Microsoft’s Reported Nscale Exit Puts Britain’s AI Power Bottleneck in Focus

Microsoft has reportedly withdrawn from Nscale’s planned Loughton data-centre agreement as grid delays threaten the site’s timetable, while Nscale says its wider Microsoft relationship remains intact. The development highlights why power delivery, alongside chips and financing, determines when AI capacity becomes usable.

A data centre can have an ambitious customer agreement and a detailed equipment plan while still facing a basic unanswered question: when can it obtain enough electricity to operate? Sunday’s reporting about Nscale’s Loughton project in Essex brings that question into focus for the United Kingdom’s AI infrastructure ambitions.

Reuters reported on October 18, citing the Telegraph, that Microsoft had terminated its agreement for the site. Reuters said Microsoft had not immediately responded to its request for comment. The report linked the project’s difficulties to repeated delays and said grid officials did not expect a power supply until the 2030s.

Nscale told Reuters that its relationship with Microsoft remained strong, other contracts were unaffected, and it was exploring alternative power sources while remaining committed to Loughton and the UK. Those statements matter: the reported withdrawal concerns a particular agreement, while the company’s response describes a continuing relationship. Neither establishes that an alternative electricity solution has already been delivered.

What the original Loughton announcement promised

The most useful baseline is Nscale’s own September 16, 2025 announcement. It described an AI campus delivering 50 megawatts of capacity, scalable to 90 megawatts, with an initial 23,040 NVIDIA GB300 GPUs scheduled for the first quarter of 2027. The facility was intended to support Microsoft Azure services in the UK.

That was a forward-looking deployment plan. It should not be confused with an operational inventory of installed, powered and available servers. The same announcement described other projects, including a separate Stargate UK initiative involving OpenAI and NVIDIA. Combining every item in that package into the Loughton agreement would obscure the scope of Sunday’s news.

The distinction between a commitment and a functioning service is central to evaluating infrastructure announcements. A proposed GPU count describes potential computing equipment. A power figure describes an electrical requirement or capacity. A delivery quarter sets an expectation. Customers ultimately need all three to align with construction, networking and commissioning before the advertised resource can become dependable capacity.

One site is part of a wider network of commitments

Nscale’s later announcement of approximately 200,000 GB300 GPUs contracted with Microsoft described several locations. It listed approximately 104,000 GPUs at a roughly 240-megawatt Texas campus, approximately 12,600 in Sines, Portugal, and an earlier agreement involving approximately 52,000 in Narvik, Norway. It also repeated the Loughton plan.

These were company-announced commitments and projected delivery arrangements, not independent confirmation that every phase is now operating. They nevertheless explain why a change at one campus cannot automatically be extended to every Microsoft–Nscale contract.

Geographical diversification can spread development risk, but it does not remove the need to examine each project separately. A connection date in one jurisdiction does not validate a connection date elsewhere. Likewise, a customer’s ability to procure capacity in another location does not prove that relocation has happened. The evidence needed is specific: the site, the contracted service, its delivery conditions and its actual operational status.

Britain’s connections queue already prompted regulatory action

The broader issue predates Sunday’s report. In a July 29, 2026 consultation announcement, Ofgem said contracted demand connection offers had increased from 41 gigawatts to 125 gigawatts between November 2024 and June 2025, driven largely by data-centre applications.

The regulator proposed a commitment fee and project milestones designed to distinguish viable developments from speculative applications. Its proposed fee would be refunded at energisation and forfeited if a project left the queue early. Ofgem said responses would inform final policy decisions; a consultation proposal should not be presented as an automatically effective rule.

This is not evidence that Ofgem classified Loughton as speculative. It is evidence of a system-wide planning problem. A queue containing projects that may never proceed can make demand appear larger than deliverable demand, while delaying schemes that are ready to build. Equally, removing an application from a queue does not by itself construct the physical network needed by the next applicant. Administrative readiness and engineering delivery are related but distinct challenges.

The regulator separates three different jobs

Ofgem’s February demand-connections call for input discusses financial mechanisms, readiness requirements and the possibility of ramped or non-firm connection arrangements. These options concern how users enter the system and how available capacity might be used, rather than a promise of unrestricted electricity for every applicant.

