Anthropic, OpenAI Seek Smaller Data Center Deals in AI Capacity Race

The companies are seeking compute capacity deals for deployments of roughly 20 to 30 megawatts, the sources said. Both have signed huge agreements in the past year for multi-hundred-megawatt and gigawatt facilities.

Anthropic has sounded out agreements in that smaller range across the U.K. and the Nordics, according to four people familiar with the conversations who asked to remain anonymous when discussing private business dealings. OpenAI has been exploring opportunities for those smaller capacity deployments in the Nordics, two of the sources said. One source also said they were familiar with talks involving Anthropic and OpenAI about U.S. capacity deployments at that scale.

An OpenAI spokesperson told CNBC that the company is "building a diversified compute portfolio to meet growing demand for AI around the world." The spokesperson added: "Different workloads need different infrastructure, so we have conversations with a range of partners and assess opportunities based on our requirements, performance, reliability, timing and cost. We don't comment on specific commercial discussions." Anthropic did not comment when approached by CNBC.

Both AI labs typically rent compute capacity from data center operators and neoclouds, and they have sought large-scale, long-term agreements. Anthropic signed a roughly $45 billion cloud deal with Nscale to rent around 460 MW of compute capacity at a data center development in West Virginia, two people familiar with the matter told CNBC in August. OpenAI has said it surpassed its original commitment of 10 GW to its Stargate AI infrastructure project in April and has since committed to developing a further 3 GW in Georgia and 8 GW in Ohio.

Large data center projects in the U.S. and abroad are increasingly facing pushback from local communities. The sector is also under pressure in much of Europe, where available land and power are in short supply. Smaller capacity deals are often attractive because of "speed to usable capacity," Jabez Tan, head of research at Structure Research, told CNBC.

"Securing a few megawatts at an existing powered site can be more practical than waiting for a much larger block in one location," Tan said. "For workloads that can operate across separate sites, a collection of smaller deployments can add up to substantial capacity."

Training AI models requires large amounts of computing power to process huge quantities of data, but deploying those systems day-to-day, a process known as inference, can be done with smaller clusters of chips. "Training a large model typically requires many chips working closely together," Tan said. "Many inference workloads can instead serve separate requests across multiple smaller clusters, opening up more locations."

The shift matters as more AI compute moves from training models to serving them in production. The amount of capacity used to serve inference is expected to rise. The proportion of total data center capacity used for inference workloads is expected to overtake training workloads in 2027, according to a report by real estate company JLL. In 2025, inference made up 9% of global workloads in data centers compared with 14% for training, the report said. By 2030, inference is projected to use 37% of that capacity, compared with just 13% for training.

In February, it was announced that Nvidia would collaborate with several data center stakeholders to study smaller-scale data centers designed for distributed inference. U.S. company Crusoe, which built a huge data center complex in Texas used by OpenAI, is now investing in smaller data centers, The Wall Street Journal reported on Thursday. Those facilities will be faster and cheaper than larger builds, which are facing delays across the U.S., the Journal said. Crusoe did not respond to a request for comment. Crusoe announced on Thursday that it had raised a $3.9 billion funding round at a $30.9 billion post-money valuation.