Six Chinese AI models carry the fingerprint of Nvidia Blackwell chips, one analyst reports
The method reads the random starting weights a chip leaves behind. It shows where a model began, not where it ran.
Will DePue of Unemployment Labs published a method on 12 November 2026 that infers which Nvidia chip generated a model’s starting weights. PyTorch fills those weights with random numbers on the GPU, split across threads whose count depends on the chip’s number of streaming multiprocessors (SMs): 132 SMs on Hopper chips such as the H100, 148 on Blackwell B200 and B300, 152 on GB200 and GB300. Some weights, such as those for unused tokens, barely change in training, so the seed and thread layout that reproduce them can be recovered from published weights.
On six Chinese models the match points to Blackwell: DeepSeek V4-Flash (148 SMs), DeepSeek V4-Pro (152, a partial match), Moonshot’s Kimi K3 (152), Kimi K3.1-Pro (same seed as K3), MiniMax M3 (148) and Ant Group’s Ling-3.1-flash (152). The strongest matches reach 1,385 sigma for V4-Flash and 305 for Kimi K3, against a bar of about 8 sigma for a chance match across roughly six trillion combinations. DeepSeek-V3, Kimi K2 and K2.5, MiniMax M2.x, Qwen and GLM read as Hopper, though a Hopper reading cannot tell an H800 from an H100 or H200.
The author checked the method on real hardware: unmodified PyTorch on a B300 reproduced 461 of 469 selected rows of DeepSeek-V4-Flash, and a rented GB200 reproduced all 13 preserved rows of Kimi K3. All 24 models with disclosed hardware agree, including Llama 3.1 (Hopper) and Nvidia’s Nemotron-3 Ultra (152 SMs). Some Chinese models leave no usable signature, among them Meituan’s LongCat-2.0, DeepSeek-V4.1-Flash and Baidu’s ERNIE-5.5.
What it means
US export controls bar Nvidia’s Blackwell chips from China, so the readings are evidence that recent Chinese frontier models were at least started on hardware those rules cover. The report does not say how the chips were obtained, and the author does not claim a violation.
Caveats
- One author published this and it is not peer reviewed. The code is public and runs on a laptop in seconds, so others can check it.
- The method identifies the chip that generated the starting weights, not every chip used in training. A lab could initialise on one system and train on another.
- The report does not establish where the chips were located, how access was obtained or whether export controls were broken. The author writes that the fingerprints support Blackwell use but do not prove it.
- AMD’s MI300X is the one other chip that could produce similar patterns. The author rules it out for V4-Flash with a per-value test and calls it very unlikely for the other five.
- Kimi K2.7-Code reads as a mix of about 80% Hopper and 20% 148-SM weights, which the author takes to be a merged checkpoint.