Microsoft has pulled out of its agreement to anchor Nscale’s planned AI data centre in Loughton, Essex, after power-connection delays pushed the expected electricity supply into the 2030s. The move, reported by Reuters, matters well beyond one site: it shows that in the current AI buildout, a signed customer contract and a headline GPU count mean far less than a firm date for megawatts.
The immediate question for cloud buyers, developers and investors is simple: when does announced capacity become usable capacity? At Loughton, the answer appears to be not when the land is identified, the chips are named or the customer is announced, but when the grid can actually deliver power.
A contract is not a powered data centre
Nscale announced the Loughton AI Campus in September 2025 as a 50-megawatt facility, scalable to 90 megawatts, with 23,040 Nvidia GB300 GPUs due in the first quarter of 2027 to support Microsoft Azure services in Britain. Reuters says the site had been due to begin operating early in 2027.
That timeline is what has now broken. Reuters reported that electricity-grid officials told Nscale they did not expect to supply power to the site until the 2030s. The Guardian separately reported that U.K. Power Networks had said it could not supply enough power until the early to mid-2030s.
Those dates make the original business case unusable for an anchor customer that expected near-term Azure capacity. Microsoft has said the decision concerns a single site and that it remains committed to its broader partnership with Nscale. Nscale says that relationship remains strong, that other contracts are unaffected, and that it is exploring alternative power sources for Loughton and other U.K. expansion sites.
That distinction matters. This is not public proof that the Essex campus is dead, or that every Microsoft-Nscale agreement is in trouble. It is proof that one flagship site no longer matched the timetable that sold the project in the first place.
The real dependency chain runs through the grid
AI infrastructure is often discussed as a race for chips, capital and customers. Loughton is a reminder that the practical sequence is harsher than that. A project needs at least four things to line up: land and planning; electrical infrastructure; power and cooling equipment; and live customer demand. If the grid connection slips by years, the rest of the stack stops being commercially coherent.
That has knock-on effects in both directions. An anchor customer helps a developer finance construction, reserve equipment and underwrite utilisation. But the customer can also walk away when the connection date no longer supports its own service plans. Once the anchor leaves, the project can look riskier to lenders, suppliers and future tenants. What began as a power delay becomes a financing and business-model problem.
That feedback loop matters especially in the neocloud market, where companies can report large GPU plans and large contracted values long before all sites are energised. Reuters noted that Nscale filed for a U.S. IPO in September and that Microsoft and Anthropic were expected to be among its major customers, according to the prospectus. Reuters also said Nscale’s total contracted value had grown from $100 million to more than $103 billion in two and a half years. Those numbers may describe commercial momentum, but they are not the same as operating megawatts, recognised revenue or delivered cloud service.
For Azure customers, another uncertainty sits underneath the headlines: Microsoft has not publicly disclosed whether the workload planned for Loughton will shift to another U.K. site or be served from elsewhere. So the key operational question is not who signed what in 2025. It is where firm capacity will actually be available when customers need it.
Why Ofgem’s queue reforms suddenly look more concrete
This is also a policy story. In Ofgem’s July consultation, the regulator proposed a commitment fee and queue-management milestones designed to push the connections process toward more credible projects. The proposed tests include evidence of a real end user, procurement of long-lead electrical equipment such as switchgear, and financial and technical capability.
Loughton gives that debate a very practical edge. A major AI campus can have an identified site, published technical specifications and a blue-chip customer, yet still fail to produce usable capacity on the promised date if the power path is not secure. At the same time, the episode should not be twisted into a verdict on every project in the queue or on the U.K.’s ability to support AI growth. Ofgem’s reforms are proposed guardrails, not a retroactive judgment on this specific site.
What the case does show is how much value is packed into one phrase that often gets glossed over: “power connection.” That can hide several different constraints, from substation work and local network reinforcement to transmission access, permitting and the contract date for firm delivery. Public reporting does not establish which combination produced the 2030s timetable at Loughton, and that gap is itself revealing. In this market, one of the most important project risks is often the least legible from the outside.
What buyers and investors should ask now
The most useful takeaway is not a broad claim that AI data-centre projects are overhyped. It is a stricter checklist.
Before treating a signed deal as capacity that can actually serve customers, buyers and investors should ask for the firm connection date and the responsible network operator; evidence that substation work and long-lead equipment procurement are moving; the share of capacity that is firm versus aspirational; and, if on-site generation is being proposed as a workaround, the permits, fuel source, emissions profile, outage plan, cost and amount of dependable power it can provide.
They should also ask how anchor-customer termination works, what replacement rights or fees exist, where GPU delivery and acceptance milestones sit against the energisation schedule, and how much of any reported contracted value has turned into operating revenue rather than future intent. Those are no longer specialist diligence questions. They are the difference between a project that photographs well and one that can deliver inference and training jobs on time.
Nscale may yet find an alternative route for Loughton, whether through on-site generation, another connection strategy or a revised tenant mix. Microsoft may replace the capacity elsewhere in its network. Both outcomes remain possible on the public record. But this week’s break is still a sharp warning for the wider market: in AI infrastructure, the scarcest asset may be neither money nor GPUs, but a credible date when the electricity arrives.
By
By
By
By
By
By
