TESTIMONY: Nvidia's AI chips were ending up in China — and Washington took 18 months to plug the gap
On May 29, 2026, an anonymous memo titled "The floodgates quietly opened" circulated in Washington. Its central message was as simple as it was devastating: the Trump administration's enforcement decisions had effectively left the United States with no restrictions on Chinese companies' access to the most advanced artificial intelligence chips. From the Pentagon to Capitol Hill
- On May 29, 2026, an anonymous memo titled "The floodgates quietly opened" circulated in Washington. Its central message was as simple as it was devastating: the Trump administration's enforcement decisions had effectively left the United States with no restrictions on Chinese companies' access to the most advanced artificial intelligence chips. From the Pentagon to Capitol Hill
- TESTIMONY: Nvidia's AI chips were ending up in China — and Washington took 18 months to plug the gap
- Introduction: The breach in the technological wall
Facts, quotes, and cited links remain in the body. Interpretations are framed as analysis or opinion according to the format.
TESTIMONY: Nvidia's AI chips were ending up in China — and Washington took 18 months to plug the gap
Introduction: The breach in the technological wall
An anonymous memo that set everything in motion
On May 29, 2026, an anonymous memo titled "The floodgates quietly opened" circulated in Washington. Its central message was as simple as it was devastating: the Trump administration's enforcement decisions had effectively left the United States with no restrictions on Chinese companies' access to the most advanced artificial intelligence chips. From the Pentagon to Capitol Hill, alarm bells rang. Two days later, on a Sunday morning, the Bureau of Industry and Security (BIS) published emergency guidance to close the breach.
But the problem goes beyond a simple regulatory gap. It reveals how Beijing exploited the blind spots of American trade policy for 18 months — through subsidiaries in Singapore, Malaysia, and Taiwan — to acquire Nvidia Blackwell and AMD MI350x processors, the engines of next-generation AI. In the Sino-American technology war, this is a battle Washington nearly lost without even realizing it.
The structural flaw: the subsidiary outside China
The mechanism is elegant in its brutality. Chinese technology companies created wholly or majority-owned subsidiaries in Asia-Pacific — outside Chinese territory. Those subsidiaries bought advanced AI accelerators from American suppliers or regional distributors. The argument: the transaction fell outside the jurisdiction of American export controls because neither the seller nor the immediate buyer was physically located in China.
According to the BIS, that argument had never had a legal basis. Since November 2023, the rule was clear: the license obligation traces the corporate structure up to the ultimate parent company, regardless of the delivery address. But between stating that the rule exists and actively enforcing it, there was a chasm that Beijing crossed at full speed for 18 months.
The May 2025 decision: when Trump opened the door
Abandoning the Biden AI diffusion rule
In May 2025, the Trump administration decided to stop enforcing the Biden AI Diffusion Framework — finalized in January 2025 in the final days of the previous administration. Trump justified the decision by citing "burdensome regulatory requirements" and a negative impact on diplomatic relations. Nvidia, which had lobbied heavily against the rule arguing it "threatened innovation and international cooperation," got what it wanted.
The problem: that decision created a grey zone around whether the pre-existing controls, restructured within the Biden rule, were also suspended. Officially, the BIS says no — the November 2023 rule was never repealed. In practice, Trump officials privately believed, according to Bloomberg, that deliveries to overseas Chinese subsidiaries were legal until the Sunday guidance. An internal confusion within the American administration that Beijing turned to its advantage.
Chinese companies in Malaysia and Singapore
Between May 2025 and May 2026, Chinese-backed technology companies operating in Malaysia and Singapore received Nvidia Blackwell chips without obtaining American export licenses. These transactions were not covert operations conducted in the shadows: they moved through ordinary commercial channels. Distributors in Taiwan and South Korea, contract server manufacturers, venture capital investors holding stakes in AI infrastructure operators — all potentially face retroactive liability dating back to November 2023.
The real scale of the deliveries remains unknown. But Chris McGuire, a former State Department official and now senior China researcher at the Council on Foreign Relations, warned: "Chinese companies were buying these chips, probably in large quantities." He added that the BIS guidance implicitly acknowledged those deliveries had occurred, since companies that acquired chips through this channel were not required to immediately stop using them.
