Leading open-weight artificial intelligence models developed across China were successfully trained using Nvidia hardware that Washington strictly restricted from reaching its geopolitical rival. This revelation highlights massive compliance vulnerabilities in the global semiconductor supply chain.
Analyzing Model Weights to Track Restricted Hardware
Will DePue, who leads the nonprofit Unemployment Labs and previously worked at OpenAI, analyzed the weights of foundational models released by DeepSeek, Moonshot AI, MiniMax, and Ant Group. By examining the unique structural configurations of these models, the investigation uncovered clear fingerprints indicating they relied heavily on Blackwell architecture accelerators during training.
Executive Fallout and Regulatory Scrutiny
The fallout from these findings reached the highest levels at Nvidia, where long-time leader Jensen Huang stepped down after admitting to a relationship with Alice Huang. She is an executive at Megaspeed, a firm implicated in covertly moving advanced hardware into mainland markets, and is further alleged to be an intelligence agent for Beijing.
Offshore Data Centers and Smuggling Routes
While direct shipping of Blackwell units remains blocked, researchers note that overseas infrastructure in locations like Malaysia has become a primary conduit for procuring computing power. Epoch AI estimates that roughly two million advanced processors entered regional hubs between 2024 and 2025 through alternative logistics networks designed to obscure final destinations.
Stockpile Dilemmas and Official Accountability
Meanwhile, hundreds of thousands of lower-tier H200 chips remain stalled in Hong Kong warehouses without official import clearance. The Commerce Department, overseen by Secretary Howard Lutnick, faces intense scrutiny over whether regulatory oversight failed to catch these unauthorized transactions.
