CoreWeave disclosed a $1.4 billion capacity contract on Thursday covering five years of GPU capacity at a campus outside Reno, Nevada, that will not deliver its first megawatt until the fourth quarter of 2027. The customer is described in the filing only as a leading artificial intelligence research organization. The campus is planned at 480 megawatts across three phases, powered by a mix of grid supply and an on-site gas generation facility with a stated intent to transition to geothermal, which northern Nevada has in unusual abundance. The contract funds the construction, which is the point and which is how essentially all of this capacity now gets built.
How the financing works
CoreWeave does not build data centers on speculation. It signs a long-term capacity contract with a creditworthy customer, then borrows against that contract to fund construction and hardware. The debt is secured by the contract receivables and, in CoreWeave's case historically, by the GPUs themselves. Lenders including Blackstone, Magnetar, and a syndicate of banks have provided this financing across several rounds, and the structure has let the company grow far faster than its equity base would otherwise permit.
The risk sits with whether the customer pays for five years. If demand for AI compute stays strong, the contract is money-good and the model works. If a customer's funding falters or its compute needs change, CoreWeave holds a partially built campus and a depreciating hardware asset. The company's disclosed customer concentration, with a majority of revenue from a small number of counterparties, is the single number analysts watch and the single number the company most wants to reduce.
Why Nevada
Power availability drives site selection above everything else, and northern Nevada has three advantages. NV Energy has capacity and a regulatory framework that permits large industrial interconnection without the multi-year queue Virginia and Texas now impose. The state has geothermal resources supplying about 10 percent of its electricity already, with more capacity identified. And the climate permits substantial free cooling for much of the year, which cuts water and power consumption.
Water remains the political question. The Truckee Meadows region has been careful about industrial water allocation, and CoreWeave's plan uses closed-loop liquid cooling with an air-cooled heat rejection stage, which the company says will consume roughly 90 percent less water than an evaporative design. The county commission approved the water permit in June after a hearing where residents raised exactly this. Nevada's experience with the Tesla gigafactory shaped how local government approaches these deals, generally in the direction of extracting firmer commitments.
The story is rarely the launch. It is what breaks, what ships, and who owns the mess at 2 a.m.
The gas question
Phase one will run substantially on on-site natural gas turbines, because grid interconnection at the required scale takes longer than the customer will wait. That is now the standard pattern across the industry and it is the largest reason data center construction has become an emissions story. xAI's Memphis facility drew sustained criticism for exactly this, and several projects in Texas and Louisiana have similar arrangements.
CoreWeave's stated plan is to transition to geothermal and grid supply as interconnection completes, with the turbines retained for backup. Whether that transition happens on schedule is the thing to watch, since the incentive to keep running paid-for generation is considerable. The company has signed a power purchase agreement with a geothermal developer for 180 megawatts starting in 2029, which is a real commitment and covers less than half the campus.
The competitive position
CoreWeave competes against the hyperscalers, who have vastly more capital, and against a set of specialized providers including Lambda, Crusoe, and Nebius. Its advantage has been speed: it deploys new GPU generations faster than Amazon or Google, because it is not integrating them into a broad service portfolio and because Nvidia has historically allocated to it favorably.
That advantage compresses as the hyperscalers catch up on deployment cadence and as Nvidia's allocation normalizes. CoreWeave's answer has been to move up the stack, adding managed Kubernetes, storage, and inference services that look increasingly like a cloud provider's portfolio. That is the right direction and it puts the company in competition with much larger firms on their own terms.
What this contract signals about demand
A five-year commitment for capacity delivering in 2027 tells you the customer believes its compute needs will grow substantially and that it cannot get the capacity elsewhere on acceptable terms. Both of those are informative. The AI labs have been signing long-dated capacity agreements across multiple providers for two years, and the aggregate committed capacity now exceeds what the industry can deliver on the announced schedules.
The counter-scenario, which several analysts have been arguing, is that inference efficiency improvements and smaller specialized models reduce compute demand growth below current projections, leaving a great deal of contracted capacity underused. Nobody knows. What is observable is that the people spending the money are behaving as though demand keeps growing, and that they are contractually committed for years regardless of what happens. The people lending against those commitments are making the same bet with less upside, which is why the terms on this financing have tightened over the past year and why several smaller providers have found the debt markets closed to them entirely. Capacity contracts are only bankable if the counterparty is, and the list of AI labs that qualify is short.
Skarvonix will keep following this beat with reporting grounded in how systems behave outside the launch keynote.
- Kubernetes
