Export-banned Nvidia chips found in Chinese AI model weights

By Zak and the True Work Office team | Published: 17 November 2026 | Category: blog | 3 min read

Export-banned Nvidia chips found in Chinese AI model weights

Key points
  • Unemployment Labs examined released weights from DeepSeek, Moonshot AI, MiniMax and Ant Group and reported fingerprints consistent with Nvidia Blackwell training.
  • Nvidia's Blackwell chips are banned for export to China, while the H200 remains allowed with a 25% tariff and an indirect shipping route.
  • Epoch AI estimates about two million advanced chips entered China during 2024 and 2025, plus 275,000 accelerators imported to Malaysia.
  • Researchers quoted in the coverage argue export controls should follow users rather than ownership alone.

Several of China’s leading open models appear to carry traces of hardware that US export rules forbid shipping to China. Unemployment Labs, a nonprofit run by the former OpenAI researcher Will DePue, analysed released weights from DeepSeek, Moonshot AI, MiniMax and Ant Group. It reported fingerprints consistent with training on Nvidia’s Blackwell GPUs. The named releases include Moonshot’s Kimi K3 and DeepSeek’s V4, as TechCrunch’s report on the fingerprinting analysis sets out.

Blackwell cannot be exported to China. The weaker H200 is still permissible under the current administration, subject to a 25% tariff and an indirect shipping route. White House adviser Michael Kratsios said earlier this year that Moonshot had used banned chips. Epoch AI estimates that about two million advanced chips entered China across 2024 and 2025, with a further 275,000 accelerators imported to Malaysia.

Weights as evidence

The substance here is methodological. Export controls have worked as a logistics instrument, policed at ports and through licences. Weight inspection makes them an evidentiary one. A released model can be examined after the fact for signs of the compute that produced it. That moves the question from who bought what to what can be demonstrated, and it leaves a trail that outlives the shipment.

The limits matter as much as the finding. Fingerprinting of this kind rests on inference from statistical artefacts. Independent replication would be needed before any of it hardens into settled fact. Released weights usually arrive with no documented account of the compute behind them, so the audit depends on researchers reconstructing what developers decline to disclose.

Accountability reaches past the models. Researchers quoted in the coverage argue that controls should follow users, not ownership alone. Commerce Secretary Howard Lutnick faces scrutiny over how export licences were overseen. Nvidia’s own leadership has changed. Chief executive Jensen Huang has departed, and reporting around the exit has cited potential corporate liability for export breaches among the factors at play.

For universities and research groups weighing open-weight models, the practical lesson is narrow but useful. Provenance claims about training infrastructure are testable. Procurement and ethics review can request them, rather than accept a model card’s silence. The same evidence standard that exposes a smuggling route also disciplines the quieter habit of adopting a capable model without asking where it came from.

Whether weight-level forensics can sustain an enforcement regime, or only produce occasional warnings that get argued over and then absorbed, is the part still unresolved.

Frequently asked questions

What evidence supports the claim that Chinese models were trained on banned chips?

The claim rests on analysis of released model weights, which reportedly show statistical fingerprints consistent with training on Nvidia Blackwell hardware.

Do the findings prove that export controls have failed?

They do not establish that on their own. The method relies on inference from statistical artefacts, and independent replication would be required before the results are treated as settled.

โ† Back to Blog