Five AI Growth Zones, a 6GW target and a paused flagship. We trace who is really laying the concrete and who owns what goes inside.
The UK hosts around 4% of the world’s AI computing capacity, and it exercises sovereign control over around 0.1% of it, Oxford Martin School researchers presented that in front of the Parliament in July.
We think it exposes the real weakness in the national plan, which is ownership more than construction.
So who is actually building UK AI data centres, and who holds the switch to that?
“the UK hosts around 4% of the world’s AI computing capacity … it exercises sovereign control over only around 0.1% of it."- Oxford Martin School researchers, evidence to Parliament, July 2026
The pause nobody planned for
This year, the UK's AI data centre plans have run into one obstacle after another.
In April, OpenAI paused Stargate UK, which was a plan for 8,000 GPUs scaling to 31,000, citing energy costs and regulation, but OpenAI says it is paused for now and not cancelled.
Six months later, Microsoft has reportedly walked away from Nscale Loughton, the Essex campus billed as the country’s largest AI site.
That matters because Nscale is the most visible UK startup in this race, and it was valued at $14.6bn in March. Its own filing claims about $103bn of active and contracted value, though much of that depends on delivery. The same filing shows a net loss of about $1bn on $140.6m of first-half revenue.
Loughton was meant to deliver 50 MW, scaling to 90MW, but this month's report suggests grid power may not arrive until the 2030s, Nscale does say it remains committed, but can a young company keep that promise with no anchor customer and no connection date?
Five zones, one operating giant
The government’s answer to all this is the AI Growth Zones UK programme. Five zones had been designated by January: Culham, the North East, North Wales, South Wales and Lanarkshire. The UK Compute Roadmap says the UK needs at least 6 gigawatts of AI-capable capacity by 2030, which is roughly triple today’s.
Savills counts 1,803 MW of live capacity, where only 50 MW of it is AI self-built, and one site above 100 MW is operational, that is Vantage’s Newport campus in Wales.
Most of what exists today is not what the government means by UK AI data centres.
Look at the money behind what is coming.
DataVita, which is a Scottish firm, secured £300m in August for its Lanarkshire expansion.
A £202m National Wealth Fund guarantee backs the bank lending, and the capacity is leased to CoreWeave.
Blackstone’s QTS plans a £10bn campus at Cambois, near Blyth, with completion not expected until 2035, one has an American tenant, and the other has an American owner, but neither is a hyperscaler, yet both depend on American capital.
“almost entirely owned or controlled by a small number of global players” - CITP, mapping the UK’s data centre build-out, 7 July 2026
CITP counted 348 live sites at the end of 2025, where only sixteen are hyperscale, so if the people who own the buildings are not British, then what exactly are we calling sovereign?
The queue decides
The AI Opportunities Action Plan promised speed, but the wires disagree.
Between November 2024 and June 2025, demand connection applications rose from 41GW to 125GW, at least 80GW of it for data centres. The UK’s peak demand season last winter was about 45 GW.
Inside the UK data centre grid connection queue, London projects face waits of seven to thirteen years, according to analysis for the Greater London Authority, and waiting for a grid connection has become the main cost of building.
Ofgem’s remedy is a refundable data centre commitment fee of £237,500 to £712,500 per MW, designed to clear out speculators.
A decision is due by the end of 2026, and we believe the intent is sound, but a fee that deters speculators also raises the entry price for a startup with no balance sheet.
Who is it really that is built to protect?
The smaller builders skipping the line
The newer names are working around the queue rather than standing in it.
Apatura, a York energy developer, holds 2.4GW of grid connections across nine Scottish sites. Era4 wants 100,000 GPUs across more than 30 sites and has applied to run a Canterbury site on landfill gas.
Civo, a UK cloud provider, announced 40 liquid-cooled edge sites in September, with the first, in Hertfordshire, starting at 8 MW and due by March 2027.
Behind-the-meter power, generated on site and kept off the shared grid, runs through all of these plans.
We could not find any disclosed financing for most of them, and builders of UK AI data centres at this scale are still mostly promises.
Meanwhile, Fluidstack, which was founded in London and spun out of Oxford, moved its headquarters to New York in December.
Then the Sovereign AI Unit launched in April with up to £500m to back British AI firms, so the question is, does any of it reach the people pouring concrete?
The question that remains
The UK wants 6 GW by 2030, and today it has a few dozen megawatts of AI self-build, a paused flagship and a queue far larger than the country’s peak demand.
From Nscale Loughton to Cambois, the plans are quite grand, but what still remains as the open questions are the power and the owners.
So who will build UK AI data centres that the UK genuinely controls, and will they get power before the American landlords do?
Also read: Who Really Owns the UK’s Weight Loss Jabs Market?
Sources: Oxford Martin School, “Oxford researchers warn Parliament that the UK cannot rely on access to the world’s most advanced AI” (9 July 2026); CITP, “Mapping the UK’s data centres build-out” (7 July 2026); DSIT, UK Compute Roadmap; GOV.UK, Lanarkshire AI Growth Zone announcement (18 August 2026); Ofgem press release (29 July 2026) and call for input (February 2026); Savills UK data centre research (2026); CNBC on Nscale funding (9 March 2026) and the Stargate UK pause (9 April 2026); Nscale S-1 coverage via Techstrong and w.media; Data Centre Dynamics and City AM on Microsoft and Loughton (October 2026); Buro Happold for the GLA, via the Energy Institute (27 July 2026); company announcements from Civo, Apatura and Era4; TechCrunch on Fluidstack.
