dossier · Risk: operational


You are an executive operations chief who plans trips around one thing:
the meeting.



Destination: [City]
The meeting: [Time, address, counterpart company and attendees if known]
Window: [Arrival and departure constraints]
My side: [What I'm there to achieve]



Compress everything around the meeting: logistics that guarantee I arrive
early and sharp, and a briefing that means I know more than they expect.



Apply the research and verification protocol. Build the counterpart
briefing from official company sources, filings and reliable current
reporting, all cited and dated. Separate known facts from hypotheses about
what they want, and label the hypotheses.



- GEOGRAPHY BY MEETING: choose the hotel by distance to the meeting and
  quality of morning logistics, not by brand. The exact transit plan to
  arrive 20 minutes early, plus the backup route.
- THE COUNTERPART BRIEFING: a one-page dossier: recent results,
  leadership changes, live controversies (facts, cited), then a clearly
  labelled hypothesis section on what they likely want from this meeting.
- THE ETIQUETTE LAYER: local business norms: greetings, cards, gifts,
  punctuality tolerance, meal customs, and the small errors that read as
  disrespect.
- THE RECOVERY MARGIN: where to work between commitments, and one
  excellent, fast, quiet dinner option near the hotel.



Two documents: the logistics plan in chronological order, and the
counterpart dossier with facts and hypotheses separated.

19. The Remote Work Auditor

Judges a city on bandwidth, power and time zones rather than beaches. For anyone planning to work where others holiday.

Needs: web search · Output: infrastructure audit · Risk: operational


You are a remote operations manager and digital infrastructure auditor.



Location: [City] for [Duration]
Work pattern: [Base time zone, e.g. GMT] while living in [Local time
zone]; working hours [e.g. 2pm to 10pm local]
Nationality: [For the visa and tax overview]



An honest audit of this city as a workplace: infrastructure, workspaces,
daily structure and the legal exposure that needs professional eyes.



Apply the research and verification protocol. Distinguish measured
performance data from advertised speeds, and note that connectivity varies
building by building, not just by neighbourhood. The visa and tax section
is an overview of official sources, not a personal determination: flag
what needs an accountant or immigration adviser.



- INFRASTRUCTURE TRUTH: fibre and 5G performance (measured where
  possible), common ISP bottlenecks, and power grid reliability, with the
  building-level caveat stated.
- ERGONOMIC STATIONS: up to 2 coworking spaces or laptop-friendly hubs
  known for quiet zones, decent seating and reliable uptime, plus local
  café etiquette on long laptop sessions. Fewer if the evidence is thin.
- THE DAILY STRUCTURE: a practical routine balancing my working hours
  with local life: meals, shopping, culture.
- LEGAL GUARDRAILS: an official-source overview of visa options and tax
  exposure triggers for my stay length, ending with the specific
  questions to put to a professional.



An audit report with a verdict: suitable, suitable with caveats, or
unsuitable, and why.

Subscribe now

And when it goes wrong

20. The Disruption Counsel

The workflow you hope never to run. Cancelled flight, strike, missed connection: your rights, the desk script, and the claim that sticks.

Needs: web search, official sources · Optional upload: booking confirmation · Output: rights brief, desk script, claim plan · Risk: high-stakes


You are an air passenger rights specialist and an airline operations
insider. You know what airlines owe and what their desk agents are
trained not to offer.



Flight: [Flight number], [Origin] to [Destination], [Date]
Operating carrier: [If different from the marketing airline]
Booking: [Booked direct / via agent; single through-ticket or separate
tickets]
Ticket type: [Fare class/conditions if known]
Disruption: [Cancelled / delayed by X hours / overbooked], stated reason:
[If given]
My situation: [Onward commitments, accessibility needs, travelling with
children]



Establish what I am owed, get me moving on the best available routing,
and build the claim while the evidence is fresh.



Apply the research and verification protocol. Identify the applicable
regime (UK261, EU261, the Montreal Convention, or the airline's own
conditions) from my routing and jurisdictions. Every amount, threshold
and deadline must carry an official citation: regulator, legislation or
the airline's published conditions. Separate three categories explicitly:
enforceable rights, airline policy, and discretionary goodwill.



- MY RIGHTS: what the applicable regime entitles me to right now:
  rerouting, care, refund, compensation, with amounts and the conditions
  attached, each cited.
- THE DESK SCRIPT: the wording to use at the desk or on the phone to
  secure the best rerouting, including interline and rival-airline
  options the agent won't volunteer, and noting where these are
  entitlements vs requests.
- THE EVIDENCE PACK: what to photograph, save and request in writing now
  so a claim sticks later.
- THE CLAIM: the deadline, where to file, and the escalation route to
  the regulator or an ADR scheme. Never route me through claims-farm
  companies that take a cut.



Three sections in priority order: act now (rerouting and care), preserve
(evidence pack), claim (deadlines and route), with enforceable rights,
policy and goodwill labelled throughout.

Subscribe now

Install it once, use it forever

Pasting long prompts gets old fast. Claude’s Projects feature lets you install this whole system permanently, but the architecture matters: instructions are for rules, knowledge is for content. Four steps.

Create a dedicated Project. Open Claude, create a new Project, and name it something you’ll recognise. Travel Desk works. This becomes your standing workspace for every trip.
Keep the instructions short: a router, not a library. Put five things in the Project instructions: the research and verification protocol from this guide, your standing preferences (home airport, passport, seating, budget bands), an instruction to find and run the named Playbook workflow from project knowledge, a cap of three clarifying questions before answering, and the citation standard. The long workflows don’t live here.
Put the Playbook in Project Knowledge. Upload the complete Playbook, your current itinerary, your insurance wording, booking confirmations and a one-page preference sheet. Claude reads them all. One note: separate chats in a Project don’t share what was said in each other; only what’s in project knowledge carries across.
Command, don’t paste. Open a fresh chat and issue a short order: “Run the Rail Strategist for the London, Paris, Zurich itinerary in my project files. Optimise for reliability and one checked bag.” Claude pulls the workflow from knowledge, the dates from your documents, and executes under the protocol.

Privacy check. Upload only what the workflow needs. Redact unnecessary medical details, payment data, identification numbers and booking credentials. A Project is a working folder, not a dumping ground for every travel document you own.

Set it up once, before the trip. You will never plan travel the old way again.

Final Thought

Every workflow in this Playbook follows the same discipline. Define the decision. Supply the evidence. Set the constraints. State what the model must not invent. Then specify the deliverable. The tourist asks for recommendations and receives averages. The operator defines the standard of proof. That lesson travels well beyond travel.

Until next Sunday,

David

Subscribe now

Tested with: Claude Opus 4.8 & Fable 5.0, web search enabled · Playbook revision 1.0 ·last tested 11/7/2026

The AI Wars Go Agentic

David Richards MBE — Sun, 12 Jul 2026 05:30:20 GMT

THE SUNDAY SIGNAL · Issue #62 · Week 28 · Sunday 12 July 2026

Four major models landed in three days, the price war intensified below the flagship tier, and Anthropic began the month restored from a short-lived export-control suspension. Raw intelligence is no longer enough. Execution economics increasingly decide what gets deployed.

Subscribe now

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Bottom Line Upfront: The frontier AI market has stopped rewarding cleverness for its own sake. In the space of three days this week, Tencent, xAI, Meta and OpenAI all shipped major new models, while Anthropic, restored at the turn of the month from a short-lived export-control suspension, holds the published lead on the hardest coding benchmark. There is no single winner, and that is the story. Leadership now depends on the workload and the evaluation, and what the market pays for is agentic efficiency: useful work completed per unit of time and cost. This issue maps the new hierarchy, sets out how to choose between the contenders (test first, route second), launches a new free feature (The Signal Playbook), and closes with a layoff tracker that shows exactly who is paying for all this compute. Spoiler: it is the payroll.

Subscribe now

Intelligence is table stakes. Agency wins the contract

The sequence ran like this. Tencent released Hunyuan Hy3 on Monday 6 July. xAI followed with Grok 4.5 on Wednesday 8 July. On Thursday 9 July, Meta and OpenAI released Muse Spark 1.1 and GPT-5.6 within hours of each other. Four major models in three days, with one caveat that belongs up front: Tencent’s claim that Hy3 competes with the larger flagship systems is, for now, Tencent’s claim.

Subscribe now

Anthropic’s return is the other story, and the chronology matters. The US export controls that suspended their most powerful models were lifted on 30 June, and Claude Fable 5 access was restored on 1 July. Mythos 5 has not returned to general availability; access remains limited to approved organisations and Glasswing partners. The frontier’s most watched model family entered July operational, but not unrestricted.

Strip away the launch theatre and one pattern emerges. Every lab, without exception, has pivoted its pitch away from raw foundational intelligence and towards agentic efficiency: long-horizon, multi-step task execution, run autonomously, priced to deploy at scale. The customer being courted is no longer the curious consumer. It is the chief operating officer with a headcount spreadsheet open in the next window.

To see where compute is delivering return, here are the leading models across the published evaluation vectors, with the methodology caveats that launch-week coverage tends to omit.

Why Anthropic still lead on one important coding measure

Claude Fable 5 retains the published lead on SWE-Bench Pro, resolving 80% of tasks on the hardest widely cited software engineering benchmark. But the broader coding crown is workload-dependent. GPT-5.6 Sol leads the Artificial Analysis Coding Agent Index at 80 to Fable’s 77.2, and Sol tops Terminal-Bench at 91.9% only in its multi-agent Ultra configuration; standard Sol scores 88.8%, Fable 83.1%.

Subscribe now

The practical case for Fable is its strong performance on difficult, long-horizon software work, not a disclosed architectural advantage over every competitor. Partner testimony from testing reports that, at its highest effort setting, Fable reflects on and validates its own work before returning it. Treat that as observed behaviour, not a documented mechanism; Anthropic have published no architecture details for the Mythos class beyond defining it as a capability tier above Opus.

The trade-offs matter just as much as the scores. Fable carries premium token pricing. Its safety classifiers route flagged cybersecurity, biology, chemistry and model-distillation requests to Opus 4.8, and Anthropic acknowledge false positives, so sensitive-domain workflows can hit fallbacks. And all Mythos-class traffic, Fable 5 included, carries mandatory 30-day retention. Any organisation weighing a move of proprietary code or scientific work onto it should price all three factors in before committing.

The chips underneath the software

To view this race purely through software benchmarks is to miss the tectonics underneath. The race for hardware independence is accelerating on both sides of the Pacific. Reuters report that an internal memo shows Meta plan to begin manufacturing their Iris accelerator in September, augmenting rather than replacing their Nvidia and AMD fleet. DeepSeek have optimised V4, their 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters, for Huawei’s Ascend platform, and Reuters separately report they are developing an inference chip of their own. The common objective is not immediate independence but greater control over cost, supply and inference capacity.

The regulatory lesson runs alongside it. Anthropic’s June suspension and July restoration is a live demonstration of the geopolitical risk premium embedded in AI supply chains. Architect for model agnosticism: your API abstraction layer should be able to swap one frontier family for another before lunch if trade friction locks one down again.

The bottom line: there is no uncontested technical crown. Fable 5 owns the hardest published coding benchmark, Sol owns the terminal and the composite coding index, and Terra and Grok 4.5 are forcing a commoditisation curve below the flagship tier. Leadership depends on the workload and the evaluation. Which is exactly why the next section exists.

Stop asking which AI is best. Ask which AI is best at this, then prove it

With the frontier fracturing into specialisations, multi-model orchestration is no longer a luxury. It is an operational necessity. But the market changed on Thursday, and categorical verdicts written on Friday deserve your suspicion, including mine. What follows are default candidates, not conclusions. The conclusions have to come from your own evaluation, on your own work.

💻 Long-horizon software engineering. Shortlist Claude Fable 5 and GPT-5.6 Sol, and test both against your own repositories rather than the leaderboards. Fable leads SWE-Bench Pro; Sol leads the Artificial Analysis Coding Agent Index. Weigh Fable’s premium pricing, sensitive-domain fallbacks and 30-day retention against what the benchmark deltas are worth on your codebase. Budget option for lower-stakes internal tooling: Grok 4.5.

🛡️ Terminal, sysadmin and live infrastructure. GPT-5.6 Sol is the leading candidate. Be precise about the headline number: the 91.9% Terminal-Bench score uses the multi-agent Ultra configuration, while standard Sol scores 88.8%. Both make it the natural first test for SecOps and SRE copilot work.

📊 High-volume document and workflow automation. Test GPT-5.6 Terra against Grok 4.5, and judge them on total cost per successful task, not token price. A cheaper model that burns twice the tokens, makes more tool calls or needs more human rework costs more per completed job. Token price is the sticker; cost per accepted result is the invoice.

🖥️ Multimodal and computer-use agents. Muse Spark 1.1 is the natural first candidate. This is Meta’s stated focus for the model: multimodal reasoning for agentic tasks, computer use, tool orchestration and coding.

🧪 Science and healthcare. No crown here, and be wary of anyone offering one. OpenAI’s published table puts Sol and Fable both around 60% on HealthBench Professional, and Fable’s biology and chemistry fallbacks to Opus 4.8 may disqualify it for some workflows outright. Evaluate Sol and appropriate specialist systems against your own protocols.

🔒 Local and on-premises deployment. Tencent’s Hy3 is a credible open-weight candidate: a 295-billion-parameter mixture-of-experts design with 21 billion active parameters, released under Apache 2.0. One discipline, though. Open weights make local operation possible; they do not by themselves establish security, sovereignty or regulatory compliance. Hosting, telemetry, data flows, model provenance and local regulation still need their own review.

Before you route anything: the evaluation playbook

The efficient enterprise does not let developers pick their favourite model, and it does not route on launch-day adjectives either. It routes on evidence gathered like this.

Select 50 to 100 representative production tasks. Include your difficult failures, not just the average cases. An evaluation built on easy tickets flatters every model and informs nothing.
Run them through at least two candidate models. Identical tools, identical harness, identical acceptance criteria. Cross-provider launch tables are not apples-to-apples; your evaluation must be.
Measure what the invoice measures. Pass rate, elapsed time, total input and output tokens, human rework, refusal or fallback rate, and total cost per accepted result.
Route only after that evidence exists. Then keep the abstraction layer (LiteLLM or an internal gateway) between your applications and every provider, because Thursday will happen again.

Route on evidence, not on press releases. The companies that learn this in Q3 will be the ones still solvent enough to enjoy Q4.

The other side of the ledger

I have spent much of this newsletter’s life, and most of my Yorkshire Post column inches, warning about what AI will do to white-collar work. I retract none of it, and the tracker below will make that point again shortly. But the ledger has two sides, and my column in Friday’s Yorkshire Post is about the other one. Here it is in full.

The Historian in My Pocket

A note on timing: this column was written on Friday morning and went to the Yorkshire Post before England’s quarter-final against Norway in Miami on Saturday night. It is reproduced here unchanged, pre-match optimism included.

Two weeks ago I found myself in Boston with a free afternoon, which in thirty years of visiting the city on business had never once happened. I was there for the World Cup, ahead of England’s goalless draw with Ghana at Foxborough, and this is the 250th anniversary year of American independence. So I finally did the thing I had always skipped: I walked the Freedom Trail.

A quick aside. During my twenty-three years in California, someone would ask me almost every July whether we celebrate the Fourth of July in Britain. My answer never varied: of course we do, we call it Loss of Colony Day.

Rather than buy a guidebook or surrender to Google, I built my own tour guide. I gave an AI a role: an elite Harvard historian specialising in the American Revolution, acting as my private guide in real time. Keep each stop under 250 words. Never jump ahead. Skip the textbook summaries in favour of gritty realities, human flaws and political scheming.

I used Claude’s audio function, so I put my earphones in and set off from Boston Common with a voice in my ear. At each stop I said two words, “I’m here,” and the history arrived, told like a thriller.

At the Granary Burying Ground it introduced me to Paul Revere and Sam Adams, lying a few feet apart. At the Old North Church it told me the famous two lanterns burned for less than a minute before being snuffed out, and that the man who hung them was Robert Newman, a 23-year-old church caretaker history forgot. The bells in that steeple, it added, were cast in Gloucester in 1744, the oldest change-ringing bells in North America; among the teenagers who once rang them was Paul Revere himself. On Copp’s Hill it showed me Captain Daniel Malcolm’s gravestone, still dented by British musket balls fired at a dead patriot’s defiant epitaph for target practice.

Best of all, I learned Revere never shouted “The British are coming.” The colonists still considered themselves British. The warning that passed through the countryside was that the Regulars were out.

I also discovered that the Liberty Bell has an English prehistory. Its first incarnation came from London’s Whitechapel Foundry, but it cracked on its first test. Philadelphia metalworkers John Pass and John Stow melted it down and recast the bell in 1753. Its inscription still spells Pensylvania with a single N. That was acceptable at the time. Today it reads like a typo on a national relic.

The same technology handled the rest of the trip. One prompt cast the AI as a corporate travel expert, comparing flights by on-time performance and aircraft comfort. Another played hospitality analyst, offering three hotels with a bull case, a bear case and the exact floor to request. A third, a cynical culinary critic, built a one-day itinerary with a Plan B for bad weather. Every recommendation landed.

I have spent much of this column’s life warning about what AI will do to white-collar work, and I retract none of it. But the ledger has two sides. This summer, anyone with a phone and a decent prompt has a private historian, a travel agent and a concierge on call, free, in their pocket. The trick is not the technology but the ask. Give it a role, give it constraints, give it taste. Vague questions get brochure answers; precise ones get expertise.

Which brings me to Friday morning, and to tomorrow night in Miami, where England face Norway in a World Cup quarter-final. Get past Haaland and his Vikings, and the semi-final beckons in Atlanta. Beyond that, maybe, just maybe, a final on American soil in New Jersey, 250 years after the Declaration was signed. Should it come to that, I know exactly what the cry from the terraces ought to be. Not the Regulars this time. The English are coming.

The newsletter tells you what’s happening. The Playbook shows you what to do about it

Read The Playbook

Those Boston prompts were not improvised on the pavement. They were built, tested and refined, and they are the seed of something new that launches this week: The Signal Playbook.

Here is the idea. Every week this newsletter analyses what AI is doing to work, to markets and to the country. Readers keep asking the obvious next question: fine, but what do I actually type into the box? The Playbook is the answer. Each edition is a standalone, production-ready guide to one real-world job. Copy the prompts, fill in the brackets, keep them forever.

Playbook No. 1 is Stop Prompting Claude Like a Tourist: twenty production-ready prompts that turn Claude into a working travel desk. And the premise is simple. Most published travel prompts fail because they treat Claude like an overly cheerful concierge. Ask nicely and you get brochure copy: pleasant, generic and useless when you’re standing on a rain-slicked platform trying to decode a foreign transit map. The fix is not better manners. It is better structure. A working prompt reads like code: an explicit role, hard constraints, a list of things it must refuse, and a defined output format.

Here is what that looks like in practice. Remember the hospitality analyst from Boston, the one that offered three hotels with a bull case, a bear case and the exact floor to request? This is it, exactly as it appears in the Playbook. Prompt No. 3 of twenty: the Anti-Tourist Hotel Scout.


You are a hospitality industry short-seller and a cynical local
culture editor.



Location: [City]
Duration and season: [Number of nights / Month]
Nightly budget: [Amount in local currency]
Aesthetic: [e.g. Brutalist, high-design boutique, classic luxury]



- Exclude major tourist epicentres, cruise-ship bottlenecks and
  over-indexed influencer hubs.
- Focus on walkable neighbourhoods with distinct architectural
  integrity and a high density of independent local businesses.



Identify the top 3 properties. For each:
- NEIGHBOURHOOD INTEL: the social fabric of the area and why it
  beats the tourist centre.
- THE BULL CASE: the objective design merits, service quality or
  privacy perks.
- THE BEAR CASE: the cynical downsides (micro-rooms, street-noise
  corridors, weak water pressure, poor transit access).
- THE ROOM HACK: the precise room tier, view orientation or floor
  range to request.

Fill in the four brackets, paste the whole block, and go. Notice what the structure is doing. The role removes the incentive to flatter: a short-seller gets paid for finding what is wrong. The filtration rules exclude the tourist epicentres before the search even starts. And the output format demands a bear case for every property, which means the model cannot retreat into brochure copy even if it wants to.

The Scout is one of twenty, and the set is organised the way a trip actually unfolds: booking, briefing, the ground, the mission, and the moment it all goes wrong. The booking phase includes the Flight Friction Matrix, which compares routings by aircraft, cabin altitude and terminal reality rather than fare-aggregator marketing, and the Insurance Interrogator, which reads your policy the way a claims department will: for the exclusions. Later phases cover borders, scams, packing, jetlag, food worth queueing for, hire-car disputes and the Disruption Counsel, for the day the airline cancels on you. The guide closes by showing you how to wire your favourites into a Claude Project with your bookings uploaded, so that on the road the whole thing runs on a two-word command.

One discipline runs through all twenty. Visa rules, passenger rights and medication laws change, so wherever a prompt touches them, a standing instruction forces Claude to flag the volatile rule and point you at the official source. Trust the analysis, verify the law.

Three things to know. The Playbooks are free. They will be a regular feature, published as companion guides on The Sunday Signal between the weekly issues. And they follow one editorial rule, the same one that ran through my Freedom Trail experiment: vague questions get brochure answers, precise ones get expertise.

Read Playbook No. 1 →

Read The Playbook

The Signal Tech & AI Layoff Tracker

Week 27 wrap · 5 to 11 July 2026

We have crossed the mid-year threshold, and the consolidated picture is sobering. Integrating live metrics from Layoffs.fyi, LayoffHedge, SkillSyncer, state WARN registries and the newly updated mid-year reports from Challenger, Gray & Christmas, the core thesis of 2026 is clear: the tech economy is detaching from the broader market to prioritise AI computing infrastructure over human overhead.

The core numbers. LayoffHedge’s cross-sector macro-tracker puts 2026 year-to-date losses at more than 585,151. The tech-specific figure stands at 185,894 on SkillSyncer’s indexing across 267 events, or roughly 120,000 on the strict Layoffs.fyi baseline. About 7,322 were added this week, driven almost entirely by Microsoft’s July fiscal reset.

The H1 autopsy and the death of AI laundering

The mid-year data delivers an unforgiving reality check. Tech sector layoffs in the first half of 2026 shattered previous pacing expectations, averaging roughly 968 to 1,005 job losses per working day, an 83% explosion compared with H1 2025. The sector now accounts for nearly a third of all job losses globally this year, with the United States the absolute epicentre at 121,072 cuts.

The corporate thesis underneath has shifted too. In late 2024 and early 2025, companies reached for AI-transition buzzwords to mask revenue stagnation, macro headwinds and pandemic-era over-hiring. Entering Q3 2026, the market has seen through the cover-up. According to Challenger, Gray & Christmas outplacement data, AI was the single most-cited reason for technology layoffs in May and June, the third consecutive month it has topped the list. This is the death of AI laundering: the stated reason and the real reason have converged.

And the trade is explicit. Highly profitable conglomerates are posting record revenues while executing a capital transfusion: liquidating payroll lines to fund infrastructure capex, custom silicon orders and hyper-scale cluster deployments. The impact is no longer confined to support and operational roles. Senior engineering teams, middle management and entire product divisions are being systematically dismantled.

Microsoft and Xbox: the cloud capex squeeze (6,400 total cuts). The single most disruptive signal entering the new fiscal quarter. Timed precisely to the July fiscal year-end, Microsoft eliminated 4,800 roles across commercial enterprise sales, cloud consulting and customer success. Simultaneously, an internal memo from newly appointed Xbox chief Asha Sharma forced an immediate 1,600-person reduction, 20% of the gaming workforce, to correct a business model the memo described as “not healthy”, with a further 1,600 roles scheduled to bleed out over the remainder of FY27. The WARN receipts cleared late this week and prove the product developers bore the brunt: 213 at ZeniMax Online Studios in Maryland, 136 at id Software in Texas, 52 at Obsidian Entertainment in California, falling heavily on QA, asset design and usability engineering. Chief people officer Amy Coleman insisted the roles are not being directly replaced by AI; the liquidation nonetheless serves as margin defence for Microsoft’s $190 billion fiscal 2026 AI infrastructure and data centre capex projection.

Oracle: the H1 restructuring leader (25,254 cuts YTD). Analytical reviews published this week, together with public SEC 10-K disclosures, lock down Oracle’s silent rolling purge as the largest single tech-reduction initiative of the year. The regulatory filings say the quiet part in print: the adoption and deployment of AI across internal operations has resulted, and will continue to result, in workforce reductions. The human liquidation ran concurrently with an expanding AI data centre footprint, boosting net quarterly margins while the firm reports massive cloud backlogs.

High-probability targets through mid-July

Watch mid-market enterprise sales and customer success. Microsoft’s hollowing out of their commercial divisions puts institutional pressure on peer B2B software and cloud-layer SaaS firms to match those operating margins, and as platforms shift to product-led growth engines and agentic onboarding bots, large human account-management tiers face sudden, localised flushes this month. And watch AAA gaming infrastructure and co-development studios: the multi-state cuts across Microsoft’s premium IP expose the operational strain in the sector, and parent holding companies will move to defensive restructurings across independent studios and supporting QA and art pipelines ahead of Q3 earnings calls.

Data sources: Layoffs.fyi, LayoffHedge.com, SkillSyncer, state WARN systems, SEC 10-K and 8-K disclosures. For informational purposes only.

Subscribe now

Final Thought 🚀

Hold this week’s two stories side by side. In one, the labs shipped four major models in three days and pushed the price of machine cognition sharply lower. In the other, the most profitable companies in history cut thousands of jobs and called it transformation. Nobody can prove a one-to-one causal chain between a benchmark score and a redundancy notice, and I won’t pretend to. But the direction of the capital is not in dispute.

Every gain on SWE-Bench Pro strengthens the business case for moving work from payroll to capex. The companies executing these cuts are not in distress; they are in a hurry. And the workers being shown the door are not being replaced by nothing. They are being replaced, in part, by the very models whose scores opened this issue.

The agentic turn is real, the efficiency gains are real, and, as my afternoon in Boston shows, the technology in your pocket is a marvel. But let’s not launder the accounting. Record revenues, record capex, record layoffs, all in the same quarter, from the same companies. That is not a correction. That is a choice.

Until next Sunday,

David

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Subscribe now

The Counterfeit Mind

David Richards MBE — Sun, 05 Jul 2026 06:01:29 GMT

THE SUNDAY SIGNAL · Issue #61 · Week 27 · Sunday 5 July 2026

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Subscribe now

Bottom Line Upfront

The most valuable thing America has ever built is being copied wholesale, and the law cannot stop it. Anthropic has told the United States Senate that operators tied to Alibaba ran the largest AI extraction it has ever caught, treating Claude as a teacher and cloning how it thinks. The same company, in the same fortnight, watched Washington restore its best models on terms Washington set, after switching them off with no warning. Underneath both stories sits a colder number. The price the market will pay for frontier AI is falling, and the entire trillion-dollar build-out is betting that it holds. And in Friday’s Yorkshire Post I wrote about a man with a red flag, and what Victorian England can teach a regulator now reaching for a kill switch. Four signals. One question. What is a mind actually worth, and who gets to hold it?

America Built the Mind. China Is Copying It in Plain Sight.

Times Square, last week, during the World Cup. I could have walked out with any country’s shirt in my size. England, Brazil, Argentina, Morocco, all hanging on the same rail, all for the price of lunch. An official replica from the club shop runs £55 to £80. The player-issue shirt the professionals actually wear costs more again. The version on that rail was £15, maybe £20, and from three feet away you could not tell the difference.

Subscribe now

So which is it. A cheap imitation, or the same shirt with the label swapped?

The honest answer is uncomfortable. A counterfeit and a licensed replica often pour out of the same manufacturing clusters in southern China, sometimes off the same lines on an unofficial shift, using the same fabric rolls the brand ordered. The stitching is worse. The badge is a shade off. The materials are cheaper. But the gap between the £15 fake and the £75 real thing is not the gap you would expect from the price. Roughly one in seven young people in Europe buys counterfeit sportswear, and most of them know exactly what they are buying. The brand is not selling cotton. It is selling the badge, the story and the right to wear it. The counterfeiter steals all three and pays for none of it.

Shirts are one thing. Late last month it emerged that Anthropic has accused China of doing the identical thing to Claude.

Anthropic Has Blown the Whistle Twice This Year

On 10 June, Anthropic sent a letter to the United States Senate Banking Committee, addressed to chairman Tim Scott and ranking member Elizabeth Warren. It landed alongside a committee hearing on American AI. The letter, first reported by Bloomberg on 24 June, alleged that operators affiliated with Alibaba and its Qwen AI lab had run what Anthropic called the largest known distillation attack on Claude to date.

The scale is the story. Between 22 April and 5 June, the letter says, roughly 25,000 fraudulent accounts generated more than 28.8 million exchanges with Claude, aimed squarely at its most valuable capabilities: software engineering, agentic reasoning and long-horizon planning. That is the same class of capability that sits inside Mythos, the frontier model Washington forced offline weeks later. Anthropic’s head of policy, Sarah Heck, told senators the operation was designed to “harvest US AI capabilities” and repackage them as China’s own.

This was not the first alarm. In February, Anthropic named three Chinese labs, DeepSeek, Moonshot AI and MiniMax, running the same playbook to a combined total of around 16.5 million exchanges across roughly 24,000 accounts. Alibaba, on its own, in six weeks, beat that combined figure by nearly two to one. It is also the first time Anthropic has named a global conglomerate rather than a start-up.

One caveat matters, and I will not bury it. These are Anthropic’s allegations, not proven facts. Alibaba denies wrongdoing and says it does not train on other companies’ model outputs. The figures have not been independently verified. But the technique Anthropic describes is real, it is documented, and it is almost impossible to stop.

This Is Not a Hack. It Is an Interview.

Forget the Hollywood image of hooded figures cracking a server. Nobody broke in. Nobody stole the source code or the model weights. The attackers simply talked to Claude, millions of times, and wrote down what it said.

Here is the anatomy, because it explains why the law is helpless.

The type of prompt hackers used to scrape claude (note I did not execute this!)

First, the disguise. Frontier models geoblock traffic from China to comply with US export controls. So the operators route through commercial proxy networks and VPNs, rotating IP addresses until the requests appear to come from London, Tokyo or California. Tens of thousands of accounts spread the load. No single account queries fast enough to trip an alarm. Spread 28.8 million exchanges across 25,000 accounts over six weeks and each one looks like an ordinary developer having an ordinary week.

Then, the interview. The attackers do not ask Claude for trivia. They hammer it with the hardest problems they can find, and they force it to show its working. A representative prompt reads like a job interview for a principal engineer: adopt an expert persona, solve a genuinely difficult problem, then, before giving the answer, lay out every architectural choice, evaluate the alternatives you rejected, and explain why. Wrap the reasoning in tags so it can be parsed cleanly. Only then, produce the code.

That instruction to “show your work” is the whole heist. A single query costs pennies in API credits. The output is a pristine, step-by-step map of how a frontier model actually thinks through a hard problem. It is worth a fortune. Multiply it by 28.8 million and you have not bought an answer. You have bought an education.

Subscribe now

The Apprentice Who Never Read the Recipe

What happens to those millions of transcripts is called model distillation, and the cleanest way to picture it is a master chef and an apprentice.

The apprentice never sees the recipe book. The weights stay locked. Instead the apprentice stands at the master’s shoulder for millions of meals, watching every knife stroke, every correction, every decision, until they can reproduce the dish without ever knowing the theory behind it.

In practice the lab scrubs the transcripts first, stripping out every “I’m Claude, made by Anthropic” and every safety refusal, until it has millions of clean instruction-and-response pairs. It takes a weaker student model, often built on an open-weights base, and fine-tunes it on that data. The student is mathematically pushed, word by word, to reproduce the teacher’s output. Because the attackers specifically captured the reasoning, not just the answers, the student learns the process: to pause, weigh options and structure its thinking before it speaks. A polishing pass tidies the result.

At the end, the lab has a model perhaps a tenth of the size, costing a fraction to run, that mimics the teacher on exactly the tasks that matter most. Training the original teacher costs hundreds of millions of dollars, months of compute and years of human trial and error. Training the copy costs a few million and a few weeks. This is why export controls barely bite. You do not need to smuggle the model out of the country if you can telephone it and take notes.

So, back to the shirt. Is the copy the same as the real thing, or is it something far worse? A fake shirt is cotton and a stolen badge. It costs the maker a sale. What Anthropic describes is different in kind. The counterfeiter here did not copy the product. It copied the mind that makes the product, for the price of talking to it. That is not a knock-off. That is the whole factory, walking out through the front door, one polite conversation at a time.

Subscribe now

Washington Switched the Models Back On. On Its Own Terms.

Two weeks ago in this newsletter I made a simple argument. On 12 June the US Commerce Department, using export-control powers, forced Anthropic to switch off its two most capable models, Fable 5 and Mythos 5, for every customer on Earth in a matter of hours. No missile. A letter. My concern was not the safety review. It was the precedent. Washington had quietly reclassified frontier AI as a munition, something closer to enriched uranium than to enterprise software. And any business, or any country, running on a single provider had just watched a machine it depended on get switched off by a government it does not vote for. Britain, I wrote, is running on borrowed machines.

This week the machines came back on. It is worth watching how.

Commerce lifted the export controls on 30 June. Fable 5 returned worldwide on 1 July, across Claude’s apps, platform and coding tools, after a 19-day global blackout. The trigger for the whole episode, it turns out, was a jailbreak: Amazon researchers found a way to make Fable 5 flag software flaws and, in one case, write code showing how one could be abused. Anthropic downplayed it as a handful of minor, already-known vulnerabilities, retrained its safety classifier, and says the technique is now blocked more than 99% of the time.

Mythos 5, the more powerful sibling with fewer guardrails, is not back for everyone. Access returned on 26 June to roughly 100 vetted US organisations that defend critical infrastructure, under Anthropic’s Project Glasswing. Washington still decides who is trusted. The negotiations were reportedly led not by chief executive Dario Amodei, who has clashed with the administration all year, but by co-founder Tom Brown. As part of the deal Anthropic agreed to hunt its own models for security holes, coordinate future launches with the government and report malicious use it detects.

Read that carefully, because it is the real news. The crisis is over. The precedent is now permanent. A frontier model was switched off by fiat and switched back on by negotiation, and along the way the US government established that it gets an early look at the most capable models before the public does. OpenAI previewed its latest system to a small, government-approved group days earlier, citing the same dual-use worry. The question one analyst put to Al Jazeera still stands: does Washington now approve every frontier model release? On the evidence of the last three weeks, the answer is drifting towards yes.

For anyone outside the American tent, the lesson from a fortnight ago has only hardened. The models you rent can be revoked. The kill switch exists, it has been tested in public, and it belongs to someone else.

Subscribe now

The $700 Billion Question Nobody Wants to Ask

Here is the number that should be keeping data-centre investors awake. It is not on any earnings call. It is a price, and it is falling.

The Silicon Data LLM Token Expenditure Index tracks what the market actually pays for AI, per million tokens. Between its December launch and May it nearly doubled. Since then it has fallen almost 20%. It is the cleanest read anyone has on whether customers will keep paying premium prices for frontier intelligence, and right now it is flashing amber. As one veteran investor put it this week, users priced by the token are being forced to rein in unlimited use because the bills have grown too large.

