NextFin News - US authorities are asking a question that Nvidia would rather not answer: with billions of dollars of its own AI processors flowing through a global smuggling network into China, why did the world's most valuable chipmaker not see the red flags? The inquiry has moved from private questioning to public pressure as prosecutors, customs agents, and regulators across three jurisdictions have exposed a diversion trade built on forged export documents, serial-number labels peeled off server crates with a hair dryer, and suitcases of chips carried onto commercial flights by students.
Nvidia has not been charged with any wrongdoing, and the company's formal line is that it sells to distributors and system integrators, not to end users, so the destination is the customer's responsibility. But the enforcement wave of the past year has turned that boundary into the central question of the story. When your hardware is the prize that export controls are designed to keep out of a country, how much responsibility do you carry for where it ends up?
The Network Behind the Headlines
The emblematic image is almost darkly comic. A surveillance photograph released by US prosecutors in March shows a woman in an unidentified Southeast Asian warehouse holding a red hair dryer, using hot air to lift serial-number labels from server packages so that restricted Nvidia processors can be rerouted to China. Court records and enforcement filings describe the same low-tech toolkit used at scale: thousands of fake servers staged in a warehouse to deceive compliance auditors, dummy server shells built to disguise the hardware's nature during customs inspection, and re-boxing in unmarked packaging to conceal origin and destination before final delivery to Chinese buyers.
The largest single case sits with Super Micro Computer. On March 19, 2026, US prosecutors in Manhattan unsealed an indictment charging Supermicro co-founder Yih-Shyan "Wally" Liaw and two others — Taiwan sales manager Ruei-Tsang "Steven" Chang and contractor Ting-Wei "Willy" Sun — with conspiring to violate the Export Control Reform Act, smuggling goods from the United States, and defrauding the federal government. Each defendant faces up to 20 years in prison on the conspiracy count. Prosecutors allege the group moved more than $510 million of hardware within a matter of weeks in mid-2025, with the cumulative alleged scheme reaching approximately $2.5 billion. It is the largest AI hardware export-control prosecution in US history.
The alleged mechanics were blunt. Liaw and Chang directed executives at an unnamed Southeast Asian company to place purchase orders with Supermicro as though the servers were destined for that company's own operations. The servers were then shipped to Supermicro's Taiwan facilities, transferred to the Southeast Asian company, and handed off to a logistics firm that removed identifying packaging and shipped the units to Chinese buyers. During a three-week period from late April to mid-May 2025 alone, roughly half a billion dollars of servers assembled in the US were shipped to China as part of the alleged conspiracy.
Taiwan followed in May with its first-ever criminal enforcement action against illegal AI hardware exports. On May 21, the Keelung District Prosecutors' Office executed search warrants across 12 locations and obtained court-granted detention orders for three people, including Liaw, over allegedly forged documents covering approximately 50 Hopper-generation servers bound for China, Hong Kong, and Macau. Taiwan prosecutors have since charged nine people, including former employees of Nvidia and Supermicro, with aiding the shipments on counts including breach of trust and forgery. Singapore, meanwhile, seized a $42 million mansion and $772,000 in a bank account owned by alleged AI GPU smugglers, and charged three men with fraud in a case domestic media linked to the movement of Nvidia chips toward Chinese AI firm DeepSeek.
Not all of the leakage requires forged paperwork. Authorities have documented a 26-year-old Chinese student in Singapore carrying six advanced Nvidia AI chips — each roughly the size of a Nintendo Switch game console — in his luggage on a commercial flight home. Chinese engineers have flown into Malaysia with suitcases of hard drives, trained models on rented servers packed with restricted Nvidia chips at a Malaysian data center, and flown home with the models. US officials are now reviewing how Chinese AI firms access advanced Nvidia processors by renting computing power located in other countries, a legal avenue that achieves the same outcome as a smuggled chip without a single crate crossing a border.
Why the Controls Create the Black Market They Try to Prevent
This is the first-order paradox at the heart of the story: the tighter the restriction, the wider the arbitrage, and the arbitrage funds the smuggling network. Nvidia's H20, a chip deliberately detuned for the China market to comply with US export rules, generated only around $50 million in sales in the quarter ended October 26 in the prior fiscal year, even as demand in China remained the strongest in the world. Overall Nvidia revenue from China, including Hong Kong, was $2.8 billion in that quarter, or 5% of company sales, down sharply from the 13% share the region represented in the prior fiscal year — and far below the $8.4 billion Wall Street analysts had projected.
The supply-demand gap is what makes smuggling profitable. Ray Wang, an analyst at SemiAnalysis, estimates that more than 60% of the leading AI models currently running in China use Nvidia hardware. Jensen Huang, Nvidia's founder and chief executive, has said the company's share of China's advanced-chip market fell from 95% to zero under the combined weight of US and Chinese regulatory scrutiny. A market that is 95% dependent, suddenly forced to zero through legal channels, does not simply disappear; it goes underground, and it pays a premium to get there.
The financial penalty for Nvidia when Washington moves unexpectedly illustrates the stakes. In April 2026, the Trump administration imposed a surprise ban on H20 sales to China. Nvidia said it lost $2.5 billion in revenue in its fiscal first quarter from the China restrictions, took a $4.5 billion inventory charge on H20 chips it had produced but could no longer sell, and projected an additional $8 billion revenue hit in the second quarter. China accounted for 12.5% of Nvidia's revenue in that first quarter, down from roughly 14% and 15% in the two prior periods. Huang called the export ban "deeply painful."
