AI data-center power doubles every ten months but infrastructure lags

According to Epoch AI, the electrical capacity of the largest AI data centers has doubled every ten months since mid-2024. The current record holder is Colossus 2 from SpaceXAI at 950 megawatts of IT load — nearly four times the previous peak, Google's Papillon at 237 megawatts two years ago. For context, the entire city of San Diego averages roughly 800 megawatts.

Fitting a trend line to the historical data yields 2.3× annual growth. Epoch AI's survey of planned facilities suggests that pace will hold at least through early 2027. Beyond that horizon, however, the existing project pipeline shows a sharp deceleration.

The model projects that in twenty months — two doublings at the historical rate — the largest center would reach only 2.3 megawatts, which the researchers describe as "only a bit more than double" the current leader. The figure contradicts itself: 2.3 megawatts is a fraction of 950 megawatts, not a multiple. Epoch AI acknowledges the sample is small and that new announcements or expansions of existing sites could still appear.

The deeper mismatch is physical: a ten-month doubling cycle collides with infrastructure that moves on decade timescales. The most recent capacity auction from PJM, North America's largest grid operator, cleared at 11 times the price of two cycles prior and still failed to meet reliability targets. Chips scale like software; transmission lines, transformers, and switching stations do not.

To enable continuous tracking, Epoch AI has launched the "AI Data Centers hub", a centralized repository of capacities, timelines, and plans for major facilities. The tool aims to clarify whether the emerging slowdown is real or merely a gap in currently available information.