Announced, contracted, powered: Microsoft's exit from an Essex AI data center shows which capacity counts
Microsoft left Nscale's planned Essex AI data center after the grid operator said full power may not come until the early to mid 2030s. What capacity buyers should contract for.
Listen to this article · 5 min
AI-generated narration of the full article.
A data center can be announced by a government, promoted as sovereign infrastructure, financed, signed with an anchor tenant and still have no date for the electricity it needs. That is where Nscale’s planned AI data center in Loughton, Essex, stands after the weekend of October 17 and 18.
According to the Guardian, following a report in the Telegraph, Microsoft has pulled out of its agreement to be the anchor tenant of the site, the only publicly announced customer for it. Earlier in October, UK Power Networks, which runs the grid across London and the south-east of England, told Nscale it could not supply enough power to the site until the early to mid 2030s. The project had been due to open next year.
From announcement to the power date
2025Early to mid 2030s
- 2025Announced as the UK's largest sovereign AI data center
- March 2026Site still an operating scaffolding yard, per the Guardian
- October 2026Grid operator: full power not before the early to mid 2030s
- October 18Microsoft exit reported
- 2030sEarliest full grid supply
What is known, and what is not
Nscale says it remains committed to Loughton and to the UK, is exploring alternative power sources, including generating power on site, and is looking at other UK locations. It calls the problem “this unfortunate London-wide grid issue” and says its relationship with Microsoft “remains strong and long-term,” with no impact on other contracts. The Guardian reports Microsoft remains a customer for an Nscale project in Norway. The reporting ties Microsoft’s decision to power supply concerns; Microsoft’s own reasoning has not been published in detail.
This is one project, not proof that AI demand is falling, and not a judgment on Microsoft’s or Nscale’s wider plans. Its value for enterprise buyers is that it makes visible a gap that usually stays inside contracts: the distance between capacity that has been announced, capacity that has been contracted and capacity that is energized and serving.
Capacity has three states
Announced capacity is a press release and, often, a policy goal. In the UK, this site was promoted as the country’s largest sovereign AI data center.
Contracted capacity is a signed commitment, with a tenant, a price and a start date. It is what most buyers think they have bought.
Energized capacity has a grid connection or its own generation, cooling and hardware installed, and is serving workloads. Only this one runs inference.
Grid connections are increasingly the slowest of these, and no improvement in chips shortens them. That is the part of the AI cost picture we left out when we looked at what cheaper inference chips mean for customers: a provider’s cost per token can fall while the date your capacity becomes real slips by years.
What to write into a capacity contract
Most enterprises do not lease data centers. Many do sign multi-year commitments for GPU capacity, reserved throughput, sovereign or dedicated cloud regions, or private AI infrastructure with a specialist provider. The Essex case suggests five questions for those contracts:
- Is the capacity energized today? If not, what are the milestones (grid connection, generation, commissioning), and what happens to your commitment if one slips?
- What are your exit rights? The anchor tenant here was able to withdraw. Check that your contract gives you a comparable option on a delivery failure, without paying for capacity that does not exist.
- What can be substituted, and where? A provider that cannot deliver in one location may offer capacity in another. For a workload chosen for its jurisdiction, capacity in a different country is a different product. A sovereign commitment that can be fulfilled abroad is not sovereign. We described how far sovereignty now reaches down the stack.
- How concentrated is the supply? Specialist AI cloud providers grow on project finance and anchor tenants. If your capacity depends on one site, one provider and that provider’s financing, name that in the risk register.
- What is the fallback? Keep at least one tested path to run critical inference elsewhere, even at a higher price, so a delayed site becomes a cost problem and not an outage.
The planning lesson
AI roadmaps still tend to treat compute as something that appears when the budget is approved. The Essex case is a reminder that it is physical: land, permits, transformers, cables and a grid operator with its own queue. For any plan that depends on capacity not yet running, the useful question is not “who is building it?” but “when will it be powered, and what do we do if it is not?”