Nvidia’s most powerful chips may be behind China’s AI models

A research group says open models from DeepSeek, Moonshot AI, MiniMax and Ant Group were trained on Nvidia accelerators that the US has banned from export to China. An analysis of the models’ weights found signs of Blackwell, Nvidia’s most powerful chips. If confirmed, the findings would indicate how widely Chinese labs have accessed restricted accelerators and could intensify questions for Nvidia about where its products end up and what customers do with them.

Chinese AI models may have been trained on banned Nvidia chips

What the model analysis found

Will Depue, a former OpenAI researcher and head of the nonprofit group Unemployment Labs, examined the weights of large language models released openly by DeepSeek, Moonshot AI, MiniMax and Ant Group. The weights reflect characteristics of the chip configurations used in training. Depue says there are convincing signs of Blackwell use in closely watched models, including Moonshot’s Kimi K3 releases and DeepSeek V4 models.

Earlier this year, White House technology adviser Michael Kratsios said, citing US intelligence, that Moonshot had used Nvidia chips banned for export. The research nonprofit Epoch AI estimates that about two million advanced chips reached China in 2024 and 2025 through schemes that concealed their final destination.

These findings do not prove that the models were trained entirely on Nvidia chips or that export controls were necessarily violated. The accelerators may have been located in data centers outside China: Chinese companies use computing capacity abroad, and Malaysia has become a popular destination. Epoch AI estimates that 275,000 Nvidia AI accelerators were imported into Malaysia in 2024 and 2025.

Export bans and overseas data centers

US export controls are intended to preserve the country’s advantage in AI. Blackwell chips cannot be shipped to China or Chinese companies. The Trump administration allows exports of less powerful H200s if the shipments take an indirect route, for which the US can charge a 25% tariff.

The bans are intended to prevent Chinese labs from training on the most powerful chips. Supporters of tighter restrictions have long argued that oversight should cover not only who owns accelerators and where they are installed, but also who uses them.

Sam Bresnick, a researcher at Georgetown’s Center for Security and Emerging Technology, told TechCrunch this summer that allowing Chinese companies to use chips in Malaysia while barring access in mainland China, Hong Kong and Macau makes little sense if the goal is to preserve the US technological advantage.

Megaspeed and questions for Nvidia

Megaspeed, led by Alice Huang, has been involved in the chip trade for several years through its subsidiary Speedmatrix. US and Singaporean authorities have investigated its activities. On Friday, Bloomberg reported that a company linked to Alice Huang had been involved in smuggling banned Nvidia chips to a data center in Shanghai.

In May, short-selling research firm Culper published a report on “Nvidia’s China problem.” The authors argued that a significant share of Nvidia’s revenue depended on Chinese demand, including demand from remote data centers and smuggled shipments that circumvented US law. The report described Megaspeed’s relationship with Nvidia and claimed that Chinese companies barred from accessing US chips were supporting Alice Huang’s businesses.

Culper’s bet against Nvidia was not based on an expectation of new action by US authorities. The company was counting on a response from China: the country could increase support for its domestic chip industry. If successful, that could wipe out billions of dollars in Chinese demand for Nvidia accelerators.

Meanwhile, around 600,000 H200 chips are reportedly being held in Hong Kong, awaiting shipment to China. Observers have not agreed on why they have not yet been cleared for import. Stanford researcher Graham Webster, who studies US-China technology policy, doubts that delaying shipments alone will spur Chinese manufacturing: demand for chips is too high. Other explanations include preparations for a possible trade deal involving large purchases of US goods, and concerns about surveillance technology in the chips.

Potential consequences for Nvidia

US companies currently have limited obligations to monitor how customers use goods subject to export controls, including AI accelerators. But Nvidia CEO Jensen Huang’s personal and business ties to Alice Huang could become a serious problem for the company if investigators confirm that smuggling occurred and that the CEO knew about it.

Earlier this week, Jensen Huang left the company after acknowledging an affair with Alice Huang. She leads Megaspeed, a company linked to the smuggling of Nvidia chips into China; she has also been accused of working for Chinese intelligence. Public attention has focused mainly on Jensen Huang’s relationships with a customer and two employees, but Nvidia’s potential liability for violating export rules likely also influenced the decision to part ways with him.

Nvidia and Huang supported Megaspeed: Nvidia representatives visited its data centers and confirmed that they were operating legally. US authorities’ actions regarding Nvidia and Megaspeed have also come under scrutiny. Commerce Secretary Howard Lutnick, who oversees the agencies that issue chip export licenses, may reportedly lose his job after reports that his investigators failed to identify suspicious buyers and were too close to Nvidia representatives.

Dataist’s take: The analysis of the weights points to possible use of Blackwell, but does not by itself establish where the chips were located or whether export rules were violated. For Nvidia, the key question remains what the company and its leadership knew about the accelerators’ end use: confirmation of smuggling and Jensen Huang’s awareness could change the legal consequences of the case.