Anthropic and OpenAI Seek Small AI Data Centers to Accelerate Deployment of 20 to 30 MW Compute

TradingKey - Anthropic and OpenAI are seeking small data centers with capacities of around 20 to 30 megawatts outside of massive AI infrastructure projects to gain available compute faster and meet growing model inference demand.

Anthropic has approached relevant projects in the UK and Nordic region, while OpenAI previously explored similar opportunities in the Nordics, with both companies also discussing equivalent capacity within the U.S., CNBC reported, citing people familiar with the matter. Compared with the hundreds of megawatts or even gigawatt-scale deals signed by both parties over the past year, such projects are significantly smaller, but typically faster to build and deliver.

OpenAI said it is building a more diversified compute portfolio to meet global AI demand. As different workloads have varying requirements for performance, reliability, cost, and delivery timelines, the company engages with multiple types of partners but does not comment on specific commercial negotiations.

Anthropic did not respond to requests for comment.

This does not mean the two companies are abandoning hyperscale data centers, but rather adding more flexible compute sources alongside large long-term projects. Anthropic previously reached a roughly $45 billion partnership with UK cloud provider Nscale to rent about 460 megawatts of capacity at its West Virginia project. OpenAI is also pushing forward several gigawatt-scale infrastructure projects through Stargate and other partnerships.

While mega-projects are well-suited for centralized training of advanced models, they often involve lengthy grid connection, permitting, and construction cycles. Some data centers in the U.S. face opposition from local communities, while traditional European markets suffer from constrained land, power, and transmission capacity. Securing tens of megawatts at existing powered sites is often more practical than waiting for a single large campus to become fully operational.

For workloads capable of running across regions, accumulating multiple small clusters can also build up substantial compute scale. This approach also mitigates the impact of individual project delays, allowing AI companies to deploy certain services ahead of large facilities' completion.

The shift of AI computing from training to inference is another reason small data centers are gaining attention. Training large models typically requires a vast number of GPUs working together at high speeds within a single cluster, placing high demands on interconnects and chip density. Inference, on the other hand, primarily handles user requests; many tasks can be split across smaller clusters in different locations without needing to concentrate all chips in a single hyperscale facility.

As chatbots, AI agents, and enterprise applications enter production environments, inference demand is becoming a major driver of data center expansion. JLL projects that global data center capacity allocated for AI inference will surpass training by 2027; by 2030, the share of inference workloads could rise to 37% from 9% in 2025, while the share of training workloads drops to 13% from 14% over the same period.

Smaller projects can also be positioned closer to users and enterprise clients, helping reduce network latency and meet regional data compliance requirements. The UK and the Nordic region not only boast abundant existing data center resources, but also possess relatively ample clean power and lower cooling costs, creating favorable conditions for distributed AI deployments.

Therefore, Anthropic and OpenAI's search for 20 to 30 megawatt projects represents a complement to their existing compute strategies rather than a reduction in infrastructure investment. Both companies still require large clusters to train frontier models, but as inference traffic grows, future AI infrastructure may no longer rely entirely on a small number of gigawatt-scale campuses.