Full Breakdown
Anthropic and OpenAI Turn to 20-30 MW Data Center Deals
By Drooid · · How we work
Shift Toward Smaller Compute Deployments
Anthropic and OpenAI are now exploring data-center contracts for much smaller capacities—roughly 20 to 30 megawatts (MW). Sources familiar with the negotiations say the two AI labs are seeking such deals across the United Kingdom, the Nordics, and the United States. The move follows a year of “hyperscale” agreements that secured facilities measured in hundreds of megawatts to gigawatts, reflecting a strategic pivot toward faster-to-market compute that can support distributed inference services.
Recent Large-Scale Infrastructure Agreements
Over the past twelve months, both companies announced massive procurements. Anthropic reportedly signed a cloud agreement in August valued at about $45 billion with Nscale, leasing roughly 460 MW for a West Virginia data-center project. OpenAI’s “Stargate” infrastructure initiative exceeded its original 10-gigawatt (GW) commitment by April, later adding 3 GW in Georgia and 8 GW in Ohio. These hyperscale projects have encountered obstacles such as local community opposition in the United States and land- and power-supply constraints in Europe, which can delay the delivery of contracted capacity.
Official Statements & Rationale
The spokesperson emphasized that “different workloads need different infrastructure,” and that OpenAI evaluates potential partners based on performance, reliability, timing, and cost. Anthropic declined to comment on the smaller-capacity talks.
Implications for AI Deployment
Pursuing 20-30 MW allocations allows the labs to bring new workloads online more quickly than the multi-year timelines typical of gigawatt-scale builds. Market-research director Jabez Tan of Structure Research noted that the appeal of smaller deals lies in “faster access to available capacity.” If successful, the shift could accelerate the rollout of inference-heavy services while mitigating the regulatory and logistical challenges that have slowed larger projects.