The £30bn AI datacentre Britain announced was mostly a number nobody had committed

Information Technology 4 min 58 sources

A Guardian investigation finds two-thirds of the "£30bn" Stargate UK deal was hypothetical — and OpenAI never visited the key site. Plus Europe's sovereign-tech push and Alibaba banning Claude Code.

Key takeaways

Britain's "£30bn" Stargate UK AI datacentre was mostly hypothetical: only £10bn was committed, by a firm building a separate project, and the other £20bn was just the estimated cost of construction relabelled as incoming investment.
Europe's Mistral and China's Alibaba are both pulling away from US-controlled software — a growing push for "sovereign tech" that governments and firms control rather than rent from a rival power.
The NHS will use AI to triage 200,000 patients through its app over the next year, part of a £10bn tech overhaul that cut appointment-line queues 29% at one trial site.

Britain’s flagship US-UK artificial-intelligence deal was announced last year as a “£30bn” investment. A Guardian investigation published Friday finds most of that money was never committed by anyone — and the company at the centre of it, OpenAI, appears never to have visited the key site.[1]

The number that was calculated backward

Stargate UK was unveiled during Donald Trump’s September visit to London — a planned datacentre at Cobalt Park in North Tyneside, designated a government “AI growth zone.”[1] OpenAI, the maker of ChatGPT, was to work with the British firm Nscale and the chip-maker Nvidia to build it. The government’s press release said the zone was “set to” attract £30bn.[1]

Here is how that £30bn breaks down. £10bn was genuinely committed — but by Blackstone, for a separate datacentre in the area that is still going ahead and has nothing to do with OpenAI.[1] The other £20bn was labelled “potential for an additional investment from future partners.”[1] When the Guardian and the group Spotlight on Corruption asked the government how the £20bn figure was reached and who the partners were, the government declined to name any. It said the number was the amount the site would need to build a datacentre and power it.[1]

In plain terms: the site was said to attract £20bn because building it would cost £20bn. That is not investment secured. It is a construction estimate, relabelled as money on its way. A freedom-of-information request showed neither OpenAI nor Nscale ever met the local authority overseeing the site; only Nvidia visited, five months after the announcement.[1]

The plans were paused in April, with OpenAI citing UK regulation and high energy costs.[1] “It was never really a thing,” one source told the Guardian. “It was effectively just a government PR stunt.”[1] It follows a March Guardian investigation that found many of the deals meant to “mainline AI into the veins” of the British economy were “phantom investments.”[1] For anyone tracking the AI-datacentre boom, the lesson to carry is concrete: a headline investment figure is a claim, not a fact, until you can name who signed for it.

Europe and China both pull away from US software

Two stories this week point the same direction: away from depending on American technology. In France, the AI startup Mistral is “caught in a whirlwind of attention” — its rumoured $3.5bn raise would nearly double its valuation to about $23bn.[2] The reason is partly politics. A US government directive recently forced the American lab Anthropic to pull its most powerful models offline, and European calls for “sovereign tech” — software a country controls rather than rents from abroad — have grown louder.[2] Mistral’s revenue climbed from $20m to over $400m in a year, and it is following what TechCrunch calls the Palantir playbook: forward-deployed engineers who help governments run AI on their own terms.[2]

From the other side of the world, TechCrunch reports that Alibaba, China’s e-commerce and cloud giant, has banned its employees from using Claude Code — Anthropic’s AI coding assistant.[3] The details are thin, but the direction is the same as Mistral’s rise: large organisations reducing their reliance on tools built and controlled by a rival power. The practical thread to watch is the cost of that independence — sovereign alternatives are cheaper politically and dearer to build and run.

AI moves into the plumbing of daily services

Away from the datacentre headlines, AI kept quietly entering ordinary systems. The NHS in England said it will use AI in its app to triage patients — sorting who needs a GP appointment and who should go to a pharmacy or A&E instead.[4] The tool is expected to reach 200,000 patients over the next year, part of a £10bn package to overhaul NHS technology.[4] A trial at one Sussex practice cut phone-line queues for appointments by 29%, the government said.[4] The same package funds AI that records doctor-patient consultations to cut note-taking time.[4]

The contrast between this and Stargate UK is the useful one. A £10bn systems-overhaul with a measured 29% result at a named trial site is a different kind of claim from a £30bn figure with no partners named. One reports what happened; the other reports what might.