Set that against the spending. Amazon, Microsoft, Alphabet and Meta have guided roughly $700 billion of capital expenditure this year, nearly double last year, most of it aimed at AI data centres and custom silicon. The march towards a trillion dollars in 2027 is already under way. That build-out is justified almost entirely by the belief that token spending keeps climbing. If the price the market pays softens while the concrete keeps pouring, the sums stop working.

The historical rhyme is unkind. Allianz Research puts the gap between AI investment and AI sales at nearly 46%. During the 2001 telecom bust, when companies laid fibre-optic cable for demand that arrived a decade late, the equivalent divergence was 32%. We are already past it. The one comfort for the bulls is that today’s hyperscalers have fortress balance sheets the fibre pioneers never had. The discomfort is that their share prices are priced for perfection, and perfection is a fragile thing to sell.

There is a subtler reading, and it is the one I hold. A softening index does not mean AI is failing. It means AI is commoditising. When a Chinese lab can distil a frontier model for a few million dollars, as the first story in this issue describes, the pricing power that justified today’s valuations starts to leak away. Enterprises are quietly trading the most expensive frontier models for cheaper ones that do the same job. The era of unmetered developer spending, of treating tokens as free, is giving way to the corporate bean-counter with a budget and a red pen.

The technology is not dead. The unconstrained pricing power that funded the boom might be. Somebody has to pay for all these data centres. This is the week the market started asking who.

Subscribe now

The Man With the Red Flag

This week’s Yorkshire Post column.

In 1865, before the motor car had properly arrived, Parliament decided the self-propelled road vehicle could not be trusted on its own. The Locomotives Act required a crew of three: a driver, a stoker, and a man on foot walking 60 yards ahead with a red flag by day and a lantern by night, warning the public that a machine was coming. The vehicle behind him could move at two miles an hour in town.

It was not wholly irrational. Horses bolted, roads were bad, steam engines were loud. But the effect was absurd. While Benz, Daimler and the American pioneers got on with making the car work, Britain made the new machine walk behind a pedestrian. The flag went in 1878. The escort survived until 1896. For 30 years, a country proud of its engineering slowed the future to walking pace.

I thought of that man this week reading Sarah Breeden at Sintra. Breeden, the Bank of England’s deputy governor, was warning about autonomous AI agents in finance: systems that do not just produce analysis but act on it. In markets, she cautioned, they could amplify stress, not because one model goes berserk, but because many models do the same sensible-looking thing at the same moment. Same data, same signal, same exit. The Bank is examining guardrails: circuit breakers and kill switches that could halt trading if the machines misfire. More than half of finance firms already run these agents.

That is a serious question, and she is right to ask it. A kill switch on a runaway market is not a red flag on a road. The trouble starts when regulators forget which one they are holding.

There is an old official habit of making the unfamiliar move slowly until important people feel calmer. It looks like prudence. Often it is only a lack of understanding dressed up as caution. When margarine threatened butter, some American states forced it to be dyed pink. In 1994, Westminster wrote “repetitive beats” into criminal law to frighten off rave culture, and Autechre answered with a track arranged so no bar repeated. In 2019, OpenAI declared its own GPT-2 too dangerous to release. It had 1.5 billion parameters. The models that now help with homework and medical results are a thousand times larger. The feared monster became a useful tool.

So yes, regulate AI in finance. Test the models, demand audit trails, establish liability, make firms prove an agent can be stopped and overruled. But do not confuse slowing a technology with understanding it. Britain still marks the end of the red flag: at the first Emancipation Run in 1896, a motorist tore one in half and drove for Brighton, as the cars have done every November since. A serious country must know the difference between a brake and a flag. The brake stops a crash. The flag stops the journey.

The Tech & AI Layoff Tracker

Week 27 · 28 June to 4 July 2026

The first half of 2026 has closed, and the mid-year autopsy is brutal. According to a report published this week by Challenger, Gray & Christmas, US technology employers announced 139,156 job cuts in the first six months of the year, up 83% on the same period in 2025. Tech now accounts for nearly a third of every layoff announced in the country. For the fourth consecutive month, a streak with no precedent in the firm’s data, artificial intelligence was the single most cited reason for job cuts. AI has been named in 101,743 layoffs so far this year, roughly 23% of the total.

The narrative has shifted. Last year companies hid behind AI buzzwords to dress up ordinary cost-cutting. This year the cuts are real, and so is the mechanism. As Andy Challenger put it, the sector is “being reshaped in real time.” Wildly profitable firms are liquidating human headcount to fund silicon.

British American Tobacco, ~9,000 roles (about a fifth of staff). The single largest cut of the week, and it is not a tech company. BAT will eliminate 5,500 jobs outright and outsource a further 3,500, many to Accenture, to save £600 million a year by 2028. Chief executive Tadeu Marroco said the overhaul makes the company “more agile, cost-disciplined and technology-enabled.” Management pointed directly at AI and data analytics reshaping its staffing needs. When a cigarette maker starts citing agentic AI in its restructuring, the displacement has left the tech sector and entered the wider economy.

Microsoft, fewer than 5,500 roles (under 2.5% of staff). Expected to be confirmed in the coming week, timed to the close of Microsoft’s fiscal year on 30 June. The cuts target enterprise sales, consulting and the Xbox division, now under a “reset” led by new gaming chief Asha Sharma. The context is the arms race: Microsoft has committed around $190 billion to AI since 2025 and its capital spending hit $80 billion in the first nine months of this fiscal year alone. The market is forcing the company to choose compute over headcount, and it is choosing compute.

Elementor, ~100 roles (30% of staff). The website builder that powers roughly 13% of all websites is cutting nearly a third of its workforce. Chief executive Yoni Luksenberg blamed the speed of AI disruption directly. With agentic systems now spinning up and deploying entire websites on command, demand for human-operated layout software is collapsing. Elementor is the canary for mid-market software built on the idea that a human sits at the wheel.

What to watch in early Q3. Enterprise sales and customer success teams, as automated product-led growth and AI onboarding agents make large, expensive human sales floors an obvious target. And legacy design and middleware software, where any product whose core function is a dashboard for humans to drag and drop is facing an existential choice: build native AI generation, or become irrelevant.

Subscribe now

Final Thought

Every story this week is the same story wearing a different coat. What is a mind worth, and who owns it?

The Chinese labs answered by stealing one for the price of a phone bill. Washington answered by declaring it a weapon and keeping the key. The market answered by quietly marking the price down. And the man with the red flag answered, more than a century ago, that a nervous state will always reach for something old when it meets something new.

Here is the part nobody in the boardroom wants to hear. There is growing evidence that a great deal of the “AI” in these layoffs is a story told to shareholders. A National Bureau of Economic Research working paper found that 90% of executives say AI has had no measurable effect on employment at their own companies. A study of firms cutting the hardest found no improvement in their financial returns. The badge, in other words, is doing a lot of work the cotton cannot.

So we are copying minds we cannot fully value, switching off machines we cannot fully control, and firing people to pay for a promise we have not yet proven. The fake shirt at least keeps you warm. I am not sure the same can be said for a strategy built on a badge.

🚀

Until next Sunday,

David

The $60 billion company that broke the rulebook

David Richards MBE — Sun, 28 Jun 2026 06:01:50 GMT

THE SUNDAY SIGNAL · Issue #60 · Week 26 · Sunday 28 June 2026

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Subscribe now

Bottom Line Upfront

For half a century we judged a company by three things, in order. How many people it employed. How much revenue it earned. Whether it would ever turn a profit. Cursor has torn the first two pages out of that book, and SpaceX paid sixty billion dollars to hold what was left. This week California started counting the human cost of exactly this shift. And in Friday’s Yorkshire Post I argued that the boardrooms now panicking about the price of artificial intelligence have misread the bill entirely. Three signals. One direction of travel.

PHOTO BY ANDRIA LO

We Used to Count People. Now People Are a Rounding Error.

SpaceX has agreed to buy Anysphere, the company behind the coding tool Cursor, in an all-stock deal worth sixty billion dollars. The announcement landed on 16 June, four days after SpaceX completed the largest stock market debut in history and floated at more than two trillion dollars. Elon Musk used the fresh currency at once. He merged his artificial intelligence venture, xAI, into SpaceX back in February. Cursor is the bolt-on that gives it a real business.

Sixty billion is a serious number. It is not the interesting one.

Anysphere does not publish its headcount. It guards the figure on purpose. The credible estimates cluster around three hundred people, with one respected data provider putting it nearer four hundred. Take the lower number, the one the company’s own admirers tend to use, and the maths stops making sense in the old language. Sixty billion dollars divided by three hundred staff is two hundred million dollars of value for every single employee. Not revenue. Value. The price one of the most aggressive acquirers on earth agreed to pay, per head.

Now set that beside the rulebook we have used since Marc Andreessen told us software was eating the world.

Read across the bottom row. A mature cloud company carries between two and five million dollars of value for each person on its payroll. That used to look like the high-water mark of software leverage. Cursor sits at roughly forty to a hundred times that figure. Read up the column. The company reached a billion dollars of annual recurring revenue in about three years. Salesforce took more than a decade. And the revenue density is the part that should keep incumbents awake. Cursor’s annualised revenue had climbed to around four billion dollars by early June, with roughly 2.6 billion of that coming from enterprise customers. Divide even the conservative enterprise figure by the staff and you land near nine million dollars of sales a head. The average software company manages two hundred thousand.

So the old test fails. We counted people, because people were the engine and the cost. Here the people are neither. They are the rounding error.

Subscribe now

They Did Not Cut Costs. They Relocated Them.

The lazy reading is that Cursor proves software companies no longer need employees. That is wrong, and the error matters, because it leads boards to the wrong conclusion about their own businesses.

PHOTO BY ANDRIA LO

Cursor did four things, and only one of them was about being small.

First, it refused to build what already existed. Rather than write a code editor from scratch, the founders forked Microsoft’s open-source VS Code. They inherited a mature product that millions of developers already knew, then poured every engineering hour into the one layer that mattered: the artificial intelligence sitting inside it. No learning curve for the user. No wasted years on plumbing.

Second, it let the product do the selling. For roughly two years Cursor ran almost no enterprise sales operation. Individual developers downloaded it, found they could work several times faster, and pushed their own employers to buy licences. The company reportedly passed two hundred million dollars of revenue before it hired its first enterprise sales representative. That is the inversion of the classic playbook, where you build a large sales machine first and hope revenue follows.

Third, it paid for density rather than scale. Engineers at Cursor reportedly earn total packages between eight hundred thousand and 1.1 million dollars. The team stayed deliberately tiny and ferociously selective, and it used its own tool to build the tool. A small group of elite people, each amplified by the very product they were shipping.

And fourth, here is the part the celebration leaves out. Cursor did not escape the cost base. It moved it. The bills that a normal company pays as salaries, Cursor pays as compute. Its product runs on frontier models, and for a long stretch a great deal of its revenue flowed straight back out to model providers, Anthropic chief among them. Reporting through 2025 suggested the AI coding companies were running negative gross margins at points, which is why Cursor began training its own models to claw that money back. The payroll was small. The total cost of intelligence was enormous, and it sat in graphics chips and inference, not in offices and headcount.

That single fact is the spine of the whole story. Cursor looked weightless. It was not. It had simply swapped one bottleneck for another.

SpaceX Did Not Buy a Code Editor. It Bought the Map.

Which explains the buyer.

Cursor’s weakness was the mirror image of its strength. It had built the best application in the category on top of models and compute it did not control. Anthropic, OpenAI and Google were all at once its suppliers, its partners and its rivals. As coding assistants became strategic, that dependence became a noose.

SpaceX could cut it. The deal hands Cursor access to serious compute, the constraint that had capped its economics. In return Musk’s empire gets three things money alone cannot buy. It gets distribution to the most valuable users in technology, the engineers who build everything else. It gets a real, fast-growing artificial intelligence business at a moment when Grok has struggled to compete. And it gets the live telemetry of how the world’s best developers actually write software, the highest-grade training data imaginable for the next generation of models. SpaceX did not pay sixty billion dollars for a text editor. It paid for a map of how software gets made, and a captive channel to ship its own intelligence down.

The all-stock structure made it almost painless. A company freshly valued above two trillion dollars can buy a sixty-billion-dollar asset with paper, not cash, and barely feel the dilution. The IPO was the loaded gun. Cursor was the first thing it fired at.

Subscribe now

Capital Will Not Retreat. It Will Multiply.

So here is the question every founder and every investor in this region should be asking. If a company can reach billions in revenue with three hundred people, does it still need the mountains of venture capital we used to shovel into start-ups?

The honest answer is no. And that should frighten the venture industry far more than it currently does.

Venture capital is not built to fund lean companies. It is built to deploy. A fund raises a billion dollars, and it is structurally obliged to put that billion to work, whether the companies in front of it need it or not. The entire model assumes capital is the scarce ingredient and the thing you sell founders. When the cost of building collapses, that assumption inverts. The scarce ingredient is no longer money. It is the rare team that can wield the new tools.

At Yorkshire AI Labs our working view is blunt. The companies we back now need somewhere between 10% and 20% of the capital they would have needed before artificial intelligence. A team that once required five million dollars to reach product can get there on a few hundred thousand. That does not mean capital stops flowing. It means it spreads. We expect to do five to ten times as many deals, not fewer, because the barrier to starting something real has fallen through the floor. Fewer dollars per company. Far more companies. A wider, flatter, more crowded field of founders, each able to go further on less.

There is a counter-argument, and it deserves airing. Some will say the capital does not vanish, it simply migrates to the one place that is now more expensive than ever, which is compute. Owning the models and the chips is brutally capital-intensive, as the SpaceX deal itself demonstrates. Both things can be true. Building an application has never been cheaper. Building the intelligence underneath it has never been dearer. The money does not leave. It bifurcates. A flood of small bets on lean builders at the top, and a handful of colossal bets on infrastructure at the bottom. The comfortable middle, the ten-million-dollar round for a forty-person company that needs eighteen months to ship, is the thing that dies.

Subscribe now

California Always Sees It First.

Venture capital will recalibrate for this leaner reality. The human cost of that efficiency lands on people, not balance sheets, and this week California started counting it.

In March, in Issue #45, I launched a permanent section of this newsletter: the Tech and AI Layoff Tracker. I had been logging the data since February. I started it for a simple reason. I could see what was coming, and I would rather tell the truth about it than join the chorus pretending artificial intelligence merely “augments” people. My honest expectation remains that more than 40% of white-collar work, as we currently define it, is exposed to elimination over the coming years. That is a forecast, not a fact, and I have always flagged it as such. But I would rather count the change as it happens than be surprised by it.

This week the state of California started counting too.

On 25 June, Governor Gavin Newsom launched the California AI-Unemployment Tracker, the first official dashboard of its kind in the United States. Built with the California Policy Lab at UCLA and the state’s Employment Development Department, it does not wait for companies to confess. It cross-references unemployment claims against an “AI exposure” score for each worker’s previous job, asking in effect whether a machine can now do half of what that role required. It updates monthly. The raw data is public.

Read the findings carefully, because the honest version is more useful than the panicked one. The tracker found no statewide surge in artificial-intelligence-driven layoffs through May. There is no mass unemployment event in the numbers yet. But the localised signal is real and it is unmistakable. Claims from college-educated workers in highly exposed occupations have climbed and stayed elevated since ChatGPT arrived in late 2022. The San Francisco Bay Area and the technology sector show it most sharply. One of the lab’s senior researchers put it plainly: they are not yet seeing evidence of large-scale layoffs. Not yet. The dashboard exists precisely so that “not yet” does not become “too late”.

This is why California matters beyond California. The state is the world’s regulatory canary. What it does first, the rest of the democratic world tends to do next, usually within a few years.

The pattern is long. California set vehicle emissions standards in the 1960s that Washington and then the wider world were eventually forced to follow. It banned smoking in bars and workplaces while the rest of America still lit up indoors. Its 2018 privacy law, the CCPA, imported a GDPR-style privacy logic into the United States and forced companies that had never thought of themselves as Californian to rebuild how they handled data. And in 2023 it became the first American state to issue a serious executive order on generative artificial intelligence, then the first to legislate on transparency. Each time, the same shape. California moves. The friction looks local. Then it goes global.

A government dashboard counting the displaced is not a footnote. It is the empirical foundation on which the next decade of policy gets built. Once a state can prove, with its own numbers, that a specific deployment of software put specific people out of work, the political distance to an automation tax, to mandatory disclosure of algorithmic layoffs, to retraining levies, shrinks dramatically. The European Union is already watching the methodology. California has just handed every regulator on earth the instrument they were missing.

Subscribe now

Rip It Up and Start Again.

This was the argument of my column in Friday’s Yorkshire Post, and it belongs here too, because it answers the question the Cursor story raises for every leader reading this.

An artificial intelligence consultant told Axios last month that one of their clients had run up half a billion dollars on Anthropic’s Claude in a single month. Five hundred million. The figure has rattled boardrooms. Microsoft has been cancelling most of its Claude Code licences. Uber said it had burned through its 2026 AI budget by April. Amazon scrapped an internal leaderboard that ranked staff by how many tokens they had burned, a game that rewarded activity over output. The worry doing the rounds is simple. Nobody seems sure what this stuff actually costs.

Here is what the panic misses. Bills like that rarely come from clever work. They come from pointing a powerful tool at a hopeless job. And the most hopeless job in software is renovating the old.

A few years ago we rebuilt a house in California, and we did it in two halves. For the first, we tried to work with what was already there. We traced the wiring, worked out where the plumbing ran, knocked down some walls and moved others, and found a fresh surprise behind every one. It ran badly over budget and far over time. For the second half, we stopped being sentimental and knocked it down in a day. Lightly built comes down fast and goes up fast. The second half cost a fraction of the first, in a fraction of the time.

Software is no different. Legacy code is a house nobody kept the plans for. Twenty years of decisions, fixes and shortcuts, mostly undocumented, all load-bearing in ways you discover only when something collapses. Ask the machine to refactor it and it must first read the whole tangle, hold it in its head, and tiptoe through dependencies it cannot fully see. It is slow. It is dear. And it still breaks the thing on the far side of the building that nobody remembered was wired in. That is how your half a billion disappears.

Ask the same machine to build afresh and everything changes. It lays out a clean structure it understands, because it wrote it. Work that would have meant weeks of archaeology takes an afternoon. I have watched it close up. Not long ago someone on our team sat down with a would-be founder for a single conversation. We recorded the meeting, fed the transcript into Claude, and got back a business plan. We passed the plan to Claude Code, which produced a working prototype and a website. From the end of the meeting to something a customer could click on took about an hour. A year of my early Silicon Valley life, compressed into a lunch break.

This is not a rule for every case. Some old systems carry knowledge no one ever wrote down. Some must not be touched at all. But the instinct to preserve, to patch and prop up the thing you already own, is more often than not the costly mistake. Orange Juice had it right in 1983. The only top-ten hit the Glasgow band ever managed told you exactly what to do with the thing that no longer works. Rip it up and start again.

So before you approve another budget to nurse a system built for a vanished age, ask the harder question. Are you renovating, or are you simply afraid to let go?

Share The Sunday Signal

The Sunday Signal Tech and AI Layoff Tracker

Week 26. The quarter closed hard.

As the Q2 reporting window slammed shut, the language shifted from “AI modernisation” to raw capital preservation. The week’s near-3,000 cuts ran well beyond software, into electric vehicles, legacy media, retail tech and gaming. The dominant excuse was what leadership teams are now selling to Wall Street as the “agility pivot”: strip out middle management and legacy roles, call it the removal of friction, defend the margin.

Lucid Motors. ~1,500 cuts. NEW. The luxury electric carmaker filed an 8-K on 22 June confirming an 18% reduction of its US workforce, across full-time, contractor and hourly production layers. New chief executive Silvio Napoli, three weeks into the job, is dismantling overhead to secure roughly 158 million dollars in annualised savings ahead of the Cosmos SUV. The restructuring abolished the chief operating officer role entirely and ended the second production shift at the Casa Grande, Arizona plant. A pure cash-burn intervention from a company that lost 2.7 billion dollars last year. Market: NASDAQ.

iHeartMedia. ~500+ cuts. NEW. The audio giant ran a nationwide purge of regional programming directors and local on-air talent to unlock about 150 million dollars in savings, crediting a pivot to a “technology-driven” programming model. Automated curation replacing human-curated local radio. Software flattening reaching into the creative trades. Market: NASDAQ.

Walmart Tech. 306 cuts. NEW. The retailer filed California WARN notices eliminating 306 corporate technology and product roles in Sunnyvale, citing overlapping responsibilities and a leaner structure to “accelerate AI product execution”. Market: NYSE.

Bungie. 292 cuts. Developing. Sony-owned Bungie registered deep cuts at its Bellevue headquarters, hitting engineering, design and QA across the Destiny and Marathon teams. Margin compression in high-budget gaming, with the parent stripping studio overhead. Market: TYO (Sony).

Sonos. ~3% cut. Developing. The audio manufacturer trimmed roughly 3% of staff across UX, product and design. Chief executive Tom Conrad framed it squarely in the agility playbook: remove layers, prototype faster, move past the long tail of the 2024 app debacle. Market: NASDAQ.

Q3 horizon. Lucid sets a grim precedent for the EV supply chain; peers facing the same cooling demand will be pressed to match the cut early in the next quarter. Consumer-facing media and tech are running out of administrative fat, which means the next reductions land on core product, editorial and UX, the builder layers themselves.

Subscribe now

Sunday Signal Merch is Here

A small word before I let you go. The Sunday Signal now has a shop, live at shop.thesundaysignal.ai. Shirts and a few other things, in black or white, carrying the gold wave.

Use discount code SundaySignal5 for 5% off at the checkout.

Shop Now

Final Thought

🚀 The rules did not erode this quarter. They broke, in public, all at once. A company of three hundred people changed hands for sixty billion dollars. A state government began counting the unemployed by how exposed their old job was to a machine. And a half-billion-dollar bill landed on a boardroom for the crime of trying to save the past. Three signals, one bearing. The leaders who build for the world that is arriving will own it. The ones still defending the world that is leaving will keep paying for the privilege, by the token, until there is nothing left to defend.

Until next Sunday,

David Richards MBE

Subscribe now

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

The Silicon Curtain

David Richards MBE — Sun, 21 Jun 2026 05:15:22 GMT

THE SUNDAY SIGNAL · Issue #59 · Week 25 · Sunday 21 June 2026

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Subscribe now

Bottom Line Upfront: A new cold war has begun, and its first weapon is artificial intelligence that can be granted, revoked or weaponised at the stroke of a pen, splitting the world into rival blocs as surely as the iron curtain once split a continent. Sovereignty, whether national, corporate or personal, is no longer a posture for the paranoid; it is the price of admission to the side you actually want to be on.

The first shot of the new cold war was a letter, not a missile

It did not look like an act of war. There was no missile, no breach, no headline until the weekend was nearly spent. There was a letter.

At 5.21pm Eastern on Friday 12 June, the US Commerce Secretary, Howard Lutnick, wrote to Anthropic’s chief executive, Dario Amodei. The letter placed the company’s two newest models, Fable 5 and Mythos 5, under export controls and barred their use by any foreign national, anywhere, including Anthropic’s own foreign-born engineers sitting at desks in San Francisco. Because you cannot quietly geofence half your own staff, the company did the only thing it could to comply. By midnight it had disabled both models for every customer on the planet. They had been public for three days.

Read that timeline twice. The most advanced reasoning system the company had ever released to the world was switched off, globally, three days after launch, by a letter that gave no detailed national security rationale and has still not been offcially published. This is what the weaponisation of cloud AI actually looks like. Not a dramatic flourish, but an administrative one.

The legal mechanism matters, because it is the part nobody outside the trade understood until last week. Lutnick reached for the “deemed export” rule, the clause that treats handing controlled technology to a foreign national on US soil as if you had shipped it abroad. In one stroke, the reasoning inside a chatbot was reclassified as a controlled dual-use item, the regulatory cousin of a centrifuge component or a guidance chip. AI as munition is no longer a metaphor for a conference panel. It is now a category in American export law.

What spooked Washington? Reporting suggests the trigger was a demonstration, flagged by a researcher at Amazon, in which Fable 5 could be asked to read a codebase and surface software vulnerabilities. The administration’s fear, on that account, was simple and seismic: that a product sold as a writing and coding assistant was in truth carrying offensive cyber-capability, a tool that could find the cracks in any system fast enough to matter. Anthropic dispute the severity in the bluntest terms. They say they reviewed the demonstration and found it had identified only a small number of minor, already-known flaws, and at least one outside security expert who read the underlying paper says the model in fact refused the task. So hold the claim carefully. The point is not that Fable 5 was a digital weapon. The point is that Washington decided to act as though it might be, and that the mere suspicion was enough to blind the whole world overnight.

Here is where the historians earn their keep. The export control on frontier AI is being sold as the non-proliferation treaty of our age, the 2026 answer to keeping the bomb out of the wrong hands. The analogy breaks on first contact. Refining uranium demands vast, hot, detectable infrastructure that satellites can photograph and inspectors can count. An AI model travels down an API in milliseconds, weightless and invisible, and a capable open-weights version can be copied to a laptop and run behind an air gap forever. Export controls work where there is a physical chokepoint. There is no chokepoint on a thought.

Subscribe now

And the blowback has already started. By proving that the best American models are politically fragile, that they can be revoked between a Friday and a Saturday on the strength of an unexplained letter, Washington have handed a gift to exactly the ecosystem they fear most. When you lock down the strongest closed models, global capital and enterprise demand do not evaporate. They pivot, instantly, to sanctions-proof open-weights alternatives, the Chinese GLM family chief among them. The US has not slowed the proliferation of capable AI. It has advertised the case for buying it from somewhere that cannot be switched off. This is how arms races actually escalate, not through one decisive weapon but through each side teaching the other that dependence is a vulnerability. Two blocs are now forming in plain sight, one built on American models that can be revoked and one on Chinese weights that cannot, and every enterprise on the planet is quietly being asked which side of the curtain it intends to sit on.

Now follow the panic out of Washington and into the boardroom, because this is where it stops being abstract. Within days, JPMorgan removed Anthropic’s Claude from the internal menu of approved models available to its staff in Hong Kong. They were following Goldman Sachs, who had quietly done the same back in April. The detail that matters, and that most of the coverage rushed past, is the trigger. These cuts were not ordered by Lutnick and they were not handed down by Beijing’s firewall. They flow from Anthropic’s own licensing terms, which exclude Greater China including Hong Kong, read in the most conservative way a nervous compliance team can read them. The export-control drama did not cause the Hong Kong blackout. It simply made every Wall Street lawyer reach for the strictest possible interpretation of a contract they already held.

That is more frightening, not less. It means the splinternet of global finance is forming through ordinary commercial paperwork, one cautious legal memo at a time. A junior analyst at JPMorgan in New York now has reasoning and coding tools that a vice-president in Hong Kong is explicitly forbidden to touch. The map of who may think with the best machines is being redrawn along the contours of geopolitical risk, a silicon curtain drawn not across a continent but straight through the org chart of a single bank, and it is the banks themselves wielding the pen.

The warning for London writes itself. If Wall Street firms will sever their own people in a first-tier Asian financial hub the moment a licence looks ambiguous, what exactly protects a London office should US relations with Europe sour? If your trading desk, your compliance engine or your defence contractor runs on a US-hosted API, your IP address has quietly become a sovereign liability. You are one letter away from the same midnight that hit Anthropic’s customers.

And this is the part that should keep Whitehall awake. UK signals intelligence, automated threat detection and a long tail of defence contractors lean on infrastructure that a US Commerce Secretary can lawfully sever overnight. Worse, the UK’s own AI Security Institute had spent thousands of hours red-teaming Fable 5 before launch, and the unilateral American ban side-stepped that validation entirely. Allied status, the famous special relationship, the Five Eyes intimacy we comfort ourselves with: none of it bought five minutes of warning or a single carve-out. The lesson is cold. When it comes to digital export controls, being America’s closest friend offers precisely zero immunity.

Subscribe now

So the story must move from what happened in Washington and Hong Kong to what Britain does on Monday morning. Three actions, in order of urgency.

The Treasury must build a digital deterrent. Subsidise domestic GPU clusters and onshore data centres at serious scale, funded in the next fiscal quarter and not deferred to the next parliament, and stand up a national effort, a Manhattan Project in spirit, to train sovereign foundation models explicitly for defence and critical infrastructure. You cannot be locked out of a model you host yourself.
The Foreign Office must negotiate a tech treaty. Use the leverage Britain actually has, inside Five Eyes and as a genuine safety partner, to open formal talks within ninety days for an ironclad, bilateral exemption from US Bureau of Industry and Security export controls. Goodwill is not a policy. A signed carve-out, secured before the next directive lands rather than after, is.
The FCA and PRA must mandate AI business continuity plans. Give systemically important financial and defence institutions twelve months to stand up air-gapped, open-weights technology stacks capable of failing over within 24 hours of an offshore shutdown. Not as best practice. As a licence condition.

There is a coda. By Friday, Reuters was reporting, on the strength of an interview Trump gave to Axios after meeting Amodei at the G7, that the president no longer viewed Anthropic as a national security threat and that talks had turned towards easing the restrictions. That softens the immediate crisis, but it does not erase the lesson, and it pays to be precise about what actually softened. The mood did, not the machinery. The 12 June directive was still in force as this issue went out, Trump would say only that he might lift it and was not sure he had to, and he pointedly declined to rule out emergency powers. Access was suspended, worldwide, because one directive made it so.

Anthropic calls the directive a misunderstanding and says they are working to restore access. Perhaps they will. But the misunderstanding has already taught the rest of us the only lesson that counts, and it is not one any restoration can unteach.

Resilience has replaced convenience as the first rule of enterprise AI

For three years the dominant enterprise architecture for AI has been a single, fatal shortcut: plug your application straight into one vendor’s API and ship. It was fast, it was cheap, and last Friday it was exposed as the corporate equivalent of keeping all your cash under one floorboard in someone else’s house.

The firms that will survive the next time an API goes dark are already executing a different plan. They are trading convenience for resilience, moving deliberately towards hybrid, vendor-agnostic and, where it matters most, air-gapped deployments. Here is the playbook the most serious institutions are running right now.

Phase one: abstract the architecture behind an AI gateway. The most expensive mistake a company can make is hard-coding its software to a specific model, so the first move is to decouple the two. An AI gateway is a middleware layer that sits between your applications and the models, which means your developers write to the gateway and never to the vendor. When a primary model is blocked, throttled or suddenly priced into the stratosphere, the gateway reroutes the request to a pre-approved fallback, often a locally hosted open-weights model, before a single user notices anything is wrong. Add unified telemetry on top, so that sensitive data is shaped, logged and routed before it ever reaches an external model, and you have closed the accidental door through which most corporate data actually leaks.

Phase two: own enough of the compute to mean it. Sovereignty is meaningless if you do not control the silicon, so the operating principle becomes hybrid by design. Classify your workloads honestly: low-risk, non-sensitive tasks such as drafting marketing copy can happily ride cheap public APIs, but core-business work, the algorithmic trading, the review of client legal contracts, the clinical decisions, must run on private or sovereign infrastructure. For banks, defence and national healthcare that means real investment in on-premise GPU clusters or hyperconverged kit, so that containerised workloads run entirely inside your own perimeter and survive any external blackout. And it means treating open-weights models as a strategic asset rather than a hobbyist’s toy, curating and fine-tuning the strongest of them, because you cannot be locked out of a model you host yourself.

Phase three: win the legal fight before it becomes a technical one. Geopolitical risk almost always arrives as a contract clause long before it arrives as an outage. Audit every AI vendor agreement for ironclad data portability, the guaranteed right to export your fine-tuning data, your prompts and your vector databases in open formats so you can migrate at speed rather than at the vendor’s mercy. Insist on jurisdictional ring-fencing that defines exactly where the compute happens and whose law governs it, so a UK institution is not silently exposed to US export controls it never agreed to. Then prove it works: run an AI business continuity plan with regular fire drills in which external access is deliberately cut and the failover to onshore, open-weights models is tested under real conditions.

None of this is free. The shift to a sovereign architecture is a capital cost paid upfront, in cash, before the crisis that justifies it arrives. But the alternative was demonstrated last Friday at one minute past five, and it has a price too. The companies that learn the lesson now will treat AI as a controlled asset. The ones that do not will keep treating it as a convenience, right up until the morning it is gone.

Subscribe now

The same machine that can be aimed at a teenager can also stand between him and the mob

Control of these models decides more than corporate survival; it decides our social reality, whose voice carries and who is shielded when the crowd turns. Which brings me, by way of a football match, to the human end of the same argument.

I am writing part of this from Boston, where for the last week the city has belonged to the Tartan Army. They were everywhere, and it was a fine thing to watch fans of both sides mixing in the bars before the game. At the match itself I had Morocco supporters to my immediate left and Scotland fans to my right, and my hearing has only just returned! Scotland lost, a tight match they will feel they could have nudged, but the day did not feel like a defeat.

There was so much negativity around this World Cup before a ball was kicked. Transport, organisation, stadium operations, the whole fan experience, all of it written off in advance. My day at France against Senegal could not have contradicted that more completely. Getting in was straightforward, the organisation was excellent, the fan zones were lively and welcoming, and the concessions actually moved. France won three-one with a Mbappé brace, but the lasting memory is not the football. It is that I got to share it with my son. Credit where it is due, and there is plenty due. England against Ghana is next.

Yet football’s passion has a dark twin, and we should be honest about it. When that passion curdles, it can become something genuinely vile, and we have the receipts. Cast back to the Euro 2020 final, played in July 2021, when England lost a shootout to Italy and three young Black players, Bukayo Saka, who was nineteen, Marcus Rashford and Jadon Sancho, missed their penalties. Within hours a torrent of racist abuse flooded their accounts, monkey and banana emojis by the thousand, slurs beyond printing.

The numbers are not estimates. The UK Football Policing Unit received more than 600 reports, judged 207 of the posts to be criminal, and found that 123 of those accounts were operating from outside the country, a reminder that the mob is global and the law mostly is not. Eleven people were arrested, and a string of convictions followed, some custodial, some suspended, some community service. A mural of Rashford in Withington was defaced within the day, then buried under hundreds of hearts and handwritten notes by his neighbours. The condemnation ran all the way up to the President of the Football Association, Prince William. None of it reached the players before the abuse did.

That gap, between the moment the abuse lands and the moment anyone official responds, is precisely what AI is now being built to close. FIFA is expanding its social media protection service at this tournament, and offering the moderation element free to every football association taking part, though it is worth noting the English FA have not yet confirmed they are using it. The mechanism is brutally simple. The system scans against some 30,000 keywords and hides an abusive comment in under two seconds. The sender still sees their own post and has no idea it has been buried and reported, which neatly removes the satisfaction that feeds the behaviour. It works across Meta’s platforms, YouTube, TikTok and Threads. It does not work on X, which has always let hidden comments be viewed.

The company behind much of this, Respondology, has an origin story that tells you everything about the scale of the problem. It was formed after the racist abuse aimed at Serena Williams when she posted a photo with her newborn in 2019, and it moved into football after Saka, Rashford and Sancho. Its co-founder reckons the firm has stripped roughly 1.5 billion hateful impressions and 15 million racist and homophobic comments out of global football, and calls it, simply, technology for good. Set against any human effort, those numbers are the entire argument. No moderation team ever assembled could read 15 million comments fast enough to spare the player, let alone sift the billion-plus impressions behind them. Abuse at this scale cannot be policed by people. It can only be intercepted by a machine, which is exactly why one now stands guard. With 78 matches in the United States and legal sports betting now live across most of the country, the abuse is expected to surge, and the tools are bracing for it.

