The UK's AI Data Centres Are Announced Faster Than They Can Be Powered

Britain's flagship AI data centre has planning permission, Nvidia backing and Microsoft as anchor tenant. What it does not have is enough power to switch on.

Britain's flagship AI data centre cannot get the power to switch on

Britain's flagship AI data centre project has almost everything it needs. Nscale's £2bn facility in Loughton, Essex has full planning permission, backing from Nvidia, and Microsoft signed on as anchor tenant. It was announced during Donald Trump's UK state visit and positioned by the government as a centrepiece of Britain's AI infrastructure strategy. Designed to support 50MW of AI and HPC capacity scaling to 90MW, the site could eventually house up to 45,000 Nvidia GB200 GPUs.

What it does not have is power. National Grid has told Nscale it cannot connect the site in time for its planned opening, according to reporting from Armason's UK Data Centre Intelligence briefing. The initial late-2026 opening date was already missed because of planning delays. Now the grid connection, not the building or the chips, is what stands between a fully funded, fully permitted facility and actual operation. Nscale is reportedly in talks with California-based Bloom Energy to bridge the gap using solid oxide fuel cells rather than waiting for a permanent grid connection, echoing installations Bloom has already delivered at a Manchester data centre and the Wytch Farm oilfield in Dorset.

This matters well beyond one project. Nscale is the most visible example of a pattern playing out across the UK's AI infrastructure pipeline, and it is a useful reminder for any UK business planning to rely on domestic AI compute capacity over the next 18 to 24 months: the constraint has moved. It is no longer primarily about chip allocation or planning consent. It is about whether the electricity is actually there when the contract says it will be.

The initial late-2026 opening date was already missed because of planning delays. Now the grid connection is what stands between a fully funded, fully permitted facility and actual operation.

Grid queues are now the real capacity constraint, not chip supply

The Nscale story sits inside a much larger capacity problem. More than 125GW of planned UK energy and data centre projects are currently stuck waiting for a National Grid connection, according to the same Armason briefing. Sightline Climate data cited in the report suggests 26% of UK data centre capacity was delayed during 2025, and the firm estimates that between 30% and 50% of the global data centre capacity expected to arrive in 2026 could slip because of power constraints, equipment shortages, or local opposition.

The knock-on effect is already visible in how operators are responding. More than 100 UK data centre projects have reportedly said they will turn to gas generation as grid connection backlogs stretch to a decade or longer in some regions. That is a striking admission from an industry that has spent the last two years positioning itself around renewable-powered, low-carbon AI infrastructure. When the choice is between missing delivery dates on contracts with hyperscale anchor tenants or running on gas, commercial pressure is winning.

Separately, industry analysis from Capacity puts the scale of the shortfall in context: Britain currently has roughly a third of the data centre capacity per person that the United States has, and the gap is projected to widen rather than close without intervention. The government's own target is at least 6GW of data centre capacity by 2030, up from around 2GW today, a threefold increase inside four years, against a grid connection queue that is already the primary blocker for projects that have cleared every other hurdle.

More than 100 UK data centre projects have reportedly said they will turn to gas generation as grid connection backlogs stretch to a decade or longer.

What this means for workload placement decisions right now

For a UK business weighing where to run AI workloads over the next two years, this changes the calculus in a specific way. Sovereign or UK-hosted AI capacity has been treated by many buyers as the safer, more compliant default: data stays in the country, procurement is simpler, and the story to the board is straightforward. The Nscale case is a reminder that UK sovereign AI capacity and available UK sovereign AI capacity on the date you need it are not the same thing.

What this means in practice is that workload placement decisions now need a power and delivery timeline question built in, not just a data residency and compliance question. If a supplier or colocation partner is promising GPU capacity tied to a facility that has not yet secured its grid connection, that promise carries genuine delivery risk regardless of how strong the underlying funding or planning position looks. Ask directly: has this site's grid connection been confirmed and dated, or is it still sitting in the National Grid queue alongside 125GW of other demand?

The practical response is not to abandon UK sovereign hosting plans, which still make sense for regulated data and for firms with genuine data protection or contractual residency requirements. It is to stop treating announced capacity as delivered capacity, and to build contingency into procurement timelines, either through hybrid arrangements that can flex to established cloud regions in the short term, or through contract terms that protect the business if a supplier's own capacity commitments slip past their stated date.

