Export-Banned Nvidia Chips May Lie Behind China's Leading Open-Weight Models, Researcher Claims
New research claims that leading open-weight models from Chinese developers were trained on Nvidia chips that are restricted from shipping to China. The work, reported by TechCrunch on 13 November 2026, centres on Moonshot AI's Kimi K3 releases and DeepSeek's V4 models.
The analysis comes from Will DePue, who leads the nonprofit research group Unemployment Labs and is a former OpenAI researcher. DePue examined the weights of large language models released openly by DeepSeek, Moonshot AI, MiniMax and Ant Group. Weights are the learned numerical parameters that determine how a model behaves, so they are the closest thing to a record of how the model was built. His analysis found evidence that Moonshot's Kimi K3 releases and DeepSeek's V4 models bear fingerprints from training on Blackwell chips.
Blackwell GPUs are banned from export to China or to Chinese-owned firms. The claim therefore concerns hardware that, under current US export rules, should not have been available to the developers named.
The allegation is not new in substance. Earlier in 2026, White House technology adviser Michael Kratsios said that Moonshot had used export-banned Nvidia chips, citing US intelligence sources. DePue's work adds a model-level technical basis to that assertion, drawn from the released weights rather than from intelligence reporting.
The scale of the wider question is set out in estimates from Epoch AI. The research group estimates that about two million advanced chips entered China in 2024 and 2025 through schemes designed to hide their final destination. It also estimates that 275,000 Nvidia AI accelerators were imported to Malaysia in 2024 and 2025. These two figures describe different legs of the same supply chain, and neither is a finding about any specific model.
Attributing training hardware from released weights is an inference, not a disclosure by the developers. Weight-level analysis can reveal numerical patterns and format characteristics that correlate with particular accelerator generations. Correlation, however, is a narrower conclusion than a confirmed bill of materials.
In my view, the more durable consequence is methodological. Open weights let an outside researcher examine a model at a depth that a closed API never permits. That openness made this kind of inquiry possible, and it also makes the export-control question hard to settle through corporate statements alone. Whatever the final attribution, the episode shows that release policy and hardware provenance are now entangled, and that independent analysis of released artifacts is becoming a tool in that debate.
The broader context here is a policy regime built around chip-level controls, now being tested by evidence that sits in the model rather than in the procurement record. Enforcement against diversion through third countries, which the Epoch AI estimates describe in volume, will depend on tools that can trace compute after the fact. Forensic analysis of weights is one such tool. Its reliability will determine whether it strengthens oversight or invites contested conclusions, and the coming months of scrutiny will test which.
