Anthropic and OpenAI pursue smaller data center deals

Anthropic and OpenAI are pursuing smaller data center capacity deals even as they commit to gigawatt-scale infrastructure projects, according to sources familiar with the companies' strategies.

Both AI labs are exploring agreements for compute deployments in the 20-30 megawatt range, a significant departure from the multi-hundred-megawatt and gigawatt facilities they've announced over the past year. Four sources told CNBC that Anthropic has been discussing deals of this size across the United Kingdom and Nordic countries, while two sources said OpenAI had explored similar opportunities in the Nordics. One source indicated both companies are also pursuing U.S. capacity at this scale.

The shift reflects practical constraints and evolving technical requirements. While Anthropic recently secured a roughly $45 billion cloud deal with Nscale for around 460 MW of capacity in West Virginia, and OpenAI has committed to more than 21 gigawatts through its Stargate project and other initiatives, smaller deals offer what Jabez Tan of Structure Research calls "speed to usable capacity."

"Securing a few megawatts at an existing powered site can be more practical than waiting for a much larger block in one location," Tan explained. Large data center projects increasingly face local opposition, and in Europe, available land and power remain scarce.

Why it matters

The move toward distributed, smaller-scale capacity signals a fundamental shift in how AI companies are thinking about infrastructure. As more compute moves from training models to serving them in production—a process called inference—the technical requirements change. Training demands tightly coupled clusters of chips processing massive datasets, but inference workloads can operate across separate, geographically distributed sites. This architectural flexibility makes smaller deployments viable and often faster to bring online than massive single-site builds.

The inference inflection point

The economics of AI infrastructure are shifting as inference overtakes training. According to real estate firm JLL, inference accounted for just 9% of global data center workloads in 2025 compared to 14% for training. By 2027, inference is expected to surpass training, and by 2030, it could consume 37% of total capacity while training drops to 13%.

"Training a large model typically requires many chips working closely together," Tan noted. "Many inference workloads can instead serve separate requests across multiple smaller clusters, opening up more locations."

This trend extends beyond Anthropic and OpenAI. Crusoe, which built a large Texas facility used by OpenAI, is now investing in smaller data centers that can be deployed faster and more cheaply than delayed large-scale projects, the Wall Street Journal reported Thursday. Crusoe announced a $3.9 billion funding round at a $30.9 billion valuation the same day.

Nvidia announced in February it would collaborate with data center operators to study smaller-scale facilities designed specifically for distributed inference workloads.

An OpenAI spokesperson confirmed the company is "building a diversified compute portfolio to meet growing demand for AI around the world," noting that "different workloads need different infrastructure." Anthropic declined to comment.

These details were first reported by CNBC.