The Geography of AI Chips
AI was supposed to make borders irrelevant. The model training clusters would sit anywhere with cheap power, the code would travel at light speed, and the physical location of intelligence would become, economically speaking, a rounding error. That was the theory.
The practice looks different. Epoch AI published a report this week estimating that between 290,000 and 1.6 million H100-equivalents of compute were smuggled into China by the end of 2025, with a median estimate of 660,000 Epoch AI. That is roughly three percent of the world's total AI compute stockpile, comparable to what xAI had at the time, sitting in gray-market data centers that do not appear on any export declaration.
The number is jaw-dropping. The method is more interesting.
Epoch AI built two independent estimates and checked them against each other. The first traces diversion — how chips leave legitimate supply chains through pass-through entities, staged dummy servers, and repackaged shipments. The second traces resale — the active marketplace inside China for banned chips, documented through investigative reporting, vendor listings, and transaction records. A smuggled chip has to clear both stages. Epoch AI found substantial overlap between the two estimates, which is the statistical equivalent of two witnesses agreeing on the same alibi without knowing what the other said.
The underlying evidence is not speculative. The DOJ indicted Supermicro cofounder Yih-Shyan Liaw and two associates in March for allegedly diverting $2.5 billion in Nvidia servers to China between 2024 and 2025 DOJ. The indictment describes a scheme that ran through Taiwan, Malaysia, and Singapore, with dummy servers staged to fool compliance inspectors and identifying markings removed before final delivery. In one three-week window alone — late April to mid-May 2025 — roughly $510 million in servers were allegedly diverted through this route. The Financial Times reported separately that more than $1 billion in Nvidia chips moved through gray markets in the three months after the H20 export ban Reuters/FT. Nvidia itself disclosed $5.5 billion in charges related to that same ban NPR.
The H20 is worth dwelling on. Nvidia designed it specifically to comply with earlier export controls — a downgraded version of its flagship AI chip that would pass regulatory scrutiny and still sell in China. It was the compliance chip, the one that let everyone claim the rules were working. Then in April 2025, Commerce restricted it anyway. The gray market responded within weeks.
This is the pattern. Every round of export controls creates a profit opportunity. The bigger the restriction, the wider the margin between what chips cost in Shenzhen and what they sell for in a Malaysian data center with a shell-company end user. Epoch AI’s Monte Carlo simulation essentially models this as an economics problem: given the documented cases, the estimated detection rate, and the spread between restricted and unrestricted market prices, what total volume is consistent with the evidence? The answer is 660,000 H100-equivalents, with a wide confidence interval — but the interval crosses every other independent estimate, including the CNAS figure of roughly 140,000 chips for 2024 alone CNAS.
The wider confidence interval is real and should not be collapsed. The 660,000 median is higher than most prior estimates. Epoch AI notes that its upper bound — 1.6 million — would mean the majority of China’s AI compute was smuggled, which is a striking claim that requires more scrutiny than a Substack post can give it. The detection-rate assumption, which Epoch AI pegs at a median of 24.5 percent with a wide range, is the model’s most speculative input. If most smuggling goes undetected, the true number is certainly higher than the documented allegations. If detection is higher than assumed, the lower bound may be closer to reality. Nobody knows.
What is not in dispute is that the mechanism works in both directions. US export controls created the price gap that funds the smuggling industry. The CNAS report noted that the profits from three reported smuggling cases exceeded the entire annual budget of the Bureau of Industry and Security — the agency responsible for enforcing these controls. The regulatory infrastructure designed to contain the flow has less funding than the incentive to circumvent it.
The methods documented by investigative reporting include falsified end-user certificates, repackaging with false descriptions, shell companies in Singapore and Malaysia, and the classic — relabeling shipments as tea or toys. This is not sophisticated tradecraft. It is basic logistics with a profit margin built in.
The historical parallel is the COCOM regime that governed Western technology exports to the Soviet bloc for four decades. The US and its allies restricted computing equipment, aerospace technology, and precision manufacturing to the USSR through a formal multilateral framework. The Soviets acquired much of it anyway — through front companies, sympathetic intermediaries, and legitimate trade channels that happened to have Soviet end users several steps removed. COCOM collapsed in 1994, by which point the strategic technology gap the US sought to maintain had largely been crossed on schedule, if not ahead of it. The technology reached its destination; it just took longer and cost more.
AI is repeating the pattern at significantly higher speed. The chip restrictions began in 2022. By 2024, investigative reporters were documenting hundreds of active distributors on Chinese online marketplaces. By 2025, the Supermicro indictment was describing a professional operation running billions of dollars through Southeast Asia. The infrastructure of circumvention built up faster than the infrastructure of enforcement.
For the AI industry, the question is not whether China has smuggled chips. The evidence says it has. The question is what the aggregate volume means for competitive timelines. If China’s effective compute stock is significantly larger than official figures suggest, the assumption that US labs maintain a durable training compute advantage — an assumption embedded in billion-dollar investment decisions and strategic planning — may need recalibration. This does not mean China has closed the gap. It means the gap may be smaller than the public record implies, and that the export control regime that was supposed to preserve it has, in the short run, made the problem worse.
The policy implication is uncomfortable: each cycle of restriction reinforces the smuggling infrastructure, which makes the next cycle of restriction harder to enforce, which enlarges the gray market. This is the opposite of what the controls are supposed to do. The Biden administration restricted the H20 in April 2025. The gray market responded with $1 billion in documented flows within three months. The Trump administration lifted the restriction in July, then China reimposed its own import limits. The result is not clear sailing for US chips — it is a market with no stable rules and maximum uncertainty, which is exactly the condition that maximizes smuggling margins.
Epoch AI’s estimate is not confirmed fact. It is the best attempt yet to put a number on something that resists precise measurement. But the two-model cross-validation approach — diversion and resale agreeing on the same order of magnitude — is more rigorous than anything published before. And the fact that every other independent estimate falls within its confidence interval suggests the number is directionally correct, even if the exact figure is uncertain.
What is certain is that AI’s physical infrastructure — the chips, the power, the cooling — is more geographically concentrated and more strategically contested than the technology’s borderless reputation suggests. The code may be everywhere. The compute is somewhere. And the somewhere matters more than the industry wanted to believe.
Sources: Epoch AI Substack (Apr 29, 2026); DOJ indictment press release (Mar 20, 2026); CNAS report by Grunewald and Fist (Jun 2025); Reuters/FT (Jul 2025); NPR (Apr 2025).