Anthropic and OpenAI are now chasing 20-30MW data center deals alongside the gigawatt ones.

CNBC reported September 18 that both labs are pursuing much smaller sites in parallel with Anthropic's $45 billion Nscale agreement for 460MW in West Virginia and OpenAI's Stargate commitments. The stated reason is speed to usable capacity rather than cost, since a small site comes online far sooner than a campus. The structural driver is the inference shift, with inference expected to pass training in 2027 and reach 37% of data center capacity by 2030 against 13% for training, and inference distributes across small sites in a way training runs don't.