02 · Lesson · why it matters

The number that was really a wish wearing a suit

A figure you can't trace to a signature is a hope, not a fact — and the AI boom is full of them.

Start with the arithmetic

A government press release said a site would attract £30bn in investment. That single number did a lot of work. It anchored a US-UK “AI growth zone,” it accompanied a presidential visit, it made headlines. Then someone did the subtraction.

Ten billion of it was real — committed by Blackstone, for a different datacentre, in the same postcode, unconnected to the deal being announced. The other twenty billion had no name attached. When asked who would provide it, the government said the figure was the amount the site would need to be built and powered.

Read that again slowly. The site was said to attract £20bn because building it costs £20bn. The cost of the thing was reported as the funding for the thing. That is not a lie exactly. It is a wish dressed as a fact — and the difference is worth learning to see, because it is everywhere right now.

How a cost becomes a promise

There is an honest sentence buried in here: “this site would need £20bn to build.” That is a construction estimate. It is true and useful.

Then a small edit happens. “Would need” becomes “potential for.” “Potential for” becomes “set to attract.” By the time it reaches a headline, the estimate of what something costs has become a forecast of money arriving. Nobody had to lie. Each step was only slightly softer or firmer than the last. The number never changed; only the verb around it did.

This is how most manufactured certainty works. It is rarely one big fabrication. It is a chain of small, defensible rewordings, each plausible on its own, that together carry a claim a long way from what anyone actually knows.

The tell is always the same

You do not need to be an economist to catch this. There is one question that does the work: who signed for it?

A real investment has a name behind it — a firm, a fund, a board that voted. You can point to the party who is out the money if it fails. Blackstone’s £10bn passes that test. The £20bn does not; no partner would be named, because there was none. The moment a number can’t be traced to someone on the hook for it, treat it as an aspiration, not an event.

The same test cuts the other way and rewards the honest claim. The NHS said its AI triage tool cut appointment-line queues by 29% at a named practice in Sussex. That number has a place, a trial, a measured before-and-after. It might not scale. But it reports something that happened, and you can check it. One kind of number tells you what occurred; the other tells you what someone would like you to expect.

The wish leaves marks even when it never comes true

Here is the part that should make you hold your conclusions loosely. The Stargate deal was paused. The money mostly never existed. And yet the announcement did real things while it stood. Land was designated an “AI growth zone.” A region reorganised its planning around it. Ministers built policy on top of it. A source said the partner firm was “pretty much told to back” a project it was unaware of, to give the government “a big announcement” before a state visit.

An unreal number can still redraw a map. The forecast that never lands still shapes where roads get built, which region gets the growth-zone label, whose energy grid gets planned for a datacentre that may never come. The gap between the announced and the real is not empty space. It is where a lot of decisions quietly get made — decisions you and I live inside, whether or not we ever saw the press release.

You are downstream of numbers you didn’t check

It is easy to read this as a story about one government and one datacentre. It is not. Every large AI figure you’ve seen this year — the trillion-dollar plans, the sovereign-compute pledges, the growth zones — is a mix of committed money and hoped-for money, and the two are almost never separated in the telling. The people announcing them have every reason to blur the line. The people reporting them rarely have time to unblur it.

So the humility here is not “distrust everything.” It is smaller and more usable: when a number is big enough to move a headline, it is big enough to have been rounded up by hope. Find the signature, or file it as a wish. You are standing downstream of hundreds of these figures — in the news you read, the policy that shapes your town, the confidence of the market you save into. Most you will never trace. The most any one of us can do is know that the gap is there, and stop mistaking the map’s most confident numbers for the ground.

03 · Lab · your turn

Audit the Headline

Break a "£30bn investment" into its parts and ask of each: who signed for it? Watch the real, traceable total shrink to a fraction of what was announced.

04 · Hope · carry this

The £20bn that was never real got caught — not by magic, but by reporters and a watchdog group patiently asking who signed for it. The same scrutiny that exposes a phantom number is available to any of us, and it still works.