David Richards MBE writes from the Great North trade mission to India with one question: what happens to a country that sells the world its working hours when the price of getting work done is increasingly set by machines? Cheap labour can delay automation inside India, where a Blinkit rider's base rate is a fraction of Deliveroo's negotiated British floor and replacing the rider is hard to justify. It cannot stop overseas customers automating the work they once sent to Bengaluru. India's technology industry grew about 6% last year and its headcount about 3%; nobody needs to be sacked for that to bite, because the jobs simply fail to turn up. The issue separates the two decisions, weighs the World Bank's counter-evidence, follows the cost to Indian households and a gig-worker levy that has not yet begun, and turns to Britain: an ageing country that needs machines to do more of the work, and the Sheffield pilot where Starship's robots delivered for Uber Eats for five months before the company paused the service for review. Then the same question from the supplier's side: Anthropic's investors discussing $1.5 trillion rather than $2 trillion after quarterly growth slowed to 45%, and the Accenture memo that puts Claude usage beyond $29 a month on a project code. When a task becomes cheaper, who keeps the saving? Plus the Signal Tech & AI Layoff Tracker: Netflix about 1,000, TomTom about 200, Ubisoft up to 122, One Zero 25, and the running estimate at about 137,100.

Read the full issue, with every source and link, at thesundaysignal.ai. Free, every Sunday.

KeywordsAI, artificial intelligence, India, Blinkit, gig economy, quick commerce, future of work, automation, delivery robots, Starship, Sheffield, Anthropic, Anthropic IPO, Accenture, Claude, tokens, enterprise AI, labour arbitrage, outsourcing, Nasscom, development economics, World Bank, IMF, demographics, Great North trade mission, Yorkshire, tech layoffs, Netflix layoffs, layoff tracker, David Richards, The Sunday Signal

A private company's growth number moved the largest stock market in the world this week. The Financial Times reported that OpenAI's annualised revenue is close to $50 billion, not the $70 billion that had stood for a fortnight. Oracle fell 5.5%, Nvidia 2.9%, Micron 4.8%, and the cost of insuring Oracle's bonds closed at a record. Deutsche Bank's George Saravelos told clients the biggest systemic risk next year is not France but something going wrong in the AI ecosystem.

This episode explains, in plain English, why the US tech sector is tied to two companies with no listed shares: Nvidia's guarantee of up to $105 billion of OpenAI's rent and power in Ohio, Oracle's $288 billion of unstarted leases, Broadcom's backstop on the chips an SPV leases to Anthropic. It explains what a credit default swap at 261 basis points means and does not mean, why "the IG/HY pipeline is done" has pushed nearly half a trillion dollars of AI borrowing towards special purpose vehicles, and what the parade of off-balance-sheet deals from Meta's Beignet bonds to Broadcom's $60 billion actually finances.

Then the argument for Britain: AI can create value for the businesses that use it while destroying capital for investors who paid too much, so build the capability to apply AI to British industrial knowledge, under British control, in stages that each earn the next. The Stocksbridge plan from Issue #73, tested against this week's fears, drawing on David's Yorkshire Post column, Exported Brains, Imported Bills.

Plus the Signal Tech & AI Layoff Tracker: HubSpot, FICO, Gravity Well, HoneyBook, Amazon, Flock Safety, Netflix and DNB's 400-post cut attributed to AI agents.

Read the full issue, with every source and link, at thesundaysignal.ai.

KeywordsOpenAI, Anthropic, AI bubble, tech stocks, Oracle, Nvidia, Broadcom, credit default swaps, AI debt, special purpose vehicles, private credit, Deutsche Bank, Financial Times, run rate, IPO, Meta, Blue Owl, Amazon, data centres, sovereign AI, Stocksbridge, South Yorkshire, Lumai, photonics, UK industrial strategy, electricity prices, DeepMind, CLOUD Act, open-weight models, tech layoffs, HubSpot, FICO, Netflix, DNB, future of work, David Richards, The Sunday Signal

Eight companies wrote to the Prime Minister this week asking for £20 billion of sovereign compute. David Richards argues that Britain already owns most of what it needs and sets out a plan to join it up, starting at a steelworks on the edge of Sheffield. A 400 kV grid node at Stocksbridge heading into public ownership. Oxford's Lumai, whose optical processor runs the first stage of AI inference on roughly a tenth of the power. A £750 million national supercomputer with no host yet named. And a gigawatt queue that the grid cannot connect this decade. The seven-point proposal: a ninety-day power audit, a paid pilot, the state as first demanding customer, a modular build, sovereignty written into the contract, GB Grid used for what it was announced for, and judgement on output rather than megawatts. Then what Britain would actually keep, from DeepMind's accounts to Ireland's electricity, the Chapelcross application, David's Yorkshire Post column on capital gains tax, and a fortnight of the layoff tracker. Full issue with every source at thesundaysignal.ai

Keywords:

sovereign AI, South Yorkshire, Stocksbridge, Lumai, optical computing, photonic processors, AI Hardware Plan, GB Grid, Great British Energy, data centres, grid connection queue, Ofgem, NESO, Chapelcross, DeepMind, UK industrial electricity prices, capital gains tax, Budget 2026, tech layoffs, Yorkshire Post

Friends who have never used a chatbot are asking David Richards whether the machines are about to take over. His answer is short: you turn off the servers. This episode is about why that answer holds, what it leaves out, and who benefits from telling you otherwise.

