Anthropic isn’t waiting for its IPO to sort out its compute problem. The Claude maker has struck a roughly $45 billion deal with British cloud startup Nscale to rent capacity from a data center in West Virginia, Bloomberg reports. Starting late next year, Anthropic will tap 460 megawatts of Nvidia’s next-generation Vera Rubin chips there, according to two people familiar with the arrangement. That’s a serious power draw — comparable to what a mid-sized city consumes — and it signals just how aggressively Anthropic is scaling inference and training capacity.
The six-year commitment is the latest in a frantic compute land-grab ahead of Anthropic’s planned fall IPO. The company already has deals with Amazon and SpaceX, plus a staggering commitment to lease Google’s AI chips for more than $150 billion. Bloomberg also reports Anthropic held early talks with Meta about a possible multibillion-dollar arrangement. The pattern is unmistakable: Anthropic is spending like a company that believes the AI race will be won by whoever can secure silicon and power first, not whoever has the cleverest model architecture.
Nscale, for its part, is no passive landlord. The London-based startup raised $2 billion in March at a $14.6 billion valuation, led by Norwegian energy giant Aker, and is preparing its own IPO. The West Virginia site it’s building in Mason County is planned for 1.35 gigawatts of capacity and carries a $69 billion price tag, according to Nscale. That’s the kind of number that makes you blink twice — and local residents have reportedly pushed back on the project.
Here’s what makes this deal interesting beyond the sticker price: Nvidia’s Vera Rubin generation represents a significant architectural shift, and Anthropic is committing to it before the chips have shipped at scale. If Rubin under-delivers or arrives late, Anthropic is on the hook for billions in capacity it can’t fully use. The company is essentially making a leveraged bet on Nvidia’s roadmap while simultaneously diversifying across three cloud partners. That’s not desperation — it’s the new economics of frontier AI, where compute contracts are signed years in advance and the price of falling behind is existential.
