Study finds Blackwell fingerprints in China's leading open-weight models

A new research effort claims that the leading open-weight models built by Chinese companies were trained on Nvidia chips that the U.S. government restricts from shipping to China, TechCrunch reported on Friday.

The analysis was carried out by Will DePue, a former OpenAI researcher who leads the nonprofit Unemployment Labs. Rather than tracing supply chains or customs records, DePue examined model weights, the released parameters of open models from DeepSeek, Moonshot AI, MiniMax and Ant Group.

The premise is that weights retain traces of the chip configurations used to train them. DePue said he found compelling evidence that some of the most closely watched releases, including Moonshot's Kimi K3 models and DeepSeek's V4 models, bear fingerprints from training on Blackwell chips.

That matters because Blackwell GPUs are barred from export to China or Chinese-owned firms. Under current policy, the Trump administration allows Nvidia to sell the less powerful H200 to China, but only along a circuitous shipping route that lets Washington claim a 25% tariff.

Similar allegations have surfaced before. White House technology adviser Michael Kratsios said earlier this year, citing U.S. intelligence, that Moonshot had used export-banned Nvidia chips. Epoch AI, a nonprofit research organization, estimates that about two million advanced chips entered China in 2024 and 2025 through schemes designed to hide their final destination, with roughly 275,000 Nvidia accelerators imported to Malaysia over the same period.

The fingerprints are not proof of a violation. As TechCrunch notes, the finding does not mean these models were trained entirely on Nvidia hardware, nor that export law was broken. The chips could sit in an offshore data center, a route that has become popular enough to drive a compute boom in Malaysia.

That is precisely the loophole critics want closed. Sam Bresnick, a researcher at Georgetown's Center for Security and Emerging Technology, told TechCrunch it makes little sense to bar Chinese companies from chips in the mainland, Hong Kong and Macau while allowing them to "do whatever you want" from Malaysia. Hawks have long argued that controls should follow who uses a chip, not just who owns it or where it is installed.

The commercial chain is under scrutiny too. Megaspeed, the Singapore-based infrastructure company led by Alice Huang, and its subsidiary Speedmatrix have been investigated by both the U.S. and Singapore over suspected diversion; Bloomberg reported Friday that a company linked to Alice Huang was involved in smuggling banned Nvidia chips to a data center in Shanghai. In May, the short-selling research firm Culper published a report arguing that Nvidia had a "China problem" because a significant share of its revenue depended on Chinese demand.

What to watch next is policy, not just detection. About 600,000 H200 chips are reportedly stockpiled in Hong Kong awaiting shipment into China, and observers disagree on why they have not been cleared. Stanford research scholar Graham Webster told TechCrunch he doubts that withholding chips will spur domestic production, given the scale of demand. Under current rules, U.S. companies have limited obligations to track what customers do with controlled goods, and the fingerprints left in a model's weights are, in part, a byproduct of that gap.