OpenAI and Anthropic Explore Smaller Data-Center Deals, CNBC Reports

OpenAI and Anthropic are reportedly looking for data-center capacity in much smaller pieces: deployments of roughly 20 to 30 megawatts. The searches would sit alongside their enormous infrastructure commitments, offering a potential way to bring computing online faster as more AI work shifts from training models to answering users’ requests.

CNBC, citing people familiar with private discussions, reported that Anthropic has sounded out agreements in that range in the U.K. and the Nordic countries. OpenAI has explored similar opportunities in the Nordics, the people said; one source also described talks involving both companies about U.S. capacity at that scale. The report does not establish that any of the discussions have become deals.

A second route to usable compute

The distinction is technical but important. Training a large model typically requires many chips working closely together. Inference—the work of running a trained model to generate responses—can often handle separate requests across smaller clusters. That makes it possible, in principle, to spread some serving work across more sites.

Jabez Tan, head of research at Structure Research, told CNBC the appeal is “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 place, he said. That advantage depends on the workload being able to operate across separate locations.

Small capacity does not replace the giant commitments

Neither company appears to be abandoning large projects. OpenAI said in April that it had surpassed Stargate’s original 10-gigawatt U.S. infrastructure commitment and added more than 3 gigawatts in the preceding 90 days. It says site choices depend on power, land, permits, transmission, workforce, community support and partner readiness—conditions that can slow the path from a plan to operating computers.

Anthropic likewise announced in April an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity expected to start coming online in 2027, with the vast majority located in the United States. Separately, CNBC reported in August that Anthropic had agreed to a roughly $45 billion cloud deal with Nscale for about 460 megawatts at a West Virginia development.

Flexibility has limits

A portfolio of smaller deployments could add up to substantial capacity, but it is not a universal substitute for a tightly connected training cluster. The reported talks instead point to a split infrastructure strategy: large facilities for concentrated computing and smaller allocations for work that can be distributed. OpenAI said it evaluates opportunities on requirements, performance, reliability, timing and cost; it declined to discuss specific commercial conversations. Anthropic did not comment to CNBC.