Epoch AI

Epoch AI is a research nonprofit that describes itself as “data-first”, investigating the future of artificial intelligence. It started in 2021 as a group of volunteers collecting and analyzing data on hundreds of machine learning models, and the group founded Epoch AI to scale up that work after its 2022 paper on trends in training compute.¹ It builds and hosts public databases on the inputs to AI development, among them machine learning hardware, model training compute, GPU clusters and AI data centers. It also runs independent evaluations of new models, reviews prominent benchmarks and creates new ones, such as FrontierMath, its program for testing AI on advanced mathematics.¹⁷ Its work is free to reuse under the Creative Commons Attribution license, so MetaPrompt publishes its figures with credit.¹

Its AI data centers hub covers 94 sites, built from satellite imagery, permits and public documents, and Epoch estimates that the set covers 44% of global deployed AI compute.² Every power figure there is an estimate: Epoch expects 80% of its IT power estimates to fall within a factor of 1.4 of the true value.³ Those estimates rank the sites in Largest AI Data Centers by Estimated Power Draw (October 2026), and Epoch’s dated satellite notes set the status of each site in Stargate's Sites: Operating vs Under Construction vs Planned; its April 2026 overview counted seven US Stargate sites with over 9 GW of planned capacity.²⁴ Its hardware dataset is a cross-check in AI Accelerators With the Most HBM per Chip (October 2026), which uses the makers’ own figures where the two differ.⁵ And in a review dated September 3, 2026, Epoch rated SWE-bench Verified “Flawed”, judging it unlikely to meet its bar of fewer than 20% of tasks having accuracy-affecting issues, as covered in SWE-bench Verified, Pro and What Replaced Them: By the Numbers.⁶