Anthropic and OpenAI are shifting toward small-scale data centers with capacities of 20 to 30 megawatts (MW) to secure computing power more quickly, supplementing their hyperscale infrastructure deployments. This shift stems from the transition in AI computing from training to inference, for which smaller facilities are better suited for distributed deployment. Inference workloads are projected to surpass training loads by 2027, rising to 37% by 2030. The competition for computing power is increasingly focusing on speed, flexibility, and distributed capabilities.
Anthropic and OpenAI are redirecting part of their computing demand toward data centers in the 20–30 MW range to accelerate access to available capacity. This contrasts with their previous investments in hyperscale infrastructure, which typically involved hundreds of MW or even gigawatt (GW)-scale projects.
On September 18, CNBC reported, citing insiders, that Anthropic has initiated discussions regarding 20–30 MW projects in the United Kingdom and Northern Europe, while OpenAI is seeking similar opportunities in Northern Europe. Both companies are also engaged in related negotiations within the United States.
This shift occurs as AI computing demand gradually transitions from model training to inference services. Smaller facilities can be deployed more rapidly and are better suited for distributed inference tasks. Structure Research forecasts that by 2027, inference workloads will exceed training workloads in terms of share of global data center capacity, reaching 37% by 2030.
Beyond mega-projects, AI companies are beginning to adopt a "multi-site deployment" strategy.
Over the past year, both companies have signed major infrastructure agreements. According to a CNBC report last August, Anthropic reached a cloud computing agreement worth approximately $45 billion with Nscale, a cloud service provider, to lease about 460 MW of computing capacity at its data center in West Virginia, USA. Meanwhile, OpenAI’s Stargate project has surpassed its initial commitment of 10 GW, with additional development plans of 3 GW in Georgia and 8 GW in Ohio.
Currently, both companies are beginning to supplement their capacity with more flexible sources. Citing four informed sources, CNBC reported that Anthropic is seeking resources in the 20–30 MW range in the UK and Northern Europe; two of these sources indicated that OpenAI is also exploring similar projects in Northern Europe, while another noted that both companies are involved in relevant negotiations within the United States.
In response, OpenAI stated that the company is "building a diversified computing portfolio to meet the growing global demand for AI." Different workloads require different infrastructure, and the company takes into account demand, performance, reliability, timelines, and cost in its comprehensive assessment.
"Speed" becomes a new consideration in computing power procurement
Jabez Tan, Research Director at Structure Research, stated that the core advantage of smaller projects is faster access to available capacity. Securing a few MW of capacity at existing, powered sites is often more practical than waiting for large-scale capacity to become available at a single location; for workloads that can operate across multiple sites, several smaller facilities can be aggregated to form substantial total capacity.
The construction of large-scale data centers faces extended timelines. Major projects in the United States and other markets have encountered resistance from local communities, while many regions in Europe are grappling with shortages of land and power resources. For AI companies, securing distributed resources first can help supplement computing capacity before large-scale projects are fully implemented.
According to a Thursday report by The Wall Street Journal, Crusoe, which previously built a large data center campus for OpenAI in Texas, is increasing its investment in smaller data centers. Such facilities can be constructed more quickly and at lower cost, helping to mitigate the impact of delays in large-scale projects. Crusoe declined to comment, but on the same day announced the completion of a $3.9 billion funding round, bringing its post-money valuation to $30.9 billion.
Inference demand is driving computing power toward distributed architectures
Changes in the structure of AI workloads are another key driver behind this trend. Jabez Tan noted that training large models typically requires a vast number of chips operating collaboratively within a single cluster, whereas inference tasks can be split across multiple smaller clusters to handle independent requests separately, making them better suited for distributed deployment.
Data from JLL shows that in 2025, inference workloads accounted for 9% of global data center workloads, while training workloads accounted for 14%. Inference workloads are projected to surpass training workloads by 2027, and further rise to 37% by 2030, while the share of training workloads will drop to 13% over the same period.
This trend is also beginning to influence data center construction models. In February this year, NVIDIA announced collaborations with several industry participants to research small-scale data center solutions tailored for distributed inference. As inference demand continues to grow, competition in AI infrastructure is shifting from a sole focus on hyperscale capacity to also emphasizing the speed of acquiring computing power, deployment flexibility, and distributed capabilities.