Six of China’s leading open AI models carry the signature of Nvidia’s export-banned Blackwell chips, according to a forensic analysis of their published weights. The finding, released on 12 November, covers models from DeepSeek, Moonshot AI, MiniMax and Ant Group.
The Fingerprint in the Weights
The method reads the starting values of a model’s weights, set by a process called Philox. Philox splits the work across a chip’s processing units, so the pattern depends on the chip. Each seed and chip layout leaves a unique fingerprint that survives training.
Researcher Will DePue, a former OpenAI engineer who now runs the nonprofit Unemployment Labs, searched more than six trillion seed and layout combinations. The strongest matches run from 49 to 1,385 standard deviations above chance, far beyond what random luck could produce.
The result was checked on real hardware. PyTorch on a Blackwell B300 reproduced 461 of 469 rows from DeepSeek-V4-Flash, 445 bit for bit, while a Hopper H100 reproduced the older V3. A rented GB200 reproduced all thirteen preserved rows of Kimi K3.
Six Models, Four Labs
Two models match the 148-unit Blackwell B200 and B300 chips: DeepSeek-V4-Flash and MiniMax-M3. Four match the 152-unit Grace Blackwell GB200 and GB300: DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group’s Ling-3.1-flash. All six were released between April and November 2026.
The same labs’ earlier models read as Hopper, Nvidia’s previous generation. DeepSeek’s V3 from December 2024 and Moonshot’s Kimi K2.5 from January 2026 both match the older chips; the switch to Blackwell begins with DeepSeek V4 in April and Kimi K3 in July.
The code and the preserved weights are public. Transclew ran the published checks on 14 November and reached the same answer. The Register and TechCrunch both reported the findings on 13 November; neither Nvidia nor any of the four labs has responded publicly.
What It Does Not Prove
The fingerprint identifies the hardware that created a model’s starting weights, not the chips that did the training. A lab could generate weights on one system and train on another. The report is explicit: the evidence does not show where the chips were.
Blackwell systems can be rented from data centres outside China, so the finding proves neither smuggling nor an export-control violation. An AMD chip could mimic the pattern, but the report rules it out for DeepSeek-V4-Flash and calls it very unlikely for the rest.
The report estimates the five largest training runs used about 12,000 Blackwell chips together, worth more than half a billion dollars. DePue warns he is not confident in those figures, which rest on unpublished assumptions about run lengths and efficiency.
Our Read
This is the first evidence anyone can check: earlier claims rested on unnamed officials, leaving outsiders nothing to verify. And the offshore-rental loophole is the policy problem, since nothing stops a Chinese lab from renting the same hardware in Malaysia or Singapore.
The timing sharpens the story. Nvidia reports earnings on Tuesday under interim chief Colette Kress, days after Jensen Huang’s resignation amid the Megaspeed investigation. Export-control questions will be unavoidable on that call, and this report gives investors a checkable fact to press.
Why It Matters
Why it matters: export controls are the centrepiece of Washington’s effort to slow China’s AI advance, and this report suggests the centrepiece is leaking. If the chips are reaching the labs the rules were written to stop, the debate shifts from tightening controls to enforcing them.
Background
Background: Washington has alleged Blackwell use before without public proof. The Information reported in December 2025 that DeepSeek used smuggled chips, a senior official told Reuters in February that its latest model trained on Blackwell, and the White House said in July that Moonshot had access to GB300 chips.
Source: the full forensic report and The Register’s report.


0 Comments