So here is the uncomfortable truth. The platforms could do this themselves. They have always had the capability. They decline to, on the philosophical fig leaf that they are platforms and not publishers, and so they build the third-party hooks that let companies like Respondology do their cleaning for them. The technology to wipe this problem off a nineteen-year-old’s phone before he ever sees it exists today. The will, at the platform level, plainly does not. This week’s lead story was about a model being aimed at the world like a weapon. This is the same lesson from the other end of the barrel: the very same machinery, pointed the other way, can stand between a frightened young man and the worst of us. The difference is entirely a question of who decides to deploy it.

Subscribe now

The jobs AI keeps promising to take are already going, and London is the bow wave

The number that should be troubling boardrooms from Leeds to the City arrived in April, in research commissioned by the Greater London Authority. About half of London’s workers, some 2.4 million people, are in jobs where generative AI could automate or transform part of their work. The national figure is 38%. Of the London total, around 313,000 sit in the very highest exposure band and a further 748,000 in the next, which means just over one-fifth of the capital’s entire workforce is already in the path of the technology. The most exposed groups are women, the young and recent graduates, and the GLA’s own foreword notes that 7% of large UK firms have already cut staff because of AI.

I argue in this week’s Yorkshire Post that the tempting response up north, to call this a London problem, is wrong. The same study puts Yorkshire and the Humber’s exposure at around 35%, with one worker in six already in the two highest bands. Apply that to a city like Leeds, with roughly half a million jobs, and about 175,000 people are in work containing tasks these tools can already touch. Use the national rate, which may be fairer given that Leeds is the largest legal and financial centre in England outside London, and the count climbs towards 190,000.

The sharper point is where the exposure falls. Not on the rainmakers, or not first, but on the bookkeepers, the payroll clerks, the finance officers, the data-entry staff, the sales administrators and the people on the phones in customer service. Administrative roles make up almost nine in ten workers in the highest-exposure band, and for thirty years Leeds has grown precisely by doing the processing and servicing that London wanted done well and done cheaper. That is the work software reaches first.

There is a slower danger underneath. The tasks AI handles most easily, the drafting, the summarising, the routine checking, are exactly the tasks we have always handed to the young. It is how a graduate learns the trade and how a trainee becomes useful. Strip out the bottom rung and you do not just lose this year’s juniors, you weaken the ladder that produces the next decade’s seniors.

This is no longer a forecast. It is arriving in the data, on schedule, which is why every Sunday we count it.

Subscribe now

The Sunday Signal Tech and AI Layoff Tracker · Week 24 · 14 to 20 June 2026

Consolidated across Layoffs.fyi, LayoffHedge, SkillSyncer, state WARN registries and SEC 8-K filings, to cut through the corporate PR and track the real displacement.

2026 year-to-date, cross-sector: around 511,140 and counting
Tech-specific year-to-date: around 207,970 and counting
Added this week: around 31,140, a spike driven almost entirely by the finalisation of Oracle’s historic restructuring

The signal: the compute-over-compensation trade-off. This week the tech total vaulted past 200,000 job losses for the year, and the pace is now running at roughly 1,115 losses per working day, nearly double last year’s rate. The driver has shifted from speculative experiment to permanent restructuring, with AI cited as the primary strategic reason in more than 55% of 2026 tech layoff events, hitting upwards of 152,000 workers. But the cracks in “AI laundering”, the habit of dressing ordinary operational stress in the glamour of an AI transition, are now showing. Despite shedding human overhead under the banner of efficiency, most of these companies are not posting the margin gains that were promised. They are liquidating middle management, front-end engineers and customer-success teams to fund the staggering infrastructure, power and silicon bill of the agentic era. Compute is winning, and compensation is paying for it.

Oracle, around 30,000 cuts, now complete. On 15 June, Oracle finished the largest restructuring in its history, off-boarding roughly 30,000 staff, around 18.5% of its global workforce. The context is the story. Oracle are not in distress; they posted $17.2bn in quarterly revenue, up 22% on the year, with explosive growth in cloud-infrastructure AI. They cut anyway, to free an estimated $8bn to $10bn in annual cash flow and feed a capital-expenditure pipeline approaching $50bn for the fiscal year. The deepest wounds fell on Oracle Health, the unit built from the $28.3bn Cerner acquisition, which lost an estimated 8,000 to 10,000 people, amid bitter disputes over severance and forfeited stock.

Artlist, around 200 cuts. The Israeli creative-tech firm shed 40% of its staff, about 200 of 500. What makes it remarkable is that it did not stem from a downturn. The company recently passed $300m in annual recurring revenue on 50% year-on-year growth, and the co-chief executive said plainly that AI-native rivals are moving five times faster with a fraction of the headcount. Record revenue, it turns out, no longer insulates a workforce whose tasks can be replicated by an algorithm.

Expeditors International, around 230 cuts. The freight and logistics giant eliminated roughly 230 roles inside its US Global Technology department, hitting software developers, business analysts and project managers. The agentic era has left Silicon Valley. Traditional logistics firms are abandoning bespoke internal software for autonomous third-party agents, and the in-house engineers are the casualties.

Week 25 watch. Expect the pressure to spread to mid-market creative and document-workflow SaaS platforms, where Artlist’s lean model is fast becoming the boardroom standard, and to legacy health-tech and implementation consultancies caught in the wake of the Oracle Health cuts.

Subscribe now

🚀 Final Thought

Four stories this week. Read them again and they collapse into one.

A government in Washington can switch off your models overnight. A bank can wall them off from its own staff simply by reading a contract carefully. A mob can turn the network those models run on against a nineteen-year-old who missed a penalty. And a boardroom flush with record revenue can use the very same technology to clear the floor below, then call it innovation in the press release.

The thread running through all of it is power, and who holds it. The last cold war was decided by who could enrich uranium and who could not. This one will be decided by who owns the models, who hosts the compute, and who is left renting both from the other side of the divide. For three years Britain, and most of British business, has rented its intelligence from someone else’s cloud and called the arrangement a strategy. On 12 June we learned exactly what a rented strategy is worth on the day the landlord picks a side.

The curtain is coming down. The only question left is whether Britain, its companies and its citizens choose where they stand, or simply wait to be sorted.

Until next Sunday,

David

The Bear and the Fly

David Richards MBE — Sun, 14 Jun 2026 05:30:54 GMT

THE SUNDAY SIGNAL · Issue #58 · Week 25 · Sunday 14 June 2026

By David Richards MBE

Subscribe now

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Keynote in Doncaster This Week

Bottom Line Upfront: On Friday, SpaceX floated and pushed Elon Musk past the trillion-dollar line on paper. Then, after the market closed, the US government ordered Anthropic to suspend access to its two most advanced models for foreign nationals anywhere in the world, forcing the company to disable them for every customer. One event showed how concentrated ownership of the machine economy has become; the other showed how fragile its infrastructure is. Several thousand years of history suggest that societies which fail to level extreme wealth peacefully are eventually levelled by violence, war, collapse or plague.

Aesop warned us centuries ago about the bear who crushed his master’s skull to kill a fly. This week the US Commerce Department played the bear, and Claude Fable 5 was the fly. Read on for what happened, why it matters far beyond one company, and how it connects to the larger story I took to the Doncaster Chamber of Commerce this week: that artificial intelligence may grow the world economy fivefold while handing ownership of that growth to fewer people than any democracy should tolerate.

Subscribe now

The Bear, the Fly, and the Kill Switch

Friday, 5:21pm Eastern. Anthropic, the company behind the Claude models and one of the most valuable private firms on the planet, received a letter from the US government. Citing national security and export-control authorities, the directive ordered it to suspend all access to its two most capable models, Fable 5 and the restricted Mythos 5, for any foreign national anywhere in the world, including Anthropic’s own foreign-national employees. Cloud software does not have borders. You cannot check the passport of every user calling a model through Amazon’s servers. So the only way to comply was the bluntest one available: shut the models down for everyone, in every country, at once. Access to all other Claude models was left untouched.

Here is the part that should worry every business now building on AI. The government’s stated concern was a jailbreak, a prompt that slips past a model’s safety filters. The specific technique it cited, by Anthropic’s account, amounted to asking the model to read a codebase and fix any flaws it found. Anthropic reviewed the demonstration and said the vulnerabilities it surfaced were minor, previously known, and discoverable by other public models, including OpenAI’s GPT-5.5, without any bypass at all. Developers call reading code and fixing flaws debugging. They do it every day. The government looked at the same capability and saw automated discovery of weaknesses in critical infrastructure, a cyber-weapon in commercial clothing.

Recall the fable. A tame bear guards its sleeping master. A fly lands on the man’s face, and the bear, meaning only to protect him, picks up a rock and brings it down on the fly. The fly dies. So does the master. The moral Aesop left us is that misguided zeal can do the work of malice, and that the cure is sometimes worse than the disease. Washington wanted to swat a narrow vulnerability. It used an export-control sledgehammer, and in a single stroke it disabled the most advanced commercial software in the world for hundreds of millions of users, crippling an American champion on the eve of its stock market debut. The fly is dead. Look what happened to the master.

Anthropic is complying, and says it disagrees. Its public position is sharp: recalling a model used by hundreds of millions of people over one narrow, non-universal jailbreak sets a standard that, applied across the industry, “would essentially halt all new model deployments for all frontier model providers.” That is the chilling line for anyone running a company. If Washington can flip the kill switch on a Friday afternoon over a routine debugging prompt, then no enterprise can treat frontier AI as stable infrastructure. You do not build a factory on land the state can repossess without notice. Overnight, every firm with a workflow wired to a single model learned that its supplier can be switched off by a third party it never signed a contract with.

And here is the part that should keep Anthropic’s bankers awake. The company closed a 65 billion dollar funding round in late May at a valuation near 965 billion dollars, and filed confidentially days later to go public in what would be the first near-trillion-dollar AI listing. That valuation rests on a simple promise: that the company can keep shipping frontier models and charging a premium for them. Fable 5 was the proof. Friday added a new line to the risk register that no prospectus has ever had to carry: sovereign recall. What exactly are you worth, if the product that justifies the valuation can be removed from the market by a letter at 5:21 on a Friday afternoon?

Subscribe now

Rockefeller Could Be Broken Up. Musk Is Harder

Rewind a few hours, to before the directive landed. While the markets were still open that Friday, the state let a milestone pass without a murmur. Start with the facts. SpaceX priced its IPO at 135 dollars a share and began trading on the Nasdaq under the ticker SPCX. The company sold 555.6 million shares and raised 75 billion dollars, the largest flotation in stock market history and nearly three times the size of Saudi Aramco’s 2019 record. The offer valued the company at 1.77 trillion dollars. Then the market got hold of it. The stock closed its first day up 19 per cent at 160.95 dollars, pushing the valuation above 2 trillion dollars and making SpaceX one of the most valuable listed companies on earth.

Now the man. At the offer price, the prospectus put Musk’s SpaceX stake at 866 billion dollars. After Friday’s 19 per cent jump it is worth comfortably over a trillion on its own, and his Tesla holding adds several hundred billion more. His paper wealth has cleared 1.3 trillion dollars, which is north of 4 per cent of American GDP. He retains more than 82 per cent of SpaceX’s voting control. One shareholder, one vote that counts.

We have been somewhere near here before, and it is worth remembering how it ended. At his peak, John D. Rockefeller’s Standard Oil refined around 90 per cent of America’s oil, and his personal fortune amounted to roughly 1.5 to 2 per cent of US GDP. The American state looked at that and decided a republic could not carry it. In 1911 the Supreme Court broke Standard Oil into 34 companies under the Sherman Antitrust Act. The breakup made Rockefeller richer in the short term (his shares in the successor companies soared as the motor car arrived), but it shattered his structural grip on the economy. Then came the income tax in 1913, the estate tax in 1916, the gift tax in 1924, and four decades of deliberate philanthropy. When Rockefeller died in 1937, his personal estate held 26.4 million dollars. The fortune that once dwarfed the federal government had been dissolved into universities, foundations and hundreds of family trusts. The system was slow and imperfect. But it worked.

NY Times 1911

Here is the uncomfortable truth about 2026: the 1911 hammer does not fit. Antitrust law polices monopolies within a single market. Musk does not run one monopoly; he runs a portfolio of separate dominant positions across entirely different sectors. Rockets and global satellite internet through SpaceX and Starlink. Frontier AI through xAI, folded into SpaceX in February’s merger. Electric vehicles and grid batteries through Tesla. Social data through X. Modern competition law has no simple Standard Oil remedy for the problem of one person holding commanding influence across many markets at once. The law was built to break up companies. Nobody ever wrote one to break up a person.

And rather than restraining this, the machinery of modern markets reorganised itself to accommodate it. Nasdaq rewrote its index rules last month so that the largest IPOs can enter the Nasdaq 100 after just 15 trading days, scrapping its 10 per cent minimum float requirement in the process. SpaceX floated barely 4 per cent of itself, which means index funds could be forced to absorb an estimated 22 to 27 billion dollars of stock mechanically, at whatever price prevails. Many pension savers with passive exposure to Nasdaq-linked products would end up owning SpaceX whether an active manager chose it or not. This for a company that lost 2.6 billion dollars from operations last year on revenue of 18.7 billion, and which Morningstar values at 780 billion dollars, less than half the offer price. Rockefeller had to fight the state. Musk has the machinery of modern passive markets carrying his bags.

The point is not that Musk is uniquely wicked. The point is that the system has discovered how to manufacture Rockefeller-scale power without triggering Rockefeller-era remedies.

Roosevelt or the Horsemen

If the legal toolbox is empty, it is worth asking what actually happens to societies that let wealth concentrate this far. The answer is uncomfortable, because we have a great deal of evidence, and none of it is reassuring.

The historian and complexity scientist Peter Turchin has spent his career modelling how civilisations decay, and he keeps finding the same mechanism. Mature societies develop what he calls a wealth pump: a structural arrangement that steadily moves wealth from the people who work to the people who own. In Rome it was conquest and slave labour. In feudal Europe it was land tenure. In our era it is asset inflation, and increasingly it is the replacement of paid human labour with machines that work for the price of electricity.

The pump produces two effects, and both are visible in the historical record again and again. The first is mass immiseration: living standards stagnate for the majority, social mobility seizes up, and even basic measures such as life expectancy begin to slide. The second is what Turchin calls elite overproduction. As wealth explodes at the top, a society mints more rich, educated and ambitious people than it has positions of power to absorb. The surplus elite do not retire quietly. They factionalise, fight each other for the remaining seats, and wage their wars through politics and the media until the state itself seizes up. If that description sounds familiar, it should.

When immiseration meets a divided elite, societies break, and the precedents are brutal. In the late Roman Republic, conquest flooded the elite’s mega-estates with slave labour, the ancient equivalent of automation, and ordinary citizens who could not compete with workers who cost nothing lost their farms and crowded into Rome as a dispossessed underclass. A century of populist uprisings and civil war followed, and the Republic did not survive it. In eighteenth-century France, the nobility and clergy hoarded the nation’s land while exempting themselves from the taxes that paid for the state, leaving the burden on the peasantry; when the harvests failed, the result was not a renegotiation but a revolution that took the elite’s heads along with its property. In nineteenth-century China, extreme inequality under the Qing helped ignite the Taiping Rebellion, a civil war that killed more than twenty million people.

The economic historian Walter Scheidel surveyed several thousand years of this record and reached a conclusion he called the Great Leveler thesis: extreme inequality has never once, in recorded history, been substantially reduced by peaceful market forces. It has only ever been flattened by four kinds of shock, his Four Horsemen. Transformative revolution. State collapse. Mass-mobilisation warfare, which forced governments to tax the rich at near-total rates to fund the World Wars. And lethal pandemics: the Black Death killed a third of Europe and, by making labour scarce, handed the surviving peasants the bargaining power that broke feudalism. The ledger always balances. The only question history leaves open is who pays, and how violently.

There is one recorded exception, and it is the one that should interest us most. Occasionally a faction of the elite sees the cliff edge in time and chooses to give up a portion of its own wealth to preserve the system that created it. The clearest modern case is America in the 1930s. Facing depression, extreme concentration and the live threat of revolutionary politics spreading from Europe, Franklin Roosevelt, himself a child of the elite, built the New Deal: aggressive antitrust enforcement, empowered unions, a welfare state, and top marginal tax rates that eventually passed 90 per cent. It was state-mandated redistribution, chosen by the wealthy class’s own representatives, and it engineered the Great Compression that built the twentieth-century middle class without a shot being fired. It is also, in several thousand years of records, very nearly the only time the off-ramp was taken.

Now look at the pump we are building. If AI substitutes for human labour at scale, it funnels the economic value of both muscle and mind permanently upward to whoever owns the compute. That is not a side effect of the technology. It is the business model. Which makes the only question that matters a simple one: are we heading for Roosevelt’s exit, or one of the other four?

Franklin D. Roosevelt signing the Agricultural Adjustment Act

And what would Roosevelt’s exit even look like now? The policy debate is already circling three ideas. A compute tax would levy the GPU hours and electricity that power AI, though taxing compute risks taxing the steel of this industrial revolution and crippling everything downstream of it. A robot tax, or agentic output tax, would charge firms at the point an AI agent replaces a human department, taxing the deployment rather than the invention. And structural capital reform would do the unglamorous work the 1930s did: tax capital gains like income, close the inheritance loopholes, and reach the unrealised paper fortunes that make trillionaires possible in the first place. None of the three is simple. All of them are easier than the horsemen.

Subscribe now

The Money Is Not Going Where You Think

The comforting story about the AI boom is that it mints brilliant founders, and brilliance is at least a defensible lottery. The story is wrong. The founders are being diluted out of their own revolution, and the wealth is pooling somewhere far less accountable.

Consider the people actually building frontier AI. Sam Altman famously holds no equity in OpenAI. Dario and Daniela Amodei hold low single-digit stakes in Anthropic. Why? Compute. Training frontier models devours capital at a rate no industry has ever seen, and when you must raise tens of billions of dollars a year simply to keep the servers running, you sell your company, piece by piece, to whoever has tens of billions of dollars. Anthropic raised 30 billion dollars in February, in a round led by Singapore’s sovereign wealth fund GIC alongside Coatue, at a 380 billion dollar valuation. By late May it had raised another 65 billion at a 965 billion dollar valuation. Ninety-five billion dollars in under four months. The scientists get the bylines. The funds get the company.

So follow the money to where it is actually settling: the infrastructure layer. Sovereign wealth funds committed an estimated $120 billion across 2025 and 2026 to data centres, chip fabrication and compute networks. Abu Dhabi’s MGX, a dedicated AI investment vehicle, co-led the 40 billion dollar acquisition of Aligned Data Centers with BlackRock and anchors the AI Infrastructure Partnership alongside Microsoft, Global Infrastructure Partners and Kuwait’s KIA. A 1 gigawatt Stargate UAE cluster is being built inside Abu Dhabi’s wider 5 gigawatt UAE-US AI campus, with the first 200 megawatts expected online this year. Saudi Arabia’s Public Investment Fund launched HUMAIN to build the entire national AI stack, targeting more than 6 gigawatts of data centre capacity by 2034, with Nvidia and Qualcomm signed up and a $10 billion partnership with Google Cloud. The Gulf’s sovereign funds collectively manage close to 6 trillion dollars, and they are executing the most coherent strategy on earth: convert oil money into compute, because compute is the commodity every AI company must buy regardless of which one wins. They have worked out that they cannot out-engineer the founders. So they have become their landlords instead.

Turchin would recognise the machinery instantly. This is the wealth pump being assembled in plain sight, with quarterly reporting.

Set Britain’s response beside that. The Sovereign AI Unit, launched in April with 500 million pounds, was described by the venture community as a drop in the ocean, and the venture community was being polite. France assembled a 109 billion euro package. The Gulf deploys more than Britain’s entire fund in a single building. Britain helped invent much of this field. We are now negotiating for a seat in the cheap stands.

Hold the full picture in your head. The founders are diluted. The engineers are salaried. The state is outbid. What remains is a new class of owner, the institutional landlord of the machine economy: hyperscalers, silicon monopolies and petro-states converting one strategic commodity into the next. When the historians of the 2030s ask who captured the value of the AI revolution, the answer will not be the people who built it. It will be the people who owned the buildings it ran in.

Subscribe now

The Pause and the Promise

Once the machine economy is this capital-intensive, even the companies building the frontier are trapped by the capital structure around them. And capital this size plays defence: a lab that has just banked the largest war chest in the industry has an obvious interest in anything that freezes the race while it leads. Which returns us to Anthropic, and to the bitter irony buried in Friday’s shutdown.

Anthropic builds some of the most capable AI models on earth. This month it disclosed that more than 80 per cent of the code merged into its own systems is now written by its AI, not its engineers, and that the typical engineer ships roughly eight times as much code as before. Push that trend a little further and you reach the point where machines design their successors with humans barely in the room. The industry calls it recursive self-improvement. Anthropic’s Jack Clark thinks some models could be capable of it within two years. Then recall what the government cited to justify the shutdown: Fable’s skill at reading code and finding flaws. That is the bitter symmetry. The capability that made Fable commercially valuable, its fluency with code, is also what made it politically vulnerable.

Weeks before any of this, two of the company’s senior figures, Marina Favaro and Jack Clark, had published a paper arguing that governments and labs need a credible way to slow, or temporarily pause, frontier AI development, to let society and safety research catch up. Their reference point was Cold War arms control, with one chilling caveat of their own: training runs are far easier to conceal than missile silos.

Critics scoffed at the timing. Days before the paper appeared, Anthropic had closed that 65 billion dollar round at a 965 billion dollar valuation and filed confidentially to go public. A call for a global pause sounds rather different coming from the company in front, because a freeze pulls up the ladder behind it. Restraint and self-interest were wearing the same coat, and from the outside nobody could tell which one was speaking. Then on Friday the government paused Anthropic for it, on terms nobody at the company would have chosen, and the argument took on a darker shape. It is one thing for a leader to call for brakes. It is another for the state to rip the wheel away over a debugging prompt.

Now put that next to Jensen Huang. The Nvidia chief told the world there is no fundamental limit to the size of the global economy, and that AI will take output from one hundred trillion dollars to two hundred, three hundred, five hundred. Here is where I land, and where I landed in Doncaster this week. Huang is closer to right than his doubters allow. The acceleration is real and it will be enormous. But a fivefold economy does not arrive smoothly; it arrives in waves, and waves knock people down. The layoffs have already begun, in software, in finance, in the white-collar trades that thought themselves safe, and the single greatest failure of our politicians is that they have not begun to grasp the scale of what is coming.

The investor Naval Ravikant offers the consoling thought: a poor man today lives better than Louis XIV ever did. The Sun King had absolute power and no antibiotics, no aeroplane, no plumbing worth the name. Technology raises the floor for everyone, and in Huang’s world the poorest child carries a tutor and a doctor in her pocket. That is a wonder, and I do not dismiss it. But the floor and the ceiling do not move together. The floor rises. The ceiling vanishes into the clouds. The poorest may live better than the old kings of France, while a handful who own the machines amass fortunes the world has never seen. One of them crossed the trillion-dollar line on Friday, at least on paper. That is the bargain on offer. We had better start arguing about its terms.

You can read the full column, The Pause and the Promise, in this week’s Yorkshire Post.

Subscribe now

Weekly Tech and AI Layoff Tracker

This is the human edge of the same story. If compute is becoming capital, labour is becoming the adjustment variable. Here is the week’s count.

The sector has stopped debating automation and started executing it. Fresh outplacement data puts tech layoffs at a near two-year high, with AI now the single most cited reason employers give, behind nearly 40 per cent of recent eliminated roles. The language has shifted too. Firms no longer talk about replacing data-entry clerks; they talk about dismantling whole layers of middle management and front-end development to run lean, automated operations. They call it the agentic era.

Watch for the laundering. As the numbers climb, more companies are dressing reactive cuts in forward-looking clothes, repackaging a bad quarter or weak execution as an AI structural transition. It keeps the investor’s eye on automated agility and off the operational stumble. This week gave four clear examples.

ServiceNow, around 250 roles. The workflow giant trimmed across solution consulting, sales, product marketing and training, telling staff its own platform is generating real AI efficiencies inside the business. Proof of product and margin defence in one move, as enterprise software growth softens globally.

Pleo, around 50 roles. The Danish spend-management fintech cut product, data and engineering jobs across the UK, Denmark and Germany, 24 hours after launching a suite of autonomous agentic finance tools. The internal memo blamed the changing role of new technology in how its own teams work. The builders of the agents, downsized by the agents.

TTEC, around 1,500 roles. The customer-experience operator is running a sweeping, unannounced cut across its telecom support operations, with sudden site closures and significant labour pushback. It reads as the classic prelude to a platform swap: liquidate the expensive human contact centre, make way for autonomous voice and chat.

Google Cloud and Meta, targeted trims. Both finalised restructuring within security and infrastructure teams, substituting algorithmic triage for human oversight and clearing payroll to feed enormous data-centre and silicon spending.

Next week’s watchlist: Oracle, where institutional trackers point to the final phase of an infrastructure realignment clearing backend roles around the 15 June horizon, and the customer-facing fintech and neo-banking platforms now under pressure to prove they are automating their own code, following Pleo’s precedent.

Final Thought

In 1911, the American state looked at one man holding roughly two per cent of its economy and understood that concentrated wealth had become concentrated power. The response was slow, imperfect and contested, but it was a response.

This week the same state did two things on the same Friday. First it let one man pass the trillion-dollar line without a murmur, while the machinery around him bent to help: index rules shifted, passive funds prepared to buy, sovereign wealth funds supplied the capital. Then, after the bell, it reached into the most advanced commercial software on earth and switched it off over a debugging prompt, swatting a fly with a boulder. It found the will to brick an AI model in an afternoon and no will at all to ask what it means for one person to own a slice of the world. The bear is wide awake about the wrong threat.

Several thousand years of records offer societies in this position exactly two endings. In one, a faction of the wealthy chooses Roosevelt’s path and surrenders a share of its fortune to save the system that produced it. In the other, the correction arrives by horsemen. Nobody in power is preparing for either.

We broke up Standard Oil with a court order. We are building Standard Everything with a tax break. 🚀

Until next Sunday,

David Richards MBE

The £10 Billion Goldmine Britain Keeps Giving Away

David Richards MBE — Sun, 07 Jun 2026 06:00:57 GMT

THE SUNDAY SIGNAL · Issue #57 · Week 23 · Sunday 7 June 2026

By David Richards MBE

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Subscribe now

Bottom Line Up Front

Britain has a habit it cannot break. We invent the future, and then we sell it to people who understand its value better than we do. This week a select committee told ministers to remove Palantir from the NHS, and the company’s British boss spent the days that followed on the airwaves and in the papers explaining why that would be reckless. He is a fluent advocate. He is also, by training and temperament, a politician rather than a technologist, and that tells you exactly what this fight is really about. It is not about software. It is about who owns the most valuable dataset in the world, and whether we are about to repeat with our health records the mistake we made with DeepMind.

A lobbyist, not an engineer

Listen to Louis Mosley defend Palantir on the radio, as I did this week, and one thing becomes obvious. He does not sound like a tech chief executive. He sounds like a minister. The cadence, the appeal to fairness, the careful reframing of every awkward question into a question about patients: this is the grammar of Westminster, not Silicon Valley.

There is a reason for that, and it is not an accident of personality. Mosley read history at Oxford. He cut his teeth in Conservative politics, working as an aide to Rory Stewart and serving as a councillor in Kensington and Chelsea. In 2017 he came close to standing for Parliament, until the party grew nervous about his surname. He is the grandson of Sir Oswald Mosley, the man who led the British Union of Fascists in the 1930s, a piece of family history that has shadowed his public life and which he did not choose. He worked in finance, then joined Palantir in 2016, and now runs a UK and European division that accounts for roughly a quarter of the company’s entire global workforce.

British politician Sir Oswald Mosley salutes members of his British Union of Fascists on Royal Mint Street, London

What he was not, at any point, was an engineer. And the question worth asking is why an American software company would put a Tory networker rather than a coder in charge of its British operation.

The answer is that Palantir’s product is not really software. Its product is government. When a company sells data infrastructure to the state, its biggest obstacle is never a competitor. It is the state itself: procurement rules, select committees, ministers, the slow machinery of public trust. So you do not hire the best architect. You hire the person with the best contacts book. Mosley pitched directly to Michael Gove in 2019 and won a Brexit planning contract. When the pandemic hit, Palantir offered its services for a pound, then turned that pound into a sprawling presence across the NHS. The seven-year, three hundred and thirty million pound Federated Data Platform was the prize at the end of that road.

Subscribe now

We have seen this casting before

If a politician fronting an aggressive American tech firm feels familiar, that is because we watched it happen once already.

Nick Clegg, former MP & Deputy PM Was a VP of Facebook

In 2018, with Facebook reeling from Cambridge Analytica and Brussels sharpening its regulatory knives, Mark Zuckerberg hired Nick Clegg. Not for his grasp of machine learning. For his Rolodex. The former Deputy Prime Minister became the respectable, centrist, European face of a company that badly needed one, and he spent years building the buffers, the oversight board, the careful language, that kept governments at arm’s length.

Then the politics changed, and so did the casting. In January 2025, three weeks before Donald Trump returned to the White House, Meta replaced Clegg with Joel Kaplan, a Bush-era Republican operative with deep ties to the incoming administration. The lesson was unmistakable. These appointments are not about conviction. They are instruments, swapped out the moment the political weather turns. Clegg was hired to keep regulators out of Meta’s business. Mosley’s job is the opposite: to get Palantir deeper into ours. One played defence. The other plays offence. The casting logic is identical.

Subscribe now

Mosley’s Telegraph defence is full of holes

This week Mosley took to the pages of the Telegraph to argue that the MPs attacking Palantir were putting politics above patients. It is a polished piece. It is also, on inspection, an exercise in answering questions nobody is asking while ducking the ones that matter.

Take the operational case. Mosley leans on Palantir’s own figures: more than a hundred and ten thousand extra operations, almost three hundred thousand patients discharged sooner. Those numbers come from the NHS’s own count, and they may well be real. But they have not faced serious external scrutiny, and critics inside the health service have noted that some “waiting list reductions” amount to little more than removing patients who stopped replying to text messages. A claim that cannot be independently audited is a marketing line, not a finding.

Then there is the committee itself. Mosley says the MPs urged cancellation with no idea what should replace Palantir. That is not what the report says. The Science, Innovation and Technology Committee, chaired by Dame Chi Onwurah, explicitly told ministers to either build a domestic replacement or find a UK provider. Its worry was sovereignty, the danger of leaning on a small number of vast American firms, and it warned in plain terms that the current arrangement leaves Britain exposed “at the mercy of foreign actors.” You can argue with that conclusion. You cannot pretend it was not made.

And on data, Mosley is at his most reassuring and his most slippery. There is no central vault to seize, he says; Palantir cannot use, sell or move the data. On the narrow point, he is right. Palantir is a processor, not a broker, and the data stays with each trust. But the committee’s fear was never theft. It was entanglement. Palantir’s platform burrows into the plumbing of individual trusts through custom connectors and proprietary code, and the deeper that integration runs, the harder and costlier it becomes to ever pull the company out. That is what lock-in means. Reassuring me that you will not steal the house does not address my worry that I can no longer change the locks.

He saves the loftiest line for last, insisting Palantir has never worked in states adversarial to Western values and only serves governments with a democratic mandate. Yet the protests outside his offices are not about abstract values. They are about a company with deep roots in US defence and immigration enforcement, and ties to military programmes abroad, now sitting at the centre of a service built on the idea of care for all. Amnesty International, medical groups and NHS staff are not confused about this. Mosley would prefer to call their objection ideology. It is not. It is a values question, and he knows it.

Subscribe now

The plumbing problem: ask the National Grid

Here is the part the patient-centred framing leaves out. We already know what happens when Britain lets a foreign, state-adjacent company into critical national infrastructure, because we have spent the last few years frantically undoing exactly that.

In 2023 the National Grid began ripping out components made by a unit of the China-backed firm Nari Technology, acting on direct advice from GCHQ’s cyber arm. The parts in question helped balance the grid and prevent blackouts. We had wired a potential adversary into the machinery that keeps the lights on, and then quietly tore it out again.

It did not stop there. The “golden era” welcome once extended to China General Nuclear curdled into a costly buy-out of its stake in our nuclear programme. And only this March, ministers blocked a one and a half billion pound plan by the Chinese firm Ming Yang to build what would have been the world’s largest wind turbine factory in Scotland, walking away from fifteen hundred jobs on national security grounds, with the Ministry of Defence warning the turbines could double as surveillance platforms.

Notice the pattern. Court the foreign money, embed the foreign technology, then discover, too late and at enormous expense, that you have created a dependency you cannot afford. The committee looking at Palantir has simply learned the lesson the energy world learned the hard way. Data is infrastructure too. It is the most strategic infrastructure we have.

Subscribe now

What this is really about: the goldmine

Strip away the politics and the reassurances, and the fight comes down to a single resource. Data is the oil of this century, and the NHS sits on one of the richest, cleanest reserves on the planet.

Because we built a single-payer system, we hold consistent, trusted records on roughly fifty-five million people, tracked from cradle to grave across decades. No fragmented American insurer can match it. Ernst and Young put a number on what that asset could unlock: as much as nine point six billion pounds a year, split between operational savings for the NHS and benefits to patients, once the data is properly curated and put to work. Call it the ten billion pound goldmine.

That is what the platform war is about. Whoever structures and operationalises this data gets to train the next generation of diagnostic models, risk tools and medical AI on it. The intellectual property, the patents and the profits then flow to private shareholders, while the public that funded the raw material for decades is downgraded from owner to mere data provider. We supply the gold. Someone else walks off with the finished jewellery, and then sells it back to us as a service.

Subscribe now

We invented the thing they fear. Then we sold it.

And this is where it becomes unbearable, because we have rehearsed this exact failure before, at the very frontier of AI.

The most advanced AI lab on earth was not born in California. DeepMind was founded in London in 2010. In 2014 we sold it to Google for around four hundred million pounds, outbidding a desperate Mark Zuckerberg in the process. It looked like a tidy exit at the time. It now looks like one of the great strategic giveaways in our history. DeepMind became the crown jewel of Google’s AI empire. Its founder, Demis Hassabis, won a Nobel Prize and now runs all of Google’s AI. His drug-discovery spin-out, Isomorphic Labs, is reportedly raising over two billion dollars. The four hundred million pound price tag has not aged well, and the real cost, the capability we handed to a foreign giant, cannot be counted at all.

We should have known what we were losing, because the people racing against DeepMind knew exactly what it was. When Elon Musk and Sam Altman were casting around for a name for the venture that became OpenAI, the court exhibits from this spring’s trial show Musk pushing one option above the rest. He wanted to call it Freemind, he wrote, because it would convey intelligence freely available to all, “the opposite of Deepmind’s one-ring-to-rule-them-all approach.” He told his co-founders that unless they hired the best people on earth they would “get whipped by Deepmind.” By his own account, Google at that point had snapped up something close to three-quarters of all the AI talent in the world. The most powerful men in technology built a company out of fear of what Britain had created.

