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

In 2013 a group of Berkeley researchers sat in my office in Silicon Valley asking for funding for an academic project called Spark. They were not on my calendar by accident. They were there because Konstantin Boudnik, one of the early developers of Apache Hadoop and co-creator of Apache Bigtop, who ran R&D for me, insisted I meet them. A man whose work was closely bound up with Hadoop, telling his chief executive to look hard at the technology most likely to displace part of it. On Thursday that company closed a $5 billion round at a $190 billion valuation.

This episode is the story of what happened in between, and it begins with a second-hand Commodore 64 with a broken tape deck in a Swedish suburb. Four years of nobody wanting the technology. One sorting benchmark that reversed the industry's opinion overnight. A board telling them they had less revenue than the local restaurants. And a set of founders who genuinely did not know how venture capital worked.

Then the claim that got Ali Ghodsi more attention this week than the money did: that artificial general intelligence has already arrived, measured against the definition researchers used before 2022, and that the industry quietly moved the goalposts rather than admit the bar had been cleared. His argument is that AI does not have an intelligence problem, it has a context problem. Demis Hassabis at Google DeepMind says 2029. Greg Brockman at OpenAI will not pick a date. I argue they are answering different questions, and only one of those has already been settled.

Then what AGI actually means once you strip the science fiction out of it: not robots, but a fall in the price of cognition, and the ability to buy thinking without putting another person on the payroll. Goldman Sachs Research on 300 million exposed jobs, why its 16,000-a-month figure is a drag on hiring rather than a redundancy count, and the mechanism that matters more than the attribution, which is a hiring freeze at the entrance.

Then my Yorkshire Post column this week. An AI law firm at Yorkshire AI Labs applied for professional indemnity cover, and the underwriters asked every question about people. We assumed they wanted a human in charge. They were confirming the opposite. The quote came back 50% lower.

And finally the Week 33 layoff tracker update, led by Oracle, which reported headcount down 21,000 in a single fiscal year while spending $55.7 billion on capital expenditure, and told regulators in writing that AI deployment contributed to workforce reductions.

Read the full issue free at thesundaysignal.ai

Keywords:

Databricks, Ali Ghodsi, Konstantin Boudnik, Hadoop, Apache Bigtop, AGI, artificial general intelligence, Apache Spark, enterprise AI, AI agents, Demis Hassabis, Google DeepMind, Greg Brockman, OpenAI, venture capital, Ben Horowitz, Andreessen Horowitz, startup funding, valuation, Lakehouse, Lakebase, context engineering, AI layoffs, Oracle layoffs, future of work, automation, Goldman Sachs, labour market, entry level jobs, professional indemnity, insurance, human in the loop, AI regulation, UK technology, Yorkshire AI Labs, David Richards, The Sunday Signal