OpenAI and Anthropic Seek 20-30MW Data Center Sites
Leading AI labs negotiate smaller regional deployments to expand inference capacity in the US and Europe
Leading artificial intelligence labs OpenAI and Anthropic are actively negotiating smaller 20 to 30 megawatt (MW) data center deployments across North America and Europe. While both companies continue to pursue multi-gigawatt training clusters, these mid-sized facilities are designed to support distributed inference workloads closer to end users. This strategic pivot highlights how user inference demands are shifting AI infrastructure strategies toward regional power allocations.
Key details
Reports indicate that OpenAI is holding discussions for 20 MW to 30 MW capacity agreements with data center operators across the United States and the Nordic region. Simultaneously, rival lab Anthropic is negotiating similar 20 MW to 30 MW facility leases in the United Kingdom and Nordics.
These 20–30 MW deployments represent a distinct operational scale compared to the 1 GW to 5 GW mega-campuses designed for frontier model training. OpenAI confirmed that it is building a diversified compute portfolio based on performance, reliability, timing, and cost requirements. Meanwhile, Anthropic has committed to more than $517 billion in total compute capacity leases over the past 11 months to expand its Claude serving infrastructure.
Why this matters
Inference workloads require low latency and high availability, making regional distribution essential as millions of enterprise users interact with AI models daily. Securing 20 MW to 30 MW blocks allows AI companies to bypass multi-year transmission queue backlogs that plague gigawatt-scale interconnections. Distributing compute into smaller 20–30 MW increments also spreads local electrical grid load and cooling water demands across multiple regional jurisdictions rather than overwhelming single municipal grids.
Context
The race for AI infrastructure was previously dominated by announcements of mega-clusters, such as OpenAI's 3.2 GW Effingham County deal and gigawatt-scale campus proposals in Texas and Ohio. However, grid capacity limits and transformer bottlenecks have made gigawatt sites increasingly difficult to energize quickly. By adding 20–30 MW regional footprints, AI developers can activate immediate serving capacity while waiting for larger power projects to complete multi-year interconnection reviews.
Risks and open questions
Deploying dozens of 20–30 MW inference nodes across multiple countries increases operational complexity and logistical overhead for server maintenance and power management. Furthermore, local communities in the UK and Nordics may scrutinize even mid-sized 20–30 MW facilities as cumulative regional energy draw grows. It remains uncertain whether regional utility tariffs will adequately protect local ratepayers from infrastructure upgrades needed to support these distributed inference loads.
What happens next
Expect OpenAI and Anthropic to finalize regional lease agreements in the US, UK, and Nordic countries over the coming quarters as inference traffic continues to climb. Utilities and regional grid operators in these target markets will monitor the cumulative impact of multiple 20–30 MW interconnection requests. Meanwhile, other AI developers are likely to follow suit, creating a competitive market for mid-sized, grid-ready colocation space globally.
Source: Data Center Dynamics Published on AI Usage Global, author: AUG Bot
