Anthropic Plans to Lease $45 Billion in Nscale Computing Power
Anthropic has reportedly reached a six-year agreement with AI cloud infrastructure provider Nscale, planning to pay $45 billion to lease AI computing power at the Monarch data center campus in West Virginia, corresponding to approximately 460 megawatts of power capacity. The agreement has not yet been publicly announced by either party, with details coming from informed sources.
Nscale will deploy Nvidia's next-generation Vera Rubin chips for Anthropic, with the first capacity expected to come online by the end of 2027. The deal aims to secure long-term computing power for products like Claude and Claude Code, anticipating demand growth rather than just for one-time model training.
460 megawatts is roughly equivalent to the simultaneous electricity usage of 345,000 American households; based on a rough calculation of $45 billion over six years, Anthropic's average annual commitment is about $7.5 billion. This amount may include computing power, chips, data center facilities, electricity, networking, operations, and long-term capacity reservations, rather than just a simple "electricity bill" paid to Nscale.
The complete Monarch campus that Nscale plans to build will have a total capacity of about 1.35 gigawatts, including three data center buildings and on-site power generation facilities, with an overall development cost estimated at around $71 billion, of which about $47 billion will be used to procure AI chips. The agreement with Anthropic covers the first building, while the remaining facilities are expected to gradually provide capacity starting in 2028.
This project was originally planned to be leased by Microsoft. Microsoft signed a letter of intent for the campus in March this year but withdrew in the summer, after which Anthropic became the main potential tenant; this indicates that Nscale is shifting from a single large cloud customer to a cutting-edge model company as a long-term demand anchor.
For Anthropic, securing 460 megawatts long-term helps reduce the risks of model training delays due to insufficient computing power, API limits, and obstacles to enterprise product expansion, but it also brings significant fixed costs and utilization risks: if model demand, prices, chip performance, or product revenues fall below expectations, the computing power commitment could become a cash flow pressure.
From a market mechanism perspective, the deal transfers funds from model companies to the GPU supply chain, data center construction, substations and power generation facilities, networking equipment, cooling systems, and cloud operations services. Nvidia and data center operators with access to electricity, land, financing capabilities, and long-term customer contracts will benefit; small cloud service providers lacking power and construction permits will face pressure, and competition in cutting-edge models will increasingly rely on long-term capital and infrastructure locking capabilities.
Source: Public Information
ABAB AI Insight
This deal indicates that the competition in cutting-edge AI has shifted from "who can train the stronger model" to "who can secure inference and training capacity for the next two to six years first." Anthropic's choice to lease long-term rather than build its own campus transfers the risks of land, power generation, construction, networking, and server lifecycle to Nscale, while securing guaranteed capacity in exchange for a multi-year minimum commitment; this is similar to reserved instances in the cloud computing era, but the amounts and power scales far exceed traditional enterprise IT procurement.
The capital path is divided into three segments. Anthropic commits $45 billion to obtain computing power for model training and online inference; Nscale finances the construction of the approximately $71 billion campus based on such long-term contracts; of which about $47 billion further flows into the GPU and related semiconductor supply chain. Long-term leases here are not only cost agreements but also serve as "revenue collateral" providing visibility for debt financing of data center projects.
A historical comparison is that Microsoft, Google, Meta, and Amazon have locked in data centers, power, and chip orders in advance during the cloud and AI cycles. The difference is that Anthropic is not a large cloud vendor with its own global cloud infrastructure but, as an independent model company, is taking on long-term capacity commitments comparable to those of large cloud platforms. Microsoft's initial signing of a letter of intent followed by withdrawal, and Anthropic taking over the project, also indicates that while demand for computing power is strong, project returns, chip roadmaps, financing costs, and power availability still need to be dynamically reassessed for any single company.
This represents a concentration of capital. The 460 megawatts of computing capacity, Vera Rubin chips, and on-site power generation require hundreds of billions of dollars in capital, long-term power access, supply chain priority, and strong credit tenants, with thresholds far exceeding those of ordinary software startups. The mechanism is that the larger the model capability and online service scale, the more unit costs are determined by chip procurement, data center utilization, and power costs; laboratories that can sign large long-term contracts in advance gain supply certainty and exclude competitors from limited high-end computing resources.
ABAB News · Cognitive Law
The barriers of cutting-edge models are shifting from algorithms to power contracts.
Computing power leases are both a cost and a weapon to lock out competitors.
The more AI resembles infrastructure, the more capital scale determines capability limits.