source S-0045
Unemployment Labs: "Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs"
Full report retrieved (HTML archived). Self-published by the author; not peer reviewed. Contains: executive summary; Table 1 (lab/model/chip); main findings; method (Philox PRNG via cuRAND, thread layout = SMs × 2,048 threads: Hopper 132 SMs/270,336 threads, B200/B300 148/303,104, GB200/GB300 152/311,296); "Limits of our investigation"; hardware verification (H100, B300, rented GB200, physical MI300X); sigma statistics; appendices with every model read; a code repo (github.com/macrogrants/checkpoint-forensics — not retrieved here).
States a same-day revision: "Revised November 12, 2026. The first version of this report gave 2,000 to 3,000 Blackwell GPUs for the largest runs and $150 million to $200 million for the largest cluster."
Referenced by (15)
- source S-0044 Will DePue, X thread: "I have a major result to share" (Chinese labs on Blackwell)
- fact F-0042 Will DePue posts thread and report claiming Blackwell fingerprints in Chinese open-weight models
- fact F-0043 The DePue report states its own limits: initialization is not training, and location, ownership and legality are not established
- fact F-0044 The report lists no signature for DeepSeek-V4.1-Flash, in its main "Limits" section and appendix; the thread does not mention it
- fact F-0045 DePue revises his GPU-count and cost estimates upward on the same day
- fact F-0047 Reuters: a senior Trump administration official says DeepSeek trained its upcoming model on Blackwell
- fact F-0058 US imposes advanced-computing export controls on China (Oct. 7, 2022); Nvidia ships A800/H800 variants
- fact F-0064 BIS moves H200-class exports to China to case-by-case review
- claim C-0043 DePue: "conclusive forensic evidence" the top 4 Chinese labs built their newest models on Blackwell
- claim C-0044 DePue report: six models from four labs were initialized on Blackwell-class hardware (148/152 SMs); earlier models read as Hopper
- claim C-0046 DePue: the largest runs used thousands of Blackwell GPUs; over half a billion dollars, likely billions in total
- claim C-0050 DePue rebuttal: 24 of 24 known cases match; training jobs create weights on their own GPUs; renting a Blackwell would still mean using a barred chip
- claim C-0051 @dummerspast39: DePue "left out" that DeepSeek’s newest model, V4.1-Flash, showed nothing
- claim C-0053 @ariaurelium: the forensics proves "something that nobody is concealing" because DeepSeek publishes Blackwell kernels
- claim C-0069 DePue report: a DeepSeek engineer wrote "we don’t have SM100 GPUs" (June 2025); Ant Group’s public repo has GB300 benchmark logs (Mar. 31, 2026)