AI’s Race for Power Is Driving Data Centers Off the Grid

Summary

Meredith Fowlie argues that the race to train frontier AI models is pushing major companies to build their own electricity supplies rather than wait four to seven years for grid connections. Cleanview data cited in the post tracks about 2 GW of off-grid generation already powering US data centers and 90 GW of announced or planned behind-the-meter capacity. In Ohio, two 200 MW gas plants being built for Meta reportedly cost $140-$160 per megawatt-hour, compared with roughly $90-$95 for grid supply; nationally, RaboResearch estimates behind-the-meter gas at $100-$165 per megawatt-hour. xAI reportedly brought Colossus I online in 122 days after choosing to bypass the queue, and modular turbines can later be redeployed or retained as backup. Almost all new off-grid capacity is natural gas rather than solar and storage, which Fowlie suggests reflects speed, procurement, and mobility advantages, despite the smaller gas units being less efficient and more polluting than modern combined-cycle plants. She says private willingness to pay may exceed the broader social value of moving an AI race ahead by months, weakening the deterrent effect of pollution penalties. A New Jersey $1.1 million fine against DataOne for 45 unpermitted generators did not stop the generators from operating, illustrating the limits of enforcement when firms place a high value on speed. The post describes a middle path: add friction to quick, polluting options while accelerating cleaner alternatives, such as locating data centers near grid-connected gas plants and adding solar and storage. Fowlie concludes that the AI race makes this transition urgent, while noting that Claude assisted with drafting the post.