Epoch's chip imports explorer: the US holds 72% of AI compute, China 10%

Where is the world's AI compute? Until now the best answers came from company earnings calls and analyst notes. On October 27, 2026, Epoch AI published an AI chip imports explorer that works from the other end: customs records. Its headline estimate is that 27.4 million H100-equivalents of AI compute arrived in the world's countries between 2022 and June 2026, and 72% of it went to the United States.

The explorer, written by Lucio Melito and Isabel Juniewicz, is the closest thing yet to a map of the physical race. This piece covers what it measures, how it was built, and where the uncertainty sits. No named person is quoted in the post beyond its authors, so there is no pull quote here.

The country shares

Epoch counts compute in H100-equivalents, which converts every accelerator to the performance of one Nvidia H100 so that different chips can be added up. Through June 2026 the totals are:

• United States: 19.8 million H100e, or 72%.
• China, including domestic production: 2.6 million, or 10%.
• European Union: 1.4 million, or 5%.
• Malaysia: 0.7 million, or 2%.

No other country received more than 2%. Within that, the US lead is widening. Its share of each year's new AI compute was about 60 to 63% in 2022 through 2024, 72% in 2025 and 78% in the first half of 2026.

Why China looks smaller in compute than in chips

China accounts for 18% of the AI chips in Epoch's data but only 10% of compute. The difference is performance. A large part of its hardware is below the frontier, including chips built to comply with US export controls, such as Nvidia's H20, and domestic parts from Huawei and Cambricon, which Epoch says make up about 40% of China's total. Counting imports alone, China received 12% of chips and 6% of compute.

China's estimate is the least certain: Epoch's range under alternative assumptions is 1.9 to 4.2 million H100e. Domestic production is added from Epoch's separate AI Chip Sales dataset, because chips that never cross a border leave no customs record. That is a modeling step, not a measurement.

For the market-share side of the same story, see our report on Nvidia's falling China share and Huawei's lead among Chinese vendors.

How it was built

Epoch collected customs data from 16 major destinations and netted out transshipment hubs such as Singapore and Hong Kong. It estimated the AI share of server and parts trade from the remaining flows. For countries without detailed customs data, it estimated imports from what other countries report exporting. Each country's resulting share was then used to split Epoch's AI Chip Sales estimates by destination. Results are given in H100e, chip units and trade value.

The AI Chip Sales dataset behind that split covers Nvidia, Google's TPUs, Amazon's Trainium and Inferentia, AMD's Instinct line and Huawei's Ascend. Epoch's page for the dataset, updated October 21, 2026, says its estimates carry 90% confidence intervals typically about a factor of two wide, with Nvidia the most certain and Amazon Trainium the least. Power figures use thermal design power rather than measured draw. Microsoft's Maia, Meta's MTIA and Tesla's Dojo are excluded.

How much to trust it

Epoch tested the method against its own sales estimates. Cumulative spending derived from customs data lands within 4% of the sales-based figure for the same chips. Individual quarters differ by about 15% on average. Both series rose nearly tenfold from mid-2023 to mid-2026, from under $10 billion to almost $80 billion per quarter. Epoch is candid that the check is not fully independent: the chip-value estimates rely on its own prices and product mix.

The listed limitations are worth repeating. Customs codes mix AI and non-AI chips, and chips with other components. Customs data shows arrival, not whether a chip is switched on or has been resold. Countries such as the UAE and Vietnam are estimated only from partner-reported exports. Countries under 1,000 H100e are not shown. Chips made and used in the same country, such as in Taiwan, are not captured.

What it changes

The enforcement debate about chips has been run on anecdote. The US government's rules are summarized on the Bureau of Industry and Security site, which as of our reading listed guidance on licensing for entities headquartered in Country Group D:5 and an easing of controls for the UAE announced July 10, 2026. A country-by-quarter dataset lets outsiders test whether those rules show up in the flows. Our tracker of chip smuggling prosecutions covers the cases that sit outside the customs record entirely.

What to do with the numbers

For readers who cite figures from the explorer, three habits help. Quote the share rather than the raw count, since shares are less sensitive to the H100 conversion. Name the period, because the tool covers arrivals from 2022 through June 2026 and later quarters will move the totals. And say "estimated", because customs records measure what crossed a border, not what is running. Journalists and analysts who skip those steps will end up repeating a precision the data does not have. The explorer's own limitations section is the best guide to what can safely be said.

Our take

This is the most useful new dataset in the compute race this month, because it replaces vendor narratives with a method others can rerun. Cite the shares, flag the Chinese range, and treat the Gulf figures as placeholders. Watch two things: whether the US share keeps rising from 78%, and whether the next update shows Malaysia and the UAE growing as transshipment tightens.