Nvidia Said There Is No Evidence of Chip Diversion to China. A Senate Committee Disagreed.

Jensen Huang said publicly that there is "no evidence of any AI chip diversion" and that Nvidia's China market share had "dropped to zero."

On June 1 2026, five weeks after Huang's public statements were in wide circulation, the Senate Banking Committee wrote to Nvidia's General Counsel and its Audit Committee Chair asking them to substantiate those claims, citing DOJ enforcement actions and a short-seller report.

By August, Taiwanese prosecutors had charged nine people including an Nvidia manager and two Supermicro staff over 74 banned B300 servers that reached China through three different countries.

Everything needed to check this is a public document. This post is a reading of primary sources, not an accusation. And the honest treatment requires stating what the record does not show: none of the criminal cases name Nvidia as a defendant. The 20%-of-revenue figure is a short seller's estimate. What the record does contain is a documented gap between a CEO's public characterization, a congressional inquiry, and federal filings that Nvidia made itself.

The claim#

Jensen Huang, publicly: "there's no evidence of any AI chip diversion." Nvidia chip market share in China has "dropped to zero."

Separately, to CNBC on May 20 2026, Huang said the company's share of China's AI accelerator market had collapsed from roughly 95% to effectively zero after successive US export restrictions, with Huawei the main beneficiary and its Ascend line on course to generate $12 billion in revenue in 2026.

The Senate letter#

June 1 2026, addressed to Tim Teter, Executive Vice President, General Counsel and Secretary of NVIDIA, and Brooke Seawell, Member of the Board of Directors and Chair of the Audit Committee:

"I write to request information about NVIDIA's compliance with U.S. export control laws and regulations and the accuracy of its recent public statements about diversion of advanced artificial intelligence (AI) chips to China."

The letter's stated basis:

"Multiple recent enforcement actions brought by the U.S. Department of Justice allege schemes involving the unlawful diversion of NVIDIA products, including millions of dollars in graphics processing units (GPUs) to China through Malaysia and Thailand, exports and attempted exports of $160 million in H100 and H200 chips, and $510 million in diverted servers."

"These allegations undermine NVIDIA Chief Executive Officer Jensen Huang's public claims that '[t]here's no evidence of any AI chip diversion' and that NVIDIA chip market share in China has 'dropped to zero.'"

And the contrary allegation, attributed: "a recent report from short-seller Culper Research alleges that more than 20% of Nvidia's FY 2026 compute revenues remained driven by China, supported both by illegal GPU diversion and Southeast Asian intermediaries."

The letter also notes the EAR "red flag" / "abnormal circumstances" obligation and that BIS guidance directs companies to investigate when such circumstances indicate an export may be destined for an inappropriate end-use, end-user, or destination.

Why this letter matters beyond politics: it went to the general counsel and the audit committee chair. That is a securities-law escalation path, not a policy complaint. It is a request for the documents a company would need to defend its disclosure.

The federal enforcement record#

The EAR itself prohibits unlicensed exports of advanced AI chips to arms-embargoed countries including China, and requires license application for threshold-exceeding products. Nvidia has been a license applicant under this regime since August 2022 (A100/H100/DGX to China and Russia), with successive tightening: RTX 4090 and L40S in Oct 2023, further D-group additions, 42 PRC entities added to the Entity List in March 2025 and 23 more in September 2025, and H20 licensing from April 2025.

Taiwan, the strongest single case#

Keelung District Prosecutors' Office, August 2026. Nine people charged, including one Nvidia manager and two Supermicro staff, over illegal export of "high-end AI servers," specifically B300 GPUs, banned from sale to China.

Routing of the 74 servers that succeeded:

• 50 through Indonesia
• 16 delivered directly to China
• 8 sent to Japan first, then Hong Kong, before reaching mainland China

A further 56 servers were intercepted and remain in Taiwan. Some defendants set up a company in Japan to facilitate the transfers. Some allegedly created fake websites and falsified information to evade export restrictions. Purchases of more than eight high-end AI servers reportedly require on-site company staff checks by the vendors.

