Summary
- Reuters reports a six-year, $45bn agreement covering Nscale capacity representing about 460MW of power.
- Nscale is expected to deploy Nvidia Vera Rubin systems for Anthropic at its West Virginia campus.
- The UK company's wider Monarch site is designed for 1.35GW in phase one and more than 8GW over later phases.
Anthropic will reportedly spend $45bn over six years renting AI computing capacity representing around 460MW of power from UK-headquartered Nscale at its Monarch campus in West Virginia.
Reuters reported the agreement on Wednesday, citing a person familiar with the matter. Nscale is expected to deploy Nvidia’s Vera Rubin systems to support Anthropic’s workloads. Anthropic declined to comment on the reported deal.
The 460MW power commitment is substantial even by the standards of current AI infrastructure procurement. Unlike a chip order alone, it represents a long-duration requirement for an operating environment that must provide generation, cooling, networking and physical capacity around the compute hardware.
The infrastructure is in the United States, making the project a global comparator rather than a European data centre build. Its relevance to the UK and European market lies in Nscale’s role as a UK-headquartered infrastructure provider developing campuses on both sides of the Atlantic and in the contract structure now emerging around AI capacity.
Nscale describes Monarch, in Mason County, as a purpose-built AI campus with 1.35GW of computing capacity planned for its first phase and a development path exceeding 8GW. Delivery of phase one is targeted for early 2028.
The company says more than $20bn of capital investment is planned at Monarch. The site covers more than 2,250 acres and is being developed with Nvidia and Caterpillar, illustrating the physical scale required when individual AI customers begin contracting power and compute capacity in the hundreds of megawatts.
Power is central to the design. Nscale says Monarch will generate all of its electricity on site and operate independently of the public grid. Caterpillar is involved in the generation system, while West Virginia has certified the development as an AI microgrid.
That configuration is very different from the dominant European development model, where large sites generally depend on utility connections and face increasingly long queues for transmission and distribution capacity. It demonstrates one route around grid scarcity, but shifts a larger share of generation infrastructure, fuel supply, emissions exposure and operational responsibility inside the data centre project.
Nscale also says the site’s cooling systems will operate on a closed loop without drawing municipal drinking water. At gigawatt scale, cooling-water strategy and heat rejection become infrastructure questions in their own right, particularly where communities and regulators are scrutinising the resource demands of AI campuses.
The reported Anthropic contract would represent roughly a third of Monarch’s 1.35GW phase-one power envelope if the figures are directly comparable. It also demonstrates why developers are pursuing multi-gigawatt sites: a relatively small number of large AI customers can now account for hundreds of megawatts each.
Long-term compute agreements can provide the commercial foundation for that development. A six-year contract can give an infrastructure provider visibility over demand while providing the customer with reserved access to scarce hardware and power. The trade-off is a very large financial commitment to a technology stack whose performance and economics will continue to change during the contract term.
The use of Vera Rubin systems also means the physical facility has to be prepared for the electrical and thermal characteristics of a new GPU generation rather than simply providing generic rack space. Higher rack density pushes power-distribution and cooling choices further upstream in the development process.
For European developers, Monarch offers a useful comparison with the constraints shaping projects closer to home. The demand is recognisable, but the power solution is markedly different: instead of waiting for enough public-grid capacity, Nscale is building generation alongside compute.
The agreement remains reported rather than confirmed by Anthropic. If completed on the terms described, capacity representing around 460MW of power under a single six-year customer commitment would give Nscale a large anchor workload for the first phase of one of the world’s biggest planned AI campuses.
