The DOJ charged a co-founder of Super Micro Computer with diverting two and a half billion dollars in Nvidia AI servers to China through a Southeast Asian pass-through company. Workers used hair dryers to peel serial number stickers off real servers and press them onto thousands of dummy shells staged for compliance auditors.
The Department of Justice unsealed an indictment on March 19 charging three people with conspiring to divert cutting-edge AI hardware to China in violation of U.S. export controls. The central figure is Yih-Shyan Liaw, a co-founder of Super Micro Computer — one of the largest server manufacturers in the world — who served as its Senior Vice President of Business Development and sat on its board. Two co-defendants managed the logistics: a sales manager in Supermicro’s Taiwan office and a third-party broker described as a fixer.
The indictment alleges that Liaw caused two and a half billion dollars in Supermicro server sales to be routed through a pass-through company in Southeast Asia. Of that total, five hundred and ten million dollars worth of servers containing restricted Nvidia H100, H200, and B200 chips were confirmed diverted to end users in China. The scheme operated between 2024 and 2025. The charges carry a combined maximum of fifty years in prison.
The method is what makes the case instructive. The pass-through company compiled fake paperwork identifying itself as the end user. A logistics firm repackaged the servers into unmarked boxes for forwarding to China. Thousands of dummy servers — non-functional shells — were staged at the company’s storage facilities to satisfy compliance audits. And when U.S. export control officers visited to verify serial numbers against shipping records, the serials matched — because surveillance footage later revealed workers using hair dryers to peel serial number stickers off genuine servers and press them onto the dummy shells.
That detail — the hair dryer — is the image that encapsulates the entire enforcement challenge. The legal supply chain has serial numbers, compliance officers, audit procedures, and end-user certificates. The physical supply chain has heat guns and unmarked boxes.
This is not isolated. Three months earlier, the DOJ announced Operation Gatekeeper, dismantling a separate smuggling network that moved at least one hundred and sixty million dollars in Nvidia H100 and H200 chips to mainland China and Hong Kong. The operator relabeled restricted chips with a nonexistent fake brand and classified them as generic computer parts. He pleaded guilty in October 2025. Congress approved a twenty-three percent increase in the Bureau of Industry and Security’s enforcement budget for fiscal year 2026, with earmarked funding specifically for semiconductor cases.
Two major enforcement actions in three months. Each one dismantled. Each one had been operating for years before detection.
The Prohibition Premium
Every restriction on a high-demand good creates a price premium for circumvention. This is not a policy critique — it is an economic identity. Cocaine is more expensive than coffee because trafficking cocaine requires evading enforcement. The premium funds the evasion. The evasion funds the infrastructure. The infrastructure creates a durable alternative supply chain that persists even as individual operators are caught.
The same pattern has played out with every export control regime in history. In the 1990s, Philip Zimmermann published PGP encryption software as a printed book to circumvent ITAR restrictions on cryptographic export. The munitions classification made the algorithm more valuable to adversaries, not less — the classification was the signal that the technology mattered. When the U.S. eventually decontrolled strong encryption in 2000, it was partly because the black market had made the restriction unenforceable and partly because American companies were losing commercial ground to foreign competitors who faced no such constraints.
The AI chip export controls face the same structural dynamic at a larger scale. The restricted chips — H100, H200, B200 — are the most valuable computational assets on Earth. A single B200 GPU generates more revenue per hour in an AI training cluster than most capital equipment generates in a week. The demand-side incentive is not abstract. It is a specific dollar figure, and it is enormous.
The Visibility Paradox
Export controls assume two things: that the legal supply chain has defined borders, and that enforcement can monitor what crosses them. When circumvention is small and amateur, both assumptions hold. Customs catches mislabeled shipments. End-user certificates get verified. Serial numbers match.
When circumvention scales — when a co-founder of the manufacturer is directing the diversion, when thousands of dummy servers are manufactured specifically to defeat audits, when logistics firms specialize in repackaging — the enforcement surface expands faster than enforcement capacity. Each successful prosecution reveals infrastructure that operated undetected for years. The detected cases are by definition the ones that failed. The undetected ones are the ones that worked.
This creates a paradox. The harder you enforce, the more sophisticated the evasion becomes. The more sophisticated the evasion, the less visibility you have into the actual flow of hardware. The export control regime produces two outputs simultaneously: a public record of enforcement actions that demonstrate resolve, and a private reality of invisible channels that grow more professional with each prosecution.
The journal has tracked what happens on the demand side. The Endosymbiont documented how forced containment drove China’s AI ecosystem toward functional independence — building on domestic chips, optimizing for constrained hardware, developing techniques that extract more from less. The Second Track documented the fork: DeepSeek V4 running on Huawei Ascend chips manufactured on SMIC’s 7nm process, creating an independent hardware path that does not depend on Nvidia at all.
The Diversion adds the supply side. While China builds its own chip ecosystem — slowly, expensively, but with increasing capability — the restricted chips continue to arrive through channels the export control regime can see only after the fact. Both tracks run simultaneously. The official track builds independence. The unofficial track buys time while independence is being built.
What the Indictment Reveals
The most telling detail in the indictment is not the hair dryer or the dummy servers. It is the defendant’s position. This was not a rogue trader or a middleman who spotted an arbitrage opportunity. This was a co-founder and senior vice president of the company that manufactured the servers — someone with intimate knowledge of the compliance infrastructure who used that knowledge to circumvent it from the inside.
The Foundation for Defense of Democracies published an analysis the day after the indictment arguing that the case shows the limits of industry self-policing. The assessment is accurate but understates the structural problem. Self-policing assumes the interests of the company and the interests of the enforcement regime are aligned. For a server manufacturer selling to the fastest-growing market in the world, they are not. The two and a half billion dollars in diverted sales represents revenue that the compliance regime was designed to prevent. The compliance regime was defeated by the person who understood it best.
Congress can increase the BIS budget. Prosecutors can bring charges. But the fundamental tension remains: the most valuable computational hardware in the world is manufactured by a small number of companies, demanded by every nation building an AI capability, and restricted by a regulatory framework designed for a world where the supply chain had borders.
The hair dryer and the serial number sticker are not a story about criminality. They are a story about what happens when a border encounters sufficient pressure. The border deforms. The question is not whether export controls can catch individual smugglers — they can, and they are. The question is whether the enforcement surface can scale as fast as the incentive to cross it.
History suggests it cannot.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)