Jensen Huang Has Left Nvidia, and U.S. Semiconductor Export Controls Still Have Not Stopped China
Jensen Huang resigned as chief executive of Nvidia on November 12. He was the most vocal corporate critic of U.S. semiconductor export controls. Four years after the controls began, controlled technology still reaches China. Why have the controls not stopped it?
How the Controls Work
U.S. semiconductor export controls are a licensing system, not a ban. Under the Export Administration Regulations, the Commerce Department’s Bureau of Industry and Security, known as BIS, sets performance thresholds that decide which chips are controlled, and most exports to China are reviewed under a presumption of denial. The Entity List names specific companies, the Foreign Direct Product Rule reaches products made abroad with American technology, and a separate rule bars Americans from supporting advanced chip factories in China.
The stated goal is not a total blockade but a small yard with a high fence. That was the phrase used by Jake Sullivan, national security adviser in the Biden administration, and it means keeping as large a gap as possible in cutting-edge technology. The controls were meant to slow China’s access, not to keep a single chip out of China. That distinction matters, because it is the standard by which the controls should be judged.
Computing Can Be Rented Without Moving a Thing
The first place the controls leak is that they watch the movement of goods, while computing can be rented without moving anything. Renting chips in an overseas data center over the internet for training and inference is largely legal under current U.S. law. Citing a Wall Street Journal report, a May 2026 analysis by the Carnegie Endowment for International Peace found that INF Tech, a Shanghai startup, remotely used about 2,300 cutting-edge Blackwell chips in an Indonesian data center to train artificial intelligence. Tencent signed a $1.2 billion contract for access to 15,000 B200 chips from a cloud provider in Japan.
But the cloud route has not made frontier pretraining possible. Carnegie noted that 2,300 Blackwell chips are enough to fine-tune a model but orders of magnitude fewer than needed to train one from scratch. The cloud lowers the threshold, it does not defeat the controls.
Chips Designed Just Below the Line
The second leak is that the control line is drawn in performance numbers. A manufacturer can take the same architecture, lower its performance, and build a chip for China that sits just below the line. That is legal. After the October 2022 rules, Nvidia released the A800, a version of the A100 with slower interconnect speed, and the H800, based on the H100. When the October 2023 rules replaced the interconnect-speed test with total processing performance, known as TPP, and performance density, the A800 and H800 fell under control, so Nvidia released the H20 in November 2023, tuned to sit below the new line. For about a year and a half, the H20 was the most powerful chip that could legally be sold in China.
In April 2025, the H20 also came under a license requirement, and Nvidia recorded a $4.5 billion charge. That August, the Trump administration allowed H20 sales on the condition that 15 percent of the revenue go to the U.S. government, and in December, President Trump approved H200 sales on a 25 percent condition. But Beijing’s restrictions kept actual sales minimal. According to Nvidia’s filings, H20 sales came to about $60 million, and H200 shipments were less than 1 percent of data center revenue.
Smuggling, and the Foundry Channel
The third leak is smuggling. In a June 2025 report, the Center for a New American Security, known as CNAS, estimated that between 10,000 and hundreds of thousands of AI chips were smuggled into China in 2024, with a median estimate of about 140,000. That is not a count of confirmed seizures but an extrapolation from public reporting and online listings through a probability model, and the authors themselves said the uncertainty is large.
Documented prosecutions support the pattern. The Justice Department indicted a company called Earthmade in October 2026, alleging it sent more than $300 million in Nvidia servers to China through Malaysia and Singapore. Singapore police arrested nine people in February 2025 in a server fraud case built on false end-user declarations. The U.S. Attorney’s Office for the Southern District of New York said in March 2026 that it had indicted a Supermicro insider for routing about $2.5 billion in servers through a buyer in Southeast Asia. In October 2025, Alan Hao Hsu, an operator of Hao Global, pleaded guilty to exporting GPUs worth about $160 million. Apart from the Hao Global guilty plea, the other cases remain at the allegation stage, with no convictions.
There is also a foundry and equipment channel. China’s SMIC built 7-nanometer-class chips without the extreme ultraviolet lithography equipment, known as EUV, made by the Dutch company ASML, using deep ultraviolet equipment, known as DUV, in a multi-patterning process that exposes each layer several times. TechInsights, a Canadian semiconductor analysis firm, took apart Huawei’s Mate 60 Pro, released in August 2023, and confirmed that its Kirin 9000s chip was made on SMIC’s 7-nanometer-class process. Analysts estimate yields of 30 to 40 percent, below TSMC’s 80 percent or more, and costs two to three times higher.
