China is 2.7 times more exposed to a chip supply shock than the US, Epoch finds

Who is most exposed to a chip supply shock, China or the United States? Epoch AI's October 6, 2026 study, written by researcher Daniel Carey, says China, by a wide margin. Using 2022 trade data, every $1,000 of final demand in China generated $15.2 in revenue for semiconductor producers worldwide, compared with $5.7 for the same spending in America. That is 2.7 times the exposure.

For a race publication this is the physical layer of the contest. Training the next frontier model needs advanced chips, and the countries that make or buy them set the pace. The study asks what a cutoff would cost each side.

How the study works

Carey separated semiconductors from the broader electronics category in the OECD's Inter-Country Input-Output tables for 2016 to 2022, adding detail for processors and memory. He then ran a general-equilibrium trade model calibrated to 2022 and applied shocks, assuming an adjustment period of two to three years.

The exposure gap differs by supplier:

• Taiwan: China is 4.7 times more exposed than the US.
• South Korea: 4.0 times.
• Japan: 3.6 times.

The modeled shock is where the numbers get dramatic. In the full decoupling scenario, the price of advanced processors in China rises 17-fold, against roughly 20% in the US. China loses about 3% of gross national expenditure, or about 1% if the decoupling is limited to semiconductors. The US loses 0.6%, or nothing in the limited case.

Those results are model outputs, not observations. They show a direction and a rough size. They are not a prediction of what would happen.

Why the asymmetry is plausible

Other Epoch work backs the shape of this result. Its September analysis of Huawei's roadmap estimates Huawei will make about 1.5 million Ascend units in 2026 against 5.9 million for Nvidia, and less than 4% of Nvidia's compute output. Epoch says Huawei's chips trail Nvidia's by three to four years through 2030, and that Nvidia holds a roughly 2 times advantage in transistor density that will persist.

The dependency is concrete. According to that report, Huawei's leading chip last year, the Ascend 910C, used logic dies fabricated by TSMC and memory from Samsung and SK Hynix. Epoch writes that across 2024 and 2025 Huawei stockpiled and smuggled large amounts of logic wafers and high-bandwidth memory. When the stockpile runs out, it says, Chinese AI chip volumes fall back to what the domestic memory maker CXMT can produce.

Smuggling is part of that picture. Another Epoch data insight, from September 17, found Chinese customs records showed $3.75 billion in servers imported from Malaysia between April 2024 and June 2025, while Malaysia reported exporting only $0.6 billion to China. Epoch's reading is that this is "consistent with more than $3 billion of chips smuggled," about 150,000 H100-equivalent GPUs, and it is careful to say customs records are not proof of diversion. Our own coverage of a chip smuggling prosecution shows that physical routes exist.

Where the memory gap bites

The Huawei report adds a second, more concrete number. Epoch estimates the high-bandwidth memory supply gap between Nvidia and Huawei narrows from 73 times in 2026 to 11 times in 2028, and that Huawei's domestic memory alone would support roughly 1% of Nvidia's compute through 2028. Even if Huawei obtained ten times as much smuggled memory as Epoch assumes, the report estimates its 2028 output would reach only 16% of Nvidia's production that year. In other words, the exposure the new study measures across the whole economy shows up in the AI sector as a memory bottleneck. That is a narrower, more tractable target for policy than "semiconductors" in general, and it is the part of the supply chain where controls and enforcement have the most visible effect.

What the study cannot tell you

Epoch lists its own limits, and they are significant.

The data is old. 2022 predates the surge in AI server investment, which is the very trade this debate is about. Memory and accelerator flows have changed since.

Aggregation can mislead. Industry-level data may overstate China's exposure by mixing export-oriented assemblers, which consume imported chips and re-export products, with firms serving the Chinese market.

A model is not a war game. The shock assumes a clean decoupling with a fixed adjustment period. Real disruptions arrive unevenly, and stockpiles and smuggling blunt them.

The paper also does not address what the US pays in other ways, such as dependence on Taiwan for its own supply. The 0.6% figure is a trade-model loss, not a measure of military or political risk. We did not interview the author or anyone else for this piece.

What it means for policy

The study sits behind a live debate. If China is far more exposed than the US, export controls are a stronger lever than their critics say. If the exposure shrinks as domestic production grows, the lever weakens over time, and that clock matters more than the headline ratio.

It also fits a pattern in our reporting on how much compute a frontier run costs: the constraint on the race is often the supply chain, not the algorithm.

Our position: use the 2.7 times figure to argue about direction, and do not use it to argue about size. The most useful next step would be the same model rerun on 2024 and 2025 data, once the OECD tables extend that far. Until then, treat any politician quoting the 17-fold number as quoting a scenario, not a forecast.