Anthropic and OpenAI start shopping in 20–30 MW slices

Twenty to thirty megawatts is a rounding error next to a gigawatt, and that is the size Anthropic and OpenAI are now shopping for, sources told CNBC. Anthropic has sounded out deals in that range across the U.K. and the Nordics, four people familiar with the conversations said, and OpenAI has been exploring the same size in the Nordics, according to two of them.

At a glance

• One source also knew of talks involving Anthropic and OpenAI about U.S. deployments at the same scale, a fraction of the multi-hundred-megawatt and gigawatt facilities both labs contracted over the past year.
• Structure Research's Jabez Tan says a few megawatts at an existing powered site can be more practical than waiting for one big block, and scattered small deployments can add up.
• Huge data center projects in the U.S. and elsewhere face pushback from local communities, and across much of Europe land and power for new builds are in short supply.

If you have not been tracking the announcements, both labs spent the past year buying capacity at utility scale. Anthropic's roughly $45 billion cloud deal with Nscale covers around 460 MW at a development in West Virginia, two people familiar with the matter told CNBC in August. OpenAI said in April that it had passed the original 10 GW commitment behind Stargate, and has since committed to a further 3 GW in Georgia and 8 GW in Ohio.

Speed to usable capacity is the pitch

Both labs usually rent from data center operators and neoclouds, and both have chased large, long-term agreements. Deals for smaller allocations of compute let companies deploy workloads faster, CNBC reports. "Securing a few megawatts at an existing powered site can be more practical than waiting for a much larger block in one location," Tan, head of research at Structure Research, told CNBC.

"For workloads that can operate across separate sites, a collection of smaller deployments can add up to substantial capacity," he added, calling the advantage speed to usable capacity. An OpenAI spokesperson said the lab is building a diversified compute portfolio to meet growing demand for AI around the world, and that different workloads need different infrastructure, so it talks with a range of partners and weighs opportunities on requirements, performance, reliability, timing and cost.

Inference is projected to pass training in 2027

Training AI models takes large amounts of computing power to chew through huge quantities of data; inference, the day-to-day work of serving those systems, can run on smaller clusters of chips. "Training a large model typically requires many chips working closely together," Tan said. Many inference workloads instead serve separate requests across multiple smaller clusters, which opens up more locations.

The everyday version: training is one heavy load that needs a single big truck and a single big road, while inference is a pile of parcels that a dozen local vans can carry at the same time, each from a different depot.

The mix is moving. According to a JLL report, inference made up 9% of global data center workloads in 2025 against 14% for training, and by 2030 inference is projected to take 37% of capacity compared with 13% for training. The report puts the crossover, when inference overtakes training, in 2027.

Nvidia studied small distributed sites in February

The hardware side has been circling the same idea. In February it was announced that Nvidia would work with several data center stakeholders to study smaller-scale data centers designed for distributed inference.

Crusoe, which built the large Texas complex used by OpenAI, is now investing in smaller data centers, the Wall Street Journal reported Thursday. Those facilities will be faster and cheaper than the bigger builds, which are running into delays across the U.S., according to the Journal. Crusoe is one of several neoclouds riding the buildout, and it announced on Thursday that it had raised $3.9 billion at a $30.9 billion post-money valuation.

What the reporting does not give you is a single completed transaction, and the people describing the conversations spoke anonymously about private business dealings. In our view the 20–30 MW figure is the telling detail, because it is close to what an already powered site can release while land and power stay scarce in Europe.

What to watch in the Nordics No partner, site or start date has been given for the smaller deployments, and talks at this stage can end without a contract. Two dated markers are worth keeping: JLL's 2027 crossover between inference and training capacity, and Crusoe's smaller facilities arriving while large U.S. builds slip. The open question is whether any 20–30 MW agreement in the U.K., the Nordics or the U.S. surfaces with names attached.