BIS patches the breach, but the TSMC loophole stays open
The Sunday May 31, 2026 guidance
The BIS guidance of May 31, 2026 reaffirmed a rule the agency insists it always maintained: any entity whose ultimate parent is based in China, Russia, or a Group D:5 country requires an American export license to receive advanced chips, regardless of where that entity is physically located. The relevant classifications: 3A090.a, 3A090.b, 4A090.a, 4A090.b — in other words, Nvidia's Blackwell and Rubin families, AMD's MI350x, and associated high-bandwidth memory.
That is an important clarification. But it only closes part of the problem. McGuire immediately identified the remaining breach: the due diligence rule imposed on foundries like TSMC and Samsung Electronics. Nothing in the Sunday guidance confirms that rule is still being enforced. "This is a gaping hole that still needs to be plugged," he wrote. In plain terms: Chinese companies could still use third-country subsidiaries to order chip fabrication at TSMC — not just purchase finished chips. The BIS says that rule was never repealed, but refuses to confirm it in writing.
The response from the tech giants
Nvidia declared it was in compliance with the clarified regulations. AMD and Intel did not immediately respond to requests for comment. TSMC declined to comment. Those silences are eloquent. The semiconductor industry had enjoyed a comfortable grey zone for 18 months. It was hostile to the Biden rule. It had supported — publicly or tacitly — its abandonment. Now it is navigating a regulatory uncertainty and legal exposure that no one can yet fully measure.
On the same topic
BILLET: Altman and Huang Head to the Senate as…
According to Boursorama , Sam Altman of OpenAI and Jensen Huang…
OPINION: Vaccines — Trump Pushes Kennedy to Go Further,…
Nobody signs a memo. Nobody writes "move faster" in plain ink.…
EDITORIAL: Measles — America Gives Up a Twenty-Six-Year-Old Public…
There is a line , in a table the CDC updates…
The fact that Nvidia declares itself compliant is important — but insufficient to douse the political fire. In December 2025, Trump authorized the sale of H200 chips to China — chips six times more powerful than the H20, the most advanced model previously permitted. That decision was deliberate. It was already raising questions about the coherence of America's technology control strategy.
Legal action: $420 million in penalties
The legal machinery in action
The BIS and the Department of Justice did not wait to act on documented cases. Over the past 12 months, they announced nearly $420 million in combined penalties and forfeitures linked to the illegal diversion of semiconductors to China. Among the landmark cases: in December 2025, the DOJ dismantled a multi-defendant network known as "Gatekeeper," which had attempted to export at least $160 million in restricted AI chips to China and Hong Kong.
In February 2026, Applied Materials reached a settlement of $252.5 million — the second-largest standalone BIS penalty in history — for illegally shipping semiconductor manufacturing equipment to a Chinese company on the Entity List via a subsidiary in South Korea. In March 2026, the co-founder of Super Micro Computer was arrested and indicted in a conspiracy prosecutors described as "a $2.5 billion diversion of Nvidia-chipped servers to China via shell companies in Southeast Asia."
Congress and legislative pressure
On the legislative front, pressure is building. The AI OVERWATCH Act, passed by the House Foreign Affairs Committee by a 42-to-2 vote in January 2026, would require congressional review of each export license for advanced AI chips to China, removing the Commerce Department's unilateral authority over the most sensitive transactions. In May 2026, the Government Accountability Office ruled that the BIS's initial non-enforcement announcement may have constituted a rule under the Congressional Review Act — opening the door to a legislative challenge.
Maximum penalties for export control violations are currently $374,000 per violation. The BIS deputy secretary for export enforcement publicly asked Congress to raise that ceiling to $1.2 million — aligning it with the penalty under the Arms Export Control Act. In February 2026, Congress had already approved a 23% budget increase for BIS for fiscal year 2026, specifically to strengthen semiconductor-related enforcement capacity.
The Sino-American technology war: the real stakes
Why these chips are so dangerous in the wrong hands
The chips in question — Blackwell, Rubin, MI350x — are the engines of frontier AI model training and large-scale inference. Giving China broad access to them means providing the computational fuel to develop its own military AI models, surveillance systems, large-scale disinformation capabilities, and cyberwarfare tools. That is not abstract. It is the very substance of the strategic competition that Washington claims to be waging against Beijing.
