OpenAI and Anthropic are exploring smaller data centre deals as the race to secure computing capacity for artificial intelligence workloads intensifies.
CNBC citing people familiar with the discussions said that the two AI companies have signed major infrastructure agreements over the past year involving facilities with capacities running into hundreds of megawatts and gigawatts. They are now also looking at smaller deployments of around 20-30 megawatts, according to the report.
Anthropic has held discussions for capacity in that range across the UK and Nordic countries, four people familiar with the talks told CNBC. OpenAI has also explored similar opportunities in the Nordics, two sources said.
ALSO READ: Researchers Used Anthropic's Claude To Hack Into OpenAI: Report
One source familiar with the discussions said both companies have also been involved in talks over US capacity deployments of a similar scale.
An OpenAI spokesperson told CNBC that the company is building a diversified compute portfolio to meet rising AI demand globally.
“Different workloads need different infrastructure,” the spokesperson said, adding that OpenAI evaluates opportunities based on requirements including performance, reliability, timing and cost. The company declined to comment on specific commercial discussions.
Smaller capacity deals can allow AI companies to get access to usable computing infrastructure faster, particularly when larger data centre projects face lengthy development timelines.
Both OpenAI and Anthropic typically rent computing capacity from data centre operators and neocloud providers, while also pursuing large, long-term infrastructure agreements.
Anthropic recently signed a roughly $45 billion cloud deal with Nscale, under which it is expected to rent around 460 MW of computing capacity at a data centre development in West Virginia, according to people familiar with the matter cited by CNBC.
The growing focus on smaller facilities is also linked to the changing nature of AI workloads.
Training large AI models requires enormous computing power, with large numbers of chips working together. Inference, the process of running trained models to respond to user requests, can often be handled through smaller clusters distributed across multiple locations.
ALSO READ: Anthropic Launches In-House Biology Lab To Accelerate AI-Led Drug Discovery
This shift is becoming increasingly important as a greater share of AI computing moves from model training to serving AI applications in production.
According to a JLL report cited by CNBC, inference accounted for 9% of global data centre workloads in 2025, compared with 14% for training. Inference is projected to account for 37% of data centre capacity by 2030, while training is expected to account for 13%.
Nvidia also announced in February that it would work with several data centre stakeholders to study smaller facilities designed for distributed AI inference.
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.