Buildout· Epoch AI data of Nov 12, 2026
AI compute by country
Where the AI chips went. Of all the AI computing power that arrived anywhere in the world from January 2022 to June 2026, an estimated 72.0% arrived in the United States, 9.9% in China and 5.1% in the European Union. The estimates are Epoch AI’s, built from customs records and counted in H100-equivalents (27.4 million worldwide). They measure chips crossing a border, not chips installed or running.
What the estimates show
- Most of it went to one country. The United States received an estimated 19.7 million H100-equivalents, 72.0% of the world total. Across the alternative assumptions Epoch tested, its figure runs from 14.2 million to 25.0 million, and Epoch reports that it is the largest destination under every one of them.
- China has a larger share of the chips than of the compute. It received 9.9% of the compute (2.71 million H100-equivalents; range 1.96 million to 4.26 million) but 18.2% of the chips. Epoch’s explanation: the chips counted for China, which include models sold only there such as Nvidia’s H20 and chips made in China by Huawei and Cambricon, deliver less compute each on average.
- The European Union is third, at 5.1% (1.39 million H100-equivalents; range 1.03 million to 2.57 million). Its 27 member states are one customs territory, so the data cannot split it by country.
- The share of the United States has grown. Of everything that had arrived by the end of 2024 Q2, the United States had received 61.5%; by the end of 2026 Q2, 72.0%.
- Outside the United States, the figures are much less certain. South Korea’s estimate is 105k H100-equivalents, but its range runs from 37.2k to 945k. Read the order of the smaller destinations as approximate.
Arrivals over time
Running totals since January 2022, in H100-equivalents. Each point adds the chips that arrived in that quarter to everything that arrived before it; nothing is taken off for chips later retired or moved to another country. Epoch treats its method as validated from the third quarter of 2023, when AI shipments stand clear of ordinary trade; the earlier quarters are small and less certain.
Exact values by quarter
Cumulative H100-equivalents as Epoch’s file prints them. The decimals are the file’s, not a sign of precision: see the ranges in the table below.
| Through | United States | China | European Union | Malaysia | India | Japan | All other destinations | World |
|---|---|---|---|---|---|---|---|---|
| 2022 Q1 | 12,334.3 | 3,035.4 | 1,986.4 | 0 | 0 | 33.2 | 4,905.5 | 22,294.8 |
| 2022 Q2 | 48,565.4 | 5,976.2 | 2,467.9 | 0 | 570.7 | 1,056.3 | 32,070.4 | 90,706.9 |
| 2022 Q3 | 90,082.1 | 15,452 | 8,151.3 | 253.9 | 1,342.9 | 3,045.1 | 44,741.5 | 163,068.8 |
| 2022 Q4 | 145,528.9 | 18,408.7 | 11,770.1 | 1,221.4 | 4,607.3 | 6,449.1 | 54,837.2 | 242,822.7 |
| 2023 Q1 | 201,439.6 | 29,849.3 | 16,475.6 | 1,284.7 | 9,137.7 | 7,001 | 77,244 | 342,431.9 |
