Microsoft Walked Away From Britain's Biggest AI Supercomputer. The Grid Won't Be Ready Until the 2030s.

On 16 September 2025, Satya Nadella announced a supercomputer in a former scaffolding yard on the edge of north-east London. It would be the UK's largest, he said, "powered by over 23,000 GPUs," and it would run Microsoft's Azure services for the whole country.

Thirteen months later, Microsoft has walked away from it.

The Telegraph reported on Sunday that the company had recently informed Nscale, the British neocloud building the site, that it was terminating the agreement. Microsoft has not confirmed the termination and did not comment to the outlets that reported it. Neither company has given a reason. Nscale says it remains committed to Loughton and that its relationship with Microsoft is "strong and long-term" everywhere else.

The obvious frame is another wobble in the AI buildout: a neocloud losing its anchor tenant weeks before a listing, in the same month that Australia's Firmus scrapped its own float and CoreWeave posted its worst week since November. That frame is accurate as far as it goes. But it misses the specific thing that broke here, which is not a demand forecast or a credit spread. It is a date on a power connection that the local grid operator says will not arrive until the early to mid 2030s.

What Loughton was supposed to be

The project was announced as part of the UK-US Technology Partnership, alongside Stargate UK and a promise to deploy 58,640 Nvidia GPUs across Britain. Loughton was the headline: 50MW of AI capacity, scalable to 90MW, with 23,040 GB300 chips due in the first quarter of 2027.

The government treated it as proof of concept for its AI strategy. Keir Starmer called the package "a decisive step towards the UK becoming a world leader in AI." Nscale put £2bn behind the build and promised 250 permanent jobs. Planning at Langston Road was validated in May 2026, on top of an outline approval from 2024.

Then the schedule started moving. In March, the Guardian visited and found the site still operating as a scaffolding yard. Nscale said the timetable had been revised to accommodate "the latest Vera Rubin 200 technology" and that the facility would be operational in the second quarter of 2027. By September the yard had been cleared. The building had not been built.

None of that is unusual in data centre construction. What was unusual was the other clock running alongside it.

The connection date

The Loughton campus needs 90MW at full scale. That is roughly the annual electricity consumption of 315,000 homes. In July, reports emerged that Nscale had been told the 90MW grid connection would not be ready for a 2027 opening, and that the company was talking to Bloom Energy about solid oxide fuel cells to bridge the gap.

In September, the timeline got worse. UK Power Networks, which owns the distribution network across London and the south east, told Nscale it could not supply enough power until the early to mid 2030s. The delay was first reported by the industry newsletter Morning Intelligence and confirmed in the Guardian's reporting.

UKPN would not discuss an individual project, but its explanation is the important part: "In general, many datacentres, whilst they will connect to our distribution network, are awaiting works on the wider national electricity transmission network before they can begin operation."

That is the whole problem in one sentence. The local wires are one constraint. The transmission grid behind them is another, and it is the one with the multi-year lead time. National Grid says it is working with NESO, the distribution operators and customers "to identify opportunities to accelerate connections where possible."

The size of the queue explains why. Ofgem disclosed in June that 315 data centres were waiting to connect to the National Grid, representing 73GW of demand. National peak demand for the entire country is about 45GW. The queue is not waiting for the grid that exists. It is waiting for a grid that has to be built first.

Microsoft left the site, not the company

It matters that Microsoft did not exit the relationship. The two companies are still working together in Norway, Portugal and the United States, and Microsoft remains involved in Nscale's Norwegian expansion. As recently as last month, Microsoft said part of the $30bn it is investing in the UK would fund the Loughton supercomputer.

That is what makes this different from a capex retreat. Microsoft is not telling the market it has stopped needing compute. It is telling the market, without saying anything at all, that it will not wait a decade for a building.

Microsoft declined to comment to the BBC, Reuters and the Telegraph, and did not respond to Bloomberg's request. The grid delay is the obvious backdrop, and neither company has said it was the cause. Treat the causation as inference, because that is what it is, even if the inference is loud.

The structural point stands either way. An anchor tenant is not a customer in the ordinary sense. It is the commitment that makes a project financeable: the long-term take-or-pay contract that a lender will underwrite. Remove it and the building does not lose a buyer. It loses its reason to exist on the current schedule.