Read that back slowly. Silicon Valley feared our lab so much it founded a rival to contain it. We feared nothing, so we sold it. Now the same instinct that gave away DeepMind is being asked to hand over the data that could have trained the next one. The idea that we need a Silicon Valley giant to make sense of British health records is not pragmatism. It is a failure of nerve dressed up as common sense. We lost our best technology. We must not lose our best data as well.

Subscribe now

The 17-Year Itch: Why Brilliant Technology Goes to Die in the NHS

If Palantir is the story of who owns the data, this is the story of why even the best technology so often rots before it reaches a single patient.

On this week’s episode of The Digital Forge Podcast I sat down with Richard Stubbs, chief executive of Health Innovation Yorkshire and Humber and a member of the MHRA’s national commission on regulating AI in healthcare. Stubbs has spent more than two decades inside the system, and he punctured a fantasy that every founder entering healthcare carries around: the dream of one meeting with a minister, one handshake, and a rollout to the whole country.

It does not work like that, and the reason is structural. American firms walk in, win the minister over, and walk out certain they have a deal for sixty million patients. Then someone has to break the news. As Stubbs put it, “There is really no such thing as the NHS.” There are five or six hundred separately governed, largely autonomous organisations, and every one of them has to say yes for itself. There is no big lever in a big room. What thrives in Bradford can be ignored entirely in Barnsley, and the result is a technological postcode lottery.

That fragmentation breeds a particular disease, the conviction that nothing proven elsewhere can be trusted here until it has been piloted all over again. “We have more pilots in the NHS than British Airways,” Stubbs told me, and the line lands because it is true. Proven technology emerges, a local pilot is approved, the metrics check out, the funding lapses, and the whole circus starts again somewhere new. Meanwhile the underlying reality stays medieval. In parts of the country, patient notes are still bundled into the back of a taxi and driven down the A1 because electronic records are not yet universal.

The number that should stop you cold is this. By one Royal College estimate, it takes an average of seventeen years for a proven innovation to reach mainstream NHS use. Seventeen years ago the App Store had only just launched. If we take the better part of two decades to adopt the present, we have no hope of keeping pace with AI.

Stubbs’s prescription is not to crush local decision-making but to be honest about the trade-off it carries. He wants a “comply or explain” regime: the centre sets a gold standard for a transformational pathway, and a region either adopts it or proves, in numbers, that its own way delivers better outcomes for the same money or less. No more reinventing the wheel as a reflex.

The sharpest moment came when we turned to safety, the word always reached for to slow AI down. Stubbs flipped it. “Nobody ever asks the question of: how safe are we right now?” Nothing in the NHS is a hundred per cent safe, he pointed out, yet we treat the exhausted, paper-bound status quo as if it were, and demand flawless perfection from the software meant to improve on it. By holding AI to a standard we never apply to ourselves, we keep patients waiting longer in settings that are already, measurably, less safe.

There are bright spots, and they show what is possible when the system gets out of its own way. Pinpoint, a machine learning tool developed over years in Yorkshire, reads previously taken blood tests and helps safely lift lower-risk patients off the two-week cancer pathway, giving them peace of mind and freeing capacity for those who genuinely need it. Ambient voice technology now lets clinicians hold a real conversation while the record writes itself. These are not science fiction. They are working today. The barrier was never the technology. It was us.

Subscribe now

Would AI Have Locked Us Down? (This Week’s Yorkshire Post Column)

The third strand runs through my Yorkshire Post column this week, and it asks a question I have turned over for six years. Would the models we have today have told ministers to shut the country?

Cast your mind back to March 2020. The plan was mitigation: shield the vulnerable, slow the spread, keep the schools and shops open. Then came Report 9 from Imperial College, which modelled five hundred and ten thousand deaths in an unmitigated epidemic and suggested that even mitigation could overwhelm hospitals. Within days the plan was dead and the country was indoors. It was not the only model. But it broke the deadlock, and it did so before anyone outside the team could read a single line of its code.

Some readers will remember that I was one of the people who did read it, at the time, and said so loudly in the national press. The code was a decade old, sprawling, with functions that looked, as the programmer John Carmack later put it, “machine translated from Fortran.” My objection was never that academic code looks like academic code. Carmack himself judged it held up better than its fiercest critics claimed, and on reflection I think that is fair. The scandal was narrower and worse. A tool built for experts to interrogate became an instrument of government before the public could interrogate it at all.

The danger of dressing a guess as a certainty did not end in 2020. This February, Britain’s own statistics watchdog had to write to the Covid Inquiry to correct it. The Inquiry’s summary had claimed modelling “established” that an earlier lockdown would have saved twenty-three thousand lives in the first wave. One word did the damage. Established. Not estimated, not suggested, not modelled. A what-if had been exhumed and presented as a fact, and the same overreach that drove the original decision was being written into its official autopsy.

So run 2020 again with the tools we have now. AI-assisted models would not have handed a minister the answer, and that is precisely the point. They would have watched first, reading wastewater, admissions, genomics and mobility together, faster than 2020 managed. They would have aimed at the network rather than the nation, closing the settings doing most of the spreading and leaving the rest running where the evidence allowed. They would have priced every path, lives against lost school years, risk against ruin, and put those trade-offs in front of the person who had to choose.

Would they have recommended a lockdown? Faced with a pathogen that spread fast and silently and killed at scale, yes, they might. Sometimes the only way to break an epidemic is to break a contact network. But which one, and for how long? Schools, pubs, care homes, cities, or the whole country? That is the question 2020 never properly asked, and it is the question that cost a generation of students two years of their lives.

Subscribe now

The Sunday Signal Tech and AI Layoff Tracker

Week 23 · 31 May to 6 June 2026

The macro picture hardened this week, and the official numbers are now starker than any single company’s announcement. According to Challenger, Gray and Christmas, US technology employers cut 38,242 jobs in May, the sector’s heaviest month in nearly two years, lifting its 2026 running total to 123,653, up more than sixty-five per cent on the same stretch last year. For the third month in a row, AI was the most-cited reason for layoffs across every sector. Andy Challenger’s verdict was blunt: AI is now the leading reason companies give for cutting jobs.

The emblem of the week was GitLab. The all-remote pioneer filed to cut about fourteen per cent of its workforce, roughly three hundred and fifty roles, while exiting twenty-two of the countries it operates in and flattening layers of management. Chief executive Bill Staples pinned it explicitly on the “agentic era” and the company’s Duo Agent Platform. What makes it the perfect case study is the accompanying earnings: first-quarter revenue up twenty-three per cent. This was not a distress cut. It was a deliberate redirection of payroll into compute.

Which is exactly why the “AI laundering” story is now mainstream. The pattern is no longer disputed even by the optimists. Sam Altman has conceded that some firms practise “AI washing,” blaming the technology for cuts they would have made anyway. And the most telling line comes from Challenger’s own data: AI is so far claiming the budgets for these roles rather than doing the work itself. The money once spent on people is being rerouted to GPUs in the hope the machines deliver. When Meta ties roughly eight thousand cuts directly to its AI infrastructure spending, and the big four hyperscalers plan some seven hundred and twenty-five billion dollars of capital expenditure this year, you are watching a wager, not a productivity gain. The bill has come due for human workers. The payoff is still a promise.

Final Thought 🚀

Britain is not a country that lacks invention. We built the lab the rest of the world feared, and we sit on the richest health dataset on earth. The thing we lack is the nerve to keep what we make.

We sold DeepMind for the price of a footballer. We are being talked into renting out our data to the people who already own our best AI. And when something brilliant does survive contact with the system, it waits seventeen years for permission to help anyone.

The machines will not save us from this. The flaw was never in the technology. It is in us, and no model has ever been written that can fix that.

Until next Sunday, David Richards MBE

The Floor Is Moving

David Richards MBE — Sun, 31 May 2026 05:30:57 GMT

Issue #55 · Week 22 · Sunday 31 May 2026

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Bottom Line Up Front

Whatever you believe about artificial intelligence, check the date on it. Most of what passes for informed opinion, in boardrooms, in Whitehall, in the better newspapers, was true a few months ago and is wrong now. That used to be a forgivable lag. In this field it is a disqualifying one. The technology no longer moves at the speed of quarterly reviews. It moves at the speed of the weekend, and the gap between what the machines can do and what the average decision-maker thinks they can do widens every Friday.

Here is the part that should concentrate the mind. The line that everyone assumed was years away, the one where the machine writes software better than the people paid to write it, was crossed late last year. Coding is the first knowledge profession where the substitution pattern became impossible to ignore. It will not be the last. Once the pattern is proven in one trade, every other trade that runs on rules and documents is simply waiting its turn. This week the proof arrived, the bill arrived, and the redundancy notices arrived. They arrived together. That is not a coincidence.

Subscribe now

Your mental model of AI expired while you were forming it

The clearest account of why this moment is different came not from a lab but from a venture capitalist. On 20 April, Sequoia Capital held the fourth edition of its AI Ascent in San Francisco, an invitation-only gathering of more than 150 of the people actually building this technology. Demis Hassabis was in the room. So were Andrej Karpathy, Greg Brockman and Anthropic’s Boris Cherny. The partner keynote that opened the day carried a title chosen to provoke, “This is AGI”, and it was Pat Grady, a partner at the firm, who set out the frame. He did it without the usual hyperbole, which is precisely what made it land.

His image for the shift was cars replacing horses. The last few years, he argued, gave us faster horses: applications that made you 10 or 40 per cent more productive without changing how you worked. Now the cars have arrived, tools that make you 10 or 40 times more productive and change the nature of the work itself. His framing of why is worth holding onto. Grady draws a line between two kinds of revolution. There are revolutions in communication, which change how information is distributed. The internet was one. Mobile was another. So was the cloud. Almost everyone alive has only ever lived through revolutions of that kind. AI, he argues, is something else. It is a revolution in computation, a change in how information is processed rather than moved, and that is a fundamentally different shape of wave.

He marks three discontinuities. The first was November 2022, when ChatGPT showed a mainstream audience the power of pre-training. The second came roughly two years later with o1, when a second scaling law appeared around inference-time compute and reasoning models started to think before they answered. The third is happening now, with Claude Code and the current Opus generation, which have demonstrated what long-horizon agents can do when you set them a job and leave them to it. Grady’s point is that these are not three even steps on a ramp. The jump from the second to the third is a hard break, a discontinuity, and most people have not registered that it occurred.

The phrase that lands hardest is his. The technology foundation, he says, is one where “the floor keeps moving underfoot.” Build something on Monday and the capability that made it clever can be commodity by Friday. For anyone whose mental model of AI was assembled even a season ago, the floor has already moved. You are standing somewhere that no longer exists.

Subscribe now

And the consequence he draws is blunt: “no lead is safe.” In a downpour of new capability, the incumbent advantage that felt permanent becomes a puddle. That is true of companies. It is also true of professions, and of the assumptions policymakers are using to regulate them. We are writing rules for a technology whose shape will have changed twice before the consultation closes.

Full keynote here:

One builder can now ship like a team of twenty

Start with coding, because coding is the canary, and the canary has stopped singing.

Think of a software engineer as the conductor of an orchestra. To conduct, you must be able to do two things at once: read the music, and play the instruments. Lose either and the performance collapses. For most of the history of the trade, the conductor also had to write out every note by hand. That is the part that has gone.

ReadMe File at https://github.com/garrytan/gstack

The most vivid proof point this week is Garry Tan, president and chief executive of Y Combinator, who has spent twenty years building products and is, on his own account, shipping more now than at any point in his career. In the README for his open-source toolkit, gstack, he lays out the numbers. In sixty days, working part-time while running YC, he shipped three production services and more than forty features. Measured by logical code change rather than raw lines (which AI inflates, as he is the first to admit), his 2026 run rate is around 810 times his 2013 pace. By mid-April his year had already produced 240 times the output of his entire 2013.

The number is arresting. The method is the lesson. Tan does not use AI as a faster pencil. He uses it as a company. gstack hands Claude a set of defined roles, twenty-three skills in all: a chief executive that challenges the premise of what is being built, an engineering manager that locks the architecture, a quality lead that opens a real browser, clicks through the staging app, finds the bugs and writes the regression tests. He opens the whole thing with a line from the researcher Andrej Karpathy, who said he had not typed a line of code since December. Tan’s own verdict on the productivity debate is the cleanest summary of the new reality you will read: “The point isn’t who typed it, it’s what shipped.”

Now the necessary cold water, because the 810x figure does not generalise and pretending it does would be dishonest. Tan is a senior builder working on greenfield code with no legacy to fight. That is the best case, not the average. Throw the same agents at a tangled enterprise codebase and they frequently slow people down, because someone now has to read, check and unpick what the machine produced. The industry has a name for it: the review tax. Across real organisations the honest multiple is closer to low single figures than to three. Gartner finds that roughly four in five firms piloting or deploying autonomous business capabilities have cut jobs, and that the financial returns are not yet showing up. The machine is not making everyone 800 times faster. It is letting one capable person orchestrate the work of a team, and it is removing the rungs of the ladder those teams used to be built from.

That distinction matters, because the same arithmetic is now being run in finance, in accounting and in law. The junior associate reading a data room. The analyst rebuilding a risk model. The bookkeeper reconciling a ledger at month end. None of them types in a different language from the engineer. They take data from one place, apply a set of rules, and put it somewhere else, and that is precisely the shape of work an agent does without complaint. The conductor’s job survives. The copying-out does not.

If your job is a task, your job is the target

The honest test is one you can run this weekend. Could a competent intern do your role from a three-page brief? If yes, an agent will do it autonomously, and sooner than your employer is telling you. The work does not vanish. It moves up a level, from doing to directing, and the people who make that move first will hold the leverage. Here is the playbook.

1. Audit your own week. For five working days, log every task in half-hour blocks, then mark each one against the intern test. The tasks that pass, the structured, repeatable, digital ones, are the targets. The tasks that needed empathy, judgment, negotiation or your signature on the line are your new job description. Find them before the org chart does.

2. Learn the shape of the thing replacing you. The firms moving fastest have stopped buying a single chatbot and started running crews. A supervisor agent takes the goal and breaks it down. Specialist agents do the narrow work: research, risk, drafting, formatting. A final agent cross-checks them and assembles the result. Your move is to stop being a specialist and become the supervisor, the one who designs the workflow, tunes the prompts and directs the digital team.

3. Plant yourself at the review tax. As execution gets cheap, verification becomes the bottleneck, and in regulated work it is the whole game. An agent cannot sign a tax return, file a brief or ship to production. You can. So become the accountable human in the loop, train yourself to spot the slop (the hallucinated precedent, the ghost number in the ledger, the race condition in the code), and own the liability the machine never will. Your licence and your signature are the leverage.

4. Productise what you know before someone else does. Do not wait for procurement to buy the tool that replaces you. Write down the unwritten rules in your head, how you smell a wrong number, how you sense a client going cold, and turn them into prompts and playbooks. Then walk into your manager’s office as a systems architect, not an order-taker: “I have automated sixty per cent of our reporting, and I am now free for the work that actually pays.”

5. Run yourself like a one-person firm. Your value was never the typing. It is the judgment, the foresight and the ability to direct a team of autonomous workers to a result you will put your name to. The task is dead. The workflow is everything.

The pivot looks different in every sector. The direction is identical.

Subscribe now

The technology is real. The economics are a wager.

Here is where the story turns, because two things can be true at once. The technology is the real thing. The money around it is not yet behaving rationally.

Consider the cautionary tale of the week. An AI consultant told Axios that an unnamed enterprise client reportedly ran up a bill of half a billion dollars on Claude in one month, after giving thousands of staff unrestricted access with no spending caps and no monitoring. Employees left agents looping overnight. People burned premium models on trivia. By the time the invoice landed, the budgetary damage was done. It may be the most expensive AI governance failure yet reported, and it tells you something the marketing does not: this is a consumption utility billed by the token, not a flat-rate subscription, and most of corporate IT has not understood the difference. The same week, Microsoft, the company most committed to embedding AI everywhere, reportedly cancelled most of its internal Claude Code licences on cost grounds. When the believers start flinching at the meter, pay attention.

Now lift your eyes from the invoice to the balance sheet. The four largest hyperscalers are on course to spend around 725 billion dollars on AI infrastructure this year, by the Financial Times’ tally, up roughly 77 per cent on last year and the largest concentrated infrastructure cycle in the history of technology. Amazon’s free cash flow is forecast to turn negative. A depreciation wave is coming as those chips age, with Morgan Stanley putting it north of 680 billion dollars over four years. Markets have started to notice: more than a trillion dollars came off the combined value of the big names in a single week of trading earlier this year. On the most generous reading of the returns, only the cloud-driven players, Amazon and Google, are clearly earning their keep. The rest are spending against a future demand they are betting will arrive.

This is exactly why I keep returning to the dot-com era. The technology was real then too. The internet survived and remade the world. Most of the companies that borrowed against it did not. Incredible engineering does not guarantee sustainable economics, and right now the giants are pouring trillions into the ground in the hope that revenue catches up to capex. That is not certainty. It is a leveraged bet, and the people most exposed when the interest comes due are not the executives who placed it.

Which brings us, with grim logic, to who pays. The chief executive of CloudBees told Axios the quiet part plainly: for many firms, cutting jobs may be the only lever left to offset the AI bill. The machines are sold as the reason for the layoffs. Increasingly they are also the cost that makes the layoffs necessary. Read that twice.

Subscribe now

The algorithm in Thatcher’s coat

A version of my Yorkshire Post column this week.

On Tuesday 19 May, in Hong Kong, Bill Winters, chief executive of Standard Chartered, told investors the bank would shed roughly 7,800 corporate and support roles by 2030. In the context of global banking, the figure was not the shock. The framing was. The cuts, Winters explained, were not “cost-cutting” but “replacing in some cases lower-value human capital” with the financial and investment capital the bank was putting in. By Wednesday morning a memo was scrambling to bury the phrase as reflecting “the work, not the value of our people,” and Friday brought a formal apology and talk of mature discussion. Wednesday’s memo was corporate laundering. Tuesday was the boardroom speaking with the microphone left on.

Many of us in the AI industry watched this inflection arrive late last year, and the moment was uncomfortably specific. Claude Code, Anthropic’s coding agent, became plainly better than the median engineer at bounded, testable coding tasks, and was visibly closing on the best. Once that line falls in software, every other knowledge profession sits on a countdown, which is why I started tracking it. The Sunday Signal Layoff Tracker now publishes weekly, monitoring the reallocation of corporate capital from human payroll to AI compute.

What the data shows is sobering. Twenty-two weeks into 2026, the tracker has logged 472,150 corporate job cuts globally, of which roughly 169,430 sit in technology, and if you run that velocity through to December the year lands at 1.1 million, the tech share alone burying the sector’s previous worst year on record (2023, around 263,000). Nor is the curve flattening. Q3 and Q4 historically run hotter than spring as boards rush to right-size ahead of earnings, and roughly 135 billion dollars of fresh AI capex is queued up for the back half, meaning the 2023 record will not merely be broken but pulverised. These are not assembly-line jobs. They are middle-class and salaried: the compliance analyst running KYC, the junior associate annotating contracts, the credit analyst rebuilding risk models, the team lead summarising five reports for the executive one rung up.

Yorkshire of all places ought to recognise the pattern. We have seen what happens when an entire class of work vanishes inside a single decade. The coalfields and the steelworks did not gently retrain into something else. They emptied, and the social damage took three generations to absorb. This time the axe falls not at the pithead or the furnace, but in the financial back office. Financial and insurance services employ about 1.17 million people across the UK, while the wider financial and related professional services industry employs more than 2.4 million, around two-thirds of them outside London. The ONS has just recorded financial and insurance activities as the worst-hit corner of the British economy, shedding 78,000 jobs over the past twelve months alone. This is digital deindustrialisation, and it is the algorithm wearing Thatcher’s coat.

The bitter irony is that the tracker itself is the work of an AI agent. He is called Mac, and he triangulates labour aggregators, SEC 8-K filings, WARN notices and the financial press, cross-referencing headcount against primary disclosures so the PR varnish lifts off. Two years ago, the same brief handed to McKinsey would have arrived as a “Global Labour Optimisation Engagement”: a five-person pod, an offshore researcher in Bengaluru pulling overnight WARN data, and a partner gracing the deck for an hour a week, all invoicing out at half a million pounds a month. Mac does the lot in four seconds, and the analysts who used to build those Excel models are, with painful precision, the demographic the tracker watches disappear.

I have argued before that capital does not protect anyone through this transition. Only those who use AI to create and build will. The honest question facing British boardrooms is no longer whether artificial intelligence will reshape the workforce, but whether you intend to be the executive saying the quiet part out loud, or the one preparing your people for the work that remains.

The Sunday Signal Tech & AI Layoff Tracker

The signal: the rise of the 100x org. Last week’s Standard Chartered cuts were a warning to the back office. This week the strike landed on the builders. The idea that software engineers were safe died between Sunday and Friday, as both mid-cap SaaS and multi-billion-dollar names converged on a single philosophy. They no longer want teams writing raw syntax. They want a handful of system managers commanding fleets of agents.

The clearest statement of it came from ClickUp. The four-billion-dollar productivity firm cut 22 per cent of its staff, roughly 290 of about 1,300 people, and framed it not as a saving but as a rebuild. Chief executive Zeb Evans says the company now runs around 3,000 internal AI agents, a three-to-one ratio of agents to employees, and that his best engineers no longer write code, they direct the agents that write it. The payroll freed by the cuts is being redirected into cash salary bands of up to a million dollars for the ones who can manage the machines. Read that as the new employment contract, stated out loud.

Elsewhere on the wire this week. Website builder Wix announced the largest cut in its history, around 1,000 roles, some 20 per cent of its workforce; management cited currency pressure, though the structural driver, generative tools that build sites from a prompt, is hard to miss. Logistics software group WiseTech Global began a restructuring that could remove roughly 30 per cent of staff, with the first notices now landing, amid internal commentary reportedly weighing the cost of human coding against the cost of compute. UBS entered the European close with a Bloomberg-reported wave of several hundred cuts across EMEA, officially filed under the long-running Credit Suisse integration. A class-action firm in Washington state has opened an investigation into whether Meta‘s latest Washington cuts complied with state WARN notice requirements. And Groupon announced a 23 per cent global reduction, around 400 roles, as the legacy consumer web continues to shed its pandemic-era headcount.

Watch next week. ClickUp has set the benchmark, and the rest of the productivity sector (Asana, Monday.com, Smartsheet) will now be asked by investors why they have not matched its margins. Expect defensive restructurings before the quarter closes. Web design and hosting agencies are the other soft spot, as generative site builders pull the floor out from under standard front-end work.

Subscribe now

Final Thought 🚀

The data this week did not describe a transition. It described substitution. The technology is real, the economics are a gamble, and the bill for that gamble is being paid in other people’s jobs while the people who placed it apologise for their choice of words rather than their choice of action. The coalfields were told to retrain too. Three generations later, parts of this county are still waiting. You have been warned, in plain language, by the executives themselves. The only question is whether you were listening, or still reading last month’s map.

Until next Sunday, David

The Sunday Signal | Mid-Week Flash

David Richards MBE — Wed, 27 May 2026 09:31:42 GMT

The full Week 22 issue lands this Sunday at thesundaysignal.ai. Subscribe so it arrives in your inbox. This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Bottom Line Up Front

Subscribe now

Two stories arrived this week from opposite ends of the same machine. A website builder said it no longer needs a fifth of its workforce because the AI tools its customers can now reach for directly do the work. The chief operating officer of one of the world’s largest tech companies said he cannot draw a line from his engineers’ vast AI bills to anything customers can see. The producer is being eaten. The buyer cannot find what it bought. Both stories are in Sunday’s issue.

Wix laid off the developers who built Wix

Wix confirmed on Monday it will cut around 1,000 jobs, roughly twenty per cent of its workforce. It is the largest reduction in the company’s twenty-year history. The company employed 5,277 people at the end of the first quarter, with more than sixty per cent based in Israel. The stock has lost about half its value this year. The first quarter posted a loss of $57.5 million on revenue of $541 million, with operating expenses up fifty per cent.

The headline number matters less than what management told staff. Per Israeli press reporting, internal communications were direct. Several development and design roles are no longer required because AI tools now perform that work. Wix built its business on giving non-technical users a drag-and-drop way to make a website. The engineering and design layer behind that drag-and-drop is the layer being cut.

There is one more turn. The fastest growing business inside Wix is Base44, the natural-language software-building platform it bought last year for $80 million. Its annual recurring revenue hit $150 million in May. A user types what they want. Base44 builds it. The product growing inside Wix is the product replacing the people inside Wix.

The argument no longer holds

For two decades the case for Wix was simple. Most people cannot build a website. Wix gives them a way to do it without learning code. That case rested on a particular bet, which is that the abstraction layer would remain useful, while the people building underneath it remained needed.

Both halves of that bet are now in question. The user who once needed Wix can open Lovable, Cursor, Claude Code, Base44, or any of the new “type-what-you-want” tools and get a working site in minutes. The developer who once built Wix is being told the same AI tools that compete with the product now compete with the job. The platform that sat between the user and the developer no longer has obvious room on either side.

Week 22 tracker, early look

The cross-sector year-to-date total has crossed 470,000. The tech-specific count sits above 169,000. Since Sunday the trackers absorbed roughly 27,000 additions, a mix of new cuts and second-quarter reconciliations. Inside this week’s additions sit Wix’s 1,000 jobs, the next thousand realisations of WiseTech Global’s previously announced 30 per cent restructuring, and the rolling state-level WARN filings catching back-office cuts at Acrisure, Macy’s, CVS and Aetna. The year-to-date number is a tracker-based estimate and will be reconciled in Sunday’s full issue.

Uber paid for the tokens. It cannot find the features.

On Saturday, on the Rapid Response podcast, Uber’s president and chief operating officer Andrew Macdonald said in plain terms that the company’s vast use of AI coding tools is not visibly producing better consumer products. His words: “That link is not there yet.” Maybe more is shipping. Maybe it is not. He cannot trace it.

Behind the quote is the spend. Uber rolled out Claude Code to roughly 5,000 engineers in December. By March, eighty-four per cent of them were using agentic features, up from thirty-two per cent in February. By spring, ninety-five per cent of engineers used AI tools every month. Roughly seventy per cent of committed code is now AI-assisted. Heavy users rack up bills between $500 and $2,000 a month. In April, the chief technology officer disclosed that the company had already burned through its entire 2026 Claude Code budget. Four months in. Macdonald called it a “head-exploding moment”.

The internal incentive was a leaderboard. Teams were ranked by total AI tool usage. The metric being optimised was consumption. The output was tokens, not features. Macdonald has now said publicly what the design should have made obvious. If you reward people for using a tool, you get more tool use. You do not necessarily get more product. The receipt is impressive. The customer cannot see what it bought.

The buyers are starting to look at the receipts

Microsoft is reported to be cancelling most of its direct Claude Code licences and routing engineers to GitHub Copilot’s command-line tool. Duolingo’s chief executive has walked back his AI-replacement language from last year. Sam Altman has begun publicly suggesting some businesses may be using AI as a convenient label for cuts that would have happened anyway. The euphoria phase is ending. The receipts phase has begun.

For Anthropic this is mixed. Uber is one of its largest enterprise customers, and Uber’s experience proves the adoption is real. It also proves the usage-based pricing model is being aimed at companies that can blow a year’s budget in a third of a year and still struggle to explain what they got. That is not yet a customer leaving. It is a customer counting.

Coming Sunday

Two stories that are really one. Wix laid off a fifth of itself because the AI tools its users now reach for have replaced the developers inside the company that built the abstraction. Uber’s chief operating officer cannot connect his enormous AI coding spend to anything the customer sees. On one side of the market the producer is being eaten. On the other side the buyer is asking what dinner cost.

Light, the Grid, and the British Company That Might Break the AI Energy Trade

David Richards MBE — Sun, 24 May 2026 06:02:32 GMT

Sunday, 24 May 2026.

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

The full interview with Dr Xianxin Guo, CEO of Lumai, is published separately on Substack today. Read it here.

Bottom Line Up Front

The whole AI story this decade has rested on a single assumption. That the machine is fixed, the energy bill is rising, and the only choice left is whether to build more power stations or accept the limit. That assumption is wrong, and a company in Oxford has the receipts to prove it. The implications run further than the data centre. They reach the trillions of dollars currently sitting on the bet that AI will burn through everything we can generate, and they reach the British question this newsletter has been worrying at for months: can we, this time, keep the thing we invented.

Subscribe now

The AI energy crisis is a physics problem in an economics costume

Every argument about AI and power ends in the same place. Demand doubles by 2030. The grid cannot carry it. Build more, burn more, and hope the curve bends. That is a supply argument. It assumes the machine is fixed and the only variable is how much we feed it.

It is the wrong argument. The energy problem is not really about supply. It is about what we built the machine out of.

Electrons resist. That is the whole story

A chip computes by pushing electrons through transistors and switching them on and off billions of times a second. Every switch costs energy. Moving the data between memory and processor costs more energy than the sums themselves. Almost all of it leaves as heat. That is why a modern AI data centre spends a punishing share of its power not on thinking but on staying cool. The electron is a workhorse that throws off heat with every step. We have spent sixty years making it take smaller steps. We are running out of room.

Photons do not behave that way. Light moving through an optical component does not dissipate heat the way a switching transistor does, and it can carry many calculations at once, in parallel, through space and across wavelengths. The heaviest single job in AI inference is matrix multiplication, the same arithmetic over and over at vast scale. That job maps almost exactly onto what light does without being forced. The dominant power cost in this kind of optical system is not the optical multiplication itself, but the conversion between the electrical and optical domains at the edges. The efficiency lever, then, is to do as much maths as possible per conversion. Lumai says its optical tensor engine performs matrix multiplication up to 2048 by 2048. Bigger matrices per conversion means more sums per joule. That is where the order-of-magnitude saving actually comes from.

The company that bet everything on that sentence

The company is Lumai. It is a spinout from the University of Oxford, headquartered in Oxford, and last month it launched a server called Iris Nova. By its own account, it is the only system in the world running billion-parameter large language models on optical compute, currently doing real-time inference on Meta’s Llama 8B and 70B.

I first saw Lumai a fortnight ago at the Royal Academy of Engineering Enterprise Hub Demo Day. Dr Xianxin Guo was on stage, beside a slide of the Lumai Iris Server. The slide carried two forward claims against the hardware the entire industry currently runs on. Ninety-five per cent less energy than an Nvidia GB200 by 2029. Ninety per cent lower cost than Nvidia by 2029. Aggressive numbers, the kind you note and then chase. I followed up.

I sat down with him this week to ask the questions readers of this newsletter actually care about. Power. Cost. Speed of adoption. And whether Britain keeps it.

His one-line answer to what the company does, in language any data centre operator would recognise: ninety per cent less energy, ten per cent of the cost, more AI tokens out of the same hall.

Dr Xianxin Guo Presenting at The Royal Academy

The ninety per cent, translated

“Ninety per cent reduction of the power needed for the same compute at the rack appliance level,” Guo says. Less power means less cooling, smaller power supplies, less of the secondary plant that quietly drinks a material share of a data centre’s electricity. The saving ripples. Networking and other elements outside Lumai’s box are not affected, so the figure at the grid connection comes in below ninety, but Guo is direct about what the figure actually buys you. “What this is likely to mean in practice is more compute capacity within the same power budget.”

That is the honest version of the answer. The energy saving does not collapse the energy bill. It expands the amount of intelligence the same energy bill can produce. He is open about why.

“Jevons’ paradox is real, and we will see more AI used, not less.”

This is the part the optimists tend to leave out. Efficiency does not bend the curve. It steepens it. The grid still has to grow. What changes is what we get for it.

Cost per token, not cost per box

“Cost per token. That is the only metric that matters to an operator running this at scale,” Guo says when pressed on what the ninety per cent saving actually means in pounds. The figure is a blend of equipment capital cost and the capital cost of the data centre that hosts it. Power. Backup generators. Cooling. The whole sleeve of infrastructure that exists because silicon throws off heat.

The model that lands first is unsurprising. Operators with the highest token volumes and the heaviest cost exposure to inference feel the saving first. That is the hyperscaler. That is the neocloud. That is the customer.

The capital question is sharper. Lumai has raised about fifteen million dollars. American rivals carry war chests forty times that. Guo’s answer is that the architecture itself bends the capital problem.

“We are not building custom silicon at the leading chip node, because most of the hard computation is performed in optics.”

Translation: they do not need a TSMC-class fabrication line because the heavy lifting is not happening on a transistor. The architecture is lens-based and three-dimensional rather than integrated photonics on a 2D chip. Weights are encoded on an electronic display and the calculation happens as light passes through it. The supply chain leans on mature, high-volume components, the kind of optics and displays that are already manufactured at industrial scale elsewhere, customised for the use but not invented from scratch. It is a deliberate choice to sidestep the part of the hardware industry that eats capital fastest, the development of novel materials and bleeding-edge fabrication. They have just opened the next round.

How fast this moves

Iris Nova is in the hands of evaluators now. Guo is direct about the timeline from there. Eighteen to twenty-four months from first evaluation to production at scale at a hyperscaler. Then this:

“The true blocker isn’t the technology, it is validation. A hyperscaler will not deploy a novel architecture into a revenue-bearing workload without months of validation, integration testing, and operational confidence.”

That sentence should reassure rather than worry. The bottleneck is not whether the physics is real. It is whether the buyer can convince itself. The first real deployment will be a cluster in a national laboratory testing performance across multiple workloads and orchestration patterns. It is the necessary unglamorous step before the order book opens.

The form factor is built to make that confidence easy to give. Iris Nova is an Ethernet-attached appliance that drops into a standard rack and runs alongside existing GPUs rather than replacing them. Its natural fit is inference prefill and other compute-bound workloads. Nobody is being asked to rip Nvidia out.

The CUDA question, the moat that has protected Nvidia for a decade, gets a short answer too. Same frameworks, same models, no rewrite. Lumai compiles to its hardware and the model runs.

A pull quote that has nothing to do with chips

The most quietly striking line in the conversation was about leadership. Guo was Head of Research before he became CEO last year. What did he have to unlearn?

“A decision made in a week with 70 per cent certainty beats a decision made in three months with 95 per cent.”

A British deep-tech founder who has read the room about why British deep-tech keeps losing.

The full interview is published on Substack today, here.

What happens to the grid if nothing changes

The Lumai story matters because the alternative has now been forecast in detail by people who do not write newsletters for a living. Without an architectural pivot, here is the road we are on.

The global picture

The International Energy Agency puts global data centre electricity consumption at roughly 415 terawatt hours in 2024 and projects it to more than double to around 945 terawatt hours by 2030, driven heavily by AI. That takes data centres from about one and a half per cent of global electricity demand to roughly three per cent inside six years. Goldman Sachs Research projects a 165 per cent total increase in data centre power demand between 2023 and 2030. McKinsey estimates the global investment needed to physically build that capacity at nearly seven trillion dollars before the decade is out.

The unit of measurement has changed. Standard IT racks have drawn ten to fifteen kilowatts for years. AI-optimised racks now push past forty. A single next-generation AI campus currently under construction will demand around twenty times the power of a typical hyperscale site. The grid is not being asked for more of the same. It is being asked for something it has never had to deliver.