AI is not a neutral commodity. In a context where China massively funds the development of military AI, satellite image analysis, submarine detection, and autonomous weapons systems, every advanced computing accelerator that lands in a Beijing-controlled data center is an investment in China's future power-projection capacity. Allowing 18-month gaps in that control regime is a strategic error with lasting consequences.
China, the primary threat — and patient strategist
China did not steal these chips. It exploited an American legal ambiguity with the patience and methodology characteristic of its technology strategy. The subsidiaries in Malaysia, Singapore, and Taiwan are not geographical coincidences: they are structures built specifically to navigate the blind spots of American regulation. Beijing invests in the long term. Washington reacts to anonymous memos.
Saif Khan, who helped draft the original licensing rules under the Biden administration, noted that the BIS Sunday guidance closes part of the regulatory ambiguity — but not the question of the foundry due diligence rule. If TSMC manufactures chips for clients whose ultimate parent is Chinese, through a third-party structure, without rigorous verification — the problem remains intact, merely displaced one link along the supply chain.
What companies should have done — and didn't
Reversed due diligence: who is really the customer?
The compliance requirements arising from this episode are specific: every company in the sector must conduct a fresh audit of the beneficial ownership of all its clients and end users in Asia-Pacific to identify ultimate parent companies. Standard commercial databases often fail to surface this data reliably. Distributors, contract server manufacturers, financial sponsors, and private equity funds holding stakes in AI infrastructure operators are all potentially exposed.
Retroactive exposure dates back to November 2023. There is no amnesty period. Companies that delivered chips to subsidiaries of Chinese parent companies without export licenses are advised to immediately suspend similar pending transactions, mandate a legal review of past deliveries, and consult an export controls attorney regarding a voluntary disclosure to BIS. Documentation of ownership structures and regulatory justifications is now critical if a BIS or DOJ review were to follow.
The industry caught between two fires
The semiconductor industry finds itself trapped between two competing pressures. On one side, export controls aimed at preserving America's technological lead in strategic competition with China. On the other, commercial pressures from a Chinese market that represents a massive share of revenues for Nvidia, AMD, and Intel. Nvidia did not conceal its opposition to the Biden rule, calling it a threat to innovation. But when America's most advanced AI accelerators end up in data centers training models for the Chinese military, the commercial argument collides with geopolitical reality.
Discover
TESTIMONY: Assam, 700,000 Displaced and a State Rebuilding Every…
On July 20, 2026 , Al Jazeera reported that at least…
ANALYSIS: Gaza's Phase Two, a Ceasefire Stalled in Cairo
On July 28, 2026 , a Hamas delegation left for Cairo…
BILLET: Altman and Huang Head to the Senate as…
According to Boursorama , Sam Altman of OpenAI and Jensen Huang…
The uncomfortable truth: the November 2023 rule existed. It was known. Many chose not to look at it too carefully for 18 months. Now that the alarm has sounded, the question is what passed through the cracks. And according to experts, that figure is possibly massive.
US senators versus Trump: the political front opens
Senators demand accountability
While the BIS published its emergency guidance on a Sunday morning, American senators were taking direct aim at the Trump administration. According to the Straits Times, legislators sharply criticized the White House for allowing advanced AI chips to be sent to overseas units of Chinese companies. This bipartisan political front reflects a growing concern in Congress: executive commercial decisions are compromising national security.
The Foundation for Defense of Democracies (FDD) was even more direct, stating in an analysis published on June 2, 2026 that the Commerce Department had admitted its "failure to enforce AI export controls on China." That admission, if accepted, is a political bombshell: it means that for 18 months, an administration claiming to wage a technology war against Beijing was quietly leaving the doors open.
The political cost of deliberate ambiguity
The question haunting Washington is simple: was this a mistake or a choice? The mistake hypothesis implies grave incompetence in managing the country's most critical export controls. The choice hypothesis — prioritizing Nvidia's commercial interests over national security safeguards — is politically more explosive. Both parties have reason to dig into this question before the 2026 midterm elections.