Washington's answer to the enforcement gap has been to make the chipmaker part of the enforcement machinery. According to people familiar with the matter, Nvidia and Advanced Micro Devices agreed to hand over 15% of revenue from China AI chip sales to the US government as the price of licensing H20 exports — an arrangement with no modern precedent in semiconductor trade. Nvidia later said the deal had not been finalized, but the framework itself is the point: it turns the chipmaker into a revenue-collecting arm of the export regime, and it prices compliance as a line item rather than a boundary condition.
The Compliance Question: Vendor or Border Agent?
Nvidia's public position is straightforward and, on its face, reasonable. Huang told reporters in Taipei in May that Supermicro needs to "enhance and improve their regulation compliance and avoid that from happening in the future," adding that Nvidia is "rigorous" in explaining export regulations to all of its partners. The company has emphasized that its chips include export-compliance features and has publicly distanced itself from the alleged scheme. Supermicro's board said an independent investigation found no evidence that current senior management knew of the scheme or of any actual diversion of restricted products. Liaw has pleaded not guilty, and his trial has been pushed from November 2026 to March 2027.
But the compliance logic is under strain for a company that controls an estimated 97% of the server GPU market. The smuggling methods documented in court records are not sophisticated cyber-heists requiring state-level intelligence to detect; they are low-tech tells. Serial numbers stripped with a hair dryer. Shipments routed through shell companies with no operating history. A warehouse filled with thousands of fake servers. A student's suitcase. For a vendor that knows its distribution chain better than any regulator can, the question is not necessarily whether Nvidia broke the law — prosecutors have so far targeted the diverters, not the chipmaker — but whether a compliance system that treats the factory gate as the end of responsibility is adequate when the product is the single most controlled technology in the world.
There is also a second-order consequence that the market has not fully priced. The precedent being built in Washington — pre-export security reviews, revenue kickbacks, partner liability, chipmakers as de facto customs agents — raises the cost of selling to China even through legal channels. The $2.5 billion Supermicro case and the $2.5 billion quarterly revenue loss from the H20 ban are measurable. The structural cost of a compliance regime that can change without notice, and that treats the vendor as the enforcement node, is not. Analysts at Stifel, Bernstein, and William Blair estimate Nvidia could recoup as much as $15 billion from China in the second half of the year if H20 licensing resumes, taking regional revenue to roughly $20 billion for fiscal 2026. That recoupment assumes the licensing framework holds. If the 15% model expands to a broader set of products, or if pre-export reviews slow shipments, the economics of legal China sales shift again.
The Counter-Thesis, and What Would Break It
The strongest defense is also the simplest, and it deserves weight. Nvidia designs and sells semiconductors; it does not clear customs, operate cargo planes, or verify the ultimate consignee on every server. The alleged fraud was carried out by a server maker's co-founder, a sales manager, a contractor, logistics firms, and shell companies — not by Nvidia employees. Charging the chipmaker with failing to detect a criminal conspiracy it was not party to would set a dangerous standard for every exporter in every industry, and it would reward sophisticated criminals by punishing the company whose product they abused. Supermicro's board investigation cleared current senior management, and no US authority has publicly addressed whether Nvidia faces legal exposure.
There is real force in that argument, and it explains why the enforcement wave has so far stopped at customers, logistics firms, and former employees. But it collides with a practical reality: export controls only work if someone verifies the last mile, and Washington has increasingly decided that someone is the chipmaker. If Nvidia's own compliance systems cannot distinguish a shell company from a real data-center operator, the question shifts from "did Nvidia break the law" to "is the law asking the right party to enforce it" — and the answer Washington appears to be building is that the vendor is the cheapest enforcement node available.
The signal that would change the analysis is specific and observable. If the Justice Department or the Securities and Exchange Commission charges Nvidia itself, or a current Nvidia executive, with knowledge of or complicity in the diversion — rather than customers, logistics firms, or former employees — the "vendor is not liable" defense collapses and the compliance cost shifts from partner oversight to corporate liability. Until then, the company's exposure is reputational and regulatory, not criminal.
What to Watch, by Time Horizon
Short term, the risk is headline-driven volatility around court dates and new indictments. Liaw's trial is now set for March 2027, and any jurisdiction that joins the crackdown — Taiwan has already moved; Singapore is active; the US is probing a logistics firm — can produce a fresh negative headline for the ecosystem without changing Nvidia's fundamentals. Nvidia shares closed at $238.90 on October 2, 2026, up from $230.86 the prior session, though the stock has given back roughly $1 trillion in market capitalization from its peak, its lowest level since early 2019 before the generative AI boom.
Medium term, the question is whether Washington extends the 15% revenue model or the pre-export security review to a broader set of products, and whether H20 licensing actually resumes at the volume analysts expect. The base case is that smuggling remains a manageable leakage against $57 billion quarterly revenue — the company reported $57 billion in revenue in its most recent reported quarter, up 62% year over year — and that enforcement stays focused on diverters. The downside case is that compliance becomes a recurring policy tool that raises the cost and slows the timing of every China-bound shipment, legal or otherwise.
Long term, the structural shift is already locked in. Export controls have permanently fragmented the AI hardware market, and the compliance regime that grew up around them has made the chipmaker a border agent by default. That does not reverse when a single case closes. The winners over that horizon are the companies that can sell AI compute without touching US-controlled silicon; the exposed are the chipmakers whose revenue depends on a market they are legally barred from serving directly.
For now, the hair-dryer photograph remains the story's emblem. It is not a picture of Nvidia's guilt; it is a picture of a control regime that assumes the company selling the chip should have noticed the labels being peeled off.
The chips that reach China through a hair dryer and a suitcase are a rounding error for Nvidia. The precedent that the chipmaker should have stopped them is not.
更多独家洞察尽在 nextfin.ai.