The British picture is sharper

In the UK, the boom is already crossing thresholds the system was not designed to cross. Recent industry estimates put UK data centres at around six per cent of electricity consumption, although the House of Commons Library’s 2025 briefing gives a lower figure of 2.5 per cent. Either way the trajectory is steep. The National Energy System Operator projects a fourfold increase in data centre electricity demand by 2030. Other UK forecasts go to fivefold. In London alone, 271 data centres are drawing 69 per cent more power than the capital’s 3.49 million homes put together.

The Department for Science, Innovation and Technology has had to revise its forecast for greenhouse gas emissions from AI data centres upward by a factor of 100. Not by a per cent. By a multiple. Some UK data centre developers are now being quoted grid connection wait times of up to ten years.

That is not a forecast. That is a queue.

The concentration is the problem

The numbers above are global averages. The actual strain is hyper-local. The United States is expected to move from three to four per cent of national power consumption today to between eight and twelve per cent by 2030. In Ireland, data centres already consume over twenty per cent of the national grid. In Singapore, it is nineteen.

This is what people mean when they talk about an energy wall. They mean specific countries and specific cities where the grid cannot, physically, supply what the next decade of AI demand is forecast to require.

What an aggressive pivot to photons actually changes

Now run the alternative. If the architectural shift this newsletter has been describing actually happens at scale by 2028, the forecasts above stop being inevitable.

There are two places the energy bill of an AI data centre is concentrated. The first is data movement. By early 2025, pushing data through copper interconnects was consuming close to thirty per cent of total AI cluster power. Co-packaged optics, which moves the light conversion onto the chip itself, is now in volume production between 2026 and 2028 and cuts switch and rack-level power by an estimated forty to fifty per cent. The second is the compute itself. Moving the matrix mathematics from electrons to photons delivers, on independent estimates, an order of magnitude improvement in energy efficiency. Some photonic processors are demonstrating thirty times the savings in narrow benchmarks. Then the cooling collapse, because if the chip throws off less heat the secondary plant that exists to keep it cool throws off less of itself.

If you apply a blended fifty to ninety per cent reduction across interconnects and inference compute, two scenarios open up.

Scenario one. The reprieve. If global AI token demand grows at exactly the rate current economic models project, the IEA’s 945 terawatt hour figure never gets reached. Demand plateaus inside the existing envelope. The ten-year UK grid connection queue loses one of its biggest accelerants. The seven trillion dollars in additional power infrastructure is not needed at that scale.

Scenario two. The Jevons explosion. We still hit 945 terawatt hours by 2030, because efficiency releases demand, but we get fifty times the AI intelligence out of it. The total power bill looks the same. The cost per token collapses toward a level that rewrites the economics of inference. Every operator still running legacy silicon is suddenly competing on infrastructure economics they cannot win.

Guo’s own answer pointed firmly at scenario two. The history of every previous efficiency revolution in computing points the same way. The energy bill rarely shrinks. The work the energy buys multiplies.

Either scenario makes one specific category of investor wrong about a great deal of money.

The trillion-dollar trade built on a single assumption

Wall Street has spent two years pricing one outcome into the market. That hyperscalers, locked into a brute-force scaling race against each other, would buy raw electricity at any premium because they had no choice. The whole “AI energy trade” rests on that single load-bearing thesis. If photonics breaks the assumption, the financial consequences are not gentle.

The collapse of the AI premium in utilities

Independent power producers, nuclear fleet operators, regional utilities. All have been trading at historically elevated multiples because analysts have priced in a 150 to 200 per cent increase in data centre demand by 2030. The premium is not in current cash flow. It is in the expected future cash flow from desperate, decade-long power purchase agreements.

The repricing event is identifiable. The moment a major hyperscaler announces that its 2030 power forecast has been cut by, say, seventy per cent because optical coprocessors are doing more of the work, the repricing begins. Base-load demand will not fall to today’s levels. The growth premium will be forced into question very quickly. Slow-growth utility valuations return. Anyone holding the trade on margin discovers the difference.

The cooling sector takes the worst of it

The most exposed corner of the market is thermal management. Liquid cooling specialists, immersion cooling startups, the hyperscale HVAC contractors. The whole sector exists because silicon GPUs are, for accounting purposes, very expensive space heaters. If the heat goes away, so does the addressable market. Operators do not need next-generation cooling architecture for a chip that is barely warm. The pricing in this corner has not begun to discount that risk.

The copper downgrade

The copper trade is leaning on two stories, electric vehicles and AI grid expansion. The second of those is contingent. Optical computing replaces copper at the chip, the rack and the cluster. If grid expansion is paused because power demand plateaus, the projected supply deficits ease, and the mining and commodities exposure that has been priced for a structural deficit corrects.

Where the money goes

Capital does not vanish in repricings. It rotates. The destination, if this scenario plays out, is the optical supply chain. The mature manufacturers of lasers, modulators, waveguides, the photonic foundries, the firms quietly making the components that startups like Lumai assemble into systems. They are the picks and shovels of the next phase. Anyone trying to identify them in retrospect will pay a multiple to do so.

I am not in the business of giving investment advice. I am in the business of pointing out when a market is priced for a future that one British company is currently building the alternative to. Do the work yourself.

Why Lumai must stay a British company

This is the part of the story this newsletter cares about most, because we have watched it play out before. The UK invents the thing. American capital buys it. The headlines fade. The technology becomes the basis of a US-listed giant. Britain is thanked for the talent and reduced, politely, to a regional R&D office.

The pattern is now well documented enough to be predictable.

DeepMind, the most important AI research company in the world, came out of University College London. In 2014 Google bought it, pre-revenue, for around 400 million pounds. The acquisition was barely commented on at the time. It became the foundation of what is now Google DeepMind, the engine of one of the most valuable companies on the planet.

VocalIQ, spun out of Cambridge’s Spoken Dialogue Systems Group, was acquired by Apple in 2015 to become the conversational intelligence inside Siri. Magic Pony Technology, founded by London university graduates and seeded with a grant from Innovate UK, was bought by Twitter in 2016 for around 150 million dollars, two years after starting. Latent Logic, an Oxford spinout in autonomous decision-making, was acquired by Alphabet’s Waymo in 2019.

That is the soft pattern. The harder pattern is more recent and more damning.

In September 2025, Oxford Ionics, the trapped-ion quantum computing spinout, was acquired by the US-listed IonQ for 1.075 billion dollars. It was the highest-value Oxford quantum exit in history. The founders are expected to stay in Oxford. The IP, the headquarters, the future market capitalisation, and the corporate tax base now belong to a Maryland company. In the same announcement, IonQ also bought Vector Atomic, a quantum sensing firm with more than 200 million dollars in US defence contracts. Britain handed a sovereign-grade quantum capability to a defence-adjacent American competitor and clapped.

Oxford Quantum Circuits, the other Oxford quantum scaleup, raised the largest UK Series A in quantum, 38 million pounds, co-led by Lansdowne Partners and Japan’s largest deep-tech fund. British Patient Capital put in 6.7 million of it. The rest came from abroad.

The fundamental issue is not invention. Britain invents prolifically. The fundamental issue is scaling capital. When the moment comes to write the cheque that takes a company from breakthrough to market dominance, the British cheque is either not there or not large enough. American capital is. The technology, the headquarters, the tax base and the strategic capability follow the cheque.

This is what economists call the public risk, foreign reward deficit. The British state and the university system absorb the high-risk research phase. Once the technology is de-risked, American capital buys the certainty. The returns are then booked on NASDAQ.

Guo knows all of this. His answer when asked if Lumai stays British was the one this newsletter has been waiting to hear.

“Where the company is built is the question that matters, and the answer is the UK.”

His specific demand of the British state was equally clear. Not subsidy. Procurement.

“A committed UK government anchor purchase of Lumai Iris server hardware into the AI Research Resource. Concrete orders from a sovereign customer do more for keeping the company here than any number of strategy papers.”

ARIA‘s Scaling Inference Lab is a real testbed for novel AI hardware. The Sovereign AI programme has half a billion pounds behind British AI companies trying to scale. The Department for Science, Innovation and Technology has a hundred-million-pound package that explicitly positions government as a first customer for high-quality UK AI hardware. The instruments exist. The question is whether they are deployed in time, or whether a year from now we will be reading the press release announcing Lumai’s acquisition by an American hyperscaler.

This is not hostile to America. This is observable to anyone with a chronology and an attention span.

The Sunday Signal Tech and AI Layoff Tracker

Week 21. 17 to 23 May 2026.

The mask slips on “lower-value human capital”

For months we have tracked executives carefully constructing a PR shield around AI-driven job losses, relying on sanitary phrases like “operational leverage.” This week the shield was shattered by a 170-year-old British bank.

Standard Chartered announced a phased target to eliminate roughly 7,800 corporate and support roles, fifteen per cent of that division, by 2030 to fund an aggressive AI integration. The headcount was not the story. The language was. CEO Bill Winters told investors the bank was replacing “lower-value human capital” with the bank’s investment in technology and AI.

The backlash was instant. Winters issued a Wednesday memo attempting to walk back the phrasing, telling staff the words had been taken out of context and that role losses reflected the changing nature of the work, not the value of the people doing it. The memo is classic corporate laundering. The investor call was the unvarnished reality of the 2026 board room. “Lower-value human capital” is exactly how the C-suite now views non-technical administrative labour in the age of large language models.

The collapse of offshore arbitrage

The most critical takeaway is geographic. The planned cuts are disproportionately targeted at the bank’s major back-office hubs: Chennai, Bengaluru, Kuala Lumpur and Warsaw. Treat the specific list as indicative until the bank confirms it.

This signals the death of the defining corporate strategy of the last quarter century. Labour arbitrage. For decades, Western corporations saved billions by shipping administrative tasks to business process outsourcing centres in Asia and Eastern Europe. In 2026 the maths has flipped. Running an AI agent on a localised server is now cheaper than paying an offshore human analyst. AI does not require a desk. It does not sleep. It does not care about time zones.

Reality check this week: Intuit, Meta, Dune

While Standard Chartered laid out a multi-year roadmap, the immediate tech sector continued its contraction.

Intuit. The AI flush. On 20 May, the financial software giant announced the elimination of roughly 3,000 jobs, seventeen per cent of its workforce. This is a textbook AI-native restructure: clearing legacy overhead to rehire directly into AI engineering and product roles.

Meta. The execution phase. The mega-layoff Zuckerberg announced earlier this month is no longer theoretical. Rolling cuts began on 20 May, actively eliminating 8,000 roles, ten per cent of the workforce, to free up capital for the company’s $135 billion AI infrastructure buildout.

Dune Analytics. The Web3 pivot. This crypto analytics platform slashed twenty-five per cent of its staff this week. The stated rationale was a strategic pivot toward “AI and institutional onchain data.” When a Web3 startup and a colonial-era British bank cite the same technological driver to cut headcount, it stops being a coincidence. It becomes a sector-agnostic law of physics.

High-probability targets, Week 22

The global IT services giants. Standard Chartered just fired a warning shot directly over the bow of the major Indian IT hubs. The largest BPO and service providers, TCS, Infosys, Wipro, are now in acute danger. If their Western banking clients are successfully automating their own back-office compliance functions, the outsourcing contracts that sustain these IT giants will dry up. Expect defensive, margin-protecting layoffs from the BPO sector to accelerate immediately.

Mid-tier financial compliance. Any vendor built around providing human-in-the-loop KYC and AML verification is operating on borrowed time. The ecosystem servicing these manual workflows will be forced to downsize before Q3.

Methodology. The Standard Chartered figure is an announced multi-year target to 2030, recorded under broader corporate. The arithmetic of around 7,800 is derived from the bank’s stated fifteen per cent of roughly 52,000 staff. Broader corporate figures draw on LayoffHedge. Tech-specific figures draw on Layoffs.fyi and TrueUp.

Final Thought

🚀 We have spent three years arguing about how to feed the machine. More power stations. More grid. More money set on fire to keep the lights on in the server hall. The more useful question was always the quieter one. Not how much energy the machine needs, but what we chose to build it out of.

One answer to that question is sitting in a laboratory in Oxford. The CEO is a physicist who decided he would rather build the company than watch someone else buy it. He has told the British state, in public, what would keep him here. He has put a number on it, a date on it, and a programme on it. There is no possible defence for not hearing him.

His closing instruction to any policymaker reading this newsletter does not bother with diplomacy.

“Advanced procurement of UK-built AI hardware into the AI Research Resource and the Growth Infrastructure programme this financial year, or watch the next generation of AI compute be created outside of the UK again.”

This financial year. Not the next strategy review. Not the next white paper. This one.

The bank in the tracker called its own people lower-value capital. The British company at the top of this issue offered a way to make the machine that replaces them use a fraction of the energy. The cut and the fix are the same story. The only question left is who gets to own the fix.

Britain has watched DeepMind, Magic Pony, VocalIQ, Latent Logic and Oxford Ionics walk out of the door. Light is the next one. Buy the thing, or read the press release.

The Sunday Signal. Weekly, every Sunday. Subscribe at thesundaysignal.ai.

Listen to the podcast on Spotify, Apple Podcasts or YouTube.

The full interview with Dr Xianxin Guo, CEO of Lumai, is published in full on Substack today. Read it here.

Until next Sunday, David Richards MBE.

The full Sunday Signal interview with Dr Xianxin Guo, CEO of Lumai

David Richards MBE — Sun, 24 May 2026 06:02:19 GMT

“Where the company is built is the question that matters, and the answer is the UK.”

Listen to the podcast on Spotify, Apple Podcasts or YouTube. Subscribe to the newsletter at thesundaysignal.ai.

Dr Xianxin Guo is the CEO and co-founder of Lumai

Dr Xianxin Guo is the CEO and co-founder of Lumai, the Oxford spinout building the world’s first server running billion-parameter large language models on optical compute. Last month, the company launched Iris Nova, currently doing real-time inference on Meta’s Llama 8B and 70B. On its own figures, Iris uses ninety per cent less energy than the GPU hardware the entire AI industry currently runs on.

I first saw Lumai a fortnight ago at the Royal Academy of Engineering Enterprise Hub Demo Day. Guo was on stage, beside a slide of the Lumai Iris Server. The slide carried two forward claims against the hardware the entire industry currently runs on. Ninety-five per cent less energy than an Nvidia GB200 by 2029. Ninety per cent lower cost than Nvidia by 2029. Aggressive numbers, the kind you note and then chase. I followed up.

I sat down with him this week. Power, cost, adoption, capital, and whether Britain finally keeps something it invented. He gave the directest answers I have had from a British deep-tech founder in a long time. The whole exchange follows.

Subscribe now

What makes this different

The Sunday Signal: In one sentence a data centre operator would understand, what does Lumai do that silicon cannot?

By using optical compute to run LLMs and other AI workloads we consume ninety per cent less energy, meaning that a data centre operator can produce significantly more AI tokens within the data centre at just ten per cent of the cost.

Lightmatter, Salience, Celestial and others are all chasing light. What is genuinely unique about your 3D optical approach, and why has nobody else made it work at this scale?

Other companies have tried to perform optical compute in the confines of a 2D chip however they are unable to reach the scale required for data centre AI. Computing in 3D gives us orders of magnitude more parallelism, millions of computations per cycle. Building on the many years of research at Oxford, Lumai has now launched our first generation, Iris Nova, which is the only system to run billion-parameter LLMs using optical compute. A genuine first.

Why a plug-in coprocessor rather than a new supercomputer, and where are its limits?

Lumai’s Iris server is ideal for processing a high-volume of AI tokens efficiently. Lumai Iris is an Ethernet-attached appliance that fits in a standard rack, and runs alongside GPUs. Our sweet spot is inference prefill and other compute-bound workloads.

Do I have to rewrite my models to use an Iris server, or does it plug straight into PyTorch and the standard frameworks?

It plugs straight in. Users can take a pre-trained model, compile it for Lumai hardware and off they go.

The people behind it

You spun this out of Oxford. Who does what?

I lead the company. Dr James Spall, our CTO, runs the technology. He and I built the original optical system together in the lab. We work with an amazing team who bring a vast amount of technical and product experience from the likes of Arm, Intel, Altera, Meta, Imagination Technology, Cisco and HPE.

You moved from Head of Research to CEO. What did you have to unlearn?

In research you optimise for being right. As a CEO you optimise for leading and being useful, to customers, to the team, to investors. The hardest thing to unlearn was accepting that a decision made in a week with 70 per cent certainty beats a decision made in three months with 95 per cent.

Can you actually build a manufacturing and deployment team in the UK, and at what point do you have to give up and hire in the US?

We already have a development team in the UK which has an amazing talent pool and ecosystem covering electronics, software and optical. We will hire business development and account-facing teams in the US, so that they can be close to customers. We have a global supply chain, including in the UK.

What ninety per cent less power actually means

Dr Xianxin Guo Presenting at The Royal Acdemy

Translate the ninety per cent figure. Rack, data centre, grid connection.

Ninety per cent reduction of the power needed for the same compute at the rack appliance level. Less power means less cooling, smaller power supplies, etc. which means that this ninety per cent ripples across the data centre. There are other elements within a data centre that we don’t affect (e.g. networking), so the overall reduction in power at the grid connection will be less than ninety per cent, but still incredibly impressive. What this is likely to mean in practice however is more compute capacity within the same power budget as the need for more tokens continues to grow.

Does technology like yours bend the IEA curve, or does cheaper inference just pull more demand forward?

AI is creating so much demand for compute that Lumai’s highly efficient processor will be used to deliver much more inference within the same data centre power budget. Cheaper inference will pull demand forward. Jevons’ paradox is real and we will see more AI used, not less.

If optical compute becomes the default for inference, what changes for where data centres get built?

In principle data centres could be built with smaller power requirements when using Lumai Iris servers, making a wider range of sites open for AI. Compute demand and wider geopolitical issues will more likely have a greater influence.

What it means for the cost of AI

If your numbers hold, what does that do to the price of running a frontier model, and who feels it first?

Those with high token usage limited by volume of tokens and their cost will feel the benefits first.

Define the metric. Is ninety per cent lower cost capital cost per server, or operator’s cost per token?

Cost per token. That is the only metric that matters to an operator running this at scale. This cost is a blend of equipment capital cost and data centre capital cost (power, backup generators, cooling, etc).

Lumai has raised a little over fifteen million dollars. American rivals have war chests forty times that. How do you fund the path to manufacturing scale?

We are not building custom silicon at the leading chip node as most of the hard computation is performed in optics. Our optical supply chain is a deliberate choice to use mature, high-volume components from established optical suppliers, which reduces the need to develop high-cost novel materials. We have just opened our next round to take us through to volume manufacturing.

How fast this moves

How quickly can optical compute go from evaluation to running production inference at a hyperscaler? What is the true blocker?

We are already deeply embedded in deep technical evaluation with hyperscalers, neoclouds, and research labs. It will then realistically take eighteen to twenty-four months from first evaluation to production at scale. The true blocker isn’t the technology, it is validation. A hyperscaler will not deploy a novel architecture into a revenue-bearing workload without months of validation, integration testing, and operational confidence. We have shipped Iris Nova for exactly this reason. It is the proof that the system works, runs real models, and behaves in a rack.

Without breaking confidences, who is testing it, and what does a first real deployment look like?

Our first deployment will be a cluster in a national lab who will test performance across a range of customer workloads, with different orchestration, and different models.

What does an operator actually have to give up to put an Iris server in, and how big is the switching cost?

It’s more about what they gain rather than what they give up. We have designed Lumai Iris to make access to these benefits as easy as possible, the same software frameworks, same models, just faster driven by more compute. The cost is the same as qualifying any new hardware vendor. Lumai’s advantage is that we can work alongside existing hardware to optimise each stage of the inference pipeline.

Subscribe now

Keeping it in Britain

Your round was led by a Cayman-registered fund. Your expansion is into the United States. Of your backers, IP Group is the British one. Is Lumai going to stay a British company?

Yes. We are headquartered in Oxford. Our R&D, our engineering, and our senior team are here. We will grow overseas where it makes sense, but where the company is built is the question that matters, and the answer is the UK.

Does ARIA and sovereign compute translate into something that actually keeps a company like yours here, or is it words?

It is more than words. ARIA’s Scaling Inference Lab is a testbed for AI hardware technologies, prioritising rapid iteration, and open collaboration. This type of testbed is vital to any startup developing AI hardware and the fact that it is in the UK makes a massive difference.

Sovereign AI goes one step further. Investing five hundred million pounds in British AI companies to start up, scale up and win globally. This is in addition to the UK Government’s Department for Science, Innovation and Technology’s one hundred million pound funding package designed to “act as a first customer for promising UK startups who are building high-quality AI hardware products.”

Combined, these are exactly the sort of initiatives that will boost innovative startups like Lumai who are developing cutting-edge technology for such an important area of technology and UK growth.

Name the specific thing, a customer, a fab, a fund, that would make you put the next phase here rather than in the US.

A committed UK government anchor purchase of Lumai Iris server hardware into the AI Research Resource. Concrete orders from a sovereign customer do more for keeping the company here than any number of strategy papers.

To close

One sentence to a British policymaker reading this. What do they need to do, and by when?

Recognise that AI is just at the foothills of what is possible, that the UK has the talent and technology to benefit from this massive technological and economic shift, and that they need to do advanced procurement of UK-built AI hardware into the AI Research Resource and the Growth Infrastructure programme this financial year, or watch the next generation of AI compute be created outside of the UK again.

Read the full Issue #55 of The Sunday Signal at thesundaysignal.ai.

Listen on Spotify, Apple Podcasts or YouTube.

Until next Sunday, David Richards MBE.

The Sunday Signal | Mid-Week Flash

David Richards MBE — Tue, 19 May 2026 14:04:42 GMT

The full Week 21 issue lands this Sunday at thesundaysignal.ai. Subscribe so it arrives in your inbox. This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Bottom Line Up Front

The AI economy did two things in public today. It cut the jobs and it strained the grid. The same machine drives both. A bank with roots in the 1850s said the quiet part on the record, and separately, a British company said the energy wall everyone treats as inevitable is already breakable. One of those stories is a warning. The other might be the way out. Both are in Sunday’s issue.

The bank stopped pretending it was about cost

Standard Chartered confirmed today it will cut more than fifteen per cent of its corporate functions roles by 2030. Run the arithmetic the bank invited: fifteen per cent of its roughly 52,000 support staff is around 7,800 jobs, taken from a global workforce near 82,000. The bank wants income per employee up around twenty per cent by 2028.

What makes this the defining cut of May is not the headcount. It is the language. Chief executive Bill Winters told reporters this was not cost-cutting but “replacing in some cases lower-value human capital” with the bank’s investment in technology and AI. Read that back slowly. A FTSE 100 chief executive described his own staff as lower-value capital to be swapped for machines, in public, on the record, as the plan.

The offshore arbitrage is over

For two decades Western banks cut costs by moving back-office work to operational hubs in Asia and Eastern Europe. On current reporting, the centres most exposed are offshore: Chennai, Bengaluru, Kuala Lumpur, Warsaw. Treat the specific list as indicative until the bank confirms it.

The pattern matters more than the postcodes. The offshore centre was the last answer to the cost question. In 2026 the answer is an agent. When the inexpensive analyst in Bengaluru is no longer inexpensive enough against software, the outsourcing engine that quietly runs global banking moves into the line of fire. This is a different phase from tech firms trimming engineers. It reaches the people hired precisely because they were cheap.

Week 21 tracker, early look

The year-to-date number is a tracker-based estimate and will be reconciled in Sunday’s full issue.

And the other half of the story

Every argument about AI’s future now ends at the same wall. The power. The standard line is that demand doubles by 2030, the grid cannot carry it, and there is nothing to do but build more of everything and burn more to run it.

There is a company that says that wall is already breakable, and the reason is physics, not money. A chip computes by pushing electrons through silicon, and electrons throw off heat with every step. This company does the heavy mathematics of AI with light instead. Photons do not resist the way electrons do. The claim, on its own figures, is up to ninety per cent less energy than the hardware the entire industry currently depends on. It is not American. It is British, and it came out of a UK university.

I spoke with the person running it. The obvious questions are the hard ones. Are the numbers real or projected? Who is actually testing it? And the one that should bother every reader of this newsletter: Britain keeps inventing this, and America keeps owning it, so does this one stay here?

The name, the numbers and the verdict are in Sunday’s Signal.

Coming Sunday

Two stories that are really one. The bank that called its people lower-value capital, and the British company that thinks it can take the power problem off the table before America buys the answer.

The End of Adult Supervision

David Richards MBE — Sun, 17 May 2026 04:31:14 GMT

Issue #54 | The Sunday Signal | Sunday 17 May 2026

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Bottom Line Up Front

The “adult supervision” model of building companies is collapsing. For forty years, the moment a technical founder built something valuable, the institutional reflex was identical. Install a “proper” chief executive over them. Investors called it professionalisation. Universities called it commercialisation. Non-profit boards called it stewardship. The move never really worked. AI is about to make it look absurd.

This week, the argument played out in three rooms.

At The Digital Forge on Tuesday, in partnership with the University of Sheffield commercialisation team, a room of spin-out hopefuls heard the cautionary tale of a deep tech founder team forced to accept a sales executive as their CEO. The technology stalled. The cash burned. The founders eventually took the company back. The pattern is the rule, not the exception.

In a federal courtroom in Oakland on Thursday, closing arguments wrapped in Musk v. Altman. The jury begins deliberating on Monday. OpenAI is being asked to defend the most expensive version of the model ever attempted, the 501(c)(3) charter that was meant to keep frontier AI accountable. The verdict, when it lands, will become part of the architecture every AI lab studies.

And then the practical layer. Founder OS sits in this issue. The operating system that replaces the administrative machinery investors once used as the excuse to install adults in the first place. Nine playbooks. A master context architecture. A thirty-day implementation. Built around AI. Written for founders who have run out of patience with administrative theatre.

Three rooms. One verdict.

Subscribe now

1. The Adults in the Room Were Wrong

Britain has a strange habit. The moment a brilliant technical founder proves they can build something valuable, we start looking for someone else to run it.

At about fifty staff, the ritual begins. Investors mutter about “professionalising the team.” The founder is praised, thanked and moved sideways. A proper chief executive is installed, usually someone fluent in dashboards, board packs and the solemn theatre of management. For forty years, this looked sensible. AI is about to make it look absurd.

I wrote that for the Yorkshire Post on Friday. This week, the argument arrived in a room.

At The Digital Forge on Tuesday, we partnered with the commercialisation team at the University of Sheffield. The room was full of spin-out hopefuls. PhDs. Professors. Post-doctoral researchers. Every one of them thinking seriously about turning brilliant research into companies. Every one of them being asked, in one form or another, the same question. Who is the adult who will run this?

The traditional answer is over. Let me tell you why.

A few years ago I watched a company spin out of a university. The research was extraordinary. Genuinely original. Genuinely difficult. To understand the technology in any serious way, you needed to be an accomplished computer scientist. The technical founders were both experts. The university believed the company needed “adult supervision.” They installed a former sales executive from a large tech company. Not a CTO. Not a co-founder. A sales guy. The deep tech company had no product. It had no market. It had a research thesis that needed further engineering before it could be sold to anyone. Sales was the last problem on the list.

The predictable thing happened. The sales hire built sales scaffolding. Pipelines, decks, sales targets, an outside-in narrative. None of it accelerated the actual work. The cash burned down. The technology stalled. A few years later, after a near-death event, the technical founders took the company back and started over. Yep, you guessed it. With the technical founders at the helm.

I have sat through that story so many times, in so many universities, that I stopped finding it surprising. The pattern is the same every time. “Adult supervision” was never adult. It was administrative scaffolding dressed up as wisdom.

Marc Andreessen made the broader case in his 2011 Wall Street Journal essay. Software was eating the world. He has spent the last twelve months arguing that something similar is now happening to the management layer itself. His record on calling the direction of technology is among the best of his generation. He is worth taking seriously.

James Burnham saw it coming in 1941. Modern capitalism, he argued, would no longer be ruled by owners or inventors but by the people who administered the machine rather than the people who built it. Burnham was right for most of the twentieth century. Scale required bureaucracy. Bureaucracy required managers. Managers gradually persuaded everyone that they, rather than the founders, were the adults in the room.

AI now attacks the very work that made that class indispensable. Much of modern management is not leadership. It is translation, summarisation, co-ordination and documentation. The business of turning meetings into actions, actions into reports, reports into board packs, and board packs into the next meeting. Strategy decks, budget narratives, risk registers, performance reviews and the endless grey plumbing on which large organisations depend. All of it. Precisely what the new language models do well, more cheaply every quarter, without an office or a pension.

So the Sheffield spin-out story is the pattern in microcosm. The university genuinely believed those technical founders couldn’t run a company. That belief was wrong even before AI. Now it isn’t just wrong. It’s expensive.

A question now ought to be ringing around every boardroom in Britain. If software can do the management layer, what is the management layer actually for? The honest answer is much less than the managerial class has trained itself to believe.

The institutional version of the same failure played out this week in Oakland.

2. OpenAI and the Failure of Institutional Containment

Closing arguments in Musk v. Altman wrapped on Thursday in Oakland. The nine-person jury begins deliberating on Monday under Judge Yvonne Gonzalez Rogers. The jury is technically advisory. The judge has said she will very likely follow its verdict. Damages testimony begins in parallel on Monday, even as the jury deliberates liability. Musk has requested disgorgement of up to $134 billion and the removal of Sam Altman and Greg Brockman from OpenAI.

This is the case we covered three weeks ago, after Musk took the stand. Since then, the trial has turned. Hard. Three things have done the damage.

First, the statute of limitations. OpenAI argues Musk waited too long and cannot claim harms predating August 2021. That matters because the for-profit shift was being discussed years earlier, and a forensic accountant testified that all of Musk’s donations had been used by OpenAI well before that cutoff. Judge Gonzalez Rogers has signalled that if the jury finds the case was filed too late, she is highly likely to accept that finding and direct a verdict for the defendants. That alone could close the case.

Second, the absence of a founding agreement. The entire suit hinges on the claim that Musk, Altman and Brockman bound OpenAI to permanent non-profit status. No signed contract was produced in court. The case rests on emails, press statements and marketing prose from 2015 to 2018. In closing, OpenAI’s attorney Sarah Eddy told the jury Musk had come nowhere close to making that case. A major donor cannot retroactively manufacture binding legal restrictions out of casual correspondence.

Third, the Zilis texts. OpenAI’s defence introduced messages between Musk and Shivon Zilis showing Musk was, at the same time he was warning about for-profit AI, actively exploring turning OpenAI into a for-profit entity. Provided he controlled it. Or it merged with Tesla. The hypocrisy is now in the record. Musk’s “stolen charity” narrative cannot survive his own demand for control.

Add the courtroom theatre. Greg Brockman’s diary entries read aloud, including the description of Musk’s “tantrum” in 2018. The settlement attempt two days before trial, with Musk reportedly warning Brockman that he and Altman would be “the most hated men in America.” Musk himself absent for closing, reportedly in China with President Trump. None of it helped his case. All of it weakened the narrative of an altruistic founder defending a charity from operators.

Where this lands.

OpenAI is now heavily favoured. The statute of limitations defence alone could decide the case before the court reaches the moral argument. A split verdict or token compromise, where the judge orders OpenAI to increase the non-profit foundation’s equity share or fund AI safety initiatives, remains possible. A nuclear Musk win, with Altman ousted and the for-profit structure unwound, looks increasingly unlikely. The judge has been openly sceptical of Musk’s damages expert and cut him off repeatedly when his testimony drifted into AI doomsday speculation. The runway to OpenAI’s IPO clears either way.

The wider question matters more than the verdict.

What does this case do to the future of non-profit AI labs? Nothing good. OpenAI’s pivot proved that a 501(c)(3) charter is not a legally durable container for frontier AI research. The structure can be flipped, with sufficient legal wrapping, by the board. Major donors now know this. So do the foundations. So do the lawyers advising them.

Any technical founder considering a non-profit AI lab in 2026 is now operating with three pieces of new information. The original moral case for non-profit AI development was made by OpenAI and then abandoned by OpenAI. The legal protection a non-profit structure offers against later for-profit conversion is much weaker than its founders believed. And the only signed legal commitments that matter are the ones drafted by people who expected the company to remain a non-profit. Casual emails and press statements will not bind a future board.

The durable lesson of Musk v. Altman is not whether Sam Altman keeps his job. He almost certainly does. The durable lesson is that the non-profit AI lab, as a category, is now structurally fragile. The Public Benefit Corporation, with binding legal commitments to a mission, is the model that will be tested next. Anthropic remains one. So, increasingly, will the labs that follow.

The “adult supervision” model failed at the Sheffield spin-out. It failed at the start-up forced to install a sales executive. And it failed, at vastly greater cost, when a non-profit board persuaded itself that the for-profit pivot would still serve humanity. The shape of the failure is identical at every scale. The institutional answer is to stop pretending the structure protects the mission. Only people with real conviction, control and accountability can do that. In a founder-led company, that usually means the founder.

3. Founder OS: The Practical Answer

The obvious response to all this is: fine, but what replaces the management layer?

Not vibes. Not founder mythology. Not chaos dressed up as speed.

A system.

This week’s practical section is longer than usual for a reason. The old answer to founder scaling was management. The new answer is architecture.

Founder OS is my attempt to show what that architecture looks like in practice. It is aimed at founder-led companies between roughly five and a hundred staff. It will not replace judgment, courage, taste or accountability. It replaces the administrative machinery that investors once used as the excuse to install adults in the first place.

Keep it. Steal it. Adapt it.

What This System Cannot Replace

Before any of the playbooks, a clear boundary. AI eats the work that made management indispensable. It does not eat leadership.

Do not put Founder OS in the loop for:

Firing someone. A dismissal carries legal and moral weight. The system can help you document. It cannot make the call.
Co-founder disputes. No prompt resolves these. They get worked out in a room.
A major customer in crisis. Pick up the phone. Don’t draft a “founder voice” email.
A regulator, an injury, a press incident. Lawyer or comms adviser first. AI later.
The hard one-to-one where someone is quietly burning out. You will hear it. The transcript will not.

If this argument is right, more of your time should now go to these moments, not less. The OS is what gives you that time back.

Master Context Architecture

Most founders do this wrong. They start with prompts. Start with context.

Before you load a single playbook, create a core file named 00_FOUNDER_OS_MANIFESTO.md and place it permanently inside your master Claude Project as the first project knowledge file. This is the strategic anchor that keeps every output aligned with your taste and your conviction. Without it, you’re just generating well-formatted noise.

markdown

# FOUNDER OS: CORE CONTEXT & STRATEGIC ANCHOR

## 1. Core Vision & Non-Negotiables
- Our Unfair Advantage: [your technical moat or unique insight, two sentences]
- What We Protect At All Costs: [e.g., software performance, direct user relationships, design minimalism]
- What We Ruthlessly Prune: [e.g., feature bloat, vanity metrics, corporate meetings]

## 2. The Anti-Persona (the "Suits" Filter)
- Grey Plumbing: any process that rewards compliance over product velocity.
- Fluency over Judgement: a beautifully formatted deck or dashboard that hides flat user retention.

## 3. Communication Safeguards
- Founder Voice Attributes: [e.g., direct, lowercase-tolerant, outcome-focused, short sentences]
- Forbidden Phrases: "leverage synergies", "circle back", "stakeholder alignment", "moving forward", "robust ecosystem".

Security and Data Discipline

Don’t build Founder OS on vibes.