UK sovereign AI capacity and available UK sovereign AI capacity on the date you need it are not the same thing.

Regulation is tightening at the same time compute is tightening

The power squeeze is arriving at the same time as a tightening UK regulatory position on how AI systems make decisions about people, which adds a second layer of planning that businesses cannot defer regardless of where compute sits. UKSI 2026/425 came into force on 12 May 2026, placing a legal duty on the Information Commissioner to prepare the UK's first statutory code of practice on AI and automated decision-making, as detailed in analysis from The Professor. The ICO's draft ADM guidance, which previews much of what that statutory code will say, closed for consultation on 29 May 2026, with final guidance expected in summer 2026 and the statutory code itself expected to follow in 2027.

The direction of travel in the draft guidance is worth noting: human review of an automated decision has to be active and genuine, not what the ICO's own commentary calls a token gesture. A manager who glances at an AI recommendation for a few seconds before approving it does not satisfy that bar. Individuals will also gain a formal right to request human review of automated decisions made about them, which means UK businesses need documented, auditable processes for handling those requests, not a policy document that sits unused on a shared drive.

None of this is contingent on where the underlying model runs. A business that solves its infrastructure sovereignty question but has not built genuine human review into its AI-assisted decisions is still exposed. Both problems need addressing in parallel, on broadly the same 12 to 18 month timeline, rather than treated as sequential projects.

Human review of an automated decision has to be active and genuine, not what the ICO's own commentary calls a token gesture.

But isn't government money already solving this? The counterargument

The obvious counterargument is that government money is already solving the capacity problem. The UK's sovereign AI strategy includes £500 million combined with the convening power of the state, national GPU access through the AI Research Resource, regulatory fast lanes, and government procurement access, as set out in coverage from Capacity. Separately, Microsoft's participation in the UK-US Tech Prosperity Deal included a headline £22 billion investment commitment for UK data centre and AI compute expansion. On paper, that looks like a comprehensive answer to the capacity gap.

The distinction that matters for a buyer is between a funding commitment and delivered, connected capacity. Nscale's Essex site had no shortage of capital, political backing, or hyperscale demand behind it, and it still could not get switched on to schedule because the constraint sat with National Grid, not with any of the parties who put up the money. Large headline investment figures announce intent and unlock planning; they do not, on their own, move a project up a multi-year grid connection queue that already has 125GW of demand stacked behind it.

This is not an argument against UK sovereign AI investment, which remains the right long-term direction for reasons well beyond any single facility's timeline. It is an argument for reading capacity announcements the way a careful buyer would read a supplier's product roadmap: as an intention with a delivery date attached, worth verifying rather than assuming.

Nscale's Essex site had no shortage of capital, political backing, or hyperscale demand behind it, and it still could not get switched on to schedule.

What UK business leaders should do before signing sovereign AI contracts

Before committing budget or workload plans to a specific UK sovereign AI facility or provider, there are direct questions worth asking, and answers worth getting in writing rather than in a sales conversation.

First, ask for the grid connection status of the specific site your capacity will run in, with a date, not a general assurance that power is being arranged. Second, ask what contingency the supplier has if that connection slips, whether that is a genuine backup plan such as an interim generation arrangement, or simply a hope that the queue moves faster than the industry average suggests it will. Third, build contract terms that protect your business if committed go-live dates move, particularly if you are planning to migrate workloads away from an existing provider on the strength of a promised UK capacity date.

Fourth, treat the ICO's tightening ADM position as a parallel workstream, not a follow-on task once infrastructure is sorted. Document how human review actually functions in any AI-assisted decision process now, while the code of practice is still in draft, rather than waiting for the 2027 statutory version to force the issue.

Finally, keep a genuinely portable option open. A hybrid approach, sovereign UK hosting for the workloads that require it, with the ability to flex to established cloud capacity for everything else, is currently the most defensible position for a business that wants the compliance benefits of UK infrastructure without betting its 2026-2027 roadmap on a grid connection queue it does not control.

A hybrid approach is currently the most defensible position for a business that wants the compliance benefits of UK infrastructure without betting its roadmap on a grid queue it does not control.