The balance sheet built on delivery dates

Nscale's own filing shows how much of the story rests on things arriving on time. Its SEC registration statement, filed on 18 September, reports $103.4bn of total contracted value as of 31 August. By late August, roughly 2.5% of that was active.

Microsoft and Anthropic account for around 85% of the contract book. The Telegraph puts Microsoft's share above 40% of the $103bn, though it notes the UK deal is a small part of the total. TSN Media, which read the filing, reports statements of work with Microsoft worth up to about $43.8bn through December 2033 and agreements with Anthropic worth up to about $44.6bn.

The income statement is the other half. For the six months to 30 June, Nscale lost $1.02bn on revenue of $140.6m, against a loss of $368.9m on $10.4m a year earlier. Revenue grew 1,252%. So did the loss. Norway, the flagship, is six months behind, with 6.4MW of a planned 30MW running.

Now the listing. Nscale filed to list on the NYSE under NSCL, aiming to raise about $3bn. The Telegraph reported a valuation of up to $35bn. Bloomberg, writing late on Sunday, said the company has been seeking about $30bn, and that a cohort of investors has urged $16bn to $19bn instead, citing the collapse of Firmus's float among other things. The private mark was $14.6bn in March.

Read those numbers together and the Microsoft exit stops being a footnote. Nscale raised $3.36bn in convertible notes on 25 September, led by Third Point, with Nvidia contributing $1bn. That money buys time. It does not buy a customer for a site that cannot be energised.

The reform is real, and it is late

It would be wrong to write this as a story about British decline. The UK has been rewriting its connection rules at speed. NESO has cleared a queue that had grown past 700GW, four times what the country needs by 2030, and reviewed 1,500 applications in five months. The old first-come-first-served system, which let speculative projects sit on connection dates for up to a decade, is gone.

But the replacement machinery is still slipping. On 2 October, Ofgem granted NESO a second extension to issue Gate 2 Phase 1 connection offers, pushing the deadline from 30 September to 31 October, affecting about 62 offers because the transmission owner's construction offers arrived late. In February, Ofgem disclosed that 210 of 340 projects with protected connection dates, 62% of them, would have their dates or connection points changed anyway. Of those, around 135 were delayed for network reasons.

So the fix is moving. It is just not moving at the speed of a 2027 opening, and the gap between those two speeds is where a £2bn project with an anchor tenant goes to wait.

Developers are responding the way you would expect. More than 100 UK data centre schemes are reportedly weighing on-site generation. Nscale's talks with Bloom Energy point at solid oxide fuel cells, which can be running in months rather than years and run on natural gas. Scotland has halted new hyperscale AI data centre approvals. Amsterdam has restricted new builds and expansions. Denmark has proposed giving large electricity users lower priority in the connection queue.

None of these are isolated policy choices. They are the same physical constraint showing up in different jurisdictions, and each one pushes a workload somewhere else.

What the date is worth

For a decade the AI infrastructure conversation has been about chips: allocation, yields, packaging, memory. Those constraints are real and they are still binding. But the Loughton case is the first clean demonstration that the binding constraint can be a distribution network's calendar, and no amount of GPU supply fixes it.

That changes what a contract is worth. A $103bn backlog is only as good as the delivery dates underneath it, and the counterparties on the other side are now pricing that in. If Nscale reprices or postpones, it becomes the comparable that every other neocloud gets measured against, the way Firmus already has.

It also changes what "sovereign AI" means in practice. A government can designate a growth zone, lobby a planning committee and announce a partnership. It cannot make 90MW appear in Essex before the transmission works are done. Buyers planning capacity in the UK now have a specific address to ask about, and a specific date to ask for.

The wider lesson is unglamorous. In this buildout, the scarce input is a dated, enforceable connection agreement. Silicon and capital are abundant by comparison, and the companies that treat power as a procurement problem rather than an engineering one are the ones still on schedule.

If your hardware roadmap assumes compute capacity that has not been energised yet, the risk sits in the gap between the contract and the commissioning date. At DMC, we work with hardware companies modelling exactly that: sourcing strategy, cost modelling, and production ramp planning when the infrastructure arrives later than the plan. If your 2027 depends on a megawatt that is still in a queue, let's talk.