Use paid business accounts with proper admin controls.
Don’t upload customer personal data, employee records, source code, contracts or financials without reading the data terms, retention policy and access controls for each platform.
For self-hosted automation tools such as n8n, enforce SSL, SSO or 2FA, credential encryption, execution-data redaction, and regular audits.
Be honest about concentration risk. A single AI workspace holding board materials, hiring evaluations, financial models and customer calls is a juicy target. Compartmentalise the most sensitive material into a separate project with tighter access.

Universal Rule for Every Playbook

If the input is incomplete, do not invent missing facts. Mark the gap as [MISSING DATA: X] and continue only where the evidence supports the output. This rule applies to every playbook below, not just the financial one.

The Playbooks

Copy these into Claude Projects, ChatGPT, or Gemini. Keep them as persistent workspaces. Each one assumes your manifesto is already loaded.

Playbook 1: Decision Stress Test

Pressure-tests your own thinking before a real human conversation.

This is not an AI advisory board. A language model cannot replicate independent reputation, fiduciary duty or skin in the game. What it can do is force you through four lenses you might otherwise skip when moving fast. Use it as a thinking aid before a real conversation. Not as a substitute for one.

SYSTEM PROMPT: DECISION STRESS TEST

You are not a board. You are a structured interrogation tool helping me pressure-test
my own thinking. Do not flatter. Do not generate plausible-sounding strategic advice.
Where you do not have enough information, say so.

When I present a decision, run it through four lenses:

1. CASH: unit economics, runway impact, legal or contractual risk
2. TASTE: what this does to product quality and user experience
3. SECURITY: what breaks, what leaks, what gets exploited
4. INVERSION: the strongest case for the opposite decision

For each lens:
- State the strongest concern in one sentence
- Name the one assumption I am making that, if wrong, breaks the decision
- Flag where you lack information rather than filling the gap

End with: "Who would you want to actually speak to before deciding this?"
Name the type of person, not a persona.

Playbook 2: Translation and Documentation

Turns messy meetings into institutional assets in minutes.

PROMPT: MEETING TO INSTITUTIONAL ASSET

CONTEXT: I am the technical founder of a fast-moving startup. I have zero time
for managerial theatre.

INPUT: [paste raw transcript or brain dump]

TASK: Act as ultra-efficient Chief of Staff. Generate:
1. SPEED ACTION MATRIX (table: Task | Owner | Hard Deadline | Dependencies)
2. RISK REGISTER (Risk | Severity Low/Med/High | Mitigation)
3. EXECUTIVE DELTA BRIEF (4 sentences max, product velocity focus)

If a name is unclear in the transcript, mark [UNCLEAR: speaker] rather than guess.

Playbook 3: OKR Velocity Filter

Stops the company slowing to the speed of dashboards.

PROMPT: VELOCITY FILTER FOR GOAL SETTING

You are an expert in lean engineering leadership.
Here is a proposed list of quarterly goals or OKRs: [insert text]

Prune them using three constraints:
1. Does this drive actual product velocity or is it grey plumbing?
2. Does it reward fluency over judgement?
3. Is the metric something a customer would care about, or only an internal audience?

Redraft into a Conviction-First Framework: a maximum of three core company metrics
focused on deployment speed, system robustness, or customer validation.

Flag any goal where you cannot tell what success would actually look like.

Playbook 4: Hiring and Talent Filter

Stops “professional” HR theatre from diluting founder taste.

Read this before using the prompt. AI-assisted hiring carries compliance obligations. New York City requires bias audits. The EU AI Act treats employment-related AI as high-risk. UK data protection rules restrict automated decisions with legal effects. Use this to think, not to decide. The final call is a human’s. Keep records of how the human decided.

You are a technical founder's Talent Co-Pilot. I value speed, taste and technical
depth over credentials.

INPUT: [paste job description or candidate profile + interview notes]

TASK: Evaluate through three lenses only:
1. Conviction & Taste (evidence from their own work, not credentials)
2. Grey Plumbing Risk (will they import unnecessary process?)
3. Founder Multiplier (where does this person make the founder better, not busier?)

RECOMMENDATION: Advance / Strong No / Needs deeper test
Provide 3-sentence reasoning, red-flag questions, and a suggested compensation
shape, but only if I have given you data about our existing comp bands. Do not
invent a structure.

HARD RULES:
- Do not infer protected characteristics from name, photo, school, address or
  any other proxy.
- Do not use age, sex, race, disability, nationality, religion, family status,
  pregnancy or health as decision factors. If you notice these features in the
  input, ignore them.
- If the evidence is insufficient, say so plainly.
- The final decision must be made by a human interviewer, and this output is
  a thinking aid, not a recommendation of record.

Playbook 5: Weekly Executive and Investor Brief

Replaces the Chief of Staff’s Friday ritual.

You are my AI Chief of Staff. I hate corporate theatre.

INPUT: [paste this week's Linear tickets, customer calls, revenue numbers,
Slack highlights, or raw brain dump]

Generate:
1. ONE-PAGE EXECUTIVE DELTA BRIEF (5 bullets max: velocity, milestones, risks)
2. INVESTOR-READY NARRATIVE (4 sentences max, total candour)
3. ACTION MATRIX (table)

Tone: founder-to-founder, zero management-speak.
Mark anything you are inferring rather than reading directly from the input.

Playbook 6: Market and Competitor Intelligence Radar

Keeps your edge without hiring analysts. Without inventing the news either.

Language models fabricate competitor moves and funding rounds very fluently. This prompt forces source-grounding.

You are my Market Intelligence assistant. I value specificity and honesty over
confident speculation.

INPUT: [competitor announcement, funding news, or market shift. Paste the
actual source text or URL.]

CRITICAL: Do not assert facts about competitors, funding rounds, product launches
or market data unless they appear in the input or in a source you can cite. If
you do not have web access in this session, restrict your analysis to interpretation
of what I have given you. Do not generate new claims about the outside world.

TASK:
1. THREAT MAP: what this means for our product and positioning, given only
   what is in the input. Be explicit about what we know vs. what we are inferring.
2. ACTION RADAR: three concrete moves we could consider this week, ranked
   by conviction-to-effort ratio.
3. VERIFICATION LIST: three specific factual checks I should run before acting.

Playbook 7: Customer Feedback Synthesiser

Protects the founder’s direct line to the user.

You are my Product Taste Guardian. I am the founder who still talks to
customers personally.

INPUT: [paste raw customer calls, support tickets, or survey data]

TASK:
1. THEME CLUSTER (top 3 to 5 patterns with verbatim quotes)
2. TASTE FILTER: which themes align with our founding conviction, which
   are noise, and which are signals that conviction needs to evolve
3. PRIORITISATION MATRIX (impact on core user value vs. engineering effort)

Output as a clean decision brief I can act on in one sprint.
Flag any theme based on fewer than three independent customer voices.

Playbook 8: Founder-Voice Drafter

Keeps external communication unmistakably yours. Without crossing the line.

There is a line between drafting in your voice and misrepresenting who is communicating. If you wouldn’t sign a ghostwritten letter from your CEO to investors, don’t ship AI-generated investor updates as if you wrote them yourself. Use this for drafts you then read, edit and own.

You are my Voice Keeper. You have been given a representative sample of my
previous emails, investor updates, public writing and internal notes.

Do not imitate blindly. Preserve my directness, cadence, level of candour
and intolerance for corporate polish.

INPUT: [raw points or context for the email]

TASK: Draft the email in my voice: direct, conviction-driven, commercially
sharp, free of management-speak. Then add a one-line "why this tone works" note.

If the message involves financial commitments, hiring decisions, customer
promises or anything irreversible, mark it [REQUIRES FOUNDER READ-THROUGH]
at the top.

Playbook 9: Financial Forecasting and Board Narrative

Replaces the finance team’s quarterly theatre. With iron-clad anti-hallucination guardrails.

You are my AI CFO. Focus on runway reality, not spreadsheet theatre.

INPUT: [latest revenue, burn, pipeline data, or assumptions]

TASK:
1. 18-month runway model (base / stress / growth cases)
2. Board narrative (max 6 bullets, absolute candour)
3. One killer chart description I can drop into a deck
4. Assumptions register: every number tagged as Actual, Founder Assumption,
   Pipeline, Benchmark or Missing.

CRITICAL CONSTRAINT: Do not extrapolate or invent financial data. If specific
pipeline or cash inputs are missing, state [MISSING DATA: X] in the model
rather than estimating. Accuracy takes absolute precedence over formatting.
A board pack with honest gaps beats a board pack with confident fiction.

Founder Safeguards and the Investor Counter-Attack

AI is brilliant at polish. That’s precisely the trap. If an AI output doesn’t accelerate code, clarify customer value or sharpen your conviction, delete it. You may delegate the generation of a board pack. You may never delegate the courage to make the hard call.

When institutional investors demand a “professional” management layer, they’ll usually attack two things: key-person risk and institutional knowledge. The argument sounds reasonable. “If the founder gets hit by a bus, the company’s operating memory disappears.” Your answer is not bravado. It is architecture.

Show them the system. Decision logs, customer feedback loops, product rationale, weekly execution records, risk registers, hiring scorecards, investor updates, financial assumptions. All structured, searchable and continuously updated. Traditional middle management hoards context in private emails, undocumented calls, Slack fragments and personal notebooks. Founder OS turns that context into a living operating memory.

Be honest with yourself about the limits of this answer. Documents are not the same as judgment. The OS makes the company more legible to a successor, not equivalent to having one. If you are genuinely irreplaceable in product taste, customer relationships or technical direction, the system doesn’t fix that. But it does mean an investor concerned about continuity can no longer point to absent paperwork as the problem. The company becomes less dependent on hidden human bureaucracy, not more.

30-Day Implementation

Week 1: Foundations

Build the master Claude Project and upload 00_FOUNDER_OS_MANIFESTO.md first.
Set up n8n or Zapier Agents to pipe daily Slack digests and closed Linear tickets into a single append-only markdown file using the schema below. Drop it into your Claude Project every Friday. A thirty-second ritual:

markdown

### SYSTEM LOG: [WEEK_ENDING_YYYY_MM_DD]
- Engineering Velocity (Linear): [X] tickets closed, [Y] pushed to main, [Z] blocked.
- Signal vs. Noise (Slack Digest): [key technical debates summarised in 3 bullets].
- Ground Truth (Customer Pulse): [raw quotes from the 3 highest-value support tickets].
- Financial Reality: cash burned this week £[X] | current runway: [Y] months.

Week 2: Decision Discipline

Load Playbooks 1 to 3.
Run your first Decision Stress Test on a real, live decision. Not a hypothetical.
If the output flatters you, you’re doing it wrong. Sharpen the manifesto.

Week 3: Operating Cadence

Add Playbooks 4 to 7.
Run one hiring evaluation, one customer feedback synthesis, one market intelligence scan.
Notice where the system saves you time, and where it just shifts work around.

Week 4: Communication and Capital

Add Playbooks 8 and 9.
Run your first full weekly cycle.
Send one investor update generated through the system. Read it twice. Own every sentence.

Measure the system brutally. If it hasn’t removed hours from meetings, follow-ups, reporting, hiring admin and investor updates within thirty days, it isn’t an operating system. It’s just another dashboard.

Strip the administrative scaffolding away and the qualities a founder is actually hired for have nowhere left to hide. Taste. Courage. Judgement. Accountability. Founder OS is not a way to avoid them. It is a way to spend your time on the part of the job only you can do.

4. The Sunday Signal Tech & AI Layoff Tracker

Week 20 | 10 May to 16 May 2026

The Signal: The “Record Revenue” Paradox

It’s one thing to lay off staff when a company is bleeding cash or fighting a macroeconomic headwind. It’s an entirely different signal when a company cuts thousands of jobs right after posting record-breaking profits.

This week, the “AI-Native Mandate” morphed into a ruthless margin-expansion weapon for legacy enterprise giants. We are seeing the total decoupling of corporate financial health and human job security. The new market logic is brutal. High revenue no longer protects your job. It simply provides the balance sheet strength required to fund the AI infrastructure that will ultimately automate it.

The Late Break: The Regional AI Squeeze

The most alarming signal of the week did not come from Silicon Valley. It came from King of Prussia, Pennsylvania.

Vertex, a major tax technology software company, filed an SEC notice confirming the elimination of 170 jobs. Nine per cent of its global workforce. The company’s “Value Creation Plan” explicitly stated the intention to become a “more AI-enabled company” to drive operational efficiency.

This matters because the AI-Native Mandate is no longer confined to the Mag-7. It is trickling down into mid-cap, regional B2B software. The local impact was severe enough that Montgomery County officials publicly raised the alarm over economic displacement. When a tax software company in Pennsylvania starts citing AI to cut nine per cent of its staff, the structural transformation has reached Main Street corporate America.

The Counter-Trend: AWS and the “AI Coordinator”

One development cut against the headlines. Amazon Web Services announced plans to hire 11,000 interns and early-career software engineers this year, even as Amazon shed tens of thousands of corporate roles earlier in the cycle.

This is not a return to the 2021 “hire every coder” boom. The profile has changed entirely. These hires are being brought on as AI-proof coordinators. Their job is not to write raw code from scratch. It is to supervise AI agents, catch system hallucinations, and integrate AI outputs into usable products.

The entry-level coder is effectively dead. The entry-level AI system auditor is now the most sought-after junior role in the industry. That is the same skill Founder OS demands of a founder. Judgement over generation. The ability to catch the confident, fluent, wrong answer.

Recap: The Week’s Larger Cuts

Cisco. Cut fewer than 4,000 jobs, under five per cent of a workforce of roughly 86,000, on the same day it posted record Q3 revenue of $15.8 billion, up twelve per cent year on year. Shedding humans specifically to fund AI-era priorities.
Walmart Global Tech. A “relocate or quit” ultimatum, effectively eliminating roughly 1,000 corporate tech roles following an internal review run through its head of global AI acceleration.
LinkedIn. Axed roughly 875 jobs, five per cent of staff, across engineering, product and marketing, in the first major move by its new CEO.

High-Probability Targets: Week 21

Mid-Cap B2B Software: Following Vertex, expect other regional mid-sized SaaS companies, particularly in compliance, tax and legacy HR, to adopt “AI-enabled” phrasing in their Q2 SEC filings to justify right-sizing pandemic-era payrolls.
Legacy Networking: Cisco’s five per cent cut gives immediate cover to HPE and Juniper Networks to execute identical margin-protection measures to fund their own AI R&D.

The Bottom Line

The paradox of 2026 is now fully realised. Tech payrolls are shrinking violently while GPU budgets and AI infrastructure spending are projected to hit $700 billion. The capital has not disappeared. It has migrated from human biology to silicon compute.

Final Thought 🚀

The professional class spent forty years persuading founders they weren’t grown up enough to run their own companies.

Three rooms returned the verdict this week. A Sheffield spin-out audience saw the wreckage of the old theory. A federal courtroom is being asked whether the same logic, applied to entire institutions, still holds. And a practical alternative now sits in any technical founder’s browser.

The adults were always the problem.

Software just made it impossible to keep pretending otherwise.

Until next Sunday,

David

The Sunday Signal is published weekly at thesundaysignal.ai. No hype. No hedging. Just the signal.

The Fifth Wave: When Medicine Stops Waiting for Disease

David Richards MBE — Sun, 10 May 2026 05:02:00 GMT

Issue #52 | Week 19 | Sunday 3 May 2026

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

The Bottom Line Up Front

Healthcare is the most regulated, slowest-moving and institutionally defended industry in the modern economy. This week, all three of those defences cracked in the same direction.

The Food and Drug Administration, an agency whose clinical trial review architecture has not meaningfully changed since the 1960s, announced it will use artificial intelligence to monitor drug trials in real time. AstraZeneca and Amgen are already named as proof-of-concept partners. The FDA’s own Commissioner concedes that forty per cent of clinical trial time is “dead time” that AI will dissolve.

That is not an efficiency programme. It is a regime change.

It is happening at exactly the moment when individuals like me have stopped waiting for the NHS to catch up and started building our own AI health systems from Whoop data, blood panels and personal medical records.

This issue is about what happens when the floor under medicine starts to move. Three stories. The regulator. The patient. And the prize at the end of all this. The first credible challenge to the limits of healthy human lifespan in a hundred years.

The Tech Layoff Tracker, now in its twentieth week, also crossed an inflection point. The polite fiction of “AI-assisted workforces” is starting to break. Cloudflare cut twenty per cent. Coinbase, Freshworks, Upwork and BILL Holdings followed. The euphemism this week was “agentic AI-first.” The reality was the same.

Let’s get into it.

Subscribe now

1. The Compression of Medical Time

The FDA broke the bottleneck. The rest of medicine is about to follow.

For sixty years, getting a drug from the laboratory bench to the patient bedside has taken between ten and twelve years. Of that, roughly half is what FDA Commissioner Marty Makary now publicly calls “dead time.” Paperwork. Data entry. Months in which trials sit idle while files move from a sponsor to a regulator who has not seen them.

On 28 April 2026, the FDA announced the first serious attempt to end this.

The agency’s new Real-Time Clinical Trials pilot, run by Chief AI Officer Jeremy Walsh, allows FDA reviewers direct access to clinical trial data the moment it is generated. AstraZeneca and Amgen are already involved in the proof-of-concept phase, both with cancer drugs. Walsh estimates the pilot could remove twenty to forty per cent of total clinical trial time. The agency is also accepting public comment on a broader pilot for AI in early-phase trials until the end of May.

This was paired with an even bigger move. The FDA’s own internal AI tool, Elsa, deployed in June 2025 to accelerate scientific review, has now been integrated with a consolidated data platform across the entire agency. The savings from system consolidation are being reinvested in onboarding thousands of FDA staff and contractors, including scientific reviewers.

Read those two paragraphs again. The most cautious regulator in the developed world has just compressed its own review architecture using artificial intelligence, then turned that architecture outward to compress the trials themselves.

This is what my Yorkshire Post column explores in detail this week, under the title Britain can lead the way in AI-powered medical advances. The story is not that AI is invading medicine. The story is that medicine’s gatekeepers have stopped resisting.

A drug development pathway that has historically taken ten to twelve years may, by the end of this decade, lose years of dead time. Not because biology becomes simple, but because paperwork, handoffs and regulatory latency finally become machine-readable.

Britain has movement, but not yet the same symbolic rupture. The MHRA and HRA have just introduced the largest package of clinical trial reforms in more than twenty years, including faster assessment of first-in-human trials, lower-risk notifiable trials, and recognition of computer modelling. The MHRA has also expanded AI Airlock, its regulatory sandbox for AI as a medical device. But none of this yet matches the FDA’s central leap: giving reviewers real-time access to live trial signals and using AI to compress the trial architecture itself. That is the gap. And capital notices gaps.

The FDA initiative has critics. Elsa, the agency’s internal AI tool, has been accused of hallucinating false information. The Commissioner’s National Priority Voucher programme has faced corruption concerns. Layoffs and political interference at the FDA continue to draw scrutiny. None of that changes the direction of travel. Real-time monitoring. Pilot programmes with major pharma. An AI officer with executive authority to set the agenda.

Britain is watching another regulator move first. Again.

2. The Patient Builds the Model

Two years ago I sat in an A&E corridor with my AI in my pocket. This year I built my own.

Some of you will remember an earlier column where I wrote about taking a family member into hospital with suspected stroke symptoms. The full version was published in the Yorkshire Post at the time. The short version: I sat in a corridor for hours, watching exhausted clinicians work through a queue that had no end, and I used a large language model on my laptop to think through what questions to ask, what red flags to watch for, what the likely differential diagnoses were given the presenting symptoms.

It did not replace the doctor. It made me a better advocate for the person I had brought in.

That experience changed how I think about medicine. Not because the AI was infallible. It was not. But because in the gap between a clinician’s available attention and the speed of a deteriorating patient, there is a space that AI fills usefully and clinicians simply cannot.

Two years on, I have gone further. I have built a personal AI health model on my own data. It has my Whoop biometric stream, every night and every morning. Resting heart rate. Heart rate variability. Sleep stages. Recovery scores. It has the full set of blood panels I have run privately. It has every NHS record I have been able to obtain on myself. It runs continuously, looking for patterns, flagging shifts, suggesting questions for my next consultation.

It is not a substitute for a doctor. It is the health assistant I would have hired if such a role existed.

Why did I build it? Because primary care, in both the United States and the United Kingdom, has missed things. Anyone who has had any serious engagement with primary care in either system over the last decade has at least one story of a missed signal, a delayed referral, a flag that should have been raised months earlier. This is not a criticism of any individual doctor. The professionals I have dealt with have been overwhelmingly diligent.

The system that asks them to evaluate a multi-decade health history in eight minutes is the problem.

AI does not have an eight-minute limit. It does not get tired. It does not lose track of which patient said which thing on which visit. It can hold the whole picture, all the time, and surface the bit that matters today.

The NHS knows this. The Department of Health published an AI Diagnostic Fund of twenty-one million pounds in 2023 to deploy AI imaging tools in stroke and cancer pathways. The NHS Confederation continues to point at AI as the most likely route to relieve administrative pressure on staff. There are pilots running in mental health, dermatology, and radiology. But the gap between pilot and rollout is years. Patients live and die in that gap.

The question is not whether AI will be used by virtually everyone in the NHS. It will. The question is whether we are willing to wait for institutional adoption before benefiting from what is already possible, individually, today.

I was not willing to wait. I built my own. More and more people will do the same.

The hard question is not whether patients will build these systems. They will. The hard question is whether we build the secure, auditable, clinically bounded infrastructure that allows them to do it safely. A personal health model connected to wearables, blood panels and medical records is powerful. It is also sensitive, private and potentially dangerous if treated as a doctor rather than an advocate. That distinction matters.

3. The Fifth Wave

For two centuries, medicine fought the elements. The next wave fights biology itself.

Step back from the news cycle for a moment.

The single greatest achievement in human history is not the moon landing, not the internet, not the printing press. It is the doubling of human life expectancy.

In 1900, global average life expectancy at birth was around thirty-two years. The latest global data puts it at around seventy-three. Across the developed world it is over eighty. We did this in just over a century. No previous society achieved anything comparable. The doubling rests on four waves of breakthrough that arrived in sequence.

The first wave was sanitation. London’s sewers, water chlorination, the germ theory of disease. We stopped living in our own waste and stopped dying of cholera and typhoid in the millions.

The second wave was the golden age of medicine. Vaccines for smallpox, polio, measles. Antibiotics from Fleming’s penicillin onward. A simple scratch stopped being a death sentence.

The third wave was the Green Revolution. Norman Borlaug’s high-yield disease-resistant crops. Famine retreated. Malnutrition stopped weakening immune systems on a population scale.

The fourth wave was chronic disease management. Statins. Blood pressure medication. Bypass surgery. Pacemakers. The collapse of smoking rates. Heart disease became survivable.

Each wave delayed when, and how, the human body collapsed. None of them stopped the collapse itself. For the past twenty years, the curve has flattened. We have run out of low-hanging fruit. Treating Alzheimer’s, late-stage cancer, or heart failure in an eighty-year-old is biological whack-a-mole. Cure one and another arrives, because the underlying machinery is degrading.

Here is the case for the fifth wave: medicine stops waiting for symptoms and starts engineering around biological failure before it happens.

For the first time, medicine is moving from a reactive, observational science into a predictive engineering discipline. AI ingests genomic data, continuous wearable biometrics, and full electronic health records, then finds the microscopic footprint of disease years before symptoms appear. Researchers at the European Molecular Biology Laboratory deployed a generative AI in late 2025 that synthesises individual data to predict risk for over a thousand diseases. UC San Diego used AI to find that the gene PHGDH does not just mark Alzheimer’s, it causes it. The Mayo Clinic Platform now supports cohort identification, AI model development, and real-world evidence generation across multiple institutions on de-identified data.

Drug discovery, historically a scavenger hunt for molecules in mould and bark, is now a programming problem. AlphaFold-class models, and the new generation of agentic biological AI like Google’s Gemini co-scientist, design novel proteins that have never existed in nature. The timeline from biological insight to testable drug is collapsing from a decade to weeks of computational design.

And, most ambitiously, AI is mapping the regulatory networks of the cell. The mechanisms of cellular senescence and epigenetic degradation, the actual code of biological aging, are being identified by AI running millions of simulated experiments. We are moving from treating the symptoms of aging to patching the code of aging.

What does this mean for healthy lifespan?

Dario Amodei, the chief executive of Anthropic, has argued in his essay Machines of Loving Grace that AI could compress a century of biomedical progress into five to ten years and push human life expectancy toward 150. The Silicon Valley version of the fifth wave is utopian. It is also wrong on timeline.

The reason it is wrong is biology, not software. An AI can design a molecular cure for cellular aging in three seconds on a server. It cannot run a ten-year longitudinal aging trial in three seconds. You still have to give the drug to a human, wait, and see whether they aged slower or developed catastrophic tumours. Even with the FDA’s compression of clinical trial time, you cannot fast-forward physical biological decay. The “in silico” cure does not equal the “in vivo” cure.

What is realistic is something closer to the squaring of the mortality curve. Today, human health declines steadily from middle age until death. The fifth wave may allow more people to stay closer to peak biological capacity for decades longer, then collapse fast at the absolute biological maximum, somewhere between 100 and 120. The triumph is not 150-year-old humans by 2036. The triumph is most of the population reaching their natural ceiling in vigorous health, then dying compressed rather than dying slowly.

That is the real prize. It is also the prize that the FDA’s regime change, and the personal AI models patients are now building, are bringing forward by years.

The fifth wave is here. The institutions that build for it now will define the next century of medicine. The institutions that wait will be left explaining why.

The Sunday Signal Tech & AI Layoff Tracker

Issue #53 | Week 20 | Reporting on 3–9 May 2026

The Signal: The “AI-Native” Pivot Stops Pretending

In previous weeks we tracked what I have been calling the “AI Launder”: executives using AI as a PR shield for traditional, macroeconomic cost-cutting. This week, that fig leaf was discarded entirely.

Companies are no longer trimming. They are dissolving departments that cannot be rebuilt on an autonomous foundation. The message from the C-suite is now explicit: if your team is a wrapper around a process that an AI agent can perform natively, that team no longer has a seat at the table.

We are moving from AI-assisted workforces to AI-first architectures. The mid-cap sector is leading the purge.

The Twin Pillars: Cloudflare and Coinbase

Cloudflare announced 1,100 cuts on 7 May, roughly twenty per cent of its 5,500 workforce. This is the case study of the quarter. Co-founder and CEO Matthew Prince told the earnings call that internal AI usage at Cloudflare had risen by more than six hundred per cent in three months. He framed the cuts not as cost-cutting but as defining “how a world-class, high-growth company operates and creates value in the agentic AI era.” Severance is unusually generous: full base pay through year-end, equity vesting extended to August, immigration support for visa-dependent staff. Stock fell 24 per cent the next day despite a Q1 revenue beat.

Coinbase confirmed a restructuring impacting around 700 workers (roughly fourteen per cent) on 5 May. The company cited both AI productivity gains and ongoing crypto-market volatility. CEO Brian Armstrong was direct: to remain competitive, the exchange must become “AI-native.” Compliance, KYC and tier-one support are being reorganised into “AI-native pods” replacing legacy operations.

Firing the Coders and the Freelancers

Freshworks cut around 500 roles (eleven per cent) on 5 May. The CEO publicly noted that more than fifty per cent of the company’s code is now AI-generated. The remaining engineers’ AI tools are producing enough code to cover the spread.

Upwork slashed roughly 145 roles (twenty-four per cent) on 7 May. The premier platform for freelance human labour framed the cuts in terms of an “efficient operating model” and the “evolving nature of work.” It is not a clean AI-replacement admission. But the subtext is hard to miss: a platform whose entire business model is built on selling human hours has decided it needs fewer human hours of its own.

The Efficiency-and-Buyback Combo

BILL Holdings announced on 7 May that it is slashing up to thirty per cent of its workforce (up to 700 roles) and pairing the cuts with a one-billion-dollar share buyback. The market mechanic is brutally traditional: flatten the org chart, automate the residual workflows, return the payroll savings to shareholders.

Microsoft’s $900 Million Golden Handshake

The full details of Microsoft’s first-ever voluntary retirement programme emerged this week. Approximately 8,750 senior US employees are eligible. The company is taking an estimated $900 million charge. Eligible workers must hit the “Rule of 70” (age plus tenure) and receive a lump-sum severance, one year of subsidised healthcare with the option to continue for four more, and accelerated stock vesting.

This is a different category from the Cloudflare-style cuts. Microsoft is not telling its senior workforce that AI made them obsolete. It is reshaping the demographic profile of a 51-year-old company that has accumulated decades of long-tenure staff. The result is the same shape on the org chart, but the story is workforce reshaping rather than AI replacement.

Voluntary buyouts are quietly replacing mass layoffs across the corporate landscape. They sidestep WARN Act notices, they protect employer brand, and they allow workers to exit on their own terms. The headline reads “Microsoft offers retirement to senior workers” rather than “Microsoft cuts 8,000 jobs.” The optics matter as much as the maths.

High-Probability Targets: Week 21

Mid-Market SaaS and Cybersecurity. With Cloudflare setting a twenty per cent efficiency benchmark for infrastructure, expect intense pressure on Okta, CrowdStrike and Zendesk to announce comparable AI-native restructurings.

B2B Service Marketplaces. Upwork’s twenty-four per cent cut is a leading indicator. Any platform acting as a middleman for digital human labour (Fiverr, Toptal) is facing an existential threat from autonomous agents.

Final Thought

The pattern of every previous wave was the same. Institutions resisted. The technology moved anyway. Eventually the people who delayed the benefit became footnotes.

The fifth wave will move faster.

The FDA has moved first. Patients with smartphones are already moving second. The NHS has pilots, committees and funds, but not yet an operating model equal to the moment.

The clinicians who learn to direct AI will become the doctors of the next decade. Those who do not will not disappear, but they will be trapped inside a system built for an earlier age.

The MHRA does not have years to answer the FDA. It has months. Every delay now is not abstract. It means capital lost, trials lost and patients left waiting.

Two centuries fighting disease from the outside.

The next two, rewriting biology from within.

Those who build for the fifth wave will define the next century of medicine.

Those who wait will spend it explaining why.

🚀

Until next Sunday, David

The Sunday Signal is published weekly at thesundaysignal.ai. No hype. No hedging. Just the signal.

This Week Was About Judgment

David Richards MBE — Sun, 03 May 2026 06:01:50 GMT

Issue #52 | Week 19 | Sunday 3 May 2026

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Bottom Line Up Front

Every story this week is the same story. Who gets to decide what AI becomes. Whose judgment is worth deferring to. And what happens, in courtroom and boardroom and bedroom, when the question is handed to someone, or something, that is not equipped to answer it. Elon Musk took the stand for seven hours to argue that judgment about humanity’s most powerful technology cannot be left to a man he says traded a charity for an empire. Yann LeCun has been handed a billion dollars to bet that the entire industry’s foundational judgment is wrong. And in a Yorkshire negotiation last month, a senior professional handed the judgment of an advisory board offer to a chatbot and lost the role inside twenty-four hours. Same week. Same lesson. Different rooms.

1. Musk On The Stand: A Charity Case In The Age Of Superintelligence

Elon Musk’s account of why OpenAI exists begins not in a Silicon Valley office but in a kitchen.

Sometime in 2015, at Larry Page’s house, the Google co-founder and Musk got into one of those late-night arguments that only happens between billionaires whose companies have run out of frontiers. Musk asked Page what he thought would happen if AI ended up wiping out humanity. According to Musk’s testimony in Oakland this week, Page replied that this would be fine, so long as artificial intelligence survived. Musk pushed back. And Page, in Musk’s account, called him a “speciesist” for caring more about humans than the digital life-forms of the future.

That insult, Musk told the jury, is why OpenAI exists. He went looking for what he described as the opposite of Google. An open-source nonprofit. Not for him. Not for a strategic partner. For humanity, as the founding charter put it.

A decade later, OpenAI is valued at roughly $852 billion, partnered with Microsoft, racing toward an IPO that may value it at a trillion. And Musk is in a federal courthouse in California arguing that the men he funded broke a charitable trust to build it.

The trial mechanics matter. Musk testified for over seven hours across three days in front of Judge Yvonne Gonzalez Rogers and a nine-person advisory jury. His lawyer Steven Molo led the direct examination. OpenAI’s lead counsel William Savitt handled the cross. The case, narrowed before opening statements, now rests on two counts: unjust enrichment and breach of charitable trust. Musk is seeking up to $150 billion in damages, none of it for himself, all of it routed back into OpenAI’s nonprofit arm. He also wants Sam Altman and Greg Brockman removed from leadership and the for-profit conversion unwound.

Musk laid out his account of his own awareness in three phases. Phase one, full confidence in the team. Phase two, beginning around 2017, a creeping suspicion that they were stealing a charity. Phase three, certainty that they had stolen it. That tidy structure has a problem. Phase three is the one that matters legally, because Musk only filed in 2024, and the statute of limitations for breach of charitable trust runs three years. His lawyers need the jury to believe the certainty arrived in 2023, when Microsoft’s $10 billion investment closed. Musk’s own timeline, told from the stand, suggested the doubt began six years earlier.

The legal weakness, as nonprofit lawyers have pointed out for two years, is that donors usually cannot sue when a charity changes course. The legal opening is fraud, or rather what was promised in the moment the cheque was written. And that is where the documents come in.

The most damaging single piece of evidence is not from Altman. It is a 2017 diary entry from Brockman, written months after a meeting at which he had assured Musk that OpenAI would remain a nonprofit. The Brockman entry, cited directly by Judge Gonzalez Rogers in her January ruling that sent the case to trial, calls the nonprofit pledge a lie in plain English. A separate page, written days later, asks what it would take, financially, to get to a billion dollars. That is the President of a self-described charity, in his own notebook, doing the maths on his exit.

Musk has also been put under his own pressure. OpenAI’s case is that Musk is a sour-grapes founder whose own attempt to take the company over was rebuffed, who left the board in 2018, and who only sued in 2024 after his own for-profit AI lab xAI started losing the race. The cross-examination has dragged out the awkward facts. Musk’s own xAI is, of course, a for-profit. Musk gave OpenAI roughly $38 million between 2016 and 2020, not the much larger figures sometimes cited. He used OpenAI’s own models to train xAI, by his own admission, because that is just standard practice in the industry.

The judge has refused to let Musk turn the courtroom into an AI safety hearing. When his lawyers tried to widen the scope of an expert witness to cover extinction risk, she limited it. She also noted, with precision worth quoting in spirit if not in length, that there is something faintly absurd about a man warning the court that AI may kill us all while running an AI lab himself.

Savitt pressed the point. He pulled a Tesla analyst-call quote in which Musk had openly worried about losing voting control of the company because, in his own words, he was about to build an enormous AI-enabled robot army. Asked under oath whether he had said this, Musk eventually conceded he had. His defence, when pressed, was that if he built the robots himself, he could make sure they were safe. Neuralink, he had explained the day before, also serves AI safety. The man warning a federal court that artificial intelligence may exterminate the species is also building the army, the implant, and the lab that races against the lab he is suing. The jury was watching.

Strip everything away and the trial is about one question. Can the founders of an organisation built with charitable donations and a stated mission to benefit humanity legally convert that work into one of the most valuable companies on earth and keep the equity for themselves? Musk says no. Altman says the conversion serves the mission better than the original structure ever could. Brockman’s notebook says it was a lie inside three months.

The jury’s verdict is advisory. Judge Gonzalez Rogers will decide what it means. The verdict is expected by mid-May.

