Recent investigations reveal that prominent open-weight models developed by Chinese frontier labs were trained using Nvidia chips restricted by U.S. export controls. The findings highlight significant loopholes in global trade restrictions designed to safeguard American leadership in artificial intelligence and artificial intelligence hardware. The revelation comes alongside major leadership shifts at Nvidia following scrutiny over illicit semiconductor shipments into mainland China.
Will DePue, a former OpenAI researcher leading Unemployment Labs, analyzed the architectural structures and weights of open-source language models released by prominent Chinese entities including DeepSeek, Moonshot AI, MiniMax, and Ant Group. DePue discovered compelling evidence that models such as Moonshot's Kimi K3 releases and DeepSeek's V4 models bear structural fingerprints indicating they were trained using advanced Blackwell GPUs. White House technology adviser Michael Kratsios previously noted that Moonshot had utilized export-banned Nvidia hardware, referencing internal U.S. intelligence.
Key takeaways
- Chinese frontier labs trained leading open-weight models on export-restricted Nvidia Blackwell GPUs.
- Research into model weights by Unemployment Labs uncovered fingerprints from restricted hardware.
- Nvidia CEO Jensen Huang left the company following revelations regarding his relationship with Megaspeed executive Alice Huang.
- Epoch AI estimates that roughly two million advanced chips entered China illegally or via offshore hubs between 2024 and 2025.
The mechanics of chip smuggling and offshore training
U.S. export controls forbid shipping Blackwell GPUs directly to China or Chinese-owned enterprises, though the Trump administration permits exports of the less powerful H200 chip under strict tariff conditions. Despite these rules, sophisticated smuggling networks and offshore data centers have successfully supplied Chinese labs with massive amounts of high-performance compute. Epoch AI estimates that between 275,000 and two million advanced processors entered China or nearby proxies, such as Malaysia, throughout 2024 and 2025.
Megaspeed, a firm led by Alice Huang, has played a central role in these supply chains through its subsidiary Speedmatrix, according to government investigations in the United States and Singapore. Reports from Bloomberg and short-selling firm Culper detail how these channels circumvented federal oversight to supply high-end silicon to Chinese entities. These transactions raise complex questions regarding corporate liability for Nvidia, which previously vouched for the legitimacy of Megaspeed data centers after company officials conducted on-site visits.
Leadership fallout and future implications
Nvidia CEO Jensen Huang departed the company this week following revelations of his personal relationship with Alice Huang, whose firm was implicated in smuggling Nvidia chips. While the relationship drew intense public scrutiny, Nvidia's potential corporate liability under U.S. export control laws heavily influenced the decision to part ways with the longtime executive. Meanwhile, Secretary of Commerce Howard Lutnick faces growing political pressure after reports indicated that federal trade investigators failed to detect suspicious chip buyers.
The broader implications of these findings point to persistent challenges in enforcing technological boundaries. China hawks and security researchers argue that current restrictions focusing solely on direct imports and ownership fail to address offshore compute hubs in regions like Southeast Asia. As policy experts debate closing these loopholes, domestic Chinese semiconductor initiatives continue to evolve in response to ongoing market restrictions.
Comments