New Research Says China's Most Powerful AI Models Were Trained on Banned Nvidia Chips

type

Post

status

Published

date

Nov 14, 2026

slug

summary

New Research Says China's Most Powerful AI Models Were Trained on Banned Nvidia Chips

A new research effort claims that the leading open-weight AI models developed by Chinese companies were trained on Nvidia chips that the U.S. government has restricted from being shipped to its largest geopolitical rival. If accurate, the apparent breadth of Chinese frontier labs' access to Blackwell GPUs points to the magnitude of the problem now facing the leading chip designer — at an especially delicate moment for the company.

Reading the fingerprints inside model weights

Will DePue, a former OpenAI researcher who leads the nonprofit research group Unemployment Labs, examined the weights — the unique learned structure — of large language models that DeepSeek, Moonshot AI, MiniMax and Ant Group have released openly. Weights still reflect the configurations of the chips used to train them, and DePue reports finding compelling evidence that some of the most closely watched releases, including Moonshot's Kimi K3 models and DeepSeek's V4 series, bear the fingerprints of training on Blackwell chips.

That matters because Blackwell GPUs are banned from export to China and to Chinese-owned firms. The Trump administration does allow Nvidia to ship the less powerful H200 to Chinese buyers, but only along a circuitous route that lets the U.S. claim a 25% tariff.

Corroborating signals, and important caveats

The weight analysis is not the first warning sign. Earlier this year, White House technology adviser Michael Kratsios said, citing U.S. intelligence sources, that Moonshot had used export-banned Nvidia chips. The nonprofit research organization Epoch AI, meanwhile, estimates that roughly 2 million advanced chips entered China in 2024 and 2025 through schemes designed to hide their final destination.

There are caveats. None of this means the models were trained entirely on Nvidia hardware, or that export-control laws were necessarily violated. The chips could have been sitting in offshore data centers, which have become a popular way for Chinese firms to access compute — and which have helped drive an industrial boom in Malaysia. Epoch's data estimates that 275,000 Nvidia AI accelerators were imported to Malaysia in 2024 and 2025.

Still, the ostensible goal of the export bans is to prevent precisely this outcome: Chinese AI labs training on the most powerful chips. China hawks have long called for stronger controls on who uses the chips, not just who owns them or where they are installed.

"If we care about the U.S. tech edge, it doesn't make any sense to me that you would not allow Chinese companies to access chips in mainland China or Hong Kong and Macau, but then say, actually, if you just go over to Malaysia, you can do whatever you want," said Sam Bresnick, a researcher at the Georgetown Center for Security and Emerging Technology, in comments to TechCrunch this summer.

The Megaspeed connection

One company sits at the center of that trade. Megaspeed, which is led by Alice Huang, has been a hub of it for several years through its subsidiary Speedmatrix, and has been the subject of investigations by both the U.S. and Singaporean governments. On Friday, Bloomberg reported that a company with links to Alice Huang was involved in smuggling banned Nvidia chips to a data center in Shanghai.

The issue has now reached the top of Nvidia itself. CEO Jensen Huang left the company this week after admitting to an affair with Alice Huang, the Megaspeed executive who has been implicated in smuggling Nvidia chips into China and alleged to be an agent of China's intelligence service. While the CEO's relationships with a customer and two employees have drawn the most public interest, Nvidia's potential corporate liability for export-control violations likely factored into the decision to part ways with him.

Billions in Chinese demand on the line

For all the personal drama, the commercial stakes may matter even more. In May, the short-selling research firm Culper published a report arguing that Nvidia has a "China problem," because a significant share of its revenue is driven by Chinese demand — whether through remote data centers or smuggling in violation of U.S. law. Culper detailed Megaspeed's relationship with Nvidia and alleged that Alice Huang's companies are backed by Chinese firms that are barred from accessing U.S. chips.

Ironically, Culper's bet against Nvidia was not premised on a U.S. government crackdown. It was focused on the Chinese government's response to the bans: a recommitment to its domestic chip industry. If China succeeds, billions of dollars in Chinese demand for Nvidia chips could dry up.

A large cache of hardware is meanwhile stuck in limbo. Some 600,000 H200 chips are reportedly stockpiled in Hong Kong, awaiting shipment into China. Chinese companies might well want the chips, but observers have reached no consensus on why they have not been cleared for import. Graham Webster, a research scholar at Stanford who focuses on U.S.–China technology policy, said he does not buy the idea that holding the chips back would drive domestic production, since demand for chips is so large.

The limits of location-based controls

Whether or not China's frontier models were trained on banned hardware inside mainland China, the findings sharpen a question that has shadowed U.S. export policy: if a chip only has to cross a border to end up in the hands of the very labs the rules were meant to restrain, how much protection do the bans actually provide? For Nvidia, the research adds pressure on a business that has long tried to serve Chinese demand without crossing American red lines. For policymakers, it strengthens the argument of those who say controls keyed to ownership and geography, rather than end use, are falling short.

Original source: TechCrunch, https://techcrunch.com/2026/11/13/export-banned-nvidia-chips-are-behind-chinas-most-powerful-ai-models-uh-oh-jensen/