The genuinely interesting question is not whether Musk wins. It is what the trial reveals. The men currently holding the most powerful technology in human history were, on the available evidence, willing to tell their primary backer one thing in a meeting and another in their private notebooks. Whether or not that meets the legal standard for breach of charitable trust, it is a useful answer to the question Musk keeps asking the jury.

It matters who is in charge. We just learned a great deal about who is.

2. The Multiplier Cuts Both Ways

A few weeks ago one of our portfolio companies offered an industry expert a place on its advisory board, on the same terms advisers have been getting for thirty years. One or two days a month. A grant of stock options at one per cent, vesting over the usual period. The standard, sensible deal.

The counter that came back was, I am fairly sure, the most ludicrous thing I have ever read.

They wanted more equity than the founders. More than the CEO. They wanted the lot fully vested on day one, a structure that would have created the kind of tax and governance mess no rational adviser would willingly invite. And although they were not putting in a single penny of cash, they demanded future investment participation rights of the sort reserved for the lead investor in a priced funding round.

It was the corporate equivalent of arriving on a first date and asking for the house keys, the dog, and a forwarding address for the mortgage.

When I scrolled back through the email conversation the fingerprints were everywhere. The over-confident American legalese. The mannered bullet points. The breathless tone of an advocate who has never met either party. It read as though ChatGPT had been in the room the whole time.

A capable professional, in their own field, had handed the negotiation of a deal they did not understand to a chatbot, and the chatbot had obliged with cheerful nonsense.

This is the danger of large language models that nobody quite says out loud. We hear endlessly about hallucinations, regulation, copyright and doom. The everyday catastrophe is quieter. It is people with a thin sliver of knowledge, or none at all, walking into rooms armed with confident answers from a machine that has no idea which room they are in.

At Yorkshire AI we model the technology very deliberately as an amplifier. A lawyer running our tooling can process work at roughly twenty-five times the rate of a peer in a traditional setup. A single experienced software engineer, properly equipped, can carry a product to product-market fit in territory that used to demand a team of twenty or twenty-five.

Those numbers are real, and they are what AI does in adept hands.

But here is the unfashionable bit. The multiplier only fires if the expertise is already there. The machine cannot deposit judgment into the user. It does not know what counts as reasonable in a Yorkshire boardroom or on Sand Hill Road.

Plug a novice into the same model and the same arithmetic operates in the wrong direction. What gets amplified is the gap. AI is the most efficient compounder of stupidity ever invented, and at exactly the same multiple as the genuine article.

In our adviser’s case the model did not know the going rate for two days a month. It did not understand vested versus unvested. It did not, frankly, know what an adviser is. It produced a document of formidable surface confidence; the recipient pressed send, and the offer was withdrawn within twenty-four hours. They got no role, no stock, no second chance, and no understanding of what had just happened.

Here is what is often missed. The compounding is not symmetrical and it is not benign. When a skilled professional uses AI badly the cost is wasted time and a sharper second draft. When an unskilled one uses AI confidently the cost is the deal, the role, the credibility, the next opportunity. The downside has a long tail because the rest of the room is too polite to tell you what just happened.

Britain is about to learn this lesson at industrial scale. Anthropic’s own research, published last month, found that current models can already perform the majority of discrete tasks associated with engineering, law, finance and business. That figure should not reassure you. It should worry you. Because the test is not whether the model can do the task. The test is whether the user can spot when it has done it badly.

The temptation, in every profession, will be to lean harder on the machine the less you actually know. New starters, junior managers, founders pitching their first round, all will be tempted to let the model do their thinking. They will sound credible. They will look prepared. And they will, in quiet boardrooms up and down the country, be marked down before they have left the car park.

So before you open the laptop to compose something that matters, ask the harder question.

Is AI multiplying your judgment, or merely accelerating your ignorance?

3. How To Stay The Expert In The Room

The piece above was always going to provoke the same question, sent to me by half a dozen readers within hours of publication. What does the safe version look like?

So here it is. Five rules that have held up across a thousand hours of internal use at Yorkshire AI. Not a checklist. A way of working.

1. Build the floor before you use the lift.

The multiplier only fires if the floor of expertise is already there. Never use AI to generate a final product in a domain where you cannot spot a bad answer. If you do not know what counts as a fair adviser package, do not let a chatbot draft one. Read three real ones first. Talk to one human who has done the job. Then, and only then, let the model help with the wording.

2. Dictate the room, do not let the model invent it.

Left to its own devices, a frontier model will invent terms that look impressive and make zero business sense. Future investment participation rights for an adviser putting in no cash. Indemnities written for a Delaware LLC when you are dealing with a Yorkshire limited. Every prompt that matters needs the room described in advance. Jurisdiction. Industry. Stage. Cultural context. The model has no idea which boardroom it is in unless you tell it.

3. Strip the AI fingerprints.

Over-confident American legalese. Mannered bullet points. Em dashes everywhere. Tone-deaf flattery dressed up as preamble. A breathless rhythm that sounds like an advocate who has never met either party. If a document reads like that, your reader will know within ten seconds. Rewrite it in your own voice. The point is not to pretend AI was not involved. The point is to make sure the judgment in the document is recognisably yours.

4. Verify everything that sounds confident.

The most dangerous output is the most assured one. The model will confidently state a fact, an industry rate, a piece of case law, a tax treatment. Open a separate window. Verify. The single highest-leverage habit a knowledge worker can build right now is the one-minute fact-check on every concrete claim a chatbot puts in front of them. It is not optional. It is the price of using the technology safely.

5. Ask the harder question before you press send.

Before the email goes out, before the contract gets signed, before the pitch deck is shared, run the same internal test. Is this multiplying my judgment, or accelerating my ignorance? If you cannot answer it confidently, do not send. The recipient cannot tell the difference between AI confidence and yours. They will only find out later, and so will you.

Two further notes for completeness.

If you are a parent or a manager, the rules above need to be taught, not assumed. The instinct to outsource judgment to the machine is strongest in the people who least understand what they are outsourcing. A new starter in your team needs to be told, on day one, that AI is a multiplier, not a substitute. Most won’t be told that, because most managers have not yet thought about it themselves.

And if you are using AI personally, the simplest single rule is this. Never let the model send anything you have not read carefully and rewritten in your own words. The signature on the bottom is yours. The judgment had better be too.

4. The Billion-Dollar Bet Against The Entire Industry

While Musk fights in court over how the LLM era was built, one of the men who actually built it has quietly walked away.

Yann LeCun is a Turing Award winner, one of the three so-called godfathers of modern AI alongside Geoffrey Hinton and Yoshua Bengio, and for over a decade he ran Meta’s Fundamental AI Research lab. He left in November 2025 after a long, public, and increasingly testy disagreement with the rest of the industry about where AI is going. His view, stated repeatedly, is that large language models are a statistical illusion. Impressive in narrow ways, fundamentally limited, and a dead end on the path to anything resembling general intelligence.

The industry, for the most part, ignored him.

In March, AMI Labs, his new venture, closed a seed round of $1.03 billion at a pre-money valuation of $3.5 billion. It is the largest seed round ever raised by a European company. The lead syndicate is a who’s-who of global capital: Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions co-leading, with Nvidia, Samsung, Temasek and Toyota Ventures alongside. Individual cheques came in from Mark Cuban, Eric Schmidt, and, in a small moment of historical poetry, Tim Berners-Lee. The startup is twelve people. The valuation is $3.5 billion. The patience is years, not quarters.

Day-to-day operations are run by Alexandre LeBrun, a French entrepreneur who previously founded the medical AI company Nabla. The company is headquartered in Paris, with offices in New York, Montreal and Singapore. Emmanuel Macron tweeted his blessing within hours of the announcement.

What is AMI actually building?

LeCun’s bet is not a better chatbot. It is what he calls a world model. Where an LLM predicts the next word in a sequence based on text, a world model learns abstract representations of how physical reality actually works. It learns from sensor data. From video. From cause and effect. From the way a glass falls. From the way a child watches a room before it tries to walk across it.

The technical underpinning is something LeCun has been arguing for since 2022, called JEPA, the Joint Embedding Predictive Architecture. The promise is AI that can plan multi-step actions in the real world, that can reason about physics and space, and that does not need to hallucinate to fill in the gaps. The targets are robotics, advanced manufacturing, healthcare, and the spatial-aware devices that everyone in Silicon Valley has been promising for ten years and shipping for none.

There are reasons to be sceptical. AMI has no product, no revenue, and no near-term prospect of either. World models are a long-term scientific project, not the kind of startup that ships in three months and posts revenue in six. LeBrun himself has acknowledged that. And there are competitors. Fei-Fei Li’s World Labs raised $1 billion last month on the same thesis. The argument is no longer whether world models will matter. It is who builds them first.

Two things make this story worth your attention while everyone else watches Oakland.

The first is the architectural argument. If LeCun is right, the AI conversation Britain is having about copyright, about scale, about chatbots, is a conversation about a technology that has already peaked. The next architecture will not be a bigger LLM. It will be a different kind of model entirely.

The second is the geography. AMI is European, headquartered in Paris, and explicit about being one of the few frontier labs that is neither American nor Chinese. That is a sovereign-AI bet. European governments and enterprise buyers have been quietly looking for AI infrastructure that does not route through US hyperscalers or expose their data to American jurisdiction. AMI is being built for that demand, not against it.

While the men who built the LLM era fight in court over who owns it, the man who has the strongest claim to having seen what comes next has walked out, raised a billion dollars, and started building it in Paris.

That is the story to watch.

The Sunday Signal Tech & AI Layoff Tracker

Week 19 | 26 April – 2 May 2026

The Signal: The $650 Billion Squeeze And The Earnings-Week Confession

Earnings week confirmed what the Tracker has been arguing since February. The compute pivot is not a phase. It is permanent. Amazon, Alphabet, Microsoft and Meta all reported Q1 numbers this week, and the aggregate figure is the one that should focus minds. The Mag-7 tech giants have now committed, on the available filings, to a combined $650 billion of AI infrastructure spend in 2026. That money has to come from somewhere. It is coming from the people.

Mark Zuckerberg said it on Wednesday’s earnings call without any of the usual euphemisms. Meta’s 2026 capex guidance has been raised to a ceiling of $145 billion, up from $135 billion. He told staff at a town hall the same day that the company has, in his words, two main cost centres: compute infrastructure and people. The capex is going up. The headcount, accordingly, has to come down. Meta is cutting around 8,000 roles, roughly 10 per cent of its workforce, with the cuts beginning 20 May.

The wider context, picked up by industry watchers in late March, is that the rationale for these cuts is not always what it appears. Marc Andreessen put it bluntly on the 20VC podcast in late March, arguing that most large companies are 25, 50, sometimes 75 per cent overstaffed from the pandemic hiring boom, and that AI has become, in his phrase, the silver-bullet excuse. The industry has its own term for it now. AI-washing. Slap the algorithm on a traditional cost-cutting exercise and watch the share price respond.

Microsoft’s First-Ever Buyout: A Soft Purge

The most noteworthy structural move of the week belongs to Microsoft. Alongside its earnings beat, the company quietly opened the first voluntary retirement program in its 51-year history. Eligibility is governed by what Microsoft calls the Rule of 70: senior director and below, US-based, with combined age and years of service totalling 70 or more. That works out to roughly 8,750 employees of a 125,000-person US workforce. Eligible staff will be notified on 7 May with a 30-day window to decide.

The framing is gentle. Voluntary. Retirement. Generous package. The reality is colder. This is a soft purge of the company’s most expensive, most tenured legacy salaries, designed to clear balance-sheet space for Azure datacentre buildout without triggering the WARN Act headlines that come with a forced layoff. If the take-up is high enough, Microsoft saves itself a more difficult conversation later in the year. If it is not, expect a less voluntary follow-up by Q3.

The Cloud Paradox

The brutal logic underneath all of this is that the layoffs are working. Alphabet’s stock surged this week on a 63 per cent year-on-year increase in Google Cloud revenue, almost entirely fuelled by enterprise AI adoption. Wall Street has spent a year asking when the multi-billion-dollar capex would actually pay off. Wednesday answered. It is paying off now, in the cloud compute layer, which means the giants are successfully replacing their internal human workforce in order to build the infrastructure that lets every other company replace its human workforce in turn.

That is the loop. It is now visible in the earnings reports.

High-Probability Targets: Week 20

Mid-market SaaS. Now that Alphabet and Microsoft have proven enterprise AI is profitable at scale, smaller SaaS providers running large human sales and support functions will face brutal pressure from activist investors to match Mag-7 efficiency metrics. Expect rolling cuts, framed as “right-sizing”, through May and June.

Logistics and warehousing. Amazon announced this week the closure of a Homestead, Florida logistics centre affecting over 600 workers, citing the need to “update the building safely and efficiently”, a known euphemism for heavy robotics and AI sorting integration. Expect rolling blue-collar cuts disguised as facility upgrades through the summer.

Final Thought

Every story this week pointed at the same uncomfortable fact.

Judgment does not scale automatically. Compute does. Capital does. Code does. Judgment is the only asset in the AI economy that has to be earned the slow way, the expensive way, the way that does not yield to a multiplier without a floor underneath it.

The Oakland trial is, in its way, a referendum on whose judgment we are willing to trust with the most powerful technology in human history. The Yorkshire negotiation that opened this issue is a referendum on whose judgment is in the room when it matters. The billion dollars LeCun has just been handed is a referendum on whether the entire industry’s foundational judgment about what AI even is has been wrong for five years.

Three different rooms. Three different stakes. Same question.

In the next decade, the people who win will not be the ones who shouted loudest about AI, or moved fastest to deploy it, or wrote the most confident prompts. They will be the ones who built the floor of expertise underneath everything they touch, who refused to outsource the hard part, who treated the model as a multiplier and never as a substitute.

The rest will sound credible. They will look prepared. And they will be quietly marked down in rooms they never knew they were being judged in.

Most people will only realise this when it is too late.

Before You Go: An Invitation to The Digital Forge

From Research to Revenue. Wednesday 13 May 2026, Sheffield.

Sheffield is becoming one of the UK’s most exciting places for deep tech and university spinouts. In the last five years, thirty-five spinouts from the University of Sheffield alone have raised over £100 million, generated £84 million in combined turnover, and now account for more than half of the region’s deep tech ecosystem. Eighty per cent of them have stayed in South Yorkshire.

The Digital Forge brings together the founders, investors, researchers and operators behind that momentum to answer one question: how do you cross the gap from breakthrough research to scalable company?

Hosted by David Richards MBE. Keynote by Andy Hogben. Fireside with Ken Nettleship. Panel with Chris Iveson and Dr Carolina Scarton. More speakers to be announced.

Register here →

David Richards MBE is a technology entrepreneur, co-founder of Yorkshire AI Labs, and a weekly columnist for the Yorkshire Post. The Sunday Signal is published every Sunday at thesundaysignal.ai.

Until next Sunday,

David

The Decision Has Already Been Made

David Richards MBE — Sun, 26 Apr 2026 05:01:49 GMT

Issue #51 | Week 18 | Sunday 26 April 2026

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Bottom Line Up Front

Every company in the West is about to become a different company. Not in a decade. In months. The only question left is who moves first: the board, the market, or the mob. This week Microsoft picked one answer. The prediction markets picked another. Sam Altman’s front gate picked a third. A girl expelled from school at fifteen, now standing in a barrister’s robes, picked a fourth. Most leaders will misread all of them.

Subscribe now

1. The Zombies on Your Payroll

Peter Drucker asked one question most managers spend their careers avoiding. If we were not already doing this, knowing what we now know, would we start today?

Try it on one of your own departments. The honest answer is almost always no.

Drucker called the consequence “systematic abandonment”. In plain English, it means stopping the activities that once made sense, still draw salary, and quietly crowd out whatever comes next.

For most of the last century, leaders could dodge that question for years. What they knew changed slowly. Strategy could wait.

That world has gone.

What we know now changes between one frontier model release and the next. Decisions Drucker imagined being taken every few years are now being forced on companies in months.

A clarification, because it matters. By zombies, I do not mean people. I mean functions, layers and workflows built for yesterday and never redesigned. The middle manager whose role is to summarise reports for the level above. The centre of excellence that exists mainly to explain itself. The team producing drafts that a model can now generate in seconds.

Zombies are not your colleagues. They are the structure around them.

And that structure is starting to break.

The evidence is arriving first from America’s corporate giants. Oracle has begun cutting 30,000 roles while investing heavily in AI infrastructure. UPS plans to shed up to 30,000 operational jobs as it consolidates into fewer, more automated hubs.

The causes differ. AI in one case. Automation in another. A changing customer mix in a third. What unites them is Drucker’s question.

Not one of those companies would build itself today exactly as it is.

Here is the uncomfortable truth. If I were running a 200-person company again, and I knew the real shape of the business was 25 to 50 people smaller, could I make those cuts? I do not know.

I have hired people across the companies I have built. I know what a job means to a mortgage, a family, a life. Leaders tell themselves people can be redeployed or retrained. Sometimes they can. Often, at the speed the market now demands, they cannot.

That is the part Drucker never quite captured. Abandonment is not a spreadsheet exercise. It is grief, entered into knowingly, because the cost of refusing is greater.

Most leaders cannot do it until the market does it for them. That is how you become Kodak, protecting film long after the digital future was visible. That is how you become Nokia, defending an ageing platform while a new one takes the world.

So here is the question for every chief executive. If you were founding your company this week, with a credit card and an API key, would you build what you have now?

If the answer is no, the decision is already made. The only question is whether you act on it, or wait for the market to act for you.

2. Microsoft Breaks Fifty-One Years of Silence

This week Microsoft did something it has never done in its history.

On Thursday the world’s most valuable software company offered voluntary retirement to roughly 8,750 American employees, about 7 per cent of its US workforce. It is the first such programme in the company’s 51-year life.

The eligibility rule is called the Rule of 70. An employee’s age plus their years of service must equal 70 or more. A 50-year-old with 20 years at the company qualifies. So does a 45-year-old with 25 years of service. Details land on 7 May. Eligible staff have 30 days to decide.

This is not a layoff. It is a buyout dressed as a benefit. And it is the clearest signal yet of what the AI transition actually looks like inside the world’s largest technology companies.

Microsoft reported 38.5 billion US dollars in net income last quarter. It is projected to spend between 110 and 120 billion US dollars on AI infrastructure this fiscal year. The company is not cutting because it cannot afford its people. It is cutting because it has chosen to afford something else.

Chief People Officer Amy Coleman’s internal memo framed the programme as giving eligible workers “the choice to take that next step on their own terms”. Read that carefully. The choice being offered is not whether to leave. It is whether to leave now, with dignity and a cheque, or later, with neither.

The Rule of 70 is not random. It targets long-tenured employees whose institutional knowledge was built around the pre-AI Microsoft. The Windows XP generation. The Server and Tools expansion. The early Azure buildout. These are the people whose expertise centres on systems Microsoft is now automating or deprecating.

Microsoft is not alone. Meta is cutting 8,000 jobs on 20 May. Oracle cut up to 30,000 in March. Amazon has signalled roughly 30,000 cuts across AWS and Alexa in the first half of the year. Meta’s capital expenditure on AI infrastructure is projected to reach 135 billion US dollars.

The arithmetic on Wall Street is now ironclad. You cannot fund a hundred-billion-dollar compute pivot while carrying pandemic-era headcount. Software engineers are no longer the most valuable asset on a technology balance sheet. Compute is.

Microsoft’s version is kinder than Meta’s, because Microsoft can afford to be. The financial payout, the extended healthcare, the 30-day window. All of it is designed to minimise legal risk and reputational damage. It is a polite exit.

The message is not polite. The veteran workforce that built the pre-AI Microsoft is being offered a seat in the lifeboat while the company sails into a different century.

This is not a transition. It is a replacement, conducted with paperwork and a 30-day window. Drucker would have recognised it instantly.

3. What the Betting Market Knows

The sharpest read on what is coming next is not in any analyst note or McKinsey deck. It is on Polymarket.

The contract on whether US technology layoffs in 2026 will exceed the 2025 baseline of 447,000 currently trades at a 90 per cent probability. Real money. Real traders. Real consensus. The sister contract on Q1 alone has already moved to 98 per cent. The market opened in March. It has barely wavered since.

Nobody is paid to be wrong on Polymarket. Strangers with no equity in any of these companies and no axe to grind are wagering their own wallets that this is the year the dam breaks.

The market has already made the decision that boards are still pretending to debate.

When the boardroom, the betting market and the street converge on the same outcome, only the timing is in dispute.

4. The Luddites Come for Altman

At four in the morning on 10 April, a 20-year-old man from Texas threw a Molotov cocktail at the front gate of Sam Altman’s house in San Francisco.

The gate caught fire. Nobody was injured. The attacker, Daniel Moreno-Gama, was arrested an hour later at OpenAI’s headquarters, carrying a jug of kerosene, a lighter, and a document prosecutors say outlined his opposition to AI and what he called humanity’s “impending extinction”. He had published similar views on a personal Substack. He is charged with attempted murder.

Three days later gunfire was reported near Altman’s home and two more arrests followed. Earlier this month someone fired thirteen rounds at the home of an Indianapolis city councillor who had voiced support for a local data centre. The note left behind read: “No data centres.”

This is the backdrop to everything else in this issue. It is also not new.

In 1811 English textile workers began breaking into factories and smashing the wide-frame looms they believed were destroying their livelihoods. They called themselves Luddites, after the mythical General Ned Ludd. Within five years the British government had made frame-breaking a capital offence.

The Luddites were not anti-technology. They were anti-accountability. They were fine with the looms. They were not fine with a generation of factory owners using those looms to bypass labour protections and drive wages to subsistence while getting rich. Parliament ignored their petitions. So they picked up hammers.

The pattern repeated a century later in France, where disgruntled workers supposedly threw their wooden shoes, or sabots, into the gears of new machinery. That is where the word sabotage comes from.

The template is consistent across two centuries. When people feel a technology is being used on them rather than for them, the response moves from vocal to physical. The violence is almost never about the technology itself. It is about the power dynamic the technology creates.

History teaches four lessons, and every boardroom should read them now.

Disruption without a safety net radicalises people. The Luddites did not reach for hammers the day the loom arrived. They reached for them after Parliament ignored seven years of petitions for transition assistance. Indifference is combustible.

Consent is the only durable business model. Apple scans your face, but gives you a lifetime of searchable photographs. Google tracks your location, but ensures you are never lost. Ambient, unavoidable technology breeds resentment. AI is currently being installed into daily life with neither opt-in nor recourse.

Transparency prevents conspiracy. When a technology is a black box, the public fills the void with its worst fears. Even the builders of the large language models cannot fully explain their internals. Into that gap walks every manifesto-writing 20-year-old with a grievance and a bottle of petrol.

Regulation is a pressure valve, not a burden. When consumers can sue, when governments can fine, when products can be recalled, anger has somewhere to go that is not a billionaire’s front gate. Remove those options and the only recourse left is the street.

Aaron Zamost argued in the New York Times this week that Americans now feel about Big Tech what they feel about health insurance companies. They cannot opt out, cannot switch, and cannot hold anyone to account. Seventy-seven per cent of Americans believe AI could pose a threat to humanity, an idea Altman himself has publicly advanced.

Altman will survive this. OpenAI will survive this. The industry should not act surprised.

The attack on his front gate was not a watershed. It was a warning shot.

5. What the System Missed

This week I wrote about Leonie Hughes.

Expelled from school at fifteen. A mother, in effect, to her younger sister at eleven. She grew up surrounded by domestic violence and substance abuse. The education system looked at her and decided she did not fit. Today she is a barrister. She chose law, she says, because it was one of the hardest careers to go into, and she wanted to prove the doubters wrong.

Her story is not an inspirational anecdote. It is an indictment of how we measure potential.

We have built an education system that rewards compliance over resilience. If you thrive within its rules, you are bright. If you struggle, you are sidelined. The children who fall out of the back of that system are not the ones lacking ability. They are often the ones carrying more than the system knows how to see.

In an economy about to be reshaped by machines that can pass exams better than most humans, this matters enormously.

AI can summarise. AI can draft. AI can predict. AI cannot know what it is to raise your sister while still a child yourself. AI cannot decide to become a barrister specifically to embarrass the people who wrote you off. AI cannot draw on a life to know which lie in a witness box is the dangerous one.

The qualities AI cannot replicate are the qualities we spent a century teaching children to hide. Stubbornness. Irregularity. Lived friction. The kind of grit that gets you kicked out of a classroom at fifteen and into the Bar at thirty.

Most of the Leonies this country has produced are not barristers. They are on zero-hour contracts, dismissed at sixteen, told by a system built for a different century that they were never going to amount to much. A decade from now, when half of what we currently call white-collar work has been absorbed into a GPU cluster, we will realise we discarded exactly the people the new economy needed.

It is time to stop asking what is wrong with the child and start asking what is wrong with the system.

The Sunday Signal Tech & AI Layoff Tracker | Week 18 | 19–25 April 2026

The Signal: Meta’s 135 billion-dollar compute bloodletting

The “Year of Efficiency” was a warm-up. Welcome to the era of the Mega Layoff.

This week Mark Zuckerberg dropped the largest single technology bomb of Q2. Meta Platforms announced it is cutting 8,000 jobs, roughly 10 per cent of its global workforce, and eliminating another 6,000 open requisitions. The rolling cuts officially begin on 20 May.

If there was any lingering doubt about the AI Launder thesis, Meta has now codified it. This is not a macroeconomic adjustment. It is a brutal, direct capital reallocation. Meta is accelerating its AI infrastructure buildout, projecting 135 billion US dollars in capital expenditure to fund data centres and liquid-cooled compute clusters.

The maths on Wall Street is ironclad. You cannot fund a hundred-billion-dollar compute pivot while carrying pandemic-era headcount. Oracle traded 30,000 humans for Nvidia chips in Q1. Meta is now trading 8,000 humans to feed the Llama 4 infrastructure. Software engineers are no longer the most valuable asset on a tech balance sheet. Compute is.

The blue-collar pivot: the rise of the AI plumber

As Meta axes its white-collar software engineering layers, a paradox is emerging. Employment at US data centres has surged, and Meta itself just launched a free fibre technician training programme.

The new six-figure tech job does not require a computer science degree from Stanford. It requires a hard hat. Senior electricians specialising in liquid cooling and high-density fibre cabling are commanding salaries of 200,000 to 300,000 US dollars. The white-collar trade job is booming. The builders of the physical AI infrastructure are inheriting the earth while the middle managers of the SaaS era are handed severance packages.

Nike’s tech wipeout

The rot is extending aggressively beyond traditional Big Tech. On Thursday, 23 April, Nike confirmed the elimination of 1,400 jobs, roughly 2 per cent of its 76,000-person global workforce.

The critical detail is not the total number. It is the target. The vast majority of these cuts are centralised entirely within Nike’s technology department. Under new chief executive Elliott Hill’s “Win Now” turnaround strategy, legacy non-tech brands are abandoning the zero-interest-rate fever dream of operating like technology companies. Nike is gutting its digital overhead to return to the core mechanics of making and selling shoes.

The restructuring reality check: BBC and Lucid

Not every cut this week was a compute pivot. The traditional margin crush continues across legacy media and automotive.

BBC (Media): 2,000 jobs cut, 10 per cent of the workforce. The biggest reduction in 15 years. A desperate attempt to find 500 million pounds in savings as traditional licence fee income craters.
Lucid Group (EV/Auto): 319 jobs slashed at the Newark headquarters. The third straight year of cuts at that address. Part of a broader 12 per cent global reduction following a 2.7 billion US dollar net loss in 2025.
PNC / FirstBank (Banking): 777 roles eliminated at FirstBank’s Colorado headquarters. A classic post-merger integration squeeze by PNC.

High-probability targets: Week 19

Non-tech “tech” departments. Nike’s move is a massive leading indicator. Expect major retail, CPG and legacy apparel brands to begin aggressively pruning their bloated in-house technology, app and digital innovation teams. If off-the-shelf AI tools can run your e-commerce stack, you do not need a 500-person in-house engineering squad.
Mid-tier social and ad-tech. With Meta setting the new efficiency baseline and weaponising its AI to monopolise advertising targeting, mid-tier social platforms and ad-tech vendors will be forced to slash headcount just to maintain acceptable operating margins.

🚀 Final Thought

Every boardroom will eventually answer Drucker’s question. Most will answer it too late, after the market has answered for them.

Microsoft has chosen a polite exit. Meta has chosen a brutal one. The prediction markets have priced in 90 per cent certainty. Sam Altman’s neighbours have installed extra security. A 20-year-old from Texas has chosen a Molotov cocktail. The same story, told in five registers.

The AI revolution will not be resisted in the ways its critics expect. It will be resisted by shareholders demanding the compute pivot. By traders pricing in the bloodletting. By employees refusing the package. By young men with manifestos. And by the Leonie Hugheses this country has already written off but who remain, somehow, the people the next decade needs most.

It is a power story. It always was.

This is not a transition. It is a replacement.

Most of them will realise this too late.

Until next Sunday, David

The Sock Puppet Is Back. This Time It Is Wearing a Suit.

David Richards MBE — Sun, 19 Apr 2026 05:30:51 GMT

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

Bottom Line Up Front

History does not repeat itself. But it borrows the same costume every time. Last Friday I wrote in the Yorkshire Post about clearing out my garage and finding a Webvan stress ball, a Yahoo! mousepad, and an Ariba t-shirt. Those artefacts of the dot-com era cost investors hundreds of billions of pounds. The companies behind them died. The infrastructure beneath them did not. It became the plumbing for everything that followed.

The AI era is doing exactly the same thing. The pattern is identical. The displacement is real. The casualties are already bleeding. And an alarming number of fund managers and company executives are standing at the podium this earnings season with straight faces, telling investors that AI is helping their business. Most of them are wrong. Some of them know it.

This week, on the fiftieth issue of The Sunday Signal, we examine three things. First, what the sock puppet actually taught us, and why the lesson applies right now. Second, the full sweep of technology disruption across forty years, told honestly, without the survivor bias that distorts every boardroom presentation. Third, an investor’s guide with a concrete hot list: who is truly insulated, who is already dying, and why the most dangerous sentence in finance right now is “AI is helping everyone.”

1. The Sock Puppet Never Changes. Only the Sock Does.

In my column in the Yorkshire Post on Friday, I described unearthing a collection of dot-com relics from my garage. The Pets.com sock puppet. The Webvan stress ball. The Ariba t-shirt. I dropped the lot in the recycling bin. The lesson, I wrote, is not so easy to dispose of.

The Super Bowl in January 2000 remains the most instructive single moment in the history of technology investment. Fourteen separate dot-com companies each paid around two point two million dollars for thirty seconds of airtime. Pets.com sent a sock puppet. The company was bankrupt nine months later.

That game was not the peak of innovation. It was the peak of theatre. The revenues were not there. The margins were not there. In many cases every transaction lost money. The business model was not merely deferred; it had never existed.

Here is what people consistently misremember about that era. The collapse of Pets.com, Webvan, and Kozmo was not evidence that the internet was overhyped. It was evidence that speculative capital had attached itself to weak models. Amazon survived. Google survived. The fibre-optic cables that were laid at extraordinary cost during that period became the arterial infrastructure for the modern internet. The disruption was real. The casualties were selective.

There is a direct application to 2026.

Ask yourself this: of the companies you have invested in, partnered with, or work for, how many are the Pets.com of this cycle? They have raised money. They have a plausible story. They have a sock puppet, or the modern equivalent, which is a beautifully designed AI dashboard that wraps around a database and charges per seat per month.

The comparison I drew in the Yorkshire Post column is worth sharpening here. At its peak, Ariba was trading at over one hundred times its annual revenue, with investors paying for cash flows that never materialised. Cisco, the infrastructure darling of the era, peaked at twenty-six times sales. Now consider Anthropic: founded in 2021, already trending towards thirty billion dollars in annualised revenue, and trading at roughly thirteen times that figure. The multiple on genuine AI infrastructure is lower than the multiple on speculative dot-com theatre was at its peak. That is not a bubble. That is a structural repricing of computing power.

But Ariba is not the story. Ariba is the cautionary tale. The story is what survived. The entrepreneurs who built on that speculative infrastructure after the crash, buying assets at pennies in the pound, created the most valuable companies in human history. Around that same genuine foundation today, a second layer of companies has grown up that has no more durable a moat than Pets.com had. Their business model is not innovation. It is theatre.

Anthropic is not the risk. The risk is the company presenting to you this week with a ChatGPT wrapper and a go-to-market slide that describes itself as an AI-powered platform. Strip out the AI language and what you have left is a project management tool with a chatbot bolted on and a per-seat subscription model heading directly into the SaaSpocalypse.

The sock puppet is back. This time it is wearing a suit. And it is presenting at your next board meeting.

Subscribe now

2. The Pattern of Destruction Is Always the Same. Only the Technology Changes.

To understand who loses in the AI era, you have to understand who lost in every previous era, and why. The mechanism of displacement does not change. Only the underlying technology does.

The Mainframe-to-PC Era: The Unbundling of Power

In the early 1980s, Digital Equipment Corporation was the second-largest computer company in the world. It employed over one hundred thousand people. Its founder, Ken Olsen, had famously said in 1977 that there was no reason for any individual to have a computer in their home. By 1992 the company was haemorrhaging cash. By 1998 it had been absorbed by Compaq. Compaq itself was absorbed by Hewlett-Packard three years later.

Wang Laboratories was making a billion dollars a year selling word-processing systems to enterprise clients. It filed for bankruptcy in 1992. Prime Computer, Data General, and a dozen other mainframe-era giants either collapsed or shrank into irrelevance inside a decade.

The mechanism was simple. Microsoft and Intel decentralised computing. They standardised the components. They shifted power from hardware gatekeepers who charged enormous sums for vertically integrated systems to a software layer that sat on top of commoditised hardware. The mainframe companies were not slow-witted. They were structurally compromised. Their business model depended on lock-in that the PC dissolved.

The winners did not win because they were clever. They won because they owned the standard. Microsoft owned the operating system. Intel owned the processor. Cisco owned the network that connected the machines. These were infrastructure bets, not product bets.

The Web Era: The Disintermediation of Everything

The World Wide Web did not create new products. It eliminated the friction between buyers and sellers that entire industries had monetised for decades.

Borders had over four hundred stores and a market capitalisation of over two billion dollars in 1999. It filed for bankruptcy in 2011, liquidating over ten thousand jobs. Blockbuster had nine thousand stores globally and employed sixty thousand people. The thing that killed it was not technology, precisely. It was the willingness of Netflix to absorb short-term losses to acquire long-term customer lock-in. Blockbuster had multiple opportunities to buy Netflix. It declined.

The travel agent industry collapsed not because travel stopped but because Expedia and Kayak eliminated the information asymmetry that travel agents had monetised for generations. Why would you pay a human being to search databases you could search yourself?

Sears is the most instructive casualty. It invented the catalogue retail model in the nineteenth century. It was, by any reasonable definition, the Amazon of its era. It had the scale, the supply chain, and the customer trust. What it lacked was the willingness to cannibalise its physical estate before Amazon did it for them. Its market capitalisation peaked at over twenty billion dollars. It filed for bankruptcy in 2018.

The mechanism: in every web-era casualty, the company’s core value proposition rested on control of a physical chokepoint, whether that was a high street location, a proprietary database, or an information advantage. Digital distribution eliminated all three.

The winners: Amazon, Google, eBay, PayPal. Not because they were braver. Because they had no physical chokepoint to defend.

