Anthropic rents $45B of Nscale compute for 460MW in West Virginia

Anthropic has signed a deal to rent about $45 billion in AI compute from Nscale, a British AI infrastructure company founded only in 2024, a source familiar with the deal told TechCrunch. Bloomberg reported it first. Neither company has published an announcement.

The money buys power, not chips. CNBC confirmed the roughly $45 billion figure and reports that Anthropic rents around 460 megawatts at an Nscale data center development in West Virginia, running on Nvidia’s Vera Rubin chips, with the facility expected to come online at the end of 2027. Bloomberg, as relayed by TechCrunch, puts the term at six years, which works out to about $7.5 billion a year by our arithmetic.

That’s the whole verified core of the story, and it rests on unnamed people. CNBC’s two sources asked not to be named because the details are confidential. So the deal size, the megawatts and the timing are all reported, not filed or announced.

A $45 billion rental, from a company founded in 2024, for power that doesn’t switch on until the end of 2027.

Rundowns AI

That matters because the reason for the spending is on the record even when the contracts aren’t. Anthropic said earlier this year that growing demand for its Claude AI models and products had caused “inevitable strain” on its infrastructure, which impacted “reliability and performance” for users, especially during peak hours, CNBC reports. So the buying spree reads as a response to a bottleneck the company has already admitted to.

Month (2026)	Partner	Reported size
August	Nscale	About $45B, six years, West Virginia
August	Volta	$10B, six-year supply, Norway
July	AMD	$5B compute-related deal
May	SpaceX	About $1.25B of capacity a month
April	Amazon	An additional 5 gigawatts

What is actually in the rack

Vera Rubin is Nvidia’s newest rack-scale system, and the specifications are public even when the contract isn’t. Nscale’s own product page for DGX Vera Rubin NVL72 lists 72 Rubin GPUs and 36 Vera CPUs in one fully liquid-cooled rack, with 75TB of fast memory and 3,600 PFLOPS of NVFP4 inference performance. TechCrunch described Vera Rubin as six different chips working in concert, so the two counts are worth reading side by side.

The pitch behind it is efficiency per megawatt, and that matters because Anthropic has rented a fixed 460 of them. Nvidia says Vera Rubin NVL72 delivers up to 30x higher throughput per megawatt than GB300 NVL72 on agentic workloads, and up to 35x lower cost per million tokens. Those numbers come from Nvidia’s own measurements on the SemiAnalysis AgentX workload, and Nvidia notes the results are pending SemiAnalysis review. They’re vendor figures until someone else runs them, much as with Nvidia’s $105B lease guarantee for OpenAI in Ohio.

Nscale is the less familiar name here, but it’s growing fast enough to be its own story. SiliconANGLE reported on 21 August, citing Bloomberg, that the London-based company is seeking $3 billion in a US listing after a $14.6 billion valuation in March. Its West Virginia campus runs to 2,250 acres, and Microsoft is the anchor tenant, having commissioned 1.35 gigawatts there in March.

The scale of what Anthropic just added is easier to see against that. Bloomberg’s sources put Nscale’s total contracted revenue at about $51 billion as of 21 August. Five days later, one customer was reported to have committed $45 billion more, which is close to that entire prior book.

Anthropic’s side of the trade is a valuation that it has to defend. CNBC puts the company at $965 billion, with a prospectus confidentially filed with the SEC in June and early meetings with prospective investors under way. Compute commitments are the clearest evidence it can show Wall Street that the revenue it reports, including the $65 billion run rate it gave investors in July, has somewhere to run.

What’s worth watching isn’t the headline number, but whether either company confirms it. An unannounced deal for capacity that arrives in late 2027 can be resized quietly, and Nscale has an IPO to price before then.