Anthropic and OpenAI shift toward smaller, 20-30 MW data center deals as inference demand grows

Anthropic and OpenAI are both pursuing a new tier of much smaller AI data center deals, seeking capacity deployments in the 20-30 megawatt range even as they continue signing multi-hundred-megawatt and gigawatt-scale contracts, according to CNBC, which cited people familiar with the private negotiations.

What's new

CNBC reported: "The two AI labs have both inked huge deals for AI data centers in the past year for facilities of multi-hundred-megawatt and gigawatt capacity, but sources have said those companies are now also looking for compute capacity deals for much smaller deployments of 20-30 MW." Anthropic has been sounding out agreements in that range across the U.K. and the Nordics, according to four people familiar with the talks who spoke to CNBC anonymously; OpenAI had separately explored similar smaller Nordic deployments, per two of the sources. One source told CNBC they were also aware of Anthropic and OpenAI conversations about U.S. capacity deployments at that same, smaller scale.

Asked about the strategy, an OpenAI spokesperson told CNBC: "We're building a diversified compute portfolio to meet growing demand for AI around the world." Anthropic did not comment when approached by CNBC.

The smaller-deal push comes on top of the labs' existing megaprojects. Anthropic signed a roughly $45 billion cloud deal with Nscale in August covering about 460 MW of capacity in West Virginia, while OpenAI said in April it had surpassed its original 10 GW commitment to the Stargate infrastructure project and has since added commitments for 3 GW in Georgia and 8 GW in Ohio.

Context

Jabez Tan, head of research at Structure Research, told CNBC that smaller capacity deals appeal for "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. For workloads that can operate across separate sites, a collection of smaller deployments can add up to substantial capacity." Tan also pointed to a technical driver behind the shift: training a frontier model requires large numbers of chips working tightly together in one place, but day-to-day inference workloads can run across many smaller, geographically distributed clusters, since each request can be served independently. CNBC cited a JLL report finding inference made up 9% of global data center workloads in 2025 versus 14% for training, with inference projected to reach 37% of capacity by 2030 against just 13% for training. Other infrastructure players are making a similar bet: Crusoe, which built one of OpenAI's largest data center complexes in Texas, is now investing in smaller facilities, the Wall Street Journal reported, betting they can be built faster and cheaper than the large developments currently facing delays across the U.S.

Why it matters

The pivot signals that the AI infrastructure race is entering a second phase: after a year of chasing gigawatt-scale megaprojects that take years to site, power, and build, the labs serving the most inference traffic are also filling in with faster, smaller, more distributed capacity they can bring online sooner. It reflects a maturing understanding that training and inference have different infrastructure needs, and it may ease some of the community and grid-capacity pushback that has dogged giant single-site data center projects in the U.S. and Europe, since smaller deployments can plug into existing powered sites rather than requiring new large-scale power buildout in one location.