EpochAI Predicts Frontier AI Training Will Demand 4-16 GW by 2030

WillRinehart · x · 2026-08-09

A study by EpochAI and EPRI estimates that training frontier AI models by 2030 will require a massive 4 to 16 gigawatts (GW) of power.

In response to this projection, Jigar Shah offers a counter-perspective. He argues that the industry will pivot to building smaller, sub-20MW data centers ideal for distributed inference, claiming that the 1+ GW data centers already built are sufficient to handle future training needs.

Original post →

More from AGI Musings

AGI Musings channel →