The Rented Intelligence Economy

The Friday the intelligence went off

On Friday 12 June 2026, every British business, researcher and student using the most capable AI model on the market lost access to it. Not because of a cyber attack, and not because of anything Britain did. The US Commerce Department applied export controls to Anthropic's Claude Fable 5 and Mythos 5 models, requiring the company to restrict access for any foreign national anywhere in the world. Anthropic had no way to verify nationality in real time, so it did the only thing it could: it switched the models off for everyone. The New Statesman reported the notice period as around 90 minutes.

Eighteen days later the controls were withdrawn and access came back. But the episode did its work. Whitehall noticed that a decision taken in Washington, on a Friday, with no warning, could remove a general-purpose technology from the British economy in an afternoon. The Financial Times now reports that the Cabinet Office has asked the business department to assess the economic hit if Britons were blocked from frontier models in future. The government is, in effect, commissioning a study of its own dependency.

Two details make the June episode worse on inspection. First, Britain's status as a US ally counted for nothing. The UK sits in the top tier of America's chip export regime and shares intelligence under Five Eyes, and it was cut off anyway, alongside China and Russia, because the order made no distinction between allies and adversaries. Second, the dangerous capability that triggered the shutdown was not unique to the model that was banned. Anthropic's own published testing found the same vulnerability-identification capability in OpenAI's GPT-5.5 and in Kimi K2.7, a Chinese open-weight model that anyone could download throughout. The control cut Britain off from a capability that remained freely available from Beijing.

The Commons Science, Innovation and Technology Committee had seen it coming. Its report on science diplomacy, published 7 July, found the government has "no coherent strategic framework" for technology sovereignty and warned that the UK "may not be able to count even on its allies for access to technologies" critical to growth. Its chair, Dame Chi Onwurah, put it plainly: the government needs a realistic plan for sovereign capability "or risk having its access cut off at the whim of its partners."

How much of the machine does Britain own?

Two days after the committee reported, Oxford Martin School researchers presented Parliament with the single most important statistic in this debate. The UK hosts around 4% of the world's AI computing capacity. It exercises sovereign control over around 0.1% of it.

Read that again. The data centres are here, on British soil, drawing British power. But they are owned and operated by foreign cloud providers through what the industry calls a control plane, and as Oxford's Professor Amro Awad told MPs, if those companies or their governments decided to restrict access, "the physical servers in UK data centres could become unusable." Britain has built the housing for a technology it does not hold the keys to.

The market structure underneath confirms it. The Competition and Markets Authority found that Microsoft and Amazon each hold 30–40% of the UK cloud infrastructure market and that both hold "significant unilateral market power." The AI models British businesses actually use, per the ONS survey data below, are overwhelmingly American: ChatGPT, Claude, Gemini, Copilot. There is no official statistic measuring provider-level dependency, which is itself remarkable: the government has commissioned an assessment of the cost of losing access to frontier AI without holding baseline data on who uses what.

Meanwhile, the ladder

While Whitehall prices the dependency, the technology itself is already restructuring the labour market. And it is not doing it evenly. It is doing it from the bottom up.

The headline numbers, each from a named source:

• Entry-level job postings, covering graduate roles, internships and junior positions, are down roughly 32% since ChatGPT launched in November 2022, according to Adzuna. Entry-level roles have fallen from 28.9% of all UK listings to 25%.
• Indeed reported graduate postings at their lowest level since the pandemic, with roles advertised for recent graduates down about a third year on year at mid-2025.
• The Institute of Student Employers recorded 140 applications per graduate vacancy in 2026, the highest in the 35 years it has kept records.
• ONS data for May to July 2026 puts vacancies at 707,000, unemployment at 4.9%, and youth unemployment at 16.4%, an 11-year high. Just over a million 16-to-24-year-olds, 13.5% of the age group, are not in employment, education or training.

The pattern within those numbers matters more than the totals. Bank of England analysis published in January found UK job postings in AI-exposed occupations were 5.5% lower by mid-2025 than pre-ChatGPT trends implied. Related research on British firms found that companies whose workforces are most exposed to AI reduced employment by 4.5% on average, and that the effect was concentrated almost entirely in junior positions, which fell 5.8%.

The tasks going first are exactly the tasks that entry-level jobs are made of: first-pass drafting, data cleaning, routine analysis, basic client correspondence, junior code. The work a 23-year-old was hired to do while learning the job is the work the models do most cheaply. Employers have not stopped needing senior judgement. They have started questioning whether they need to grow it in-house, one graduate intake at a time, when the routine work that used to fund that apprenticeship in judgement can be done by a subscription.

Looking forward, the National Foundation for Educational Research estimates between one and three million UK jobs in declining occupations could disappear by 2035, with twelve million people currently working in occupations that are shrinking at three times the rate previously projected. Government skills projections run the other way for AI-related work: jobs directly involving AI could rise from 158,000 in 2024 to 3.9 million by 2035. Both numbers can be true at once. That is precisely the problem, because the people in the first group are not, on current policy, the people who end up in the second.

The honest counter-case

GBTT does not do single-cause stories, and AI is not the only thing hitting hiring.

In April 2025, employer National Insurance rose from 13.8% to 15% and the threshold fell from £9,100 to £5,000, a straightforward increase in the cost of every hire, and heaviest proportionally at the bottom of the pay scale where entry-level jobs sit. Employee jobs fell by 167,000 between October 2024 and April 2025. The minimum wage has risen sharply. Overall demand is weak: total vacancies peaked in 2022 and were falling before most firms had touched a language model.

