Epoch AI: the US received 72% of the world's AI compute since 2022

Epoch AI published its AI Chip Imports Explorer on 27 October, a dataset that estimates where the world's AI chips actually end up. Its headline finding is that of 27.4 million H100-equivalents of AI compute that arrived globally from 2022 through June 2026, an estimated 19.8 million, or 72%, went to the United States.

China, counting domestic production, ranks second at 2.6 million H100-equivalents, or 10%, ahead of the European Union at 1.4 million, or 5%, and Malaysia at 0.7 million, or 2%. No other country received more than 2%. India and Japan each received about 1.6%, followed by Australia, the United Kingdom and Indonesia.

The American share rose across the period. It accounted for about 60 to 63 percent of each year's new AI compute from 2022 to 2024, then 72% in 2025 and 78% in the first half of 2026. Cumulative AI chip import value exceeds $400 billion. The estimates are built from customs records covering 16 major destinations, with Chinese domestic production estimated separately. The methodology uses a 2021 baseline to separate AI-attributable chips from general server and GPU imports, then nets out re-exports so each chip counts once at its final destination. H100-equivalents normalize different GPU models, from A100s to H200s and B200s, into a comparable unit of compute capacity.

For AI teams and companies that source compute internationally, the dataset provides a single consistent baseline for where capacity is concentrated and how quickly that concentration is changing. It is relevant to supply planning, to data residency and jurisdiction discussions, and to anyone tracking how export control policy maps onto physical chip flows. The figures are estimates rather than a registry.

They track arrivals recorded in customs data and include China's domestic output, so the numbers describe shipments rather than commissioned or operational capacity, and the method notes that some chips may still be in transit. The dataset also reveals a structural split. China received 18% of chip units but only 10% of compute, meaning the chips China gets deliver less computing power on average, which is consistent with export controls limiting access to the most advanced models.

Estimates for countries outside the US carry wider uncertainty, with China's range spanning 1.9 to 4.2 million H100-equivalents under alternative assumptions. Epoch AI publishes the dataset together with its methodology documentation.

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky