Microsoft withdraws from Nscale UK data center deal over power delays

UK Power Networks told Nscale that adequate power for the 90MW Loughton connection may not arrive until the early-to-mid 2030s, versus the original roughly 2027 target ^[1]. That is the entire story mechanically: the site already had planning permission and a signed 90MW grid connection agreement in place, so the holdup sits strictly on the supply side, with the grid operator itself unable to commit a usable timeline ^[1].

This is not a Loughton-specific fluke. Ofgem data shows 315 data centers queued nationally for grid connections requesting a combined 73GW, more than the UK's roughly 45GW peak demand, with total cross-sector connection applications reaching 125GW, up from 41GW a year earlier ^[1][2]. Microsoft had agreed to take the entire 50MW-to-90MW scalable capacity and deploy roughly 23,040 Nvidia GB300 GPUs at the site ^[3]- a commitment that only made commercial sense if power actually arrived on a workable schedule, and a multi-year slip into the 2030s broke that assumption.

The Microsoft contract was worth about $43.8 billion, roughly 42% of Nscale's total $103.4 billion contracted backlog ^[4]. Losing it is not symbolic - it strikes at the centerpiece number underpinning Nscale's NYSE listing (ticker NSCL), which is seeking to raise up to $3 billion at a valuation as high as $35 billion ^[4][5].

That IPO filing already carried concentration risk even before the Loughton news: Microsoft's deal and Anthropic's separate $44.6 billion compute agreement together account for roughly 85% of Nscale's contracted revenue ^[5], echoing a warning from asset manager Sona, cited by the Financial Times, that AI infrastructure providers heavily depend on a limited number of customers ^[5]. The timing compounds an already uneven buildout record - Nscale's revenue grew from $10.4 million to $140.6 million over the six months to June 2026, but net losses expanded even faster, from $369 million to $1.02 billion ^[5]- and its flagship Glomfjord, Norway site is running roughly six months late, with only 6.4MW of a planned 30MW live ^[4].

Nscale insists the problem is narrowly confined to this one site. In a statement, the company said its relationship with Microsoft remains strong and long-term, and that this unfortunate London-wide grid issue has no impact on its other contracts, adding that it is exploring alternative power sources for the site in addition to other expansion sites in the UK ^[6].

But the episode has split financial-market commentary into two readings that the available reporting does not resolve. One camp treats it as confirmation that the UK grid, not AI appetite, is the binding constraint on hyperscaler buildouts - a straightforward supply-side failure that Nscale could not have avoided. The other reads it more skeptically: Microsoft, facing negative free cash flow and compute demand that may not yet match its own capex plans, could be using a genuine infrastructure delay as convenient cover to trim exposure to a single, highly leveraged neocloud. Neither interpretation is confirmed on the record, and claims about internal Microsoft capacity-approval tightening circulating in online discussion remain unverified community speculation rather than established fact. What the dispute does make clear is how fragile a neocloud's 'contracted backlog' can prove to be once an anchor tenant can walk away over a site-specific permitting and power problem.

The Loughton collapse lands against a backdrop Microsoft itself helped create: the company has pledged roughly £22 billion toward UK investment ^[7], yet still walked away from a specific site because the grid could not keep pace. That contrast is the real signal for the rest of the industry - capital commitments and chip supply are no longer the gating factor on AI data center expansion in Britain; grid connection timelines are.

For other hyperscalers and neoclouds evaluating UK sites, Loughton is now a concrete data point that even a project with planning permission and a signed grid connection agreement in hand can still collapse on the power question alone ^[1]. Expect future UK AI infrastructure deals to carry far more scrutiny of grid-connection timelines before anchor tenants sign, not after.