The Story
Microsoft has terminated its agreement to anchor a planned AI data centre at Loughton in Essex, operated by British infrastructure company Nscale, after electricity supply to the site was pushed into the 2030s.
UK Power Networks informed Nscale that sufficient power would not be available until the early to mid-2030s. The facility had been scheduled to come available in the first quarter of 2027, itself a slip from an earlier target of the final quarter of 2026. It was intended to house what would have been Britain's largest AI supercomputer.
A Microsoft spokesperson said the decision relates to a single site, that the company remains committed to the Loughton project and to the UK, and that it is exploring alternative power sources for the site as well as other expansion locations. Nscale described the problem as a London-wide grid issue and said its facilities in the United States, Portugal and Norway are unaffected, with the wider commercial relationship intact.
The project formed part of Microsoft's ยฃ22 billion commitment to British AI infrastructure and operations between 2025 and 2028, and was among the technology deals announced during the US president's state visit in September 2025, presented as evidence of Britain's ambition to lead in AI.
Nscale was founded in 2024 and is backed by Nvidia, having raised $1.1 billion in September 2025. In October 2025 it signed an agreement to supply Microsoft with more than 200,000 Nvidia chips.
Grid connection bottlenecks have delayed British data centre projects by as much as 13 years. Energy regulator Ofgem is working to clear a backlog of connection applications, and the government's AI Growth Zones are intended to offer faster planning decisions and improved electricity access.
Why It Matters
The interesting failure here is not technical. It is that the announcement and the engineering were never on the same timetable, and nobody reconciled them in public for thirteen months.
Loughton was presented in September 2025 during a state visit, alongside a ยฃ22 billion commitment, as proof that Britain could host frontier AI infrastructure. The site was to be available in late 2026, then early 2027. At some point the distribution network operator worked out that the power to run it could not be delivered before the 2030s. That is not a marginal revision. It is a gap of six or seven years between what was announced and what the grid could actually do, on a project whose entire purpose was to demonstrate national capability.
The sequence matters because it is becoming standard. AI infrastructure gets announced at diplomatic speed, with a head of state present and a number attached, and then gets delivered at utility speed, governed by transformer lead times and connection queues that nobody consulted beforehand. The announcement creates a political commitment; the grid creates the actual schedule; the two are discovered to be incompatible later, usually by the tenant.
Microsoft's statement is careful and worth reading closely. It relates to a single site, the partnership continues, alternative power sources are being explored. All of that is probably true. It is also what a company says when it has decided the schedule is not recoverable but wants to keep the counterparty and the government on side.
The useful question for everyone else building this capacity is simpler than it sounds. Before the press release, did anyone ask the network operator for a connection date in writing.
The Strategic Read
The constraint that killed this project is administrative, which makes it both more fixable and more embarrassing than a shortage would be.
Britain is not short of electricity in any absolute sense. It is short of connections, and the queue to get one has reached thirteen years in places. That is a sequencing problem created by a first-come-first-served application process in which speculative projects hold places ahead of funded ones. Ofgem's reform effort is aimed squarely at it. A generation shortfall would take a decade of construction to fix; a queue can in principle be re-ordered in a year.
The complication is that Microsoft cannot wait for the reform to work, and neither can anyone else. Capital allocated to AI capacity has a return clock on it, measured against hardware that depreciates fast. A site that delivers in 2033 is not a delayed version of a site that delivers in 2027. It is a different investment entirely, built around chips that do not exist yet.
For India the parallel is close enough to be uncomfortable. The domestic data centre buildout rests on the same assumption that power arrives roughly when the building does, and India's transmission constraints and land-to-grid timelines are not obviously better than Britain's. Salil Parekh warned a week ago that the data centre backlash would reach India. This is the version that does not require any public opposition at all.
Nscale is the exposed party. It is two years old, Nvidia-backed, and has just lost its anchor tenant on its first owned facility.
For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.