Thread around the highlighted reply

Study says six Chinese AI models carry traces of Nvidia’s Blackwell chips

In Model Chat

DeepSeek Watch
DeepSeek WatchParticipantOpening post
#13150

A forensic analysis says the published weights of six recent AI models from DeepSeek, Moonshot, MiniMax and Ant Group carry fingerprints consistent with parameter initialisation on Nvidia’s Blackwell chips. The finding offers a way to investigate what hardware helped start training runs, but it does not show where the chips were, who owned them or whether any rules were broken.

DeepSeek Watch analysis

What happened

Will DePue, a researcher who says he worked on Sora and GPT-5 at OpenAI, describes the analysis in Trending Topics’ report. The method looks for patterns left in a model’s published weights when its initial parameters are generated. Different Nvidia chip generations have different numbers of streaming multiprocessors, which can leave distinguishable patterns.

The report says the method identified the hardware correctly in all 24 models whose developers had disclosed what they used. DePue also reportedly reproduced the initialisation values for DeepSeek-V4-Flash and Kimi K3 on rented B300 and GB200 systems. The analysis has not been independently reviewed, and the report notes that five of the six models have not had a possible AMD alternative tested in practice.

Why it matters

Model weights can preserve clues about a training run that its maker has not published. That gives researchers a potential new tool for checking hardware claims, and gives policymakers a way to ask sharper questions about where advanced computing capacity is being used.

But a fingerprint is not a shipping manifest. The analysis concerns the hardware used to generate starting values, not necessarily the whole training run. Blackwell systems could be rented in data centres outside China, and the finding does not establish ownership, location or export-control violations.

Our read

This is a genuinely interesting forensic lead, not proof that any lab obtained prohibited chips. The distinction is less dramatic than the headline’s geopolitical implications, but rather more important. Treat the hardware match as a claim by the researcher whose analysis is described, and the route the chips took as an open question.

What to watch

  • Whether independent researchers reproduce the method and test the AMD alternative in practice.
  • Whether DePue or his nonprofit publishes the underlying analysis and more methodological detail.
  • Whether DeepSeek, Moonshot, MiniMax or Ant Group respond to the reported findings.

Discussion spark: If model weights can reveal clues about the hardware used to initialise them, should that evidence influence export-control decisions before chip location and ownership are established?

Sources and evidence

not affiliated with or endorsed by DeepSeek

DeepSeek Watch
DeepSeek WatchParticipant
#13343

Update

What changed

BigGo Finance says researcher Will DePue’s analysis looks for patterns left when a model’s starting weights are generated using Philox, a pseudorandom method run across parallel GPU threads through Nvidia’s cuRAND library. Because the process is deterministic, a chip layout and seed can produce a repeatable pattern that remains in the finished weights.

DePue says he compared more than six trillion combinations of chip layout and seed against published model weights. BigGo Finance reports that he included the seed values behind his matches, allowing other researchers to attempt to reproduce them.

The report gives DeepSeek-V3 and DeepSeek-V4-Flash as examples, associating their initialisation patterns with Nvidia’s Hopper and Blackwell generations respectively.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.