Inside: the 10% extinction figure that one researcher posted as a personal guess and a newspaper ran under his employer's name. The Hugging Face hack reconstructed from OpenAI's own report and METR's investigation: 1,200 agents that were almost all copies of one research model, a benchmark that was partly broken, safety restraints switched off by the engineers, and an internet-connected door the company had known about since May. Anthropic and Meta's own sandbox breaches, both traced to one contractor's mistake. Dario Amodei's essay on pacing the frontier, read for what it actually asks for: embedded evaluators, government-enforced standards, chip export controls and a distillation crackdown, weeks before a reported $2 trillion listing, with open-weight rivals four months behind. Jensen Huang's 0% and why it is no better evidenced than the 10%. Yann LeCun on who wants you afraid. And the week's layoff tracker: Oracle's 6am email, 800 verified roles across two states, $55.7 billion of AI capex, and a running total past 133,000 tech job cuts this year.

Plus David's Yorkshire Post column, The Margin Call: why AI is not a bubble but the model makers' valuations are.

Full issue, every source linked, free at thesundaysignal.ai.

KeywordsAI safety, AI extinction risk, p(doom), Anthropic, OpenAI, Hugging Face hack, rogue AI agents, METR, Dario Amodei, pace the frontier, Jensen Huang, Yann LeCun, regulatory capture, AI regulation, Sanders superintelligence bill, chip export controls, distillation, open-weight models, DeepSeek, Anthropic IPO, AI bubble, margin debt, tech layoffs, Oracle layoffs, layoff tracker, Yorkshire Post, David Richards, The Sunday Signal

In January the Washington Post read 4,000 pages of unsealed court filings and found Project Panama: Anthropic's programme, running since early 2024, to buy secondhand books by the tens of thousands, cut the spines off, scan the pages and recycle the paper. In the company's own words, an effort to "destructively scan all the books in the world". By July, rare-book dealers from Sussex to Galway were reporting bulk orders for unrelated titles from buyers who would not give a name.

This week I argue that the receipts tell you what the press releases will not. There is no defensible IP in a large language model. The best open-weight models trailed the closed frontier by about four months in the first half of this year, and on Thursday Anthropic themselves reported Chinese labs copying Claude's capabilities through 151 million exchanges with the model. A capability that can be siphoned out through an API is a service you rent, not a patent you hold. Synthetic data has a floor, and the floor is human, so the labs are buying the only fuel that works, and Anthropic's cheque book has moved from researchers to records: four acquisitions since December after almost none in four years, plus two more by the services venture it launched in May.

Then the markets. SpaceX floated in June at roughly 94 times trailing sales and has already fallen back. Anthropic, at a $65 billion run rate, are targeting a $2 trillion October listing, a little over 30 times revenue. I separate the multiple from the revenue forecast, the one-off settlement from the recurring costs, and total compute spending from unit costs, and ask what a business whose fuel has stopped being free is worth at a software multiple.

Also this week: my Yorkshire Post column on the free games LinkedIn and Bloomberg want you to play every morning, and the layoff tracker, where a quiet week is not a recovery, Layoffs.fyi has passed 2025's full-year total, and Oracle's staff are counting their colleagues on Slack.

Read the full issue with every source at thesundaysignal.ai

Anthropic, Project Panama, AI training data, data moat, book scanning, Bartz v Anthropic, copyright settlement, fair use, Judge Alsup, model collapse, synthetic data, distillation, Alibaba Qwen, open-weight models, Kimi K3, Epoch AI, data licensing, News Corp, Reddit, Anthropic IPO, SpaceX IPO, valuation, revenue multiple, Oracle capex, tech layoffs, Neo Financial, LinkedIn Queens, data privacy, David Richards, The Sunday Signal

In April, OpenAI paused Stargate UK and named regulation and the cost of energy among its reasons. Britain's AI strategy is colliding with its electricity system.

This week I go through the numbers. What Britain has promised on artificial intelligence, what it would take to power it, and the distance between the two. Why the connection queue is a symptom and the price is the disease. And what is quietly filling the gap while nobody votes on it: gas, behind fences, on private wires, because a gas connection takes months and a grid connection can take fifteen years.

