OpenAI and Anthropic are increasingly turning to smaller-scale data centers in the 20 to 30 megawatt (MW) range to secure usable computing capacity more quickly. This marks a notable departure from their previous strategy of pursuing massive infrastructure deals measured in hundreds of megawatts or even gigawatts (GW).

According to sources cited by CNBC on September 18, Anthropic has initiated discussions for projects in the 20–30 MW range across the UK and Northern Europe, while OpenAI is exploring similar opportunities in Northern Europe. Both companies are also engaged in related negotiations in the United States.

This strategic pivot comes as AI computing demand shifts from model training toward inference services. Smaller facilities can be deployed faster and are better suited for distributed inference workloads. Structure Research projects that by 2027, inference workloads will surpass training loads in their share of global data center capacity, climbing to 37% by 2030.

Moving beyond mega-projects, AI companies are diversifying their footprint. Over the past year, both firms have signed substantial infrastructure agreements. As reported by CNBC in August of last year, Anthropic entered into an approximately $45 billion cloud computing deal with Nscale, leasing around 460 MW of capacity at its West Virginia data center. Meanwhile, OpenAI's Stargate project has exceeded its initial 10 GW commitment and added another 3 GW and 8 GW of development plans in Georgia and Ohio, respectively.

Now, these companies are supplementing their portfolios with more flexible compute sources. CNBC, citing four informed individuals, reported that Anthropic is seeking 20–30 MW resources in the UK and Northern Europe. Two of those sources indicated that OpenAI is exploring comparable projects in Northern Europe, and one added that both companies are involved in related discussions within the US.

OpenAI responded by stating that it is building a diversified compute portfolio to meet growing global AI demand. The company emphasized that different workloads require different infrastructure, and it weighs factors such as demand, performance, reliability, timeline, and cost.

Speed has become a new priority in compute procurement. Jabez Tan, Research Director at Structure Research, noted that the core advantage of smaller projects lies in faster access to available capacity. Securing a few megawatts at an already-powered existing site is often more practical than waiting for massive capacity to come online at a single location. Additionally, multiple small facilities for workloads that can operate across different sites can aggregate into significant total capacity.

Large data center construction faces extended timelines. Major projects in the US and other markets encounter community opposition, and many European regions grapple with land and power constraints. For AI companies, securing distributed resources first provides a way to supplement computing power before larger initiatives are completed.

Crusoe, the company that previously built a large data center campus for OpenAI in Texas, is reportedly increasing its focus on smaller facilities, according to the Wall Street Journal on Thursday. These smaller sites are quicker to build and less costly, which helps mitigate the risks of delays associated with larger projects. Crusoe declined to comment on the matter but announced the completion of a $3.9 billion funding round on the same day, bringing its post-money valuation to $30.9 billion.

Growing inference demand is driving a shift toward distributed computing. Jabez Tan explained that large model training typically requires numerous chips to work together within a single cluster, whereas inference tasks can be split across multiple smaller clusters handling independent requests, making them more conducive to distributed deployment.

JLL data shows that in 2025, inference workloads will account for 9% of global data center workloads, with training at 14%. By 2027, the share of inference is expected to surpass that of training, and by 2030 it is forecasted to rise to 37%, while training's share drops to 13%.

This trend is also influencing data center construction models. In February, Nvidia announced plans to collaborate with various data center industry players to study small-scale data center solutions tailored for distributed inference. As inference demand continues to grow, the competition around AI infrastructure is evolving from a pure focus on scale to an equal emphasis on speed of deployment, operational flexibility, and distributed capabilities.