The Institute of Student Employers, who talk to more graduate recruiters than anyone, describe entry-level work as "reshaped, not replaced." In their 2026 survey, four in ten employers expect no jobs to be replaced by AI over the next three years, and while 87% expect AI to change graduate roles, most expect minor changes. Apprenticeship hiring has grown even as graduate hiring fell, which looks like employers restructuring how people enter work rather than closing the door.

So the honest statement is this: the entry-level collapse is a compound event. Employment taxes made juniors dearer at the exact moment AI made much of their output cheaper, into a weak economy. Economists studying the interaction have noted the most uncomfortable version: firms appear to be banking AI efficiency gains to offset rising employment costs rather than redeploying people, which is how you get productivity gains and falling junior hiring at the same time.

But the counter-case has limits, and the Bank of England data is the limit. The NICs rise hit every low-paid role equally. The declines are not equal. They are steepest precisely where AI exposure is highest, and within exposed firms they land on juniors specifically. A payroll tax does not know what a graduate scheme is. The pattern in the data does.

The generation that paid for the ladder

This is where the entry-level squeeze stops being a labour market story, because it is not landing on a random generation. It is landing on the one that was charged the most to climb.

From 1997, British policy made a specific offer to the young: university participation would expand massively, the expansion would be funded by the students themselves through fees and loans, and the return would come later, in the graduate labour market. Intake roughly tripled. Fees were introduced, then raised, then raised again. Governments of both parties kept the architecture: the post-2010 administrations inherited the model and entrenched it. English students graduating in 2026 carry an average of £47,730 in student debt, the highest in the English-speaking world. Even before the current squeeze, roughly a quarter of graduates were working in jobs that did not require their degree.

The offer was always a bet that the graduate premium would hold. What nobody priced in was a technology that specifically commoditises the early-career work the premium was built on. The degree was sold as the entry ticket to professional work, and the entry jobs are the ones going: the postings data shows accountancy, design, software and analyst roles, the classic graduate destinations, with some of the steepest declines.

Follow the sequence, because it is one policy regime, not a series of accidents. A generation was told the ladder required a paid ticket. They bought the ticket, at £47,730 a head. The bottom rungs of the ladder are now being removed by a technology their own government does not control, and could not have kept switched on in June. They already rent their homes at record rates. They now enter a labour market whose core productive technology their country rents too.

This is not their parents' fault, and it is not the fault of the graduates who did what every institution told them to do. Boomers entered a labour market that needed juniors and trained them on the job, free. The people now leaving university did exactly what the post-1997 settlement asked of them. The settlement never imagined that the work at the bottom of the professions, the work that turns a graduate into a professional, could be bought by the token from a company in San Francisco. And when that company's government ordered the supply cut, Britain found out exactly where it stood in the queue: nowhere.

The response, priced

Set the government's answer against the scale of the thing it is answering.

The UK's AI hardware plan, announced at London Tech Week four days before the Fable 5 shutdown, totals £1.1bn: £750m for a national supercomputer in Edinburgh that arrives in 2030, £400m for chips, £45m for skills. The Sovereign AI Fund adds £500m. US hyperscalers will spend an estimated $660–725bn on AI infrastructure in 2026 alone. Amazon by itself is spending roughly £150bn. The entire multi-year British state programme is around 0.2% of one year of American private capex. China's fourth-largest AI spender outspends it many times over.

Britain's designated sovereign frontier effort, Cosine's Lumen Sovereign model, is a company of about 30 people that has raised roughly $15m in total, training on a supercomputer with a fraction of a leading US lab's compute. It may produce something useful. It is not an answer to the dependency.

And the texture of the response repays attention. The government is procuring around £250m of "sovereign" AI compute from commercial cloud providers, which is to say, renting sovereignty from the companies whose control planes are the exposure. The state's venture vehicle for AI hardware is led by a US investment firm. The department that launched the Sovereign AI Fund in April was abolished in July. The flagship AI strategy has had no independent audit; the government marks its own homework at 38 commitments met out of 50.

None of this is a plan being executed badly. It is the absence of a decision about what Britain is actually trying to be: a country that builds frontier capability, a country that secures guaranteed access to someone else's, or a country that accepts the dependency and prices the risk. The current posture is to fund the first at 0.2% of scale, assume the second despite June, and commission a study on the third.

What the meter shows

Strip it back and the position is this.

Britain's labour market is being restructured by AI now, not in 2035. The restructuring is concentrated on entry-level work, which means it is concentrated on the generation that paid the most, in fees and debt, for entry. The technology doing the restructuring is owned, hosted and controlled elsewhere, to the point that the country holds sovereign control over one part in a thousand of the compute on its own soil. The one time access was withdrawn, allied status bought nothing. And the state's response, at £1.6bn of announced funds against a $700bn annual arms race, is smaller than the measurement error in the numbers it is competing with.

The young are not imagining it. The bottom of the ladder really is being pulled up, by a compound of tax policy and technology, and the technology half of the compound answers to a government that is not ours. Whether any of it was avoidable is a question for the historians. What the record already shows is that no government, from 1997 to now, treated the entry rung of working life, or the machinery that is now replacing it, as something Britain needed to own.