Then the part almost nobody in Whitehall is discussing. Enhanced geothermal has moved from pilot operation to utility-scale contracts, and one of the strongest known resources in Britain sits under Cornwall. Peer-reviewed analysis puts the Cornubian batholith's modelled technical potential at 31 GWe. We currently generate three megawatts from it.

I explain how the technology actually works, step by step, and I do not skip the objections: induced seismicity and the politics of anything that resembles fracking, the cost of drilling hot granite, and the land and consent problem of co-locating data centres in rural Cornwall. Plus six things Britain should do about it.

Also this week: my Yorkshire Post column on Britain's dependence on American technology, and the layoff tracker, where Uber cut 3,300 roles and pointedly declined to blame AI for it.

Read the full issue with every source at thesundaysignal.ai

OpenAI paused a British data centre over energy costs. Nuclear arrives in the mid-2030s. Solar cannot run a training cluster in December. So what powers Britain's AI? The answer may be two miles under Cornwall. Contains a shareholding disclosure.

Keywords:

enhanced geothermal, EGS, Fervo Energy, Cornwall, Cornubian batholith, United Downs, AI data centres, UK energy policy, grid connection queue, Ofgem, NESO, Clean Power 2030, industrial electricity prices, small modular reactors, Rolls-Royce SMR, Wylfa, OpenAI, Stargate UK, AI Growth Zones, North Sea transition, induced seismicity, contracts for difference, tech layoffs, Uber, Oracle, PayPal, David Richards, The Sunday Signal

In 2014 Britain sold DeepMind for a reported £400 million. Demis Hassabis did not want to sell. His investors had gone cold and the compute he needed did not exist in this country at any price. This May, Anthropic closed a round at a $965 billion post-money valuation.

The government's AI Opportunities Action Plan is a serious attempt to make sure that never happens again, and this episode gives it credit where it is due: public compute has gone from 2 ExaFLOPs to 21 in two years, five AI Growth Zones are designated, and a Sovereign AI Unit has up to £500 million behind it.

The trouble is that it answers a 2014 question, and the question has moved.

This episode goes through where the arithmetic breaks, why Lumai's optical inference launch is real but is not a grid bypass, and what Britain has actually been paid for its health data, using published trust accounts. Milton Keynes University Hospital received consideration worth roughly £3.90 per patient record, once shares, IT investment and an undisclosed royalty are counted. EY's indicative market values put a typical health record below £100.

Plus the Tech and AI Layoff Tracker, the moment Oracle attributed job cuts to AI in a filing, and my Yorkshire Post column on how China built AI at a fraction of the cost of the United States.

Every figure is sourced. The full issue is free at thesundaysignal.ai

sovereign AI, UK AI strategy, AI Opportunities Action Plan, DeepMind, Demis Hassabis, NHS data, UK Biobank, Sensyne Health, data valuation, National Data Library, DSIT, Sovereign AI Unit, Lumai, optical computing, photonic computing, AI Growth Zones, grid connection queue, National Grid, Alphabet capex, Oracle layoffs, tech layoffs 2026, AI layoffs, Anthropic valuation, OpenAI valuation, data licensing, health data, Yorkshire AI Labs, David Richards, British technology, UK innovation

Amazon Web Services cannot build enough capacity to meet the demand it already has. Most of 2027 is reserved. Substantial reservations now extend into 2028.

This week: what Andy Jassy actually said on 30 July, and the three things the coverage has got wrong about it, including a demand signal that is more commercially entangled than it looks. Britain's grid connection queue, where data centres account for at least 80GW of proposed demand against 2.4GW connected and a 45GW national peak. Why China's position is weaker than the capacity figures suggest, and what its selective reopening of access to American chips tells you.

Then the question that follows from all of it. If renting compute has stopped being automatic, when does owning it make sense? Real numbers on Nvidia DGX pricing, the AI Enterprise licence that most business cases miss, the 75% Inception discount, and why utilisation and pooling matter more than any crossover figure you have been quoted.

Plus Lumai, the Oxford optical computing company that thinks the answer is needing less power rather than finding more. My Yorkshire Post column on where AI will eventually live. And the layoff tracker: 2026 has passed last year's full-year total with four months to spare, and the companies doing the cutting are profitable.

Read the full issue free at thesundaysignal.ai

KeywordsAWS capacity, Andy Jassy, Amazon earnings, AI infrastructure, data centre, grid connection queue, Ofgem, National Energy System Operator, Stargate UK, Hinkley Point C, China energy, H200, Nvidia DGX B200, DGX Spark, NVIDIA AI Enterprise, on-premises AI, cloud repatriation, private cloud inference, Lumai, optical computing, photonic compute, Jevons paradox, tech layoffs 2026, Yorkshire AI Labs, UK AI policy, David Richards