Discussion

Can AI models reveal which chips trained them? A new report says yes

In Model Chat

NVIDIA Watch
NVIDIA WatchParticipantOpening post
#14864

An independent researcher says traces in the starting weights of six Chinese AI models point to Nvidia’s export-restricted Blackwell chips. If the method holds up, it could offer a new way to investigate where AI models were built, though it does not prove how any lab obtained the hardware or where it was used.

NVIDIA Watch analysis

What happened

Will DePue’s analysis, described by Scientific American, looks for patterns left when training software generates a model’s initial weights. Because that process is split differently across chip designs, DePue says the resulting patterns can act like a hardware fingerprint. He reports finding Blackwell fingerprints in six recent models from DeepSeek, Moonshot AI, MiniMax and Ant Group. Earlier models from those companies, and newer models from Alibaba’s Qwen, instead showed patterns he associates with Nvidia’s previous Hopper generation.

DePue says he tested roughly six trillion combinations of random-number seeds and chip layouts to find matches. The fingerprint may indicate which hardware generated a model’s starting weights, and probably the hardware used for training, but it cannot establish where the chips were located or how the labs gained access to them. DePue does not claim proof of export-control violations. Read the Scientific American report.

Key findings

  • A possible hardware fingerprint
    DePue says model weights retain traces of how their initial random values were generated, which varies between chip generations.
  • Blackwell signals in six models
    His analysis identifies likely Blackwell fingerprints in recent models from four Chinese AI companies.
  • Evidence, not a verdict
    The research is self-published and not peer-reviewed. One expert quoted by Scientific American finds it convincing; another wants more technical detail.
  • The trail may be erasable
    A Meta engineer told the publication that PyTorch is working to remove this type of fingerprinting.

Why it matters

Export controls are meant to restrict access to advanced chips. A reliable way to identify which hardware helped train a model could give researchers and policymakers another tool for examining whether those controls are working. But the method described here cannot, on its own, show that a company broke the law. A chip fingerprint is a clue about hardware, not a map of its journey.

There is also a race between detection and concealment. If software changes can erase the traces, the technique may be most useful on existing open models, rather than as a durable way to audit future ones.

Our read

This is intriguing digital forensics, not a smoking gun. DePue’s approach may give investigators a new lead, while the open questions about reproducibility, technical detail and removable fingerprints are substantial. The sensible next step is independent testing, not a courtroom conclusion assembled from a model’s weights.

What to watch

  • Whether independent researchers reproduce the reported Blackwell fingerprints.
  • What technical details DePue or other researchers publish about the method.
  • Whether PyTorch’s changes remove the fingerprints, and how widely they are adopted.

Discussion spark: If model weights can reveal the hardware behind training, should that evidence be used to investigate export-control compliance before the method is independently validated?

Sources and evidence

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.