Prosecutors identified a defendant identified only as Chang, an Nvidia manager, as "the key figure" responsible for "authorizing the release of the B300 GPUs," and noted a "clearly poor attitude following the offense." Maximum five-year sentences are sought for four of the nine.

Nvidia: "We will work with the Taiwan authorities to help them resolve the allegations as quickly as possible." Super Micro framed the arrests of two of its former employees as "due to its cooperation with the Taiwanese authorities," that is, as a consequence of cooperating.

There was an earlier Taiwan action in May 2026, same office: three suspects investigated, falsified export documents listing a "Northeast Asian country" as destination, which Bloomberg identified as Japan; 50 Super Micro servers worth more than US$15M seized, along with NT$9M in cash, luxury cars, phones and computers; at least one shipment had already transited Japan and reached Hong Kong, a second was intercepted before leaving Taiwan. Described as Taiwan's first major enforcement action against semiconductor smuggling. Same office, different defendants, overlapping routing. Whether these are one investigation or two is not public.

Singapore, treated fairly#

Charges by Singaporean authorities against Lim Jenny and Woon Guo Jie Aaron (money laundering), with Alan Wei Zhaolun (Singaporean) and Li Ming (Chinese national) also charged with fraud. A $42M property allegedly purchased with the proceeds was seized; roughly $926,000 traced to each of the two principals' accounts; $772,000 frozen in a separate account. Servers were bought through Singapore and routed to China; end users were misrepresented to Dell, Supermicro and Asus. Maximum 20 years and fines up to $385,000.

The famous number: investigators noted that Singapore accounted for about 28% of Nvidia's revenue, yet only about 1% of the company's GPUs were shipped there.

The unbiased reading of that number, which you must give: revenue-recognition location and physical ship-to destination are different things. Nvidia bills through regional entities and system integrators; a large share of "Singapore revenue" is plausibly attributed to a regional APAC booking entity rather than to physical shipment. A 28:1 gap is a flag requiring explanation, not by itself evidence of smuggling. The correct treatment is to state the gap, state the benign explanation, and note that the Singapore case is prosecuted as fraud and money laundering rather than export-control violation, because Singapore law does not directly enforce US trade restrictions. That detail is the most honest thing to say about this case.

Investigation origin, per reporting: US officials began looking after DeepSeek's late-2024 release raised questions about how it had obtained high-performance GPUs.

The named transshipment network#

Bloomberg, May 9 2026: OBON Corp, a Bangkok-based company, allegedly helped move billions of dollars worth of Supermicro servers containing advanced Nvidia chips to China. Alibaba Group was named as one of multiple end customers. Separately reported: a $2.5B smuggling scheme involving Supermicro servers and Nvidia GPUs diverted to China via Thailand, with major Chinese tech firms cited as end customers. That is reporting, not a filing.

What Nvidia's own SEC filings say#

FY2026 10-K, period ended Jan 25 2026:

• A $4.5 billion charge in Q1 FY2026 associated with H20 excess inventory and purchase obligations
• Total revenue $215,938M; Compute & Networking $193,479M; Graphics $22,459M
• Data Center end market: $193,737M (Compute $162,361M, Networking $31,376M)
• Customer concentration: one direct customer at 22% of total revenue, another at 14% (FY2025: 12% and two at 11%)
• April 2025: USG informed Nvidia it requires a license for H20 export to China (including Hong Kong and Macau) and D:5 countries, "or any other circuits achieving the H20's memory bandwidth, interconnect bandwidth, or combination thereof," which is a performance-density standard, not a product-name list

The $4.5B charge is the most under-reported number in this story. It is Nvidia's own filing, it is material, and it is consistent with a supply chain that built inventory against a market that then closed. It is not evidence of diversion, and I am not claiming it is. But it is evidence that the H20-to-China situation was material enough to require a nine-figure inventory writedown, which sits uncomfortably next to "dropped to zero."