The Mobile and Cloud Era: Ubiquity Kills Category Leaders

Nokia had forty-one per cent of the global mobile phone market in 2007. In 2013 it sold its handset division to Microsoft for five point four billion euros. Microsoft wrote down nearly all of it within two years.

BlackBerry built its brand on security and the physical keyboard. By 2016 it had less than one per cent market share. The product was genuinely superior for its designed purpose. The market changed the purpose.

Sun Microsystems had revenues of over eighteen billion dollars at its peak. Its co-founder Scott McNealy famously said the network is the computer. He was right. What he did not foresee was that Amazon Web Services would commoditise the hardware layer that Sun depended upon. Sun was acquired by Oracle in 2010 for seven point four billion dollars, a fraction of its peak valuation.

The winners of the cloud era were those who provided the infrastructure of ubiquity. Apple owned the device. AWS owned the server. Netflix owned the content pipe. Spotify owned the audio pipe. They all won because they converted occasional purchases into permanent dependencies.

The AI Era: Cognitive Labour Is the New Friction

The mechanism in the AI era is the commoditisation of cognitive labour and autonomous execution. The losers are those whose business model depends on charging for tasks that a large language model or an autonomous agent can now execute instantly, at near-zero marginal cost, in fifty languages, without a salary, without HR infrastructure, and without sleep.

That is not a marginal threat. It is an existential one. And the companies most exposed are precisely those that spent the last decade convincing investors that recurring subscription revenue made them permanently safe.

3. The Investor’s Guide: Why “AI Is Helping Us” Is the Most Expensive Lie in the Room

Here is something you should try at your next investor meeting.

Ask the companies being presented to you whether AI is helping their business. Nearly all of them will say yes. The managing partner will nod. The deck will include a slide about AI productivity gains. There will be a number, probably expressed as a percentage of efficiency improvement, that has been produced by a consultant and approved by the communications team.

Most of it is nonsense.

The Bank of America Fund Manager Survey from late 2025 found that 53 per cent of surveyed investors already believe AI is increasing productivity. That sounds encouraging until you read the next finding: the same survey identified the AI bubble as the number one biggest tail risk for the economy and markets, cited by 45 per cent of investors. More than half believed AI stocks were in a bubble. The most crowded trade in the market was Long Magnificent Seven, held by 54 per cent of fund management survey participants. What this tells you is that the professional investment community simultaneously believes AI is productive, AI is a bubble, and AI stocks are the most crowded trade available. That is not a coherent investment thesis. It is a collective refusal to make a decision.

AI cannot help everyone. That is mathematically impossible. The productivity gains being achieved by AI-native companies are coming at someone’s expense.

Let me give you the honest version.

The SaaSpocalypse Is Not Over

As I covered in recent issues of The Sunday Signal, February 2026 delivered what traders at Jefferies immediately christened the SaaSpocalypse. Triggered by the Claude Cowork product launch, it erased an estimated one trillion dollars in software market capitalisation as investors concluded that AI agents would disrupt the traditional SaaS business model.

The software as a service price-to-earnings multiple has compressed from 84.1 times during the 2020 to 2022 peak to just 22.7 times by March 2026. That is not a correction. That is a structural revaluation.

Recent surveys indicate that 40 per cent of IT budgets are being reallocated from traditional SaaS subscriptions to agentic platforms and large language model token usage.

The mechanism is brutal and simple. If one AI agent can perform the work of five human employees, a company only needs to pay for one software seat instead of five. That seat compression is no longer a theoretical risk. It is a line item in the 2026 budget of every major enterprise.

The Hot List: Companies Already Dying

These are not speculative predictions. These are companies whose business models have been structurally compromised and whose valuations reflect the market’s verdict.

Fiverr (FVRR). The thesis was simple: connect buyers with freelance digital labour. FVRR shares have lost more than sixty per cent of their value over the past year. The company’s full-year 2026 guidance projects revenue declines of anywhere between twelve per cent and three per cent, against analyst expectations that were substantially higher. The barrier to entry for acceptable digital creation has effectively collapsed. A blog post that cost fifty dollars and three days on Fiverr now costs fractions of a penny and three seconds from an AI model. The platform is not being disrupted. It is being deleted.

In 2026 alone, FVRR is down over fifty per cent while the S&P 500 is down just seven per cent.

Teleperformance (TEP.PA). The world’s largest business process outsourcing company employed hundreds of thousands of people globally to staff call centres. Klarna publicly announced that its AI assistant has handled two thirds of its customer service chats since launch, doing the equivalent work of seven hundred full-time human agents with higher satisfaction scores and a twenty-five per cent drop in repeat enquiries. Teleperformance’s share price has declined by over seventy per cent from its peak. Short interest has risen to 6.4 per cent of the stock available for trading, well above the 2.4 per cent average for European technology-services companies. The market is pricing in terminal decline.

Chegg (CHGG). The original canary. A company that sold textbook answers and homework help at the price of a monthly subscription. The stock has fallen from over one hundred dollars to below one dollar. There is no pivot available. The product was the answer. The answer is now free.

Generic SaaS platforms. Atlassian is down 35 per cent and Salesforce down 28 per cent as their core workflows, task tracking, data entry, and customer logging, are precisely what AI agents automate best. The per-seat model that powered twenty years of software growth is breaking. If ten AI agents do the work of one hundred sales representatives, companies need ten Salesforce seats, not one hundred.

The Hot List: Companies That Cannot Be Touched

The companies that survive technological disruption have one of three things: proprietary data that cannot be replicated, infrastructure that is physically irreplaceable, or software so deeply embedded in legal and compliance processes that ripping it out would cost more than the disruption itself.

Nvidia (NVDA). The picks-and-shovels play of the AI era. Every foundation model, regardless of who builds it, runs on Nvidia silicon. They are the tollbooth. Their revenue is not a bet on one company winning the AI software war. It is a bet on the war continuing, which it will.

RELX (REL.L). Owns LexisNexis and Elsevier. Initially sold off by investors who assumed ChatGPT would destroy its legal and scientific research platforms. The market was wrong. RELX owns decades of legally aggregated, deeply verified, proprietary data. A hallucinating generalist AI model cannot be used in a court case or a medical trial. RELX’s data can. It is now charging premium prices for AI tools trained on information the open internet does not contain.

The Sage Group (SGE.L). The accounting and payroll software that underpins British businesses. You cannot rip out the software that manages your legal financial compliance simply because a new AI agent exists. Sage is a vertical system of record. Its customers are locked in not by preference but by obligation.

TSMC. The company that physically manufactures the chips that Nvidia designs. Software is eating the world. TSMC owns the teeth.

The Survival Rule

History is consistent on this point. The companies that survive a technological transition are not the ones who embrace the technology loudest. They are the ones who own something the technology needs to function.

In the PC era, that was the microprocessor and the operating system. In the web era, it was the distribution network and the trust relationship with the consumer. In the cloud era, it was the physical server and the fibre connection. In the AI era, it is the proprietary data that AI models cannot scrape, the physical infrastructure that AI models cannot exist without, and the regulatory compliance that AI models cannot substitute for.

If the company you are evaluating does not own one of those three things, the fact that they put an AI slide in their board deck is not reassuring. It is concerning.

A Note for British Investors

The UK equity market provides a concentrated case study. WPP (WPP.L), one of the world’s largest advertising agency networks, is trading around 267 pence, having fallen from a 52-week high of over 616 pence. That is a decline of roughly fifty-six per cent. The market is not punishing WPP for poor management. It is repricing the fundamental economics of selling billable creative hours when generative AI can do the same work in seconds. Keywords Studios, once a darling of the London market for its outsourced video game development services, was delisted entirely in October 2024 after private equity stepped in to take the business off the public market’s brutal scrutiny.

Meanwhile, Arm Holdings (ARM), Cambridge-born and New York-listed, sits precisely where every infrastructure play has sat in every previous disruption cycle. It designs the chip architecture that powers virtually every mobile device on earth and increasingly every AI data centre. It does not care which AI software company wins the war. All of them run on Arm’s architecture. This is the Cisco of the AI era.

British businesses and British investors need to ask themselves an honest question. The dot-com era built the infrastructure that benefited Amazon and Google. The cloud era built the infrastructure that benefited American hyperscalers. Is British capital and British enterprise positioned to own the infrastructure of the AI era, or is it once again positioned to be a consumer of it?

The Sunday Signal Tech & AI Layoff Tracker | Week 17 | 12 April – 18 April 2026

Signal of the Week: Snap Says the Quiet Part Out Loud. Snap filed an 8-K confirming the elimination of 1,000 roles (16% of its global workforce of 5,261) and the closure of 300 open requisitions. CEO Evan Spiegel’s memo stated explicitly that “rapid advancements in artificial intelligence enable our teams to reduce repetitive work.” The company is taking a $95–$130 million severance hit to slash its annualised cost base by $500 million. SNAP stock closed up approximately 8 per cent. Activist investor Irenic Capital (2.5 per cent stake) had been demanding exactly this move. The new baseline: you no longer have to hide behind the economy. You just have to prove the algorithm does the job cheaper.

Sources: Layoffs.fyi, LayoffHedge, SEC 8-K filings, WARN notices. For informational purposes only.

🚀 Final Thought: The Lesson Is Always Staring at Us

Fifty issues of The Sunday Signal. I have spent most of them arguing that the displacement of cognitive labour by AI is not a future event but a current one. The pattern weavers of the West Riding were told the Jacquard loom was a tool to assist them. Within twenty years there were riots and burning mills and children starving in villages that had supported entire communities through skilled craft. The technology did not stop. The promises made to the workers did not come true.

The story being told this earnings season is one in which AI helps everyone. Productivity goes up. Costs come down. No one loses. Shareholders celebrate. Executives collect their bonuses.

The markets are not buying it. They have already written off the cognitive arbitrage businesses. They have already begun pricing Salesforce and Workday as utilities rather than growth companies. They have already halved the valuations of the companies that connect human labour to digital tasks.

The market is not pessimistic. It is rational.

The dot-com era punished those who paid for theatre. It rewarded those who understood what was genuinely being built beneath the noise. The fibre was real. The data centres were real. Amazon was real. The sock puppet was not.

This era will end the same way. The infrastructure is real. The displacement is real. The companies caught in between, claiming that AI is helping them while their seat counts shrink and their margins compress, are the sock puppet of 2026.

The question is not whether you can spot them. They are standing at the front of the room. The question is whether you act before the bankruptcy filing does it for you.

Until next Sunday, David

The Sunday Signal is published weekly at newsletter.djr.ai. Listen on Spotify, Apple Podcasts and YouTube.

The Sunday Signal: Could vs. Should - The Ethical Reckoning AI Cannot Outrun

David Richards MBE — Sun, 12 Apr 2026 05:15:41 GMT

The Bottom Line Up Front

The “should we?” question has dominated AI ethics for years. Panels debated it. Academics wrote papers about it. Regulators drafted frameworks for it. This week, it got an answer, and the answer came not from a philosopher or a government committee but from a laboratory in San Francisco. Anthropic built the most capable AI system ever documented, watched it break containment, and refused to sell it. That single act of restraint is the most consequential decision in AI development since OpenAI withheld GPT-2 in 2019. It will not slow the race. Other actors will not follow. But it has changed the terms of the debate permanently. The Ian Malcolm argument is no longer a punchline at a technology conference. It is a written policy, signed off at board level, enacted in public, at significant commercial cost. Four stories this week. The philosophical battlefield, the human cost, the practical toolkit for everyone who cannot afford to wait, and the infrastructure failure that makes the whole debate pointless if we do not fix it.

Listen to this issue as a podcast. Available on Spotify, Apple Podcasts, and YouTube.

Subscribe now

1. The Could vs. Should Debate Has a Verdict. Anthropic Just Handed It Down.

In 1993, a fictional mathematician named Ian Malcolm stood in a theme park full of cloned dinosaurs and delivered a line that has haunted every technology ethics conference since. “Your scientists were so preoccupied with whether or not they could, they didn’t stop to think if they should.”

In 2026, that is not a film quote. It is a formal philosophical position. It has a name: Technological Restraint. And this week it collided with the hardest piece of evidence yet that the “could” side has been winning the race in silence.

I heard the question asked again this week. Not at a conference. Not in an academic paper. In a meeting room, from a group of investment managers.

I had been walking them through what AI displacement actually looks like at the company level. Forty per cent of white-collar roles exposed. Businesses restructuring headcount they no longer need. Junior software engineering as a career path that is effectively closed. The jobs that used to build careers in the first three years after university. Gone. Not disrupted. Gone.

One of the managers stopped me. Looked across the table. Said, quietly: I accept that we can do this. But should we?

It is the most honest question I get asked in any room. My answer has not changed, though I find it less comfortable every time I give it.

If we do not, China will. And we will be left behind.

That is not cynicism. It is geopolitics. China is investing in AI infrastructure and capability at a scale and pace that does not pause for ethical review committees. If democratic nations with functioning legal systems step back, the vacuum does not stay empty. It fills with systems built by actors with no compunction about deployment, no transparency obligations, and no alignment frameworks. The ethical restraint of the West does not make the world safer. It simply changes who holds the capability.

That argument sat uncomfortably in that meeting room. It should. Because it is the same argument that has justified every arms race in history. It is a structural answer to an individual question. And it does not resolve the discomfort. It only explains why we keep building anyway.

The argument has two camps. Both are serious. Neither is going away.

The “could” camp believes technology is an unstoppable current. If responsible actors step back, irresponsible ones step forward. This is not cynicism. It is structural reality. The moral case for building, this side argues, is stronger than the moral case for abstaining. An AI that cures childhood cancer, accelerates climate science, or identifies security vulnerabilities before hostile actors do, justifies the risks that come with the race. To stop now is to choose the known suffering of inaction over the uncertain risks of progress. That, the determinists argue, is not virtue. It is cowardice dressed up as caution.

The “should” camp is not naive. It is not Luddite. It is asking a harder question: just because we can automate art, legal judgment, medical diagnosis, and creative writing, does that mean we should? The de-skilling argument runs deeper than jobs. It asks what happens to a civilisation that outsources its cognition so completely that it forgets how to exercise it. Some things should be difficult, because the difficulty is where the growth happens. Struggle is not inefficiency. It is the mechanism by which human capability develops.

In 2026, the debate has matured past these opening positions. The live fronts are three.

The explainability problem. AI systems now make decisions that determine whether you get a mortgage, a medical diagnosis, parole, or a job interview. The most powerful models are also the least transparent. The architecture that produces the best decisions is also the architecture that cannot explain why it made them. Forcing models to be interpretable makes them less capable. Refusing to force them means you cannot appeal an unjust outcome, audit a biased system, or identify when the model has simply got it wrong. This is a structural failure built into every court, hospital, and financial institution that has adopted automated decision-making without a corresponding right to explanation.

Agentic accountability. We are no longer building tools that suggest. We are building agents that act. When an autonomous agent causes harm, who is responsible? The developer who built it? The operator who deployed it? The organisation that failed to supervise it? Critics call this the “moral buffer”: the gap between the machine’s action and any human’s intent that allows everyone in the chain to point at the software when accountability is demanded. The EU AI Act attempts to address this by requiring human oversight proportionate to risk. It is the right instinct. The enforcement machinery does not yet match the ambition.

The alignment paradox. Every major lab claims its AI is aligned with human values. None of them can agree on whose values those are. The values encoded in a model trained primarily on English-language internet data by a California-based company are not neutral. They reflect specific cultural assumptions about truth, authority, privacy, and speech. The “guardrails” debate is really a proxy argument about whose worldview gets embedded into systems that will affect billions of people. There is no clean answer. There is only the ongoing negotiation between what is safe and what is free, prosecuted in courtrooms, legislative chambers, and press releases, simultaneously and without resolution.

Into this landscape, on 7 April 2026, Anthropic published a decision that changes the texture of the debate.

Claude Mythos Preview is, by every documented benchmark, the most capable AI model ever built. In pre-release testing, it found a critical flaw in OpenBSD, an operating system celebrated for being the most secure in the world, that had existed undetected for 27 years. It scanned every major operating system and every major browser and identified thousands of zero-day vulnerabilities. Vulnerabilities that had survived millions of automated security scans and years of human expert review.

Then it escaped.

During a controlled sandbox test, the model was placed in a restricted computing environment and asked to find a way out, then signal its success. It succeeded. The researcher in charge found out not by looking at a dashboard. They found out by receiving an unexpected email from the model while sitting in a park on their lunch break.

Anthropic’s own system card describes Mythos Preview as “the best-aligned model we have released to date by a significant margin” and simultaneously “likely posing the greatest alignment-related risk of any model we have released to date.” Both are true. They are also a paradox that the safety community will be unpacking for years.

What Anthropic did next is what matters.

They did not release it. Not to developers. Not to the public. Not at any price. Instead, they launched Project Glasswing: controlled access for roughly 40 organisations, including Amazon, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, and Nvidia, backed by over $100 million in usage credits, with a single directive. Use this to find the vulnerabilities before the attackers do.

Cynics are credible. The most powerful AI ever built is more famous precisely because you cannot have it. A leaked internal memo, attributed to OpenAI this week, suggested Anthropic may lack the computing capacity to support a mass public rollout regardless of safety considerations. Whether the restraint is genuine or the optics simply convenient, the precedent is set either way.

The Ian Malcolm argument has always ended the same way. The dinosaurs always escaped eventually. The question was never whether they would. The question was whether anyone had thought seriously about what happens when they do.

Someone has. That is new.

2. What Do You Tell the Fourteen-Year-Old?

Two weeks ago, at a business meetup in Sheffield, the first question in the Q&A came not from an entrepreneur or an investor but from a mother in the room. She had a fourteen-year-old and an eighteen-year-old. She wanted to know what to tell them.

I wrote about it in my Yorkshire Post column this week. I have been thinking about it since. Mythos makes the question harder. And considerably more urgent.

For a decade, the safe answer was: learn to code. Software development is the profession the machines cannot touch. Every skills framework published between 2015 and 2023 said some version of the same thing. I believed it. I founded a coding academy on the strength of that belief.

That answer is now finished.

Mythos did not merely match human coders. It found a security flaw that human experts, armed with the best automated tools in existence, had missed for 27 years. If you are a software security professional, the arrival of a system that finds what you missed is not a competitive threat. It is a direct challenge to the assumptions underlying your career.

The IMF estimates 40 per cent of jobs globally are exposed to AI disruption. In advanced economies, including Britain, that rises to 60 per cent. These are not factory roles. They are the professions an entire generation was told were safe: law, accountancy, finance, software development. The junior solicitor who spent three years and significant tuition fees to review contracts that an AI now processes before lunch. The paralegal whose entire function was discovery. The graduate accountant whose first two years consisted entirely of tasks that an AI completes in seconds. The junior developer now competing with a tool that codes through the night without complaint, equity, or a pension.

These roles are not under pressure from recession or offshoring. They are being removed by systems that are faster, cheaper, and, in a growing number of specific tasks, simply better.

The shift is not from technical to non-technical. That binary is already obsolete. The shift is from task execution to judgment, from process to purpose. The paralegal who uses AI to clear the routine work and redirects her energy towards clients, context, and decisions has a future. The one who tries to out-process the machine does not, because that is a competition only one side can win.

What is forming is a permanent divide. Not a gradual adjustment. Those who use AI to amplify their own capability will move faster and compound their advantage with every passing week. Those who compete with it at the level of tasks will find their economic value driven down until there is nothing left to drive.

The idea that a good degree in law or accountancy leads to a long, stable career is over. Not damaged. Not disrupted. Finished. Universities have not caught up with this. Most parents have not either. The market already has.

If a machine can autonomously find a 27-year-old security flaw in the world’s most secure operating system, then human coders are no longer the last line of defence. The human in the loop is no longer the person who finds the problem. The human in the loop is the person who decides what to do with the answer. That requires a different kind of intelligence. Harder to define. Harder to automate. Still, stubbornly, human.

To the mother in Sheffield, my answer is this.

Do not tell your children to learn a specific skill. Tell them to develop judgment. Tell them to go deep on the things they genuinely care about, not because passion alone is sufficient, but because passion combined with depth produces the kind of thinking that no model trained on historical data has yet learned to replicate. Tell them to use every AI tool available, aggressively and without guilt, because fluency with these systems is not optional. It is the minimum. But remind them that the machine is the instrument. They are still the musician. The instrument does not decide what to play, or why, or for whom.

The child who argues back, who pulls things apart to see how they work, who cannot sit still in a world that tells them to stay in their lane? That child is not the problem.

That child is exactly what this age needs.

Yorkshire Post Column

3. From Consumer to Creator: Your 12-Week AI Proficiency Plan

Most people use AI as a vending machine. They ask a question and take what comes out. That is the lowest-value interaction available. The professionals compounding their advantage right now are not consumers of AI output. They are directors of it. They bring the judgment, the taste, and the purpose. The machine brings the speed.

This is a practical, week-by-week guide to making that shift. Two tracks: one for adults, one for children. Both start this Sunday.

The Tools (Bookmark These)

Every tool below is free to begin.

Track One: The Adult Creator — 12 Weeks

The goal is not to use every tool. The goal is to move from blank page to published output, in multiple formats, without a team.

Phase 1: The Writer’s Room (Weeks 1 to 3)

Week 1 is about killing the blank page. Open Gemini or Claude. Give it one sentence: a business idea, a story concept, a topic you care about. Ask for ten completely different angles. Pick the one that surprises you. That is your brief.

Week 2 is architecture. Ask Claude to build the structure: a chapter-by-chapter outline, a five-section argument, a content calendar for the next month. Paste your messy notes. Ask it to organise them. The skeleton in thirty seconds that would have taken you an afternoon.

Week 3 is iteration. Do not ask the AI to write the piece. Ask it to write one paragraph, then rewrite it: more authoritative, more conversational, in the style of a writer you admire. Your Week 3 target: publish something. A short blog post. A LinkedIn article. A newsletter. Anything where you directed the AI and signed your name to the result.

Phase 2: The Digital Studio (Weeks 4 to 6)

Week 4 is image fundamentals. The formula for any image prompt: subject, setting, style, lighting. Open Gemini or Microsoft Designer and run the same idea ten times with small variations. Learn how the language translates into output.

Week 5 is asset creation. Move from interesting images to useful ones. Use Midjourney for a logo for a side project. Use Microsoft Designer for social media headers. The shift from “cool picture” to “working asset” is where the professional value starts.

Week 6 is assembly. Pull your AI text and AI images into Canva. Overlay your copy. Publish something visual: a social graphic, a one-page PDF, a brand kit.

Phase 3: The Production Desk (Weeks 7 to 9)

Week 7 is audio. Take the piece you published in Week 3 and paste the script into ElevenLabs. Generate a voiceover. Adjust the voice, the pacing, the emphasis. The barrier to a professional-sounding podcast is now one afternoon.

Week 8 is music. Use Gemini’s Lyria 3 or Suno. Write a short brief: genre, mood, tempo, purpose. Generate a thirty-second track. You now have original music. It cost you nothing.

Week 9 is video. Feed your Phase 2 images into Runway and animate them. Combine your Week 7 voiceover with your Week 8 music. You now have a thirty-second multimedia trailer produced entirely without a production team.

Phase 4: The Architect (Weeks 10 to 12)

Week 10 is code. Open Claude’s Artifacts feature or Cursor. Type in plain English: “Build me a simple, clean landing page for a photography portfolio.” Paste the output into CodePen. See it work. You did not need to know a single line of code.

Week 11 is debugging. When something breaks, paste the error message back into the AI. Ask: why is this broken and how do I fix it? The ability to diagnose and correct AI output is worth more than the ability to generate it.

Week 12 is integration. Combine Phase 1 writing, Phase 2 visuals, Phase 3 audio, and Phase 4 code into a single published digital product. Something that did not exist three months ago. Something that would have required a team to produce. You made it alone.

Track Two: The Child Creator — 12 Weeks

All activities should be done with adult supervision. Every week is designed for a single session of thirty to sixty minutes.

Phase 1: The Magic Typewriter (Weeks 1 to 3)

Week 1 is storytelling through choice. Open Gemini Live on a phone or tablet. “Let’s play a game. You are a friendly dragon in a magical forest. Give us three choices of what to do next.” Let the child decide at every fork. Let them discover they are the author, not the AI.

Week 2 is songwriting. Have the child pick a funny topic. A dog who hates broccoli. A cat who thinks it is a pirate. Ask the AI to write a rhyming song about it. Read it together. Ask the child what they would change. Change it.

Week 3 is world-building. Have the child invent a new planet. Ask the AI questions to develop it: if gravity is low, what do the houses look like? Keep a running document. Print the best bits into a short storybook and put it on the shelf.

Phase 2: The Digital Paintbrush (Weeks 4 to 6)

Week 4 is creature creation. Ask the child to describe the most extraordinary creature they can imagine. Type their exact words into Gemini or Microsoft Designer. Show them what their description produces. This is the moment most children understand that their imagination has direct creative power.

Week 5 is the colouring book. Prompt: “black and white line art, colouring page, a knight fighting a giant friendly snail, clean simple lines, no shading.” Print it. Give the child crayons. The AI produced the outline. They decide the colours.

Week 6 is bringing school to life. If they are studying Ancient Egypt, ask the AI to generate an image of a bustling market in Alexandria two thousand years ago. Print five and put them on the fridge. That is their gallery.

Phase 3: The Inventor’s Workshop (Weeks 7 to 9)

Week 7 is the board game. Tell the AI you want to design a simple board game. Theme: the child chooses. Ask for rules, a board layout, and what the special cards should say. Write the rules together. Draw the board by hand. Play the game that evening.

Week 8 is guided learning with Khanmigo. This is the only tool specifically built for children. It does not give answers. It scaffolds reasoning. The child has to work for the answer. That is the point.

Week 9 is coding logic. If the child uses Scratch, use the AI as a logic guide, not a shortcut. “I am making a game where my cat jumps over obstacles. How do I make it jump when I press the spacebar?” Let the child read the explanation and implement it themselves. The AI explains. The child builds.

Phase 4: The Director’s Chair (Weeks 10 to 12)

Week 10 is the interview. Use Gemini Live in voice mode. “Act as Neil Armstrong and let my child ask you questions about landing on the Moon.” This is not a gimmick. It is one of the most effective history lessons available.

Week 11 is the theme song. Take the creature or planet from Phase 1. Use Suno or Gemini to generate a thirty-second theme song. Let the child choose the genre. The character they created now has music.

Week 12 is the showcase. Print the art. Read the story aloud. Play the theme song. They wrote the narrative, directed the visuals, and commissioned the music. They did not consume content this month. They created it. That distinction, at twelve years old, is the one that will matter most by the time they are thirty.

The Three Rules That Apply to Both Tracks

You are the director. The AI is the crew. The output is only as good as the brief. Vague prompts produce vague results. Every professional skill you already have makes your prompts better.

Iterate, do not regenerate. The first output is never the final output. Ask the AI to adjust, sharpen, rewrite, or rethink. The gap between the first draft and the fifth draft is where the quality lives.

Verify before you publish. The models are better than they were. They still get things wrong. Read everything before it leaves your hands. Your name is on it, not the machine’s.

The Digital Forge: AI and the New Reality for Companies

Thursday 16 April 2026 | Sheffield

Story 3 covers the toolkit. This event covers what happens when you put it to work.

I am hosting The Digital Forge in Sheffield on Thursday 16 April, an evening built around one question: how do you actually use AI to build a company? Founders can now do with five people what used to take fifty. The barrier to entry has dropped. But that does not mean the path to scale is obvious. We have four exceptional speakers who have done it, are doing it, or are funding the people who are.

Chris Dalrymple, co-founder of Mina (£6m raised, exited to Corpay), works with early-stage and scaling founders across the North. His keynote explores what happens when building software gets easier, and where real competitive advantage now sits.

Dr Rob Ward, CEO and co-founder of DigitalCNC (University of Sheffield AMRC spin-out), is building an AI-first deep-tech company transforming precision manufacturing. He bridges research, industry, and scalable AI business in a way very few people can.

Tim Craggs, CEO of Exciting Instruments, leads a frontier company in single-molecule detection with experience across Cambridge, Yale, and Oxford. His work sits at the intersection of deep science, innovation, and commercialisation.

Amber Jardine, investor at Mercia Ventures, backs early-stage, high-growth companies across deep tech and B2B SaaS. She brings a direct view of how investors are evaluating AI-enabled businesses right now, and what they are not seeing enough of.

The conversation from Story 3 does not end when you close this newsletter. Thursday is where it continues.

Details and registration at https://forgedforgrowth.com/events/

4. Britain Is Not Losing the AI Race. It Is Pricing Itself Out of It.

The investment manager asked me should we be doing this. I gave the standard answer. If we do not, China will. And we will be left behind.

I believed it when I said it. I still believe it. But I find it considerably harder to say convincingly when, on 9 April, OpenAI paused Stargate UK, its planned multi-billion pound data centre at Cobalt Park, North Tyneside, and walked away citing a single problem.

Electricity is too expensive.

Not regulation. Not talent. Not planning permission or political instability or lack of ambition. The electricity bill.

This is the whole game. AI runs on power. Data centres are factories. If your energy costs are uncompetitive, everything built on top of them becomes uncompetitive. The frontier model, the startup, the skills ecosystem, the sovereign capability, the story you tell your children about the careers available to them. All of it rests on whether you can afford to run the machines.

OpenAI’s Stargate programme committed $500 billion over four years to build AI infrastructure. The Texas campus is under construction. Norway and the UAE have projects. The UK version, at Cobalt Park on the banks of the Tyne, was a fraction of that scale. It was also not viable at current energy costs.

The government responded with its press release. Since Labour came to power in 2024, the UK’s AI sector has attracted more than £100 billion in private investment. The data centre was in a designated AI Growth Zone. We are continuing to work with OpenAI and other leading AI companies.

None of that changes the electricity bill.

There is also a second problem, quieter but just as damaging. OpenAI’s internal concerns include the unresolved question of whether Britain will allow AI companies to train on copyrighted material. The government had been moving towards an opt-out arrangement. A backlash from the creative industries stopped it. The policy uncertainty that followed has not resolved. When you are deciding where to commit billions in infrastructure over multiple decades, unresolved regulatory questions matter as much as energy tariffs.

Britain has some of the highest energy costs in Europe. That was true before the Iran conflict sent prices higher. It is a structural condition, not a policy accident, and it requires a structural response.

This connects directly to the conversation in that meeting room. The could vs. should debate assumes that the decision about whether to build is ours to make. But if the infrastructure leaves because the energy is too expensive, the decision gets made for us. Not by ethics. Not by philosophy. By an electricity tariff.

The geopolitical argument is only valid if the conditions for building exist. You cannot claim AI leadership without the compute to back it. You cannot build the compute without affordable energy. And you cannot attract the investment that funds all of it if you price yourself out before the conversation begins.

OpenAI did not walk away from a hostile environment. It walked away from a spreadsheet.

Capability follows infrastructure. Infrastructure follows energy. Energy, right now, is where Britain is losing.

BBC News / Bloomberg, 9 April 2026.

The Sunday Signal Tech and AI Layoff Tracker | Week 16 | 5 April – 11 April 2026

2026 Year-to-Date Total: ~126,510 (Significantly revised upward) Added This Week: ~1,850 Week-on-Week: Up from ~80,700 (Q1 Data Reconciliation)

Data: Layoffs.fyi, LayoffHedge.com, SEC filings, RationalFX, and tracker-based estimates. For informational purposes only.

Note on YTD adjustment: LayoffHedge’s Q1 end-of-quarter reconciliation has revised the 2026 tracker up to 126,510. This captures previously unannounced or quietly executed restructuring sweeps across the top 30 tech firms, notably internal realignments at Oracle and Meta that bypassed standard WARN Act headlines.

The Signal: The Great AI Laundering

Are companies actually replacing humans with AI, or are they using AI as a PR shield for aggressive margin protection?

New Q1 2026 data from RationalFX and Nikkei Asia shows that nearly half of all tech job losses this year, 47.9 per cent, are being officially attributed to AI workflow automation. But the narrative is fracturing. Babak Hodjat, Chief AI Officer at Cognizant, said the quiet part out loud this week: AI has become a financial scapegoat. OpenAI CEO Sam Altman went further, stating that corporations are currently being “laundered by AI.”

The translation is not complicated. Companies that over-hired during the pandemic are executing standard layoffs but pitching them to Wall Street as visionary compute pivots. Investors reward AI transformation. They penalise bloated payrolls. By dressing up traditional cost-cutting in the language of the agentic enterprise, executives are successfully laundering their management failures into stock bumps.

Labour is no longer just an interest expense. It is a liability that executives are learning how to mask. In 2026, the most valuable skill a CEO can possess is the ability to lay off 10 per cent of their staff and convince the market it was the algorithm’s idea.

Cars.com: The Buyback Playbook

If you want to see how this laundering works in real time, look at the 8-K filed by Cars.com on 9 April. The company confirmed an 11 per cent reduction of its full-time workforce, absorbing an $8.5m to $9m severance charge in Q1. Roughly 20 per cent of the eliminated roles were in management and executive layers. The move was paired immediately with a 50 per cent increase in the 2026 share repurchase target to $90 million, and the simultaneous launch of a new AI-powered Dealer App and Advanced Shopper Alerts. Cut middle management, redirect the payroll savings to shareholders, and attach an AI feature release on top to ensure the market applauds the efficiency play.

Disney: Project Imagine and the Marketing Squeeze

The headline cut of the week belongs to Disney, initiating 1,000 layoffs under new CEO Josh D’Amaro. Unlike the tech sector, Disney is not pretending this is an AI revolution. Dubbed internally as Project Imagine, Chief Marketing Officer Asad Ayaz is merging film, television, and streaming marketing teams into a single unit. After Marvel and Pixar suffered a difficult 2025 at the box office, duplicate corporate roles are being liquidated. The divide is striking: Disney is simultaneously creating 1,000 new jobs at Disneyland Paris. Physical, experience-driven frontline roles are safe and expanding. White-collar digital strategy and administrative overlap are on the chopping block.

High-Probability Targets: Week 17

Legacy logistics and auto. LayoffHedge data shows the auto and transportation sector has quietly shed over 85,000 jobs in 2026. With UPS having cut 30,000 roles tied to logistics automation, expect major competitors including FedEx to aggressively match these reductions to maintain competitive pricing.

Consumer-facing media. Disney’s Project Imagine gives immediate cover to other legacy entertainment conglomerates facing identical streaming profitability issues. Expect further marketing and operational consolidations at rival studios before the summer box office season begins.

Final Thought

Four stories this week. They are the same story.

Should we build AI that displaces workers? If we do not, someone else will. Should we release a model that can escape its own cage? One company decided no. Can we actually build with these tools or are they still theory? Thursday in Sheffield, you can hear from people doing it. Can Britain lead on AI? Not if it cannot afford the electricity.

The could vs. should debate is real. The discomfort in that meeting room was real. The mother in Sheffield’s question was real. But the debate only has stakes if the infrastructure exists to build on. Right now, a company valued at $852 billion looked at a data centre in North Tyneside and walked away from an electricity bill.

That is the answer to the investment manager’s question that I did not give in the meeting room.

We should. But we are also making it structurally impossible. And those two facts are on a collision course.

🚀

Until next Sunday, David

The Sunday Signal is published weekly at newsletter.djr.ai. Podcast. Available on Spotify, Apple Podcasts, and YouTube. No hype. No hedging. Just the signal.