Research claims China's top AI models show fingerprints of export-banned Nvidia Blackwell chips

Model weights point to Blackwell hardware

According to TechCrunch, the nonprofit research group Unemployment Labs examined the openly published weights of large language models from DeepSeek, Moonshot AI, MiniMax and Ant Group. Because trained weights retain traces of the hardware they were built on, the group's leader, former OpenAI researcher Will DePue, was able to look for chip-specific signatures. He reports compelling evidence that the most closely watched releases — including Moonshot's Kimi K3 and DeepSeek's V4 — carry signs of training on Blackwell GPUs, which the US bars from export to China or Chinese-owned firms.

The finding echoes earlier statements from Washington. White House technology adviser Michael Kratsios said earlier this year that Moonshot had used export-restricted Nvidia chips, citing US intelligence. Epoch AI, a nonprofit research organization, estimates that roughly two million advanced chips entered China in 2024 and 2025 through routing schemes designed to hide their final destination.

What the evidence does and does not prove

As TechCrunch is careful to note, the fingerprint is not proof of a violation. The models may not have been trained exclusively on Nvidia silicon, and the chips could have sat in offshore data centers rather than in China itself — an arrangement that has made Malaysia a hub for Chinese compute, with Epoch AI counting around 275,000 Nvidia accelerators imported there across 2024 and 2025. US companies currently face only limited obligations to track what customers do with controlled chips after delivery.

That gap is exactly what China policy hawks have criticized. Sam Bresnick of Georgetown's Center for Security and Emerging Technology told TechCrunch that it makes little sense to block Chinese access on the mainland while allowing open use just over the border in Malaysia. The policy backdrop is complicated by the Trump administration's decision to permit exports of the less capable H200, routed in a way that lets the US collect a 25 percent tariff.

Nvidia and the Megaspeed tangle

The research lands during a turbulent week for Nvidia. CEO Jensen Huang left the company after admitting a relationship with Alice Huang, an executive at Megaspeed — a firm implicated in smuggling Nvidia chips into China and alleged in reports to be connected to Chinese intelligence. Beyond the personal fallout, TechCrunch reports that Nvidia's potential corporate liability for export violations likely weighed on the decision to part ways.

Megaspeed, through its subsidiary Speedmatrix, has been central to the chip trade for years and is under investigation by both the US and Singapore. Bloomberg reported Friday that a company tied to Alice Huang was involved in moving banned Nvidia chips to a Shanghai data center. Nvidia and Huang had both vouched for Megaspeed's facilities, saying company officials had visited to verify they were legitimate.

The short-selling firm Culper argued in a May report that Nvidia's heavy reliance on Chinese demand — whether through remote data centers or outright smuggling — posed a serious risk. Its bear case rested not on a US crackdown but on Beijing's response: renewed commitment to domestic chips, which could erase billions in demand. Around 600,000 H200 chips are reportedly stockpiled in Hong Kong awaiting clearance, and observers have no consensus on the delay; Stanford scholar Graham Webster, for one, rejects the theory that withholding them is meant to force domestic production.

Why it matters

If the analysis holds up, export controls are failing at their central job: keeping the most advanced training hardware away from Chinese frontier labs. Even the most benign reading — chips legally hosted offshore — still means the intended capability gap is shrinking. The story also stretches beyond one company: Commerce Secretary Howard Lutnick reportedly risks losing his job over claims that his investigators missed suspicious buyers and grew too close to Nvidia. Expect growing pressure in Washington to shift toward controlling how chips are used, not just who owns them or where they are installed.