OpenAI and Anthropic start competing for smaller orders! The battle for AI computing power shifts to small data centers of 20-30MW.

Anthropic and OpenAI are shifting some of their computing power needs to 20-30 megawatt (MW) level data centers to obtain available capacity more quickly. This contrasts with the two companies' previous ultra-large-scale infrastructure deployments of hundreds of MW or even gigawatt (GW) levels.

On September 18, CNBC reported, citing sources familiar with the matter, that Anthropic has been in talks in the UK and Northern Europe about projects of 20 to 30 MW in scale, and OpenAI is also looking for similar opportunities in the Nordic countries. The two sides also have related negotiations in the United States.

This shift occurs as AI computing demands are gradually migrating from model training to inference services. Smaller facilities can be deployed more quickly and are better suited for distributed processing of inference tasks. Structure Research predicts that by 2027, inference workloads will account for more than training workloads in total global data center capacity, rising to 37% by 2030.

Beyond mega-projects, AI companies are beginning to diversify their operations.

Over the past year, both companies have signed large-scale infrastructure agreements. According to a CNBC report last August, Anthropic reached a cloud computing agreement worth approximately $45 billion with cloud service provider Nscale, leasing about 460 MW of computing power from its data center in West Virginia. OpenAI's Stargate project has already exceeded its initial commitment of 10 GW and has further expanded its development plans by 3 GW in Georgia and 8 GW in Ohio.

Now, both companies are supplementing their computing power sources with more flexible options. CNBC, citing four sources familiar with the matter, reported that Anthropic is seeking 20 to 30 MW of resources in the UK and Northern Europe; two of these sources indicated that OpenAI is also exploring similar projects in Northern Europe, while another source stated that negotiations between the two companies are also taking place in the United States.

OpenAI responded that it is "building a diverse portfolio of computing power to meet the growing global demand for AI." Different workloads require different infrastructures, and the company takes into account factors such as demand, performance, reliability, timelines, and cost.

"Speed" has become a new consideration in computing power procurement.

Jabez Tan, research director at Structure Research, said the core advantage of smaller projects is that they provide available capacity more quickly. Acquiring a few MW of capacity at an existing, already powered site is often more practical than waiting for a large amount of capacity to be released at a single location; for workloads that can run across sites, multiple smaller facilities can also be stacked to create a considerable total capacity.

The construction of large-scale data centers faces a lengthy process. Large projects in the US and other markets have encountered resistance from local communities, while many regions in Europe also face shortages of land and power resources. For AI companies, acquiring distributed resources first can supplement computing power before large-scale projects are implemented.

According to a Wall Street Journal report on Thursday, Crusoe, which built a large data center campus for OpenAI in Texas, is increasing its investment in smaller data centers. These facilities are faster and cheaper to build, helping to mitigate the impact of delays in large projects. Crusoe declined to comment, but the company announced on the same day that it had completed a $3.9 billion funding round, bringing its post-funding valuation to $30.9 billion.

Inference demands are driving computing power toward distributed computing.

The changing structure of AI workloads is another important driver behind this trend. Jabez Tan stated that training large models typically requires a large number of chips to run collaboratively within the same cluster, while inference tasks can be split into multiple smaller clusters to handle independent requests, making them more suitable for distributed deployment.

According to JLL data, inference workloads accounted for 9% of global data center workloads in 2025, while training workloads accounted for 14%. It is projected that by 2027, inference workloads will surpass training workloads, and by 2030, they will further rise to 37%, while training workloads will decrease to 13% during the same period.

This trend is also beginning to influence data center construction models. In February of this year, NVIDIA announced that it would collaborate with several data center industry players to research small data center solutions for distributed inference. As inference demand continues to grow, the competition in AI infrastructure is shifting from simply pursuing hyperscale to simultaneously focusing on the speed of acquiring computing power, deployment flexibility, and distributed capabilities.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.