For each destination country and quarter since 2022, we estimate three cumulative indicators: AI-attributable, chip-only trade value from customs records; the number of AI chips that trade value corresponds to; and those chips’ compute in H100-equivalents (H100e).

We collect monthly customs statistics, starting in January 2021, for 16 destinations with their own customs records — the US, China, Hong Kong, the EU, South Korea, Malaysia, Singapore, the UK, Japan, Indonesia, India, Thailand, Mexico, and, via UN Comtrade, Saudi Arabia, Australia, and Canada. We focus on the four Harmonized System (HS) codes that cover AI server hardware: HS 8471.49 (complete computer systems), HS 8471.50 (servers), HS 8471.80 (other computer units), and HS 8473.30 (parts, including GPU boards and cards). All other destinations are inferred from exporters’ records: Taiwan’s exports and the re-exports and exports declared by several of those 16. We isolate the portion of each country’s imports that are AI-specific hardware by doing the following: subtracting a non-AI baseline anchored to 2021 trade levels (for all destinations, even those we inferred from exporters’ records), netting trade hubs (entrepôts) like Singapore and Hong Kong against their re-exports to make sure each chip is only counted once, and converting dollars to chip-only value using ratios measured from chip content. We then distribute the quarterly unit sales in our AI chip sales data across countries in proportion to the resulting trade value, and convert units to H100e. Each quarter’s arrivals are matched to the chips sold two months earlier, since we assume a two-month lag between a chip’s sale and its arrival.

Trade value is measured directly from customs records. Units and H100e inherit our global sales totals by construction — customs data determines only how they are distributed across countries. Trade value shares and unit shares can differ because chip mixes and prices differ across destinations (for example, we deflate China’s spending by an estimated price premium on diverted chips). The ranges given for H100e and units span the lowest 5th to the highest 95th percentile of estimates across alternative modeling assumptions.

Customs records do not identify chip models, so we estimate the split for trade value. When a designer or chip-type filter (or color-by) is active, each country’s trade value in each quarter is split across the chip types it imported, in proportion to their value at our estimated per-model prices — the prices set the mix and the customs data sets the total, so the split always adds up to the customs figures. Chips produced domestically (Huawei and Cambricon chips in China) cross no customs border and carry no trade value, though they are counted in units and H100e. The small amount of trade value in destination-quarters with no imported chips to split across (under $1M per quarter worldwide) is shown as “Unattributed”. Within each allocation pool every destination is assigned the same chip mix, so per-designer splits mostly reflect the global mix rather than destination-specific evidence (China, with its distinct eligibility rules and domestic designers, is the main exception).

Imports that we can attribute to AI-specific hardware are measured against a baseline of non-AI imports, which was built from 2021 trade, for every destination. The only purpose of the 2021 data is to set that baseline. The data starts in January 2022 because that’s when our AI chip sales data begins, and we build the estimates by spreading those sales across countries. Cumulative totals exclude chips and hardware from before 2022. The methodology is validated from 2023 Q3 onward, which is when AI shipments clearly separate from the baseline. As a result, the charts show this period by default.

“Arrival” is what we record when chips cross into the customs territory of their final destination country. Each quarter in the dataset refers to the calendar quarter of that customs territory’s arrival. For example, 2026 Q2 covers arrivals from April to June 2026. Chronologically, the sale happens an estimated two months before the arrival, and the deployment in a data center happens after. Therefore, a quarter’s arrivals are not the same as either the chips sold or the compute deployed in that quarter.

The EU-27 is a single customs territory and is treated as one destination, including on the map.

Some countries are not shown individually because our estimates for destinations with fewer than 1,000 H100e are too uncertain. The map leaves these countries blank, the table labels them as “Other”, and the chart’s destination filter does not include them. For both the map and the table, the threshold of 1,000 H100e applies to whatever the value shown is (either the selected quarter’s arrivals or the cumulative total through that quarter), with any chip designer or chip type filter already applied. These countries still count toward every total and appear in the downloadable data.

You can download both datasets in ZIP format, or grab individual CSVs from the Downloads section.
You can also explore the data using our interactive tools.
And you can view the data directly in table format.

Please send any feedback, questions, and comments to the data team at data@epoch.ai.