Junious Media NetworkMonday, November 16, 2026
Juni/Tech
Advertisement
AI

Forensic Report Links Chinese AI Labs to Restricted Nvidia Blackwell Chips

A former OpenAI engineer found hidden hardware fingerprints in AI models from four Chinese labs, suggesting export controls on Nvidia Blackwell chips are failing.

Key points

  1. A former OpenAI engineer published forensic evidence that DeepSeek, Moonshot AI, MiniMax and Ant Group trained AI models on Nvidia Blackwell GPUs restricted by U.S. export controls.
  2. The hardware fingerprint is locked into a model's starting weight values at initialization and survives the full training process, making it detectable in the published model.
  3. A separate Israeli security firm, Lava, found 32 Nvidia B300 Blackwell Ultra GPUs through exposed monitoring endpoints with IP addresses geolocated to China.
  4. At least four bills in Congress, including the AI Overwatch Act and the Cloud Security Act, are aimed at closing gaps in current chip export enforcement.

A forensic study by a former OpenAI engineer claims that at least four major Chinese AI labs used Nvidia Blackwell graphics processors to train models now available to the public. Those chips are covered by U.S. export restrictions. The report raises hard questions about whether current enforcement can keep America's most advanced AI hardware out of Chinese hands.

Will DePue left OpenAI earlier this year to start Unemployment Labs, an AI research consultancy. He published the analysis under the title "Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled Nvidia Blackwell GPUs." The report names DeepSeek, Moonshot AI, MiniMax and Ant Group as organizations whose published models carry hardware signatures pointing to restricted chips.

How Chinese AI labs leave hardware fingerprints in their models

Before training begins, an AI model must be set up with starting values called weights. Those starting values follow a pattern tied to the specific hardware that generated them. A pseudorandom number generation algorithm called Philox produces those values from a seed number. Because the algorithm always gives the same result for the same inputs, any given chip-and-seed combination leaves a predictable mark.

Using PyTorch, a common deep-learning framework, alongside Nvidia's cuRAND library for GPU-based number generation, DePue generated initialization patterns for specific GPU models. He compared those patterns against the released weights of Chinese AI models. That hardware fingerprint is locked in at initialization. It survives the full training process and stays readable in the published model.

DePue said "almost every AI model carries a hidden signature of the GPU that trained it," explaining that each chip-and-seed pair produces a unique pattern. The report includes specific seed values linking DeepSeek-V3 to Nvidia's Hopper architecture and DeepSeek-V4-Flash to Blackwell. Independent researchers can use those values to check the findings themselves.

Advertisement

Six models flagged across four Chinese AI companies

DePue tested over six trillion chip-and-seed pairings against the publicly released weights of open-source models. That work produced matches tying six Chinese AI models to Nvidia Blackwell hardware. The identified models include DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3.

Four of the matched models carry signatures consistent with Grace Blackwell hardware. Those are the GB200 and GB300 chips used in Nvidia's flagship liquid-cooled rack-scale systems. Those chips are among the most restricted items under current U.S. export rules.

Advertisement

DePue noted this marks the first time MiniMax and Ant Group have been publicly tied to Blackwell hardware use. DeepSeek has drawn prior scrutiny from U.S. regulators and policymakers. Adding two more companies to the list broadens the scope of concern. DePue also said multiple senior researchers, both independent and at frontier labs including OpenAI and Anthropic, confirmed the results.

A second firm found physical Blackwell hardware in China

DePue's model-weight analysis is not the only recent evidence. Lava, an Israel-based security company, separately reported finding 32 Nvidia B300 GPUs with IP addresses geolocated to China, uncovered through publicly exposed monitoring endpoints. The B300 is part of Nvidia's Blackwell Ultra generation and is subject to U.S. export controls.

Lava said those endpoints were part of monitoring systems intended for internal infrastructure management. Someone had left them open to public internet access. That discovery pointed to actual physical hardware inside China, not inferences drawn from software patterns.

The two sets of findings use completely different methods. One reads model weights; the other scans network endpoints. But both point to the same conclusion: rules restricting Blackwell chip exports to China are not preventing access, and Chinese AI developers appear to be training frontier models on some of Nvidia's most advanced processors.

Advertisement

Four bills in Congress aim to close the gaps

The DePue report comes as lawmakers push to strengthen export enforcement. At least four bills targeting China's access to advanced AI hardware are waiting for consideration. They are the AI Overwatch Act, the Chip Security Act, the Match Act and the Cloud Security Act.

House Speaker Mike Johnson supports tightening the existing restrictions, The Register reported. The bills reflect growing bipartisan concern that current rules carry too many loopholes. The Cloud Security Act specifically targets one of them: Chinese labs may be renting time on U.S.-based cloud systems to reach restricted chips without physically importing them.

Senators Jim Banks, a Republican from Indiana, and Elizabeth Warren, a Democrat from Massachusetts, pointed to former Nvidia CEO Jensen Huang's ties to a person described as a suspected Chinese spy. They named that relationship as a reason to advance the AI Overwatch Act. That bill would formally write the Trump administration's existing chip export rules into statute. The Trump administration said it is not certain the individual was a spy. Beijing has denied the allegation, and the FBI is investigating the matter.

Advertisement

Nvidia leadership turmoil adds to the pressure

The policy debate has intensified alongside upheaval at Nvidia itself. Jensen Huang resigned as CEO after revelations emerged about his relationship with Alice Huang. She founded Singapore-based Megaspeed International in 2023. The Register reported that U.S. authorities have opened a probe into that company over whether it facilitated chip transfers into China in violation of export rules.

Advertisement

Huang's departure drew fresh attention to how restricted hardware moves through global supply chains. The alleged use of a Singapore-based intermediary illustrates how third countries can function as transit points. That complicates enforcement efforts focused mainly on direct sales to Chinese buyers.

The case also highlights the pressure now on Nvidia as a company. Its most powerful chips sit at the center of both a geopolitical fight and a corporate governance crisis. Any new restrictions passed by Congress would directly shape how Nvidia can sell its products worldwide.

What this means for export controls and AI competition

The practical stakes are significant. U.S. export controls on Blackwell chips were designed, in part, to slow China's development of frontier AI models. The goal was to limit the hardware available for large-scale training. If Chinese labs are using the same class of chips as American ones, the competitive gap those rules were meant to create may be far narrower than policymakers assumed.

Both OpenAI and Anthropic have raised concerns that Chinese labs used distillation to absorb capabilities from American AI systems. Distillation is a process where a smaller model is trained to copy the outputs of a larger one. The U.S. Cybersecurity and Infrastructure Security Agency has also flagged that practice. Together with apparent Blackwell access, the picture suggests that neither hardware limits nor knowledge-transfer barriers are working as intended.

DePue's fingerprinting method may give regulators a new tool. Analysts could in theory screen published models for hardware signatures that reveal what chips were used to train them. This approach could act as a check that does not depend on supply-chain paperwork or physical inspections. Whether it can scale into a workable enforcement mechanism remains an open question for policymakers and researchers alike.

Advertisement

Sources