Its reform programme separates curating a credible queue, planning strategically important demand, and accelerating physical connections. That framework is useful for reading the Loughton story. A strong commercial agreement can help establish readiness. Government support can signal strategic importance. Neither automatically settles the physical connection timetable.

A ramped connection may allow a user to increase consumption in stages. A non-firm arrangement raises a different question about the circumstances in which supply can be constrained. For an AI service provider, those operating conditions matter alongside a headline megawatt figure. The relevant comparison is the service a customer can reliably obtain at a particular time, not simply the largest number associated with the campus.

Demand is growing, but forecasts are not installed capacity

The International Energy Agency’s updated assessment of energy and AI projects global data-centre electricity consumption rising from about 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030. It says bottlenecks across the supply chain are reducing the likelihood of more aggressive near-term growth scenarios, despite substantial investment and project pipelines.

The agency also identifies financing conditions, market sentiment and expectations for returns as influences on the pace of development. A demand forecast therefore does not mean that every announced campus is needed on its initial schedule, nor that every customer agreement will survive unchanged.

There is a difference between the long-term possibility of growing demand and the commercial value of a specific facility delivered several years later than expected. Hardware, customer requirements and competing capacity can all change while a project waits. The analytical question is whether the project can still provide a useful service on the timetable and terms available, rather than whether AI will continue to exist as a source of electricity demand.

Why global electricity shares can hide local pressure

The IEA’s 2025 demand analysis explains that data centres concentrate substantial consumption in particular locations. Its modelling distinguishes enterprise facilities, colocation and service-provider sites, and hyperscale operations. It also identifies the different growth rates of accelerated servers, conventional servers and supporting infrastructure.

A relatively modest share of worldwide electricity use can therefore coincide with a difficult local connection problem. Global totals answer how much electricity a sector consumes across many networks. A local connection assessment asks whether a specific part of the network can deliver a particular load within acceptable operating limits. The first number cannot substitute for the second.

This is why a project’s location is more than a property decision. The surrounding network, the timing of upgrades and the ability to coordinate equipment with energisation can determine when a facility starts earning revenue. The same distinction helps avoid the opposite mistake: a delay in one place does not demonstrate that every grid, country or data-centre project faces an identical constraint.

Alternative power raises another set of questions

Nscale’s statement that it is examining alternative power sources identifies a possible response, not a completed remedy. The IEA’s supply analysis considers a mix of renewables, natural gas, coal and nuclear electricity, with substantial regional differences. It projects renewables meeting nearly half the growth in global data-centre electricity demand through 2030 in its base case.

The relevant questions for any alternative are concrete: how much usable electricity can it provide, when, under what operating constraints, and with what supporting infrastructure? A plan must also fit the reliability requirements of the service being offered. Naming an energy source does not establish those details.

Climate implications require similar care. The IEA’s emissions analysis distinguishes data-centre electricity emissions from possible emissions savings enabled by AI elsewhere. A potential efficiency benefit in another sector does not automatically cancel a project’s electricity footprint. The two effects need separate evidence, boundaries and time periods before they can be compared responsibly.

What would demonstrate progress from here

Units also matter when comparing announcements. Megawatts describe power: the rate at which electricity is supplied or used. Terawatt-hours describe energy accumulated over time. A campus rated at a given power level does not necessarily operate at that level continuously, and a worldwide annual consumption estimate is not a connection offer. Similarly, a GPU count cannot be converted into an exact campus electricity requirement without knowing the equipment configuration, utilisation and supporting systems. Keeping those measurements separate prevents a plausible-looking numerical comparison from becoming a misleading conclusion about the capacity available to customers.

Three developments would materially improve the picture: a clarified customer arrangement, a credible and documented power-delivery timetable, and evidence that the resulting service can be commissioned. Until then, the original 2027 ambition remains historical context rather than a current guarantee.

For readers following the wider buildout, our report on Europe’s AI research infrastructure covers a separate initiative. The Loughton development illustrates a different constraint: computing ambition depends on a sequence of physical and commercial deliveries. Sunday’s reported contract change is significant, but it does not establish the fate of Nscale’s entire business, every UK AI project or the broader market for AI services.