Microsoft’s reported exit from Nscale’s Essex project shows that a signed tenant, planning permission and a grid agreement still don’t guarantee a switch
A project with everything except power !
The data centre industry has treated electricity as a line item: Secure the land, sign the tenant, order the GPUs, and the grid connection will follow – this week showed that this sequence no longer works.
Microsoft has pulled out of its data centre deal with Nscale in Loughton, Essex, according to The Telegraph, with Reuters relaying the report. The Telegraph did not say why, and Microsoft have not commented at the time of reporting, but the context is hard to ignore: Grid officials had reportedly told Nscale not to expect a power supply for the site until the 2030s. The £2bn project had been targeting a 2027 opening.
Nscale says it remains committed to Loughton and is exploring alternative power sources and it is also at pains to say that the episode does not affect its other contracts with Microsoft.
What makes Loughton instructive is how much was already in place: The project had planning permission, Nvidia backing, Microsoft as anchor tenant and a 90MW grid connection agreement – everything a conventional development checklist asks for was there – except the one item that couldn’t be bought: electricity delivered on time.
Queues, fees and registries: the free option on grid capacity is ending
Loughton is a symptom of a wider problem. In Great Britain, demand connection applications rose from 41GW to 125GW between November 2024 and June 2025, driven largely by data centres – Ofgem puts the data centre share at around 73GW.
Many of those projects will never be built, that is why Ofgem is consulting on a returnable commitment fee for data centres above 40MW, plus new progression milestones for those of 10MW and above – the regulator’s message is that a queue position should be earned, not simply held.
The same pressure is visible elsewhere – In the US, FERC has approved a PJM registry to track data centres and other large loads of 50MW or more. PJM expects data centre demand to nearly triple by 2030, to about 29GW. PJM’s latest capacity auction cleared at its $325 per megawatt-day cap for 2028/29, eleven times the 2024/25 price. When regulators start building registries and charging fees, the free option on grid capacity has ended.
Investors are pricing deliverability, not announcements
The money is beginning to ask the same question. Companies have raised an estimated $466bn of AI-linked debt so far in 2026, against $101bn in the same period last year, and the flood of supply has let credit investors demand higher yields. Issuance slowed sharply in September as investors worried about leverage and execution risk, and the Financial Times recently reported that access to electricity is among the concerns.
That is a healthy correction – A signed tenant is not a revenue stream until the building is energised. For a lender, a connection date in the 2030s on a facility due in 2027 is not a delay, it is a different asset.
Four questions every AI data centre should answer
Developers and investors should be judged on less glamorous metrics than announced gigawatts:
- Where is the power, and is it contracted rather than merely applied for?
- When does it arrive, and what happens to the business case if that date moves?
- What upgrades does delivery depend on, and who funds them?
- What is the bridging plan, and is it financeable?
That last question is already being asked. Industry data suggests more than 100 UK projects are weighing gas generation while they wait for grid connections. On-site power can be a bridge, but it adds cost, permitting and complexity of its own and it does not make the grid problem go away.
The data centre is now an infrastructure project, not a building
Power density shapes cooling. Cooling shapes the building. Electrical infrastructure shapes site layout, and security and network design now start on the construction site. The modern AI facility is an integrated infrastructure project that just happens to contain servers, and it is only as deliverable as its slowest component.
Fast only counts if it’s real
Demand for compute is not going away and nobody should read Loughton as a verdict on AI. But the industry’s obsession with speed needs a qualification: Speed is valuable only when the infrastructure underneath it is real.
The winners of this cycle may not be the developers announcing the most gigawatts – they may be the ones who can answer the plainest question in the business: can you actually deliver it?
AI may be virtual, but the infrastructure it runs on is not.
