OpenAI and Anthropic Pivot to Smaller Data Center Deals

 

By Adam Pease

Infrastructure Strategy: OpenAI and Anthropic Target Smaller AI Data Centers

The race for artificial intelligence capacity is taking an unexpected turn as major players seek smaller, distributed footprints. Leading AI labs Anthropic and OpenAI are actively hunting for smaller data center deals, often in the 20-30 megawatt range. This marks a shift from the massive, gigawatt-scale “Stargate” style projects that dominated recent headlines. These companies are exploring opportunities in regions like the Nordics and the United Kingdom, in addition to ongoing domestic US expansion. This blog overviews the shift toward smaller data center deployments and offers our analysis.

Why Did Anthropic and OpenAI Announce Smaller Data Center Deals?

Hyperscalers and frontier AI labs face an urgent reality: the massive data centers they need take years to build and power. The demand for AI compute, particularly for inference, is skyrocketing, and waiting for massive campuses is no longer a viable sole strategy. Reports indicate that Anthropic and OpenAI are now pursuing these smaller, distributed sites to accelerate their “speed to usable capacity.” Smaller capacity allocations, while perhaps less efficient for training large foundational models, offer a crucial advantage in deployment speed.

Furthermore, the power grid is under immense strain globally, making multi-hundred-megawatt grid interconnections increasingly difficult and slow to secure. By targeting 20-30 MW opportunities, these companies can bypass some of the longest utility delays and tap into existing or more readily available power pockets. This strategy allows them to distribute their operational risk and establish a presence in strategic geographic locations more quickly. It’s a pragmatic response to the physical infrastructure bottlenecks constraining the AI boom.

Analysis

This shift is more than just a real estate play; it signals a fundamental maturation of the AI market from a singular focus on model training to a dual focus on training and inference deployment. While the multi-gigawatt deals will continue for massive training clusters, inference workloads, which serve active users, do not always require maximum co-location. For serving inference, geographically distributed clusters can actually offer lower latency to end users and improved service reliability. The sheer volume of inference requests is expected to overtake training workloads by 2027, making this strategic shift critical for long-term scalability.

The pivot toward smaller deals demonstrates that even the best-funded technology companies must now align their AI ambitions with the hard realities of electrical power availability and community resistance. This development validates the importance of alternative energy solutions and decentralized infrastructure planning. It also provides a significant opportunity for smaller, regional data center operators and “neoclouds” that can offer agility and localized power access that larger wholesale providers might lack. Legacy providers will need to adapt or risk missing out on this faster-moving segment of the AI buildout.

Bottom Line

The AI infrastructure race has entered a new phase focused on speed to inference capacity through decentralized, smaller data center deployments. Major labs like OpenAI and Anthropic are adapting their strategies to overcome the physical limitations of massive centralized builds, driven by urgent demand and power grid constraints. Enterprises must recognize that inference delivery is going local and plan their connectivity and deployment strategies accordingly to leverage the oncoming wave of distributed, low-latency AI services.