US Anthropic Plans to Lease Nscale Computing Power for $45 Billion

en.Wedoany.com Reported - On August 26, a person familiar with the matter said that US artificial intelligence company Anthropic plans to reach a six-year computing power leasing arrangement with Nscale, totaling approximately $45 billion, corresponding to about 460 megawatts of AI computing power. The relevant computing power will be deployed at the Monarch Compute Campus in West Virginia, US. Anthropic declined to comment on the arrangement, and Nscale has not yet issued a corresponding contract announcement.

According to the plan provided by the source, Nscale will deploy NVIDIA's next-generation Vera Rubin chips to provide cloud computing resources for Anthropic's model training and inference operations. NVIDIA has confirmed that Nscale and Caterpillar are building a multi-gigawatt AI computing facility in West Virginia using the Vera Rubin DSX reference architecture.

The Monarch campus is located in Mason County, West Virginia, covering more than 2,250 acres. Nscale has set the first-phase AI computing capacity at 1.35 gigawatts, scheduled for delivery in early 2028, with more than 8 gigawatts of power capacity reserved for subsequent expansion. The 460 megawatts that Anthropic plans to lease accounts for approximately one-third of the campus's first-phase planned capacity.

The campus adopts an on-site independent power supply model, with Caterpillar participating in the construction of supporting power generation facilities, and does not draw electricity from the local public grid. The data center cooling system uses a closed-loop scheme and does not use local municipal drinking water during the operational phase. Computing power, power supply, cooling, and network facilities will be uniformly configured according to AI training and inference workloads.

Anthropic previously estimated that the data center scale required to train a single advanced AI model could increase from 2 gigawatts in 2027 to 5 gigawatts in 2028; during the same period, US frontier model training loads could reach 20 to 25 gigawatts, and with inference demand added, the AI industry's electricity demand could reach at least 50 gigawatts. The company has expanded GPU server and data center capacity through multiple chip and computing infrastructure suppliers.