Anthropic signed a six-year agreement with Nscale for 460 megawatts of compute capacity at a West Virginia data center, valued at $45 billion over the contract term. The deal represents roughly $7.5 billion annually in committed infrastructure spend, the largest known fixed-capacity agreement in the AI training sector by raw power allocation. Nscale, a private operator building hyperscale facilities in Appalachian power corridors, will dedicate nearly half a gigawatt to Anthropic's workloads starting mid-2025.

The 460-megawatt allocation would support approximately 230,000 to 280,000 H100-equivalent GPUs under current rack densities, enough to train multiple frontier models simultaneously or run continuous pre-training cycles without capacity rationing. Anthropic's current Claude 3.5 Opus training run is estimated to have consumed 12,000 to 18,000 H100s over four months; this deal provides 15x to 23x that scale in sustained parallel capacity. The West Virginia site sits on a 765-kilovolt transmission backbone with direct AEP grid interconnection, bypassing the 18-to-36-month queues now common in Northern Virginia and Oregon.

The commitment structure matters more than the headline figure. Anthropic is paying for reserved capacity whether or not the racks are lit, which means the company is forward-purchasing optionality on model scale rather than reacting to quarterly compute auctions. This is the infrastructure equivalent of a pharmaceutical company booking API manufacturing lines three years before Phase III trials—expensive if the product fails, catastrophic to lack if it succeeds. The deal also implies Anthropic expects to remain independent through 2031, or at minimum that its board believes compute lock-in strengthens acquisition valuation by eliminating a major buyer integration risk.

Nscale's West Virginia location is not incidental. The state offers $0.063 per kilowatt-hour industrial power rates, 40% below hyperscaler averages in Virginia, and sits within 120 miles of 68% of US internet users by fiber distance. The facility is being purpose-built with immersion cooling and rear-door heat exchangers optimized for the 700-watt TDP envelopes expected in Nvidia's Rubin architecture, shipping late 2025. Anthropic is effectively betting that the physics of model scaling—measured in FLOPs per dollar and FLOPs per watt—will continue favoring concentrated, long-duration training runs over distributed inference swarms.

Allocators should track three follow-on events in the next 90 days: Nscale's construction financing close, expected mid-Q2 2025, which will clarify whether Anthropic pre-funded any capex or simply locked tariff rates; Anthropic's next model release, likely June 2025, for evidence of multi-site training orchestration; and moves by Cohere or Mistral to secure similar gigawatt-scale deals, signaling whether this is Anthropic-specific strategy or sector-wide arms race. The $45 billion figure also suggests Anthropic's last private valuation, $18.4 billion in December 2024, will reprice upward when the Series D term sheet circulates—no board commits 2.4x enterprise value to infrastructure without revenue assumptions that justify it.

The contract runs through 2031. By then, either Anthropic operates the world's highest-margin AI product on captive infrastructure, or this becomes the most expensive unused data center lease in technology history.