Microsoft has withdrawn from its agreement to use an AI data center being developed with Nscale in Loughton, Essex, the UK's Telegraph reported on October 18. The 2025 announcement said the site would deploy 23,040 NVIDIA GPUs, with delivery planned for January to March 2027. According to the newspaper, however, a grid connection is not expected until the 2030s. The specific reason for Microsoft's exit has not been disclosed, but the plan, billed as the UK's largest AI supercomputer, is now undermined by the assumption that it could secure an anchor customer while waiting for a power connection.
GPUs planned for 2027, grid connection in the 2030s
Nscale and Microsoft announced on September 16, 2025 that the Loughton AI site would be used for Azure's UK services. At the time of the announcement, the AI capacity was initially 50 MW, expandable to 90 MW in the future, and the initial configuration was to use GB300 GPUs. If the timing of GPU delivery and the timing of receiving the required power from the grid are several years apart, it becomes difficult to provide customers with the planned computing capacity on schedule.
However, it has not been confirmed that the grid connection delay caused Microsoft's exit. Reuters reported that the Telegraph did not identify the reason for the withdrawal and that Microsoft declined to comment. The power problem is a concrete part of the background to the project, but the contract's termination conditions and the decisions the two companies made remain unclear.
Nscale told Reuters that it will continue its UK operations, including Loughton, and is considering alternative power sources and other candidate sites in the UK for expansion. It said its relationship with Microsoft is long-standing and that the London-area grid problem will not affect other contracts. The report of a major customer's departure therefore cannot be read as Nscale canceling the Loughton site itself.
On its facility overview, Nscale lists the Loughton site's features as fiber connectivity to London's financial district, support for high-density GPUs, and direct liquid cooling. Network links and cooling equipment are important elements for supporting large-scale AI processing, but even if they are well designed, a facility cannot operate without securing the power it needs. Likewise, the statement that the site can be expanded to up to 90 MW does not mean that power at that scale is available now.
Of up to 58,640 GPUs, this case covers 23,040
The figure of up to 58,640 that Nscale put forward in September 2025 is a UK-wide plan combining 23,040 for Loughton, up to 31,000 for Stargate UK, and 4,600 for NVIDIA DGX Cloud. In the announcement at the time, the sites and user companies differed, and so did how firm each plan was.
| Framework announced in September 2025 | GPUs | Positioning at announcement |
|---|---|---|
| Loughton site for Microsoft | 23,040 | For Azure. GPUs planned for delivery January–March 2027 |
| Stargate UK with OpenAI and NVIDIA | Up to 31,000 in the future | Multiple UK sites envisioned. OpenAI considering use of up to 8,000 initially |
| Collaboration with NVIDIA DGX Cloud | 4,600 | Planned deployment for UK AI developers |
Adding the maximum figures for the three frameworks gives 23,040 + 31,000 + 4,600 = 58,640. However, the 31,000 for Stargate UK indicates potential future expansion, and not every plan in the table was a firm contract with the same degree of certainty. All are forward-looking plans as of the announcement, not numbers of GPUs confirmed to have been installed and put into operation.
What was reported this time is the contract for the Loughton site that Microsoft was to use. It does not mean the entire UK GPU plan has disappeared, nor has it been confirmed that the remaining 35,600 GPUs will be realized as planned. In particular, confusing Microsoft's contract with the separate plan involving OpenAI makes it harder to tell which company is expected to use the computing resources at which facility.
The $30 billion UK investment needs to be considered separately
In a June 2026 official UK article, Microsoft also mentioned the Loughton supercomputer being built with Nscale. It was not merely a name listed in the previous year's announcement; it was a plan positioned within the explanation of Microsoft's 2026 UK business.
Meanwhile, the $30 billion investment across the UK described in the same article consists of $15 billion for cloud and AI infrastructure and $15 billion in operating costs, including employees and skills development. The capital investment targets cited included not only Loughton but also data centers in Acton, west London, and Newport, south Wales. The Loughton contract and the UK-wide investment total, which combines facilities and operations, therefore cover different scopes.
It is not possible to calculate from this report how much of the $30 billion has been withdrawn. The investment amount allocated to Loughton and the cost burden associated with terminating the contract have not been disclosed either. Even so, if one of the sites symbolizing UK AI infrastructure investment has lost its anchor customer, the impact on the plan is not small. Even when total investment is large, when a site will actually go live and which customers it can supply with computing capacity differ from site to site, and at Loughton the problem has surfaced at the execution stage.
The 73 GW awaiting connection is not actual operating power demand
According to a reform proposal document published in July 2026 by Ofgem, the UK's energy regulator, the total volume of connection contracts offered on the demand side surged from 41 GW in November 2024 to 125 GW in June 2025. For data centers alone, there are about 315 projects with a combined connection capacity of about 73 GW.
However, 125 GW is the figure for the demand side as a whole, including industrial facilities, and 73 GW is a tally of projects that have applied for or are planning connections as data centers. It does not mean that currently operating data centers are consuming 73 GW, nor is it a demand forecast premised on all plans being realized. It also cannot be simply compared with generation capacity to conclude that "the UK is short of 73 GW of power."
Ofgem points out that projects that cannot secure planning permission or funding, and therefore will not actually reach connection, remain in the connection queue and risk delaying projects that are more likely to be delivered. If plans that will never materialize pile up, it also becomes harder to judge how much transmission and distribution network needs to be built in which regions. At the same time, because building the networks and generation capacity themselves takes time, even well-prepared projects face a long wait for connection.
In other words, removing unlikely projects from the connection queue and ensuring that sufficient power can actually be delivered to facilities are separate challenges. Announcements that funding or GPUs have been secured alone do not show which of the two has been resolved, or to what extent.
Which comes first: power or customer contracts?
Ofgem's reform proposal suggests a mechanism requiring data center developers to demonstrate, as development progresses, that there are companies that will actually use the computing resources and that they have procured long-lead electrical equipment. For facilities planned to be leased or sold to third parties, the approach is to move from non-binding agreements at an early stage to legally binding usage contracts as development advances.
However, simply requiring contracts with customers does not solve the problem. In section 6.31 of the document, Ofgem expresses concern about a circular problem in which whether a usage contract can be concluded at all depends on the certainty of the power connection. Customers want to sign a contract knowing when they can use the computing resources, while the grid side wants to confirm that there is a customer who will actually use that demand. Obtaining a commitment from one side requires a commitment from the other.
The latest report on Loughton shows what problems this interdependence could lead to in the data center business. While the outlook for a power connection recedes, there is no guarantee that a customer will keep its plan to use computing resources at the same site. If the customer leaves, the developer must not only secure power but also find a new company to buy the site's computing resources.
However, Ofgem's reform proposal should not be interpreted as a penalty against Loughton. The consultation page showed "closed, awaiting decision" as of October 19. There is no information indicating that Loughton has been judged a speculative project or that it lost its connection slot because of Microsoft's exit.
The focus going forward is whether Nscale can turn the alternative power sources it is considering into a concrete plan with the required capacity and supply timing, and then sign new users on the basis of that schedule. Only by securing both power and customers can the Loughton plan turn UK AI investment into computing capacity that customers can actually use.