News

NVIDIA Blackwell Fingerprints Found in Six Chinese AI Models, Researcher Says

A forensic study links six Chinese models to Blackwell initialization, while leaving chip location and export-control compliance unresolved.

By Samantha Reed Edited by Maria Konash Published: Updated:
NVIDIA Blackwell Fingerprints Found in Six Chinese AI Models, Researcher Says
Published model weights offer evidence of Blackwell access at four Chinese AI developers. Photo: Ivan Chumak / Pexels

Key Notes

  • Will DePue reports Blackwell initialization signatures in six models from four Chinese developers.
  • A B300 reproduced 445 preserved rows from DeepSeek V4 Flash bit for bit in his test.
  • The findings do not prove that all training used Blackwell or establish where the chips were located.

NVIDIA’s Blackwell chips may have left a trace inside some of China’s leading AI models. Researcher Will DePue says numerical patterns in publicly released weights provide evidence that four Chinese developers used the advanced hardware to create their models’ starting parameters.

His November 12 report identifies six models: DeepSeek V4 Flash and V4 Pro from DeepSeek, Kimi K3 and K3.1 Pro from Moonshot AI, MiniMax M3, and Ling 3.1 Flash from Ant Group. Blackwell exports to China require a US license, but the findings do not establish that any export rule was broken.

DePue previously worked at OpenAI, where he helped develop Sora, according to his biography. He now runs Unemployment Labs, an independent research group. The investigation uses published model files rather than accounts from people inside the companies.

How Starting Weights Reveal a Chip

Before training, a model’s parameters need initial values. Pseudorandom-number generators produce those values from a seed, making it possible to repeat the sequence. The research examines how PyTorch distributes that work across parallel GPU workers: different processor configurations place the numbers in different positions.

Some parameters change little or remain untouched during training, preserving a record of that arrangement. DePue’s published verification code rebuilds selected embedding rows under different chip configurations and compares the results with released weights. It also tests a changed seed as a control, helping distinguish a matching initialization from a superficial resemblance.

PyTorch’s own documentation notes that identical seeds do not guarantee identical results across platforms or software releases. Here, those differences become evidence to investigate rather than an inconvenience to eliminate. The method still depends on recognizable behavior in the software that produced the weights.

A Bit-for-Bit Match for DeepSeek

On a physical Blackwell B300, DePue says unmodified PyTorch reproduced 445 complete rows of DeepSeek V4 Flash’s preserved parameters bit for bit. An H100 did not reproduce those rows. The repository also reports a successful check of Kimi K3’s preserved rows on a rented GB200.

DePue describes the evidence for Blackwell initialization as conclusive for DeepSeek V4 Flash and strong for the other five models. He also considers it likely that the same hardware performed training, but the fingerprint directly identifies the creation of starting weights. A developer could initialize a model on one system and move it elsewhere.

What the Weights Cannot Establish

The report says specially configured AMD hardware can mimic the layout and explains why DePue ruled it out for V4 Flash. That alternative remains a limitation for the other five. The method cannot identify chip ownership or location, and the report found no readable fingerprint in DeepSeek V4.1 Flash.

The distinction is timely. On November 16, the US Bureau of Industry and Security announced a rule addressing remote access to advanced chips through overseas data centers, extending the policy focus beyond physical shipments. We reported on that change separately; it does not establish how these earlier models obtained their compute.

Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.

AI & Machine Learning, Cloud & Infrastructure, News