Microsoft Exits Nscale’s 23,000-GPU Essex Lease

AI-cloud developer Nscale had Microsoft as its planned anchor tenant for 23,000 Nvidia chips in Essex. The facility may not receive grid electricity until the 2030s. Microsoft cancelled the tenancy before the site could deliver that capacity.

The original GPU-rental proposition answered a real shortage: customers needed access to accelerators they could not readily obtain themselves. But a GPU rental assumes a working building around the chip. In 2024, Meta described two operational clusters, each with 24,576 Nvidia H100 GPUs. Those clusters demonstrated what a complete installation could deliver; a chip count alone could not show whether another operator had finished the facility, cooling, and power supply needed to match it. By 2025, Nscale was signing large deployments, including a Microsoft agreement for chips in Texas and Portugal.

AI-cloud capacity is becoming an infrastructure-finance business before it becomes a compute business. A lender needs an energizable, completed site with a credible delivery date and customer obligations that survive delays. Customers can commit years before a data-center operator can supply the electricity on which its service depends.

Nscale’s backlog combines working business with future delivery

Nscale’s filing showed $103 billion in total contract value, with Microsoft and Anthropic accounting for 85% of it. Only $2.6 billion of contract value was active as of late August 2026. Active contract value is not recognized revenue, and total contract value is not presently available compute; the two measures answer different questions. One records the scale of agreements Nscale has assembled. The other identifies how much had entered its active-contract category by that date.

Essex shows why the distinction cannot be settled by counting chips. Microsoft’s withdrawal concerned its planned anchor tenancy there; reporting said its other Nscale arrangements were unaffected. The cancellation therefore does not prove that Microsoft has stopped buying AI capacity, or that every Nscale contract faces the same terms. It does show that a named customer and a specified GPU count did not secure that site’s delivery schedule.

Nscale faced a separate execution test in Norway. Sources reported that its flagship facility was six months behind schedule and that the company owed penalties to Microsoft and ByteDance. A delay can thus affect more than the first day of billing: depending on the agreement, it can create a payment obligation while construction remains unfinished. Nscale’s agreements may differ on cancellation rights, take-or-pay provisions and who bears delay risk. Lenders cannot determine the full terms from the available reports, so they must check the contracts before counting on future billings.

Construction finance arrives before the revenue it depends on

Nscale did convert some commitments into sales. Its first-half 2026 revenue reached $140.6 million, up 1,252% from a year earlier, while it recorded a $1.02 billion net loss. Neither figure negates the other: customers were paying for service, and Nscale was spending heavily to build beyond what it had yet delivered.

The company secured $3.36 billion in pre-IPO convertible financing, comprising $2.36 billion available immediately and a further $1 billion from Nvidia expected in November, which had not arrived as of October 19. That capital can fund work before a new facility bills a customer. It cannot by itself determine when a grid connection, a completed building, and customer acceptance will coincide. Across the sector, AI-infrastructure companies borrowed more than $100 billion in 2025; smaller companies faced higher interest rates as debt investors weighed unproven businesses.

Reported investor resistance to Nscale’s roughly $30 billion IPO target, with some urging a $16 billion to $19 billion valuation after the Essex exit, was not a final market price. It was a visible dispute over how to value obligations at different stages of completion. Nscale appeared in 29 articles in the 2024–2026 coverage record; 51.7% focused on funding—a share of coverage, not capital raised. Other developers pursuing AI-cloud project finance also face debt costs before a site can bill its customer.

Customers now specify power, but megawatts describe different things

Large buyers have begun writing the physical scale of their requirements into capacity agreements. Anthropic reportedly agreed to pay Nscale $45 billion over six years for roughly 460 megawatts at a West Virginia facility using Nvidia Vera Rubin chips. Meta was reported to have contracted for roughly 1.6 gigawatts of Crusoe computing capacity across sites in Texas and Missouri. Meta’s agreement is substantial counter-evidence to the idea that one cancelled Essex lease signals a broad collapse in demand.

Google took another position in the chain, contracting for 3,590 megawatts of electricity from Constellation Energy in the largest US power grid, with new nuclear generation accounting for about a quarter of the planned supply. Google’s electricity purchase, Meta’s computing-capacity contract, and Anthropic’s reported site agreement are not interchangeable measures of operating capacity. Each secures a different input or service, and none alone certifies that a particular new cluster has been commissioned. The distinction becomes especially costly for a facility designed around maximum power draw: its construction must support the peak requirement even while the operator waits to use it.

Factory construction cannot grant a grid connection

Crusoe is trying to shorten one part of the schedule through factory-built data centers. It raised $3.9 billion at a roughly $30.9 billion post-money valuation while pursuing that strategy. Yet Crusoe also paused a Wyoming project after, according to sources, failing to secure customers including Google amid concerns about cost and timetable. Prefabrication can change how quickly an operator assembles a facility; it cannot make a prospective customer accept the rest of the project’s schedule.

Developers have also pursued on-site power plants to get around overloaded grids, trading one dependency for permitting, equipment, and fuel-supply requirements. Texas halted approvals of data-center grid connections for audits while more than 474 gigawatts of projects sat in its queue—over five times the state’s peak demand. Queue entries are requests, not operating loads or proof that every project will be built. They do show why a developer cannot infer an interconnection date from the speed at which it can put walls around racks.

A faster regulatory answer is not the same as delivered electricity

FERC approved a process intended to handle qualifying data-center power requests within 90 days, alongside new requirements for AI hyperscalers. That target could reduce time spent waiting for an answer. It does not establish that local transmission, generation, and construction can be completed in 90 days. PJM Interconnection, meanwhile, proposed requiring large data centers to bring their own generation or curtail demand to reduce outage risk. The grid operator is asking prospective loads to account for what happens when their requested power is not available.

A project financier can test that exposure without assuming every stalled site will fail: identify when power is expected, what work remains, when customer payments begin, who pays for idle equipment, and what happens if either side misses its date. Nscale’s revenue and Meta’s large Crusoe commitment show that the demand and the means to serve some of it are real. Essex shows the limit of writing the customer’s name onto a facility before the electricity has a dependable arrival date. The cancelled lease counted 23,000 chips; the site still had to wait for the power to run them.