Anthropic nears second $45 billion Nscale deal for another 460MW West Virginia data center

Anthropic is in talks with UK neocloud Nscale over a second compute deal worth about $45 billion, matching the one the two signed in August. If it closes, Anthropic will have committed roughly $90 billion to a single infrastructure provider that has not yet gone public, at a time when investors are questioning both AI infrastructure debt and how the models running on it are tested.

• Bloomberg reported that the two companies are close to announcing the deal; a person familiar with the talks told CNBC it would cover a second 460-megawatt data center at the same West Virginia site as the first agreement.
• Nscale filed to list on the New York Stock Exchange in September, but investors are pushing back on a proposed $30 billion valuation, and Microsoft has pulled out of anchoring a planned Nscale site near London.
• The deal lands as Hugging Face’s policy chief warns that AI evaluations, including the embedded evaluators labs now promote, are still an immature science.

• ~$45B Reported size of the second deal, the same as the August agreement
• 460MW Capacity of the second West Virginia data center, per CNBC’s source
• $30B Nscale’s proposed IPO valuation, which investors are resisting, per Bloomberg
• 28% / 27% One-month share declines at rivals CoreWeave and Nebius, per CNBC

Bloomberg first reported on Monday (October 19) that Anthropic and Nscale were “on the verge of announcing yet another $45 billion compute deal,” TechCrunch noted, adding that details could still change. CNBC later reported the deal’s scope from a person familiar with the matter. Anthropic declined to comment to CNBC. Nscale had not commented at the time of writing.

A second West Virginia build on top of the August deal

The first $45 billion agreement, announced in August, covered compute capacity at a West Virginia site. According to CNBC’s source, the new deal would add a second 460-megawatt data center at the same location, with Anthropic leasing the capacity.

Anthropic has been signing compute agreements at a rapid pace. Earlier this year it said growing demand for its Claude models had put “inevitable strain” on its infrastructure, and since then it has announced major deals with AMD, SpaceX, Google and Broadcom, according to CNBC. The company is also preparing to go public, and earlier this month it committed $100 million to train engineers on Claude deployment.

Nscale’s financing picture is getting harder

Nvidia-backed Nscale builds or leases data centers and sells computing capacity to hyperscalers and AI labs. Its timing is awkward. Bloomberg reported on October 18 that investors are pushing back on the $30 billion valuation proposed for its IPO. Separately, the Telegraph reported that Microsoft has withdrawn from anchoring an Nscale data center outside London after delays connecting to the UK power grid.

An Nscale spokesperson told CNBC the company remains “committed” to the London project and is “exploring alternative power sources for the site,” adding that the grid issue “has no impact on our other contracts.” Peers are under pressure too: CoreWeave and Nebius shares have fallen 28% and 27% over the past month amid concerns about heavy borrowing.

Who checks the models all this compute will run

The spending comes as AI labs face sharper questions about testing. Speaking at TechCrunch Disrupt, Hugging Face chief policy officer Irene Solaiman said AI benchmarks are presented as neutral but rest on methods that are still “an evolving science,” TechCrunch reported. She pointed to problems with reproducibility, which depends on access to the model, with comparing different evaluations, and with tests that go stale quickly.

Labs have promoted embedded evaluators, third-party testers placed inside AI companies, as an alternative to direct government regulation. Anthropic chose consulting firm Accenture as its first embedded evaluator, a choice that drew criticism over the firm’s qualifications. Solaiman also noted that evaluation is shifting from the model alone to the whole system, including the agent harness around it, and that such agentic evaluations are “really hard and really expensive.” The debate follows the White House pact under which six AI developers agreed to independent safety audits in September.

The BlockWest read. We read a second $45 billion commitment as Anthropic locking in capacity years ahead of demand it expects, with Nscale’s IPO as the pressure point: a signed second contract is exactly what Nscale needs to defend its valuation. For investors, the concentration risk runs both ways. The scale also raises the bar on oversight. As compute commitments reach tens of billions per supplier, the question of who tests the systems, and how, will matter more to regulators and public-market investors than another benchmark score.

Next to watch: a formal announcement and contract terms, Nscale’s IPO pricing and whether the $30 billion target holds, a replacement anchor for the London site, compute commitments disclosed in Anthropic’s own IPO filing, and whether Anthropic adds more embedded evaluators beyond Accenture.

BlockWest is a news publication. Nothing here is investment advice. Read our disclaimer and editorial policy.