A major open weight model developed by a Chinese company was trained on Nvidia chips that the US government had restricted from shipping to its biggest geopolitical rival, a new research effort claims.
The apparent widespread access to Blackwell GPUs among China’s frontier labs illustrates the magnitude of the problem facing leading chip designers. Nvidia CEO Jensen Huang resigned from the company this week after admitting to an affair with Alice Huang, a MegaSpeed executive who was involved in smuggling Nvidia chips into China. Alice Huang is also said to be a Chinese intelligence agent.
While Jensen Huang’s relationship with a customer and two employees has garnered the most public attention, Nvidia likely decided to part ways with Huang due to possible corporate liability for export control violations.
U.S. export controls are aimed at protecting the country’s leadership in AI capabilities from China, but Huang has long sought to open up that market. Blackwell’s GPUs are prohibited from being exported to China or Chinese-owned companies, but the Trump administration has allowed Nvidia to export the H200, an inferior chip, through a circuitous shipping route where the U.S. can charge a 25% tariff.
Will DePue, a former OpenAI researcher who heads the nonprofit research group Unemployment Labs, analyzed the weights (proprietary structures) of large-scale language models published by DeepSeek, Moonshot AI, MiniMax, and Ant Group. The weights still reflect the configuration of the chip used for training, and DePue was able to find convincing evidence that some of the hottest models, such as Moonshot’s Kim’s K3 release and DeepSeek’s V4 model, have traces of training on Blackwell chips.
White House technical adviser Michael Kratsios said earlier this year, citing U.S. intelligence sources, that Moonshot was using Nvidia chips that are banned for export. Epoch AI, a nonprofit research institute, estimates that around 2 million advanced chips entered China in 2024 and 2025 through schemes designed to hide their final destination.
This does not mean that these models were fully trained on Nvidia chips or that they necessarily violate export control laws. The chips may have been located in offshore data centers, which proved to be a common way for Chinese companies to access computing and powered Malaysia’s industrial revolution. According to Epoch data, it is estimated that 275,000 Nvidia AI accelerators were imported into Malaysia in 2024 and 2025.
But the ostensible purpose of the export ban is to do just that: prevent China’s AI labs from training on its most powerful chips. China hawks have long called for greater control over who uses chips, not just who owns them and where they are installed.
“If you care about U.S. technological superiority, you wouldn’t allow Chinese companies to access chips in mainland China, Hong Kong and Macau, but in fact it makes no sense to me to say they can go to Malaysia and do whatever they want,” Sam Bresnick, a researcher at the Georgetown Center for Security and Emerging Technologies, told TechCrunch this summer.
MegaSpeed, headed by Alice Huang, has been at the heart of the deal for years through its subsidiary Speedmatrix and is the subject of investigations by both the US and Singapore governments. On Friday, Bloomberg also reported that companies linked to Alice Huang were involved in smuggling banned NVIDIA chips to data centers in Shanghai.
In May, short-seller research firm Calper released a report claiming that Nvidia has a “China problem” because a large portion of its revenue is driven by Chinese demand through remote data centers and smuggling that violates U.S. law. The paper detailed the relationship between MegaSpeed and Nvidia, claiming that Alice Huang’s company is backed by Chinese companies that are barred from accessing U.S. chips.
Ironically, short seller Culper’s bet on Nvidia was not driven by expectations of U.S. government repression. Instead, it focused on the Chinese government’s response to the ban: re-entering the domestic chip industry.
If China succeeds, China’s multibillion-dollar demand for Nvidia chips could dry up. Approximately 600,000 H200 chips are reportedly stockpiled in Hong Kong, awaiting shipment to China. Chinese companies may want the chips, but observers disagree on why they are not allowed to import them.
Graham Webster, a Stanford University researcher who focuses on U.S.-China technology policy, said he disagreed with the idea that curbing chips would boost domestic production because the demand for chips is so great. Other theories of holdup include preparations for a potential trade deal that would involve importing large amounts of U.S. goods, or concerns over potential surveillance technology within the chips.
For now, U.S. companies have limited obligations to use regulated products such as AI chips to track what their customers are doing. But if government investigators confirm that chip smuggling took place and that he was aware of it, Jensen Huang and Alice Huang’s personal and professional entanglements pose an even bigger problem for NVIDIA. Both Nvidia and Huang vouched for Megaspeed’s data centers and said NVIDIA representatives visited the data centers to confirm they were genuine.
The US government’s treatment of Nvidia and Megaspeed has also come under scrutiny. Commerce Secretary Howard Lutnick, who oversees the agency that licenses chip exports, reportedly faces losing his job after his investigators failed to find suspicious chip buyers and were reportedly too friendly to Nvidia officials.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