| 2023 Q2 | 290,449.2 | 54,315.7 | 27,199.7 | 5,212.8 | 10,006 | 10,794.7 | 104,075.5 | 502,053.6 |
| 2023 Q3 | 458,171.1 | 128,129.3 | 50,270.3 | 5,947.2 | 11,071.8 | 18,455.7 | 124,762 | 796,807.4 |
| 2023 Q4 | 769,974.8 | 179,243.6 | 85,074.1 | 15,466.6 | 12,500.7 | 21,500.3 | 154,003.1 | 1,237,763.2 |
| 2024 Q1 | 1,157,331.2 | 230,183.5 | 152,511.9 | 73,782.9 | 20,280.3 | 38,384.4 | 211,675.7 | 1,884,149.9 |
| 2024 Q2 | 1,713,666.4 | 353,451.4 | 236,309.2 | 77,751.3 | 34,176 | 85,035.5 | 288,210.9 | 2,788,600.7 |
| 2024 Q3 | 2,461,041.4 | 524,430.5 | 287,151.3 | 114,491.8 | 39,951.1 | 113,365.4 | 377,540.5 | 3,917,972 |
| 2024 Q4 | 3,359,164.8 | 797,286.5 | 327,571.1 | 148,684.5 | 60,821.4 | 145,335.4 | 598,185.5 | 5,437,049.2 |
| 2025 Q1 | 4,800,155.5 | 1,162,895.7 | 369,525.4 | 206,187.4 | 96,099.9 | 185,898.2 | 759,938.5 | 7,580,700.6 |
| 2025 Q2 | 6,491,838.6 | 1,571,199.4 | 476,346.9 | 358,508.9 | 123,669.6 | 258,315.3 | 970,532.6 | 10,250,411.3 |
| 2025 Q3 | 8,896,757.9 | 1,825,826.2 | 624,713.6 | 427,408.2 | 150,705.4 | 319,830.2 | 1,178,064 | 13,423,305.5 |
| 2025 Q4 | 11,917,459.1 | 2,091,784.9 | 926,855.1 | 498,739.5 | 204,162.9 | 335,588 | 1,386,008.2 | 17,360,597.7 |
| 2026 Q1 | 15,505,932.2 | 2,427,524.1 | 1,172,533.2 | 535,129 | 265,239.8 | 371,661.7 | 1,683,369.8 | 21,961,389.8 |
| 2026 Q2 | 19,711,104.2 | 2,706,192.8 | 1,385,408.7 | 661,678 | 441,776.3 | 430,892 | 2,053,578.5 | 27,390,630.5 |
Every destination Epoch lists
Estimated arrivals from January 2022 to June 2026, largest first. Compute is in H100-equivalents (H100e). The range is the lowest and the highest figure across the alternative assumptions Epoch tested; it is not a margin of error with a known probability, and the estimate need not sit in its middle. Chips is the estimated number of AI chips (Epoch’s file gives a range for it too). Trade value is the chip value Epoch attributes to AI in customs records, in US dollars. Destinations under 1,000 H100e are added into one row, as on Epoch’s own page.
| Destination | Share of compute | Compute, H100e | Range, H100e | Chips | Trade value |
|---|---|---|---|---|---|
| United States | 72.0% | 19.7M | 14.2M–25.0M | 16.1M | $288B |
| China | 9.9% | 2.71M | 1.96M–4.26M | 4.40M | $50.5B |
| European Union | 5.1% | 1.39M | 1.03M–2.57M | 1.13M | $23.2B |
| Malaysia | 2.4% | 662k | 492k–1.49M | 396k | $11.8B |
| India | 1.6% | 442k | 295k–721k | 267k | $7.58B |
| Japan | 1.6% | 431k | 188k–558k | 402k | $7.10B |
| Australia | 1.4% | 383k | 201k–570k | 210k | $6.91B |
| United Kingdom | 1.2% | 339k | 222k–496k | 224k | $6.28B |
| Indonesia | 0.80% | 218k | 181k–330k | 115k | $3.96B |
| Canada | 0.53% | 146k | 79.2k–325k | 108k | $2.66B |
| Norway | 0.47% | 130k | 102k–190k | 79.2k | $2.27B |
| Thailand | 0.45% | 124k | 92.5k–256k | 93.8k | $2.19B |
| South Korea | 0.38% | 105k | 37.2k–945k | 59.8k | $1.89B |
| Iceland | 0.35% | 95.7k | 69.2k–140k | 61.1k | $1.62B |
| United Arab Emirates | 0.33% | 90.8k | 73.8k–150k | 90.1k | $1.95B |
