📰 Key Takeaways

A new study says leading open-weight models built by Chinese companies were trained on Nvidia chips the US government bans from export to China. The analysis comes from Will DePue, a former OpenAI researcher who leads the nonprofit research group Unemployment Labs. He examined the publicly released LLM weights from DeepSeek, Moonshot AI, MiniMax, and Ant Group, which still reflect the configuration of the chips used during training. He says he found strong evidence that the most talked-about models carry the signature of Blackwell-chip training, including Moonshot’s Kimi K3 series and DeepSeek’s V4 models.

Blackwell GPUs are currently banned from export to China or Chinese-owned companies. The Trump administration does allow Nvidia to export the lower-performance H200, on the condition that shipments are routed through the US so the government can collect a 25% tariff. Earlier this year, White House tech advisor Michael Kratsios also cited US intelligence sources saying Moonshot had used Nvidia chips restricted by export controls. The nonprofit Epoch AI estimates that around two million advanced chips flowed into China across 2024 and 2025 through methods that hid their final destination.

The story also touches Nvidia’s internal leadership. CEO Jensen Huang left this week after admitting to an extramarital affair with Alice Huang, an executive at Megaspeed. Megaspeed has been accused of smuggling Nvidia chips into China, and Alice Huang has been accused of being an agent for Chinese intelligence. Outside observers speculate that possible corporate liability for export control violations was also a factor in the split.

That said, the report stresses this doesn’t mean the models were trained entirely on Nvidia chips, nor that export controls were necessarily violated, since the chips could be sitting in overseas data centers. That’s a common way for Chinese companies to get compute, and Epoch estimates 275,000 Nvidia AI accelerators were imported into Malaysia in 2024 and 2025. But the whole point of the export ban is to stop Chinese AI labs from training on the most powerful chips, and for China hawks, controls should focus on who is using the chips, not just who owns them or where they’re installed.

💬 JudyAI Lab Take

Chinese open-weight models are being tied to embargoed Nvidia Blackwell chips, and that’s worth paying attention to because it suggests compute controls may be a lot thinner in practice than they look on paper.

Will DePue, a former OpenAI researcher, examined the weights released by DeepSeek, Moonshot AI, MiniMax, and Ant Group, and found that some models carry the signature of Blackwell training, including the Kimi K3 series and DeepSeek V4. For AI builders, two things stand out. First, model weights leave traces of the training environment, so what gets released publicly is more transparent than you’d expect, and easier for third parties to inspect. Second, the report itself stresses this doesn’t necessarily mean export controls were violated, because the chips may be sitting in overseas data centers like those in Malaysia, where Epoch estimates 275,000 Nvidia accelerators were imported in 2024 and 2025. In other words, the debate has shifted from “who owns the chips” to “who is using the chips,” and the paths to compute and the lines of compliance are getting blurry. The supply chain and policy risk behind the models and APIs we use can change at any time.

Before picking an open-weight model, I’d add license terms and sourcing compliance to your evaluation checklist, and keep a swappable backup model ready.

📅 Source Info

🔗 Further Reading