Microsoft has walked away from its contract for Nscale’s planned AI data center in Loughton, England, and weak demand had nothing to do with it. The grid simply could not get the electricity there on time.
Nscale announced the site in 2025 as a 50 MW campus that could grow to 90 MW, with 23,040 Nvidia GB300 GPUs earmarked for Microsoft Azure. According to The Guardian, UK Power Networks told the company it could not supply enough power until the early to mid-2030s, long after the planned opening. A spokesperson confirmed Microsoft ended the contract, said Nscale still backs the site, and added that its other Microsoft contracts are unaffected. Nscale is exploring alternative power and other UK locations, though neither company has disclosed any financial fallout.
Analysts see a pattern here. Alex Cordovil of Dell’Oro Group says projects need a creditworthy customer, financing backed by that customer and available electricity all at once. Developers can influence the first two far more than the third, because grid operators can push connection dates back by six months or a year, sometimes longer. As he put it, capacity without power earns nothing while the bills keep arriving.
Neoclouds feel that pain more than hyperscalers. A giant like Microsoft can spread one late site across a huge fleet. A neocloud often leans on a few anchor customers and finances GPUs against specific deployments. Owen Rogers of Uptime Institute Intelligence calls the debt colossal and the demand unassured.
Nscale’s September IPO filing makes the stakes concrete. It reported $103.4 billion in active and contracted value as of August 31, 2026, against just $2.6 billion in active contracts. Microsoft’s agreements allow payments of up to $43.8 billion through 2033, subject to delivery requirements, and its Anthropic agreements allow up to $44.6 billion. The filing also said Nscale had not secured binding financing for the GPUs and infrastructure needed for the Anthropic deals. It does not break out spending tied to Loughton.
Clock pressure adds to the problem, since Nvidia ships new generations roughly yearly. Cordovil argues this is no sign that AI demand is fading, only that some schedules assume grid dates utilities cannot meet. Backup generation on site can help, but it brings cost, permits and execution risk.
