The Most Informative Part Is Who Didn't Take It
The financing numbers grab attention, but the detail worth keeping is the order in which this capacity moved: Microsoft walked away during a summer review of its data centre portfolio, Google passed after looking at its own capital spending, and Anthropic then signed for six years and $45 billion. Same racks. The hyperscalers decided the economics didn't work; the model lab decided it couldn't afford not to. That says the two sides are operating under different constraints. Hyperscalers are running portfolio management and can be choosy about timing, location and price. A model lab is racing a capacity ceiling, where being a year late means its iteration cadence and its rate limits get decided for it. Read alongside the other compute deals of recent weeks, model labs are becoming the marginal buyer of long-term capacity.
$103 Billion Is Backlog, Not Revenue
The other thing to pull apart is the definition. $103 billion is **contracted backlog** — signed but undelivered. Turning it into revenue requires Nscale to finish the campuses, install the hardware and deliver capacity on schedule, and every one of those steps can slip. The comparison makes it concrete: Nscale booked about $33 million of revenue across all of 2025, and its most recent quarter only just passed $100 million. The roughly $18.1 billion annual revenue and $13.6 billion adjusted earnings it shows investors are labeled by the company itself as illustrative rather than formal guidance. Three orders of magnitude and several years of construction sit in between. For scale on the company: founded in May 2024 by Joshua Payne and Nathan Townsend; a $2 billion Series C in March 2026 at a $14.6 billion valuation, and six months later a conversion cap at $30 billion. It has contracted roughly 194,000 Nvidia Vera Rubin chips, agreed in July to acquire Anyscale for about $1.65 billion, and recently committed at least $3.5 billion of compute to robotics company Figure while taking an equity stake.
Where Nvidia's $2 Billion Sits
Nvidia is investing about $2 billion directly, and Nscale uses the money to buy Nvidia hardware — a structure that already ran with CoreWeave and Nebius: the chip vendor invests in a neocloud, the neocloud spends it on that vendor's chips, and the vendor gets a locked-in customer plus a financial return. That isn't automatically bad, but it makes the word "demand" something to read carefully: part of the demand is created by the supplier's own capital. For teams consuming APIs, the practical implication is specific: how much capacity Anthropic can offer around 2027 rests substantially on whether this West Virginia campus is delivered on time. What will affect you is not today's figures but the queueing and rate limits then.