| Singapore | 0.19% | 51.8k | 28.0k–65.9k | 73.5k | $939M |
| South Africa | 0.18% | 50.1k | 40.4k–73.1k | 34.1k | $955M |
| Brazil | 0.17% | 46.9k | 31.2k–63.5k | 29.8k | $777M |
| Israel | 0.16% | 42.6k | 34.9k–81.7k | 37.8k | $860M |
| Switzerland | 0.14% | 37.4k | 30.8k–61.7k | 24.9k | $677M |
| New Zealand | 0.08% | 22.5k | 15.4k–32.3k | 15.8k | $401M |
| Chile | 0.07% | 19.0k | 5.82k–24.9k | 17.9k | $298M |
| Mexico | 0.06% | 15.4k | 5.90k–24.7k | 48.9k | $553M |
| Saudi Arabia | 0.05% | 13.5k | 11.2k–74.1k | 7.02k | $252M |
| Philippines | 0.05% | 13.2k | 9.55k–22.1k | 6.03k | $244M |
| Turkey | 0.05% | 12.8k | 10.6k–38.5k | 8.93k | $262M |
| Qatar | 0.02% | 6.82k | 4.47k–9.51k | 9.08k | $171M |
| Oman | 0.02% | 6.34k | 5.19k–11.1k | 5.51k | $117M |
| Kazakhstan | 0.02% | 5.55k | 4.62k–8.97k | 4.85k | $123M |
| Paraguay | 0.02% | 4.56k | 3.82k–12.5k | 3.32k | $85.0M |
| Colombia | 0.01% | 3.95k | 1.13k–5.33k | 4.56k | $76.7M |
| Argentina | 0.01% | 3.76k | 2.10k–6.98k | 6.20k | $110M |
| Ethiopia | 0.01% | 3.75k | 3.02k–11.5k | 3.81k | $69.5M |
| Uzbekistan | 0.01% | 3.54k | 2.84k–20.6k | 5.52k | $122M |
| Armenia | 0.01% | 3.22k | 2.41k–4.74k | 1.88k | $56.5M |
| Angola | 0.01% | 2.78k | 2.26k–5.71k | 2.35k | $50.5M |
| Algeria | 0.01% | 2.46k | 1.57k–3.47k | 2.37k | $44.5M |
| Ukraine | 0.01% | 2.28k | 1.29k–4.17k | 1.88k | $39.6M |
| Morocco | 0.01% | 2.24k | 1.85k–9.08k | 1.58k | $42.8M |
| Bangladesh | 0.01% | 1.76k | 1.43k–3.55k | 2.09k | $42.8M |
| Peru | 0.01% | 1.71k | 518–3.11k | 1.55k | $39.7M |
| Iraq | 0.01% | 1.69k | 1.22k–2.34k | 2.22k | $38.1M |
| Serbia | 0.01% | 1.52k | 745–2.67k | 847 | $26.9M |
| Kenya | 0.01% | 1.44k | 1.15k–2.10k | 925 | $25.6M |
| Belarus | 0.01% | 1.40k | 677–7.22k | 2.70k | $59.8M |
| Djibouti | <0.01% | 1.27k | 1.01k–3.19k | 1.00k | $20.5M |
| Bahrain | <0.01% | 1.23k | 609–1.73k | 1.46k | $30.8M |
| Mongolia | <0.01% | 1.17k | 964–1.82k | 895 | $23.5M |
| Cayman Islands | <0.01% | 1.17k | 907–2.17k | 1.21k | $23.2M |
| North Macedonia | <0.01% | 1.16k | 791–1.75k | 588 | $21.7M |
| Uruguay | <0.01% | 1.16k | 816–2.11k | 763 | $22.9M |
| Kyrgyzstan | <0.01% | 1.03k | 846–2.83k | 959 | $24.3M |
| Ecuador | <0.01% | 1.01k | 813–3.53k | 1.51k | $24.7M |
| 167 other destinationseach under 1,000 H100e | 0.09% | 26.0k | Not given | 26.5k | $589M |
| Worldall 220 destinations | 100% | 27.4M | Not given | 24.2M | $427B |
How to read these figures
What an arrival is. Epoch records an arrival when chips cross into the customs territory of the country they end up in. That is after the sale (Epoch assumes two months after, except Nvidia’s H20, which it counts in the quarter it was sold) and before the chips are installed in a data center. So a quarter’s arrivals are neither the chips sold nor the compute switched on in that quarter. Chips made in China by Huawei and Cambricon cross no border; Epoch still counts them for China, in chips and in compute, with no trade value.