Equipment detours have also been punished. In February 2026, the Commerce Department fined Applied Materials $252.5 million, alleging it sent ion implanters assembled by a South Korean subsidiary to SMIC. After TechInsights found a TSMC 7-nanometer die inside Huawei’s Ascend 910B AI chip in October 2024, questions were raised about a detour through a front company, and TSMC halted the relevant shipments and notified the Commerce Department. That allegation has not been confirmed.
Open Source, Outside the Controls by Design
The open-source route is one the controls were never designed to cover. The Chinese AI company DeepSeek trained its V3 model on H800 chips that had been legally sold before the October 2023 ban, and disclosed training costs of $5.576 million. That counts only the final successful run, and SemiAnalysis estimates the full infrastructure cost at about $1.6 billion. Efficiency innovations that DeepSeek published, including multi-head latent attention and FP8 training, show that a shortage of computing can be offset with algorithms. Publicly released model weights were also outside the diffusion rule, which covered only the weights of closed models.
The Limits of Enforcement
Enforcement resources also fall far short of the scale of trade. According to CNAS, BIS has fewer than 585 employees and a budget of about $191 million. A single export control officer covers all of Southeast Asia and Australia. An analysis of BIS data found about 38,000 license applications in fiscal 2023, while end-use checks in 2024 numbered only about 1,440. Investigations inside China are limited by Chinese law and U.S.-China agreements. Controlled chips carry no requirement for location tracking or remote deactivation.
Cooperation with allies also has gaps. The Netherlands introduced a licensing system for cutting-edge deep ultraviolet equipment in 2023, and Japan brought 23 categories of equipment under control. But the January 2023 agreement among the United States, the Netherlands and Japan was an informal understanding, not a treaty, and each country implemented it at a different speed and scope. Malaysia, Singapore and Thailand are not U.S. allies and have no obligation under their own laws to enforce American export controls.
How Far Has China Come
Even so, China’s domestic capability has grown faster than the controls assumed. Combining Huawei’s published specifications with analyst estimates puts the Ascend 910C at about 76 to 81 percent of Nvidia’s H100. But by DeepSeek’s testing standards, real-world training efficiency is closer to 60 percent, according to many assessments. With no standardized benchmark, every figure is an estimate. Mass production of SMIC’s 5-nanometer process has not been confirmed, and industry forecasts put volume production of EUV at 2028 to 2030 at the earliest, even on optimistic views. China still depends on Nvidia hardware for cutting-edge pretraining, but it has built a self-reliant structure for inference and for post-training, the work of refining an already trained model.
Insufficient, in Three Different Ways
So the verdict of insufficient needs to be split three ways. There is leakage the controls were never designed to stop, including publicly released open-source models and algorithms, legally sold low-end chips, and overseas cloud rentals. There is legal adaptation the controls allow, including chips designed just below the threshold and equipment stockpiled before the controls. And there is genuine failure, including smuggling through false end-user declarations and paper companies, and foundry orders routed through front companies. Insufficient is the right word only for the third category. The cause there is not the rules but enforcement resources, verification capacity, and gaps in allied legal systems.
Tightening the controls also has a cost. Huang said in May 2026 that Nvidia’s market share in China had fallen from 95 percent to zero. The $4.5 billion H20 charge was recorded, and ASML’s share of revenue from China has also fallen. The tighter the controls, the harder it becomes to keep allies cooperating.
Huang’s Resignation and the Controls
There is no verified causal link between Huang’s resignation and export control policy. Nvidia announced on November 12 that Huang had resigned as president, chief executive and director with immediate effect, and that Colette Kress, chief financial officer, would serve as interim chief executive. The official announcement gave no reason for the resignation. Espionage allegations raised in some reports around the resignation have not been independently confirmed by the Associated Press or Reuters, and China has denied them. No indictment has been confirmed for a large smuggling network mentioned in some reports, and the company named in those reports has denied smuggling. That Huang was the leading corporate lobbyist for easing the controls is a separate fact from the cause of his resignation, and it is not yet known how the controls will change now that he is gone.
Three things are worth watching. Nvidia reports earnings on November 17. A replacement for the diffusion rule has not been published since a draft was withdrawn early in 2026. And the Chip Security Act, which would require location verification for chips, has not yet become law.
The controls have largely succeeded in slowing China’s access to the cutting edge, but they have failed at a total blockade, and both assessments now coexist. China has commercialized 7-nanometer-class chips, offset a computing shortage with efficiency innovations, and built a self-reliant structure for inference and post-training. The longer the controls last, the more China has shifted from buying to making, and that drive for self-reliance is itself seen by many as a paradoxical result of the controls.