In that context, the BIS guidance of May 31 looks like a damage control operation as much as a serious policy correction. The speed of the reaction — a Sunday morning after an anonymous memo — betrays institutional panic, not a considered strategy. The real test will come when a comprehensive replacement framework is finally published — if it is, before the current term ends.
What Saif Khan and Chris McGuire say must happen now
Two experts, one shared conclusion
Chris McGuire at the Council on Foreign Relations and Saif Khan, architect of the original rules, arrive at the same conclusion: the BIS May 31 guidance is a necessary first step, but insufficient. The foundry loophole — the ability for Chinese companies to commission chip fabrication at TSMC or Samsung via third-party entities — remains technically open, insofar as the BIS refuses to confirm in writing that the foundry due diligence rule is still in force and being enforced.
Without written confirmation from the BIS, export controls attorneys will advise their clients to treat that zone as legally ambiguous. And Chinese companies will continue exploiting that ambiguity with the same methodology that produced the crisis of the past 18 months. History repeats when you don't give it a definitive answer.
A replacement framework still missing
More than a year after revoking the Biden AI diffusion rule, the Trump administration has still not finalized a replacement framework. That regulatory void is itself an invitation to confusion and exploitation. A replacement regulation should, according to experts, explicitly address the global licensing architecture for advanced computing, close the foundry loophole, and provide sufficient legal clarity for American companies and their regional partners to comply with certainty.
In the meantime, the Sino-American technology competition continues to play out in the gaps of poorly enforced regulations, legal grey zones, and policy decisions made under pressure from an anonymous memo on a Sunday morning. That is a dangerously reactive way to manage what may be the most important strategic challenge of the 21st century.
Conclusion: 18 months behind in the artificial intelligence war
The cost of regulatory complacency
For 18 months, America's export control architecture contained a flaw that China exploited methodically. Nvidia Blackwell chips — the most powerful ever built — potentially powered Beijing-controlled data centers through subsidiaries in Malaysia, Singapore, and elsewhere. The documented figures: a Gatekeeper network operation at $160 million, a Super Micro case at $2.5 billion, an Applied Materials fine at $252.5 million. And those are only the documented cases.
The BIS guidance of May 31, 2026 closes one breach. But the replacement framework for the repealed rule still does not exist. The foundry loophole remains open. And Beijing, which has demonstrated its capacity to exploit regulatory ambiguities for years without Washington reacting, continues to watch for its next opportunities. In the century's technology war, 18 months of delay is an eternity.
What we must take away
The lesson of this episode is not that export controls don't work. It is that they only work when enforced with consistency and without deliberate ambiguity. When a political administration decides that Nvidia's commercial pressure weighs more heavily than national security, it creates the conditions for its own failure. China does not need spies when it can simply buy what it needs through the holes in its adversary's regulations.
These 18 months of this breach may be the most embarrassing chapter in the history of American technology policy. Not because no one knew. But because many knew — and chose to look the other way.
By Maxime Marquette, columnist
Columnist's transparency note
My biases and who I am
I am Maxime Marquette, columnist. I regard China as the primary long-term technological and strategic threat to the West. My analysis is pro-Western and favors rigorous controls on sensitive technologies. I acknowledge that the semiconductor industry makes legitimate commercial arguments — but I do not find them sufficient to justify technology transfers with clear military implications.
What I do not know
I do not know how many chips were actually delivered to Chinese-backed companies during those 18 months. The real figures remain opaque and undocumented publicly. Nor do I know whether the TSMC/Samsung foundry loophole was exploited with the same intensity as the subsidiary channel. I grounded this piece in verifiable sources: TechTimes, Al Jazeera, and the public statements of the BIS, McGuire, and Khan.
Sources
Primary sources
Secondary sources
Get the geopolitics analyses
Conflicts, powers, alliances: the MadMax thread without the noise.
Cite this article
Maxime Marquette (2026). TESTIMONY: Nvidia's AI chips were ending up in China — and Washington took 18 months to plug the gap. MadMax. https://mad-max.co/en/article/temoignage-les-puces-ia-de-nvidia-finissaient-en-chine-et-washington-a-mis-18-mo
Enjoyed this piece? Get the next one.
One chronicle a week, straight to your inbox. No noise.
This article was generated with AI assistance, under human supervision.
Comments
Be the first to weigh in.