Published by Epoch AI in April 2025 (authored by Konstantin F. Pilz, James Sanders, Robi Rahman, and Lennart Heim), this study analyzes a dataset of more than 500 AI supercomputers built between 2019 and 2025. It finds computational performance grew roughly 2.5 times per year across that period, while power requirements and hardware costs both roughly doubled annually.
As an open-access empirical dataset rather than a survey or forecast, the report has become a widely cited primary source for tracking the physical scaling trajectory of AI training infrastructure.