Anthropic sounds out 20–30 MW deals in U.K. and Nordics

Anthropic and OpenAI are looking for compute in blocks of 20-30 MW, a fraction of the multi-hundred-megawatt and gigawatt deals both labs have signed over the past year, people familiar with the talks told CNBC in a report published on September 18. Smaller blocks can come online faster, and they suit inference, which is taking up a growing share of data center capacity.

Four of the people said Anthropic has sounded out agreements in that range across the U.K. and the Nordics. Two said OpenAI had been exploring similar deployments in the Nordics, and one said they also knew of talks involving both labs about U.S. capacity at that scale. The sources asked not to be named because the dealings are private.

The Nordics are not new ground for Anthropic. In August it committed $10 billion over six years to Volta, a capacity deal covering a data center in Norway.

Set against the labs’ recent big leases, the new talks are small. Anthropic’s six-year, roughly $45 billion contract with Nscale covers around 460 MW at the Monarch Compute Campus in West Virginia. A single block that size equals somewhere between 15 and 23 of the deployments now under discussion. The 460 MW is only the first of three planned buildings at Monarch; Nscale designed Phase 1 for 1.35 GW and targets delivery by early 2028, and Microsoft has also signed for 1.35 GW at the site. OpenAI, for its part, has said it passed its original 10 GW commitment to the Stargate project in April and has since committed to developing a further 3 GW in Georgia and 8 GW in Ohio.

Big builds are getting harder to land. Large data center projects in the U.S. and elsewhere face growing pushback from local communities, and in much of Europe land and power are in short supply.

Jabez Tan, head of research at Structure Research, put the appeal of small deals down to “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.”

The qualifier about workloads is where training and inference split. “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.” A training run needs one tightly connected cluster, while served requests are independent of one another and can be spread across sites.

A report by real estate company JLL projects that inference will overtake training as a share of total data center capacity in 2027. In 2025, inference accounted for 9% of global data center workloads and training for 14%. By 2030, JLL expects inference at 37% and training at 13%. Over those five years training’s share edges down while inference’s roughly quadruples.

Others are making the same bet. In February, Nvidia announced it would work with several data center stakeholders to study smaller facilities designed for distributed inference. And Crusoe, which built the large Texas complex OpenAI uses, is now investing in smaller data centers, The Wall Street Journal reported on September 17; the Journal said those sites would be faster and cheaper to build than large projects, which are running into delays across the U.S. The same day, Crusoe announced a $3.9 billion funding round at a $30.9 billion post-money valuation.

An OpenAI spokesperson told CNBC the company is “building a diversified compute portfolio to meet growing demand for AI around the world,” and that “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.” On the talks themselves: “We don’t comment on specific commercial discussions.”