SI Agent News
RumorOpen modelsChipsPolicyResearch

Research suggests Chinese AI models may have used banned Nvidia chips

Analysis of model weights suggests Chinese firms might be accessing restricted Blackwell processors through offshore data centers in regions like Malaysia.

275,000Nvidia AI accelerators imported to Malaysia in 2024 and 2025

The short version

Why it matters

This discovery suggests that current U.S. export controls may be easier to circumvent than intended by using third-party locations. It raises questions about whether geographic restrictions can effectively limit China's access to high-end AI hardware.

Evidence in the model weights

Will DePue, a former OpenAI researcher and leader of the nonprofit Unemployment Labs, conducted an analysis of openly released Chinese AI models. He examined the model weights, which are the numerical parameters that define how large language models operate. By looking at these weights, he identified specific hardware-related patterns in several recent releases.

The models highlighted in the research include the Kimi K3 releases from Moonshot AI and the V4 models from DeepSeek. Other companies whose models showed similar patterns included MiniMax and Ant Group. These findings suggest that these specific models may have been trained using Nvidia’s most advanced Blackwell processors.

It is important to note that this analysis does not prove where the training actually took place. It also does not establish which specific chips were used at every single stage of the development process. The research provides a technical clue about the hardware used rather than a definitive record of a legal violation.

Bypassing export controls

The U.S. government has placed strict export controls on Nvidia’s Blackwell processors to prevent them from reaching China. These restrictions are designed to limit the computing power available to Chinese firms for developing advanced artificial intelligence. However, the current controls focus primarily on where the chips are shipped or installed.

Chinese companies have increasingly used offshore data centers to obtain the computing capacity they need. Malaysia has emerged as a primary destination for these facilities. This setup allows companies to potentially access restricted hardware without the chips ever entering mainland China. This creates a significant challenge for U.S. policy intended to maintain a technological edge.

Sam Bresnick, a researcher at Georgetown University, noted the logic gap in these restrictions. He suggested it does not make sense to block access in mainland China while allowing the same access if a company simply moves operations to Malaysia. This highlights how third-country data centers can serve as a workaround for trade barriers.

The scale of the trade

The scale of the hardware moving into these regions is substantial. Estimates from the nonprofit Epoch AI suggest that approximately two million advanced chips entered China during 2024 and 2025 through arrangements meant to hide their final destination. At the same time, the same organization estimated that 275,000 Nvidia AI accelerators were imported to Malaysia over that same period.

Specific companies have been linked to the movement of this computing capacity. Megaspeed, led by Alice Huang, has faced scrutiny from authorities in the U.S. and Singapore regarding its role in these trades. Reports have also suggested that companies connected to Huang were involved in moving restricted chips to a data center in Shanghai.

While these reports draw attention to the flow of hardware, they do not establish that Nvidia knowingly violated export laws. Nvidia has previously defended the legitimacy of Megaspeed’s data centers, stating that their own personnel visited the sites and found them to be legitimate. The situation remains a point of high scrutiny for both regulators and investors.

Corporate and political risks

The findings add to the ongoing scrutiny of Nvidia’s exposure to Chinese demand. If investigators eventually prove that restricted chips were smuggled and that company officials were aware of it, Nvidia could face serious corporate risks. Currently, U.S. companies have limited obligations to monitor exactly how their customers use the controlled goods they sell.

The U.S. government’s oversight of the trade is also under pressure. Commerce Secretary Howard Lutnick could face political consequences if it is found that investigators were too close to Nvidia officials or failed to identify suspicious buyers. These reports come as a large stockpile of about 600,000 Nvidia H200 chips reportedly awaits approval in Hong Kong.

There is no consensus on why these chips have not been cleared for shipment into China yet. Some observers believe the delay might be tied to trade negotiations or concerns over surveillance. Meanwhile, short-selling firms have warned that China's push for domestic chip production could eventually reduce the long-term demand for Nvidia products.

What to watch

Try it

Look up the current U.S. export control list for Nvidia Blackwell processors to see which specific models are restricted.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by SSBCrack · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news