The United States held 72.2% of the world’s AI chip compute that had arrived since 2022 as of the second quarter of 2026, up from 57.5% in the third quarter of 2023, according to Epoch AI’s customs-based data, by our calculation of the shares.
China’s share fell from 16.1% to 9.6% over the same period, and the US now holds 2.6 times the compute of every other country combined, by our arithmetic.
Short explainer · US Holds 72% of World AI Chip Compute
A vertical news Short narrated from this piece’s headline and opening line.
Business Pill · Sovereign AI
A one-minute explainer of sovereign AI: why countries want AI capacity under their own control. It teaches the general idea only and says nothing about any country’s figures in this piece.
The key insight: On Epoch AI’s customs-based estimates, the US held 72.2% of the world’s AI chip compute that had arrived since 2022 by mid-2026, 2.6 times every other country combined, while China’s share fell from 16.1% to 9.6%, by our arithmetic.
A chart of these shares, titled “Compute capacity (H100e)”, was posted on X on 27 October by @SamuelAlbanie with the comment “maybe the most important graph in the world”. The post had about 846,000 views when we read it on 30 October.
Its footnote matches word for word the note on Epoch AI’s Data on AI chip imports page: “Estimated cumulative chip arrivals since 2022, based on customs records. Arrivals mark entry into the final destination country, after sale and before deployment.” We downloaded Epoch’s country file and computed each destination’s share of the cumulative total ourselves; the figures below are those shares.
Share of cumulative AI chip compute arrivals since 2022, in H100-equivalents: United States 57.5% (2023 Q3) to 72.2% (2026 Q2), China 16.1% to 9.6%, rest of world 26.4% to 18.2% (including the EU, 6.3% to 5.1%, and Malaysia, 0.8% to 2.4%). Source: Epoch AI, Data on AI chip imports; shares computed by FourWeekMBA from Epoch’s country data.
How Epoch Counts the Chips
Epoch says it collects customs statistics from over a dozen national sources, isolates “the AI-attributable slice of server and GPU-board imports above a non-AI baseline built from 2021 trade”, and nets out re-exports “so each chip is counted once, in its final destination country”. Compute is expressed in H100-equivalents, or H100e.
Domestic chips are included for China: Epoch says “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”. In Epoch’s breakdown file, Huawei’s Ascend 910C is the largest chip type in China’s cumulative compute at 746,582 H100e, ahead of Nvidia’s H100 and H800 at 603,199 H100e.
Who Gained and Who Lost Share
Between the third quarter of 2023 and the second quarter of 2026, the US share rose 14.7 percentage points and China’s fell 6.5 points, by our arithmetic. The European Union, which Epoch treats as one customs territory, went from 6.3% to 5.1%, and the United Kingdom from 2.2% to 1.2%.
Most of the US gain came recently: its share rose 8.6 points in the four quarters since the second quarter of 2025, while China’s fell 5.4 points, by our arithmetic. Malaysia, which went from 0.8% to 2.4%, now holds about half the EU’s share; in 2023 the EU’s was 8.4 times Malaysia’s.
The pool itself grew 34.4 times, from 796,807 H100e to 27.4 million, by our arithmetic from Epoch’s totals. Epoch’s central estimate for the US is 19.8 million H100e, within a published range of 14.2 million to 25.0 million.
As we read Epoch’s series, most of the shift is recent: the US share rose 8.6 points in the four quarters to mid-2026, against 6.0 points in the seven quarters before, by our arithmetic.
By our arithmetic, the EU’s share was 8.4 times Malaysia’s in 2023 and is 2.1 times it now, as Malaysia went from 0.8% to 2.4%.
As we read the data, a falling share is not a falling stock: the world pool grew 34.4 times over the period, so China’s cumulative compute still rose in absolute terms, to Epoch’s central estimate of 2.6 million H100e.
Epoch AI, Data on AI chip imports
“Therefore, a quarter’s arrivals are not the same as either the chips sold or the compute deployed in that quarter.”
Three Implications
FOR READERS OF THE VIRAL CHART Its footnote is Epoch AI’s, and the shares can be recomputed from Epoch’s public country file: 72.2% for the US at mid-2026, by our arithmetic.
FOR EUROPE The EU and UK together held 6.3% of the cumulative compute at mid-2026, down from 8.5% in 2023, by our arithmetic from Epoch’s data.
FOR CHINA WATCHERS Epoch counts Huawei and Cambricon chips made in China, and Huawei’s Ascend 910C is the largest chip type in China’s total in Epoch’s breakdown file.
The analysis puts it this way: “Layer 9 answers to geopolitics, not to engineering.”
As we read it, Epoch’s data shows where the compute layer of that stack physically arrives, and by mid-2026 72.2% of it had arrived in the US, by our arithmetic. The analysis does not discuss Epoch’s figures.
What Is Not Established
Epoch says “a quarter’s arrivals are not the same as either the chips sold or the compute deployed in that quarter”: arrivals come after a sale and before chips are installed in a data centre. The figures are estimates from customs records, with Epoch’s range for China running from 1.9 million to 4.2 million H100e at the end of the period.
The shares count chips by where they arrive, not by who owns or uses them. Epoch adjusts China’s spending for “an estimated price premium on diverted chips”.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
On Epoch AI’s customs-based estimates, the US held 72.2% of the world’s AI chip compute that had arrived since 2022 by mid-2026, up from 57.5% in the third quarter of 2023, while China’s share fell to 9.6% and the EU’s to 5.1%. The US share is 2.6 times that of all other countries combined, by our arithmetic.
A note on sourcing. Compute figures are Epoch AI’s customs-based estimates from its Data on AI chip imports, accessed 30 October 2026. The country shares, differences and ratios are our arithmetic from Epoch’s cumulative H100e data.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
