Epoch AI GPU clusters

Citation

Epoch AI. “Data on GPU clusters.” Data explorer, CSV updated Mar. 13, 2026. https://epoch.ai/data/gpu-clusters. Read 2026-10-28.

Bottom line

A public database of “over 500 GPU clusters and supercomputers”, with performance, power, hardware cost, owner and sector for each.
The data is free under a CC-BY license.

Background

Epoch AI renamed the dataset: it “was previously called ‘AI Supercomputers’ but was renamed to account for its broad coverage of GPU clusters”. The page shows it is still published, under the new name.

Key claims

Fields: “Max performance (OP/s)”, “Performance (16-bit OP/s)”, “Power capacity (MW)”, “Hardware cost (2025 USD)”, “Hardware quantity”, and color by hardware type, supplier, sector, owner and users (Settings).
Coverage: “an estimated 10–20% of existing global aggregate GPU cluster performance as of March 2025” (Disclaimer).
Method: clusters come from “machine learning papers, publicly available news articles, press releases, and existing lists of supercomputers”, from “Top500 and MLPerf, and GPU rental marketplaces”.
Estimates: when an owner does not report performance, cost or power, Epoch AI computes it from the chip count and per-chip figures (FAQ).
Owners: “Historically, government and academic research organizations … owned many of the top supercomputers. In recent years, most AI supercomputers are owned by cloud computing providers … or AI laboratories” (FAQ).
License: CC-BY, with data on clusters in China rounded and anonymized (FAQ).

Evidence and method

A research nonprofit’s own dataset page. A separate paper and a documentation page give the method.

Limits

The copy has the page text only.
Coverage is 10 to 20 percent of world GPU cluster performance. Academic clusters in California may be missing.
Some figures are estimates, not reported values.