Anthropic and OpenAI are pursuing smaller AI data center deals as they look to ramp up capacity quickly, according to people familiar with the talks. The push for smaller AI data center deals, in the 20 to 30 megawatt range, comes alongside the companies’ previously announced multi-hundred-megawatt and gigawatt-scale commitments.
People with knowledge of the discussions said Anthropic has explored agreements in that size range across the U.K. and the Nordic region. Two sources said OpenAI has been evaluating similar opportunities in the Nordics.
Another person familiar with the matter said both firms have also discussed deployments at that scale in the United States. The individuals requested anonymity to discuss private negotiations.
Both companies have unveiled a series of infrastructure agreements over the past year as they train and serve increasingly large models. Securing smaller allocations can help them bring workloads online faster amid surging demand.
An OpenAI spokesperson said the company is building a diversified compute portfolio to meet growing global demand for AI. The spokesperson added that different workloads require different infrastructure and that OpenAI regularly assesses partners based on requirements, performance, reliability, timing, and cost.
The company declined to discuss specific commercial talks. Anthropic declined to comment.
Smaller AI data center deals and speed to capacity
The firms typically rent capacity from data center operators and neocloud providers and have sought long-term, large-scale arrangements. In August, people familiar with the matter said Anthropic struck a roughly 45 billion dollar cloud agreement with Nscale that would provide about 460 MW of capacity at a development in West Virginia.
OpenAI has said it exceeded its original 10 GW commitment to its Stargate infrastructure project in April and has since committed to an additional 3 GW in Georgia and 8 GW in Ohio.
Massive projects in the United States and abroad increasingly face resistance from local communities, while much of Europe is constrained by limited land and power availability.
Jabez Tan, head of research at Structure Research, said smaller capacity deals are often attractive due to speed to usable capacity. Securing a few megawatts at an existing powered site can be more practical than waiting for a larger block in one place, he said.
For workloads that can operate across separate sites, multiple smaller deployments can add up to substantial capacity.
Shift to inference
Training AI models demands large, tightly coupled clusters of chips, but running those systems day-to-day, known as inference, can often be handled by smaller clusters distributed across more locations. Tan noted that many inference workloads can serve separate requests across multiple smaller sites.
As more compute shifts from training to production use, inference capacity needs are expected to rise. A report from real estate firm JLL said the share of total data center capacity dedicated to inference is expected to surpass training in 2027.
In 2025, inference accounted for 9% of global data center workloads versus 14% for training, according to the report. By 2030, inference is projected to use 37% of capacity, compared to 13% for training.
In February, Nvidia said it would work with data center stakeholders to study smaller-scale facilities designed for distributed inference.
U.S. firm Crusoe, which built a large data center complex in Texas used by OpenAI, is now investing in smaller facilities, the Wall Street Journal reported Thursday. The Journal said these builds aim to be faster and cheaper than larger projects that are facing delays across the United States.
Crusoe did not respond to a request for comment. The company announced Thursday it raised 3.9 billion dollars at a 30.9 billion dollar post-money valuation as demand for neocloud providers grows with the AI buildout.
Earlier this year, concerns about how rapidly these systems are being deployed led to OpenAI misalignment reports prompt tighter AI oversight as regulators and researchers scrutinize the technology’s broader risks.