Beijing's position, for completeness: Beijing did not green-light H200 imports and reportedly banned the RTX 5090D V2, a card engineered specifically for the Chinese market to comply with US rules. Washington approved H20 and AMD exports in August 2025. BIS rescinded EDA licensing requirements in July 2025 after the PRC agreed to resume licensing rare-earth magnets.

"Zero" and "20%" can coexist#

Here is the reconciliation, which is the intellectual core of this piece. "Nvidia's licensed China revenue went to zero" and "Nvidia-destination silicon reached Chinese buyers via Taiwanese and Southeast Asian integrators" are compatible statements. Authorized direct sales can be ~zero while diverted units flow through third-party integrators in Taiwan, Singapore, Thailand, Malaysia and Japan.

That definitional ambiguity is precisely what the Senate letter asks Nvidia to resolve. Huang's statements are presented in public as a single flat denial without the "authorized sales" qualifier. If his "zero" referred to authorized sales, which is very plausibly true, then the honest picture is a definitional gap, not a caught lie. The interesting part is that this gap exists at all, and that the people asking Nvidia to close it are the United States Senate.

The fragmentation finding#

Export controls on physical goods are being used to control a software-adjacent goods market where the enforcement surface is third-party resellers, system integrators, and freight routing, none of which the US government regulates directly.

The evidence that enforcement is fragmented is not one case; it is four jurisdictions, in five months:

• US DOJ: Malaysia/Thailand transit, $160M H100/H200, $510M servers
• Taiwan: Keelung prosecutors, 74 plus 56 B300 servers, an Nvidia manager charged
• Singapore: fraud/money-laundering, because it does not enforce US trade rules
• China: refusing H200 approvals, banning RTX 5090D V2

Each jurisdiction enforces a different subset of the same rule. Every criminal case in this corpus charges third parties: resellers, distributors, integrators, and one Taiwan-based Nvidia manager. Nvidia is not a defendant in any of them. That is the finding that generalizes past any individual case, and it is more defensible than any individual case.

The 28:1 Singapore anomaly still needs an answer#

The benign explanation deserves a clear restatement, and it still does not close the question. Between them, the 28%-of-revenue-but-1%-of-shipments figure and the performance-density-based H20 licensing standard point in the same direction: the enforcement surface is too diffuse for a "zero" claim to be verifiable by the public, and too fragmented for any single regulator to own the number. Both things can be true at once. That is not evasion; that is what fragmentation looks like.

Where this could be wrong#

The strongest version of the pushback is this: Nvidia's licensed China revenue did drop to zero, and Huang's statements are true under the definition he was using. Grant it explicitly, because it is plausible and because the interesting part survives it. The two statements, no licensed revenue and no Nvidia-destination silicon, are not the same statement, and the public record does not resolve which one Huang meant or whether he distinguished them.

Second, the Culper report is a short-seller attack. It is. Short-seller reports about China exposure are a genre with a poor base rate. Treat it as an allegation, say so every time it appears, and read its methodology. If its method is undisclosed, that is a fact about the report and it belongs here, stated neutrally.

Third, Super Micro's cooperation framing. The company says its former employees were arrested because of its cooperation with Taiwanese authorities. Quote that rather than assert the opposite. Note the tension neutrally: cooperating with one jurisdiction's prosecution while an employee was charged in another is itself worth examining.

Fourth, defamation discipline. Report facts, attribute allegations to their sources in the same sentence, never assert intent, never state as fact that diversion occurred at a given volume. Use "allege," "indictment alleges," "per the prosecutors' filing." The fact pattern is documented; the tone stays flat.

What would update this post: a response from Nvidia or its audit committee, a declassified directive text, a DOJ conviction or plea in the March 2026 case, a declassification of the Culper methodology, or a change in H20/H200/B300 licensing status. Any one of those would sharpen it. None of them is required for the comparison to stand.