Epoch AI 推出 AI Chip Imports explorer 估算各国 AI 算力进口

Today, we’re launching the AI Chip Imports explorer, which estimates how much AI compute has arrived in different countries, quarter by quarter, since 2022. It is the newest piece in our project to map the AI chip supply chain, which already provides data on the components that go into AI chips, how many chips are sold, and the companies that own and use them. It also complements our AI data centers hub, which details individual facilities. This newest explorer estimates country totals, including chips in facilities not currently included in the AI data centers hub. Together, these datasets aim to give the most complete public picture of how much AI compute exists, who controls it, and where it is.

Where AI compute is located matters

AI is increasingly important to economic competitiveness and military power. A country that hosts large AI data centers can train and serve frontier models within its borders, and gains significant influence over who uses those models and for what purposes. That is why US export controls aim to keep advanced AI chips out of China, and why governments from the Gulf to Europe and India are competing to host AI data centers. Yet chip designers and cloud companies rarely disclose which countries their chips are deployed in.

We build the AI Chip Imports explorer from customs records, which we use to estimate the share of server and parts trade related to AI. Almost every advanced AI chip goes through advanced packaging in Taiwan, even if the components come from elsewhere, and nearly every chip crosses at least one border before reaching the data center where it runs. Most countries publish records of those crossings monthly. We collect customs data from 16 major destinations, net out chips that only pass through hubs like Singapore and Hong Kong, and estimate how much of the remaining server and parts trade is AI chips. For the destinations where we have their customs data, we rely primarily on their own import records, and we estimate remaining countries using export data from these records. We then use each country’s share to split our global chip sales estimates by destination. Chips made in China for Chinese data centers never cross a border, so we add China’s domestic production from our AI Chip Sales data. The explorer reports arrivals in H100-equivalents (H100e), chip units, and trade value.

The AI Chip Imports explorer is intended as a resource for researchers, policymakers, and anyone tracking the geography of AI compute. Below, we highlight our initial findings.

The US received nearly three-quarters of the world’s AI compute

Of the 27.4 million H100e of AI compute that arrived from 2022 through June 2026, we estimate that 19.8 million, or 72%, went to the United States. China (including domestic production) comes second at 2.6 million (10%), followed by the EU at 1.4 million (5%) and Malaysia at 0.7 million (2%). No other country received more than 2%.

The US lead has widened. It received about 60–63% of each year’s new AI compute from 2022 to 2024, then 72% in 2025 and 78% in the first half of 2026.

Including domestically produced chips, China accounts for 18% of the AI chips in our data but only 10% of the compute, at 2.6 million H100e. Much of its hardware is lower-performance. This includes chips designed to comply with US export controls, like Nvidia’s H20, and domestic chips from Huawei and Cambricon, which make up about 40% of China’s total. Counting imports alone, we estimate China received 12% of the chips and 6% of the compute.

This estimate is uncertain. Under alternative assumptions about what the trade contains, it ranges from 1.9 to 4.2 million H100e. It includes chips brought into China in violation of US export controls to the extent they show up in China’s own customs records. Our piece on chip smuggling via Malaysia used this data to document one such route. Shipments that no one declares, however, leave no trace.

Customs data closely match our chip sales estimates

Our AI Chip Sales estimates are built from company disclosures and analyst estimates. Customs records come from governments, so they offer a separate check. It isn’t fully independent: estimating how much of each server’s value is the chip itself uses our chip prices and product mix. But except for China, the trade values themselves come only from customs data, and nothing forces the two to agree.

Both series grew by almost an order of magnitude from mid-2023 to mid-2026, from under $10 billion a quarter to almost $80 billion. Over that period, cumulative customs-derived chip spending comes within 4% of the value of the corresponding chips in our sales data. Individual quarters are noisier, about 15% off on average. Customs values ran mostly below our sales estimates through 2024 and mostly above them since mid-2025. That pattern could reflect changes in how long chips take to arrive after they’re sold, or chip prices that change over a model’s life while our estimates hold them fixed.

Limitations

• Customs codes generally mix AI and non-AI chips, as well as chips and non-chip components. The assumptions that we make to estimate the AI chip share of customs data are documented in the full methodology.
• Customs data shows when chips arrive, not when they end up running. Some chips may not be operational yet, and some may later move to another country.
• For countries without detailed customs data, such as the UAE and Vietnam, we estimate figures only from what other countries report sending. Countries with fewer than 1,000 H100e aren’t shown individually because those estimates are too uncertain.
• Chips made and used in the same country, as in Taiwan, never cross a border and aren’t captured. For China, we add domestic production from our AI Chip Sales data and assume it all remains within China. We do not have an estimate of chips retained by Taiwan.

To learn more, visit the AI Chip Imports explorer to find the methodology, full dataset, interactive visualizations, and a map of chip destinations.

Thanks to Daniel Carey, Erich Grunewald, Amelia Michael, Konstantin Pilz and Josh You for their feedback.