AI training to have whopping power needs, report says

• It's a joint project of the Electric Power Research Institute and Epoch AI.

State of play: The power required for the largest frontier training runs will surge this decade, potentially reaching 1-2 gigawatts by 2028 and a whopping 4-16 GW by 2030 (though that high end is considered unlikely).

Stunning stat: "This demand would be highly significant, with the high end for a single model approaching 1% of total U.S. power capacity," the report finds.

• "The energy demands of training cutting-edge AI models are doubling annually, soon rivaling the output of the largest nuclear power plants," said Jaime Sevilla, director of Epoch AI, in a statement.

The big picture: Total U.S. power needed to serve AI is estimated at 50 GW in 2030, a huge uptick from 5 GW today, the report finds in tallying inference and training.

• "Forecasts suggest AI could consume over 5% of U.S. generation capacity by 2030," it states.

The bottom line: The white paper acknowledges plenty of uncertainty, but it provides more data points that can help policymakers and hyperscalers plan.