Power Demands Of Frontier Ai Training

In an August 2025 white paper written with the Electric Power Research Institute (EPRI), Epoch AI researchers Josh You and David Owen forecast that the largest single frontier training runs in 2030 will likely draw 4-16 gigawatts, enough to power millions of US homes, according to Epoch's summary. At the time, the largest runs exceeded 100 megawatts, and the power needed for frontier training had been growing 2.2x per year.

The forecast starts from training compute, which has grown 4-5x per year. The authors expect that to continue despite efficiency gains and the shift to reasoning models, which still reward more compute for reinforcement learning. Two factors offset compute growth when converting it to power. Leading AI GPUs have become about 40% more energy efficient each year, and longer training runs spread the same compute over more time. Run lengths grew 26-50% per year historically, but with many runs now past 100 days the authors expect only 10-20% annual growth through 2030, because fast-moving research rewards short iteration cycles. Combining these, they put power growth at 2.2-2.9x per year. The upper end assumes caps on run length push labs toward bigger clusters; those caps could instead slow compute scaling, which is why the authors also show a plain extrapolation of the historical trend.

The authors do not model whether that much power can be supplied, and say the upper end in particular may not be feasible by 2030; the lower end is consistent with already-announced multi-gigawatt data centers. Training across several sites, which Google DeepMind has already done, may let runs grow beyond single-site power limits, a question Epoch examines in a separate report on decentralized training. Looking beyond single runs, and counting several companies plus inference and experiments, they find more than 100 GW of AI capacity worldwide and over 50 GW in the US plausible by 2030, approaching 5% of US generating capacity. When they wrote, AI power use was split roughly equally among training, experiments and inference.