The hunt for computing power has a new shape. Anthropic and OpenAI, two of the companies that signed the largest data center contracts of the past year, are now shopping for deals a fraction of that size. The two developers are pursuing artificial-intelligence data center projects in the 20-to-30-megawatt range, according to people familiar with the matter, a scale far below the hundreds-of-megawatt and gigawatt-level agreements they have struck over the past twelve months.
The smaller projects are an attempt to bring capacity online faster. A 30-megawatt site can be permitted, wired, and stocked with servers more quickly than a campus that requires years of planning, power procurement, and construction. For a model developer racing to serve customers, speed can matter as much as total capacity, and a modest project offers the shortest path from signature to live computing.
The geography of the search is telling. Anthropic has approached potential partners in Britain and the Nordics, the people said. OpenAI is exploring similar opportunities in the Nordics, and both companies are in talks on projects of the same size in the United States. The locations point to regions with available power and, in the Nordics, comparatively cheap and clean electricity.
The shift also reflects a change in what the computing is for. Training a frontier model requires tens of thousands of chips working tightly together inside a single cluster, connected by networks fast enough to behave as one machine. Inference, the work of answering queries once a model is trained, can be spread across many smaller, separate clusters. As AI products move into everyday use, the balance of demand is tilting from training toward inference.
Analysts described the smaller deals as a natural response to that tilt. A distributed set of modest sites is well suited to inference workloads, which benefit from being close to users and do not demand the extreme interconnect of a training run. Industry executives said such projects can be leased and populated on a schedule that matches a fast-moving product roadmap rather than a construction calendar.
The numbers bear out the direction. JLL, the commercial real estate firm, expects that by 2027 global data center capacity devoted to AI inference will exceed capacity devoted to training. The forecast would have seemed backwards two years ago, when the industry’s capital was consumed almost entirely by ever-larger training clusters.
Neither company is abandoning the big deals. Both still need enormous training capacity to build the next generation of models, and neither has signaled any retreat from the multi-year campus agreements already signed. The smaller deals are additive, a layer of capacity that can be brought online in months rather than years.
The change in posture comes as both companies push to expand the supply of compute at a moment when chips and power, not ambition, are the binding constraints. Signing a gigawatt deal locks in scale but also locks in a timeline measured in years. A 25-megawatt hall, by contrast, can be financed, built, and generating revenue before a larger campus would finish its later phases.
For the data center developers on the other side of these talks, the smaller deals are attractive for the same reason. They are quicker to permit and quicker to lease, and they do not require the same degree of conviction about demand five years out. A developer can take a smaller bet on a shorter horizon and keep its larger sites for the customers who still need them.
The tilt toward inference is visible in how the companies deploy once a model ships. A frontier model spends months in a training run and years answering questions, and it is the answering that ultimately reaches customers. Every chatbot query, every code suggestion, and every document summary is inference, and it is this load, spread across millions of users, that the smaller sites are meant to absorb.
Power is the other constraint pushing the companies toward smaller projects. The largest campuses need dedicated grid interconnections that can take years to secure, while a modest site can often slot into existing capacity at the edge of a grid. The 20-to-30-megawatt size range sits in a sweet spot: big enough to be worth building, small enough to find power for quickly.
The regions named in the talks fit that logic. The Nordics offer abundant renewable power, a cool climate that lowers cooling costs, and a surplus of grid capacity relative to local demand. Britain and the United States offer proximity to customers and, in the U.S., the largest concentration of AI users in the world. Each location is a different compromise between cost, speed, and distance from the people the models serve.
Neither Anthropic nor OpenAI commented on the discussions. The talks are early, the people said, and could change or fall apart. But the pattern is already clear enough: the companies that once measured their ambitions in gigawatts are now also bargaining over tens of megawatts, because the fastest route to more computing no longer runs only through the biggest buildings.
