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

Two numbers carry this story: 20 and 30. Anthropic has been sounding out compute deals of just 20–30 MW across the U.K. and the Nordics, four people familiar with the conversations told CNBC, and OpenAI has explored deployments of the same size in the Nordics.

At a glance

• One source also described talks involving both labs about U.S. capacity at that 20–30 MW scale, a fraction of the multi-hundred-megawatt and gigawatt facilities the two have been contracting for all year.
• Structure Research's Jabez Tan says the draw is speed to usable capacity: a few megawatts at an already powered site beats waiting for one big block, and separate sites add up.
• The catch is stage: these are conversations and sounded-out agreements rather than signed contracts, and inference was still only 9% of global data center workloads in 2025, against 14% for training.

If you have not been tracking the buildout, the pattern of the past year has been one large infrastructure deal after another, with both labs renting capacity from data center operators and neoclouds under long-term, large-scale agreements. That part has not stopped. The 20–30 MW conversations sit alongside it.

Anthropic's 20–30 MW talks span the U.K. and the Nordics

Four people familiar with the conversations, who asked to remain anonymous discussing private business dealings, told CNBC that Anthropic has sounded out agreements in that range across the U.K. and the Nordics. Two of the sources said OpenAI had been exploring the same small deployments in the Nordics. One said they were familiar with talks involving both labs about U.S. capacity at that scale.

An OpenAI spokesperson told CNBC the company is building a diversified compute portfolio to meet growing demand for AI around the world, and does not comment on specific commercial discussions.

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.

Next to the small blocks sits a $45 billion Nscale rental

Two people familiar with the matter told CNBC in August that Anthropic signed a roughly $45 billion cloud deal with Nscale, renting around 460 MW of capacity at a data center development in West Virginia. OpenAI said it surpassed the original 10 GW commitment to its Stargate project in April, and has since committed to developing a further 3 GW in Georgia and 8 GW in Ohio.

The large end is getting harder. Big projects in the U.S. and further afield increasingly face pushback from local communities, and in much of Europe land and power are in short supply. Jabez Tan, head of research at Structure Research, told CNBC that small deals are often attractive 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, and for workloads that can run across separate sites, a collection of smaller deployments adds up.

JLL expects inference to overtake training in 2027

Training a large model requires many chips working closely together, Tan said, which pins the capacity to one place. Many inference workloads can instead serve separate requests across multiple smaller clusters, which opens up more locations.

Think of training as a rowing eight: everyone has to sit in the same boat and pull in time. Inference is closer to a row of ticket windows, where each request is handled on its own, and the windows can stand in different buildings without anyone noticing.

A report from real estate company JLL puts the crossover in 2027. In 2025, inference made up 9% of global workloads in data centers, while training made up 14%. By 2030 the report projects inference at 37% of that capacity, with training at 13%.

Crusoe raised $3.9 billion at a $30.9 billion valuation

Crusoe, which built the huge Texas data center complex 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., according to the Journal.

Crusoe is one of several neoclouds whose business has boomed during the buildout, and it announced on Thursday that it had raised $3.9 billion at a $30.9 billion post-money valuation. In February, Nvidia was announced as a collaborator with several data center stakeholders on a study of smaller-scale facilities designed for distributed inference.

Talks are not capacity. The reporting describes sounded-out agreements and conversations, and OpenAI's own framing goes no further than assessing opportunities against its requirements. In our view the 20–30 MW size is the telling detail, because it is small enough to fit into power that already exists rather than power somebody still has to build, which is the exact shortage the reporting names in Europe.

Watch the 2027 crossover

The checkpoint is 2027, the year JLL puts the inference crossover. Until then the open question is whether sounded-out blocks turn into signed capacity: the sources named no operator, no site and no start date, so there is nothing on the calendar to check against. The other thread to follow is Crusoe, which the Wall Street Journal says is now putting money into smaller builds while the large ones slip.