How the estimate is built. No customs code identifies AI chips. Epoch takes monthly customs records for the four product codes that cover servers and their parts, removes a baseline of ordinary trade set from 2021 levels, and nets out re-exports through trading hubs such as Singapore and Hong Kong so that each chip counts once, in its final destination. It then shares out its own estimate of worldwide chip sales across countries in proportion to that trade value. Sixteen destinations are measured from their own customs records; every other one is inferred from what exporters recorded. Epoch’s methodology has every step.
What an H100-equivalent is. A unit of computing power: the number of Nvidia H100 chips that would match it on paper. Epoch divides the peak number of 8-bit operations a chip can perform by the H100’s and multiplies by the number of chips, so a faster chip counts as more than one and a slower chip as less. It is a yardstick taken from specifications, which Epoch calls an imperfect proxy, not a measure of work done (Epoch’s definition).
What the range is. Epoch repeats its estimate under 144 alternative sets of assumptions about what the trade contains, the baseline and the mix of server products. The low figure is the lowest 5th percentile and the high figure the highest 95th percentile across those sets. In Epoch’s words: “We don’t assign probabilities to the alternatives, so the structural range is not a 90% interval, and it only covers the alternatives we tested.” The ranges of different countries cannot be added together, and wrong customs declarations are not reflected in them.
What the data cannot tell you.
- Whether a chip is installed or running. Arrivals stand in for deployment; they are not a count of working capacity.
- Where a chip is today. Nothing is deducted for chips later moved abroad or retired, and hardware routed around export controls can be recorded under another destination.
- Anything before 2022. The totals leave out chips that arrived earlier.
- Which chips a given country bought. Within each group of eligible countries Epoch gives every destination the same mix of chip models, so a country’s chips and compute mostly follow the global mix (China is the main exception).
- Who owns or uses the chips. The data is by country, not by company.
Why these figures are not added to the rest of this site. The AI lab compute page shows Epoch’s estimates of the compute each lab used at a year end, from a different Epoch dataset. The buildout page lists the power capacity, in megawatts, of named data-center sites. This page counts chips arriving in a country over 4.5 years. The three differ in what they count, when, and for whom, and one chip can appear in all of them. So this site never adds a country’s arrivals to a lab’s H100-equivalents or to a site’s megawatts, and never converts one into another.
Source, licence and files
- Source
- Epoch AI, Data on AI chip imports (documentation). Epoch builds it from the customs statistics of the national agencies it credits there.
- Licence
- CC BY 4.0. The figures are Epoch’s, unchanged. The shares, the combined rows, the rounding for display and the wording are this site’s. To reuse this page, credit: AI Compute Tracker (aicomputetracker.com), CC BY 4.0; includes data from Epoch AI (epoch.ai), CC BY 4.0.
- Release
- Updated by Epoch on Nov 12, 2026; first released Oct 27, 2026 (Epoch’s changelog).: Nvidia’s H20 is now assumed to arrive in the quarter it was sold; other chips still arrive two months after sale. About 216,000 H20s in China move from 2025 to 2024, China’s share of cumulative arrivals through June 2026 rises from 9.6% to 9.9% and the United States’ falls from 72.2% to 72.0%.: First release: quarterly estimates since 2022 of AI chip arrivals by destination, in trade dollars, chips and H100-equivalents.
- Retrieved
- (UTC).
- Country file
- ai_chip_trade_by_country.csv, 979 KiB, 3,960 rows. Every figure on this page comes from it.SHA-256
4df64317375d4557f84923f96ef62a43b660b623a4ae3c2091a4fd159688139b - Chip-type file
- ai_chip_trade_by_country_breakdown.csv, 1,883 KiB, 13,758 rows. Used only to check that its rows add up to the country file’s totals; nothing shown here is read from it.SHA-256
fa4c0d7564a8b4a9d078896e4ff0615fb62da5bdab1ae2a5f7164b33447ff8c7 - Updates
- Epoch states no update schedule for this dataset. This page changes only when the two files are downloaded again by hand, and the dates and hashes above change with them.
- Cite the source as
- Epoch AI, ‘Data on AI chip imports’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/data/ai-chip-imports’ [online resource]. Accessed Nov 18, 2026.