Six Chinese AI Models Show Blackwell GPU Fingerprints

A forensic analysis officially released on November 12, 2026 by Will DePue's Unemployment Labs reveals that six Chinese AI models carry mathematical fingerprints consistent only with training initialization on export-controlled Nvidia Blackwell GPUs. The investigation identifies DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax M3/M3.1-Flash, and Ant Group's Ling-3.1-flash as containing weight patterns matching Blackwell B200/B300 and GB200/GB300 chips with 148 or 152 streaming multiprocessors. Statistical confidence levels range from 49 to 1,385 sigma across the six models.

Release Date and Technical Evidence

Officially released on November 12, 2026, the Unemployment Labs report provides the first publicly documented technical evidence of Blackwell hardware access by Chinese AI developers. While U.S. officials and press reports had previously alleged that DeepSeek and Moonshot AI were using Blackwell chips, no technical proof existed until this forensic weight analysis. The evidence comes entirely from the models' published weights rather than from leaks or insider accounts, making the findings independently verifiable by other researchers.

The analysis recovers traces of the models' initial weight values and identifies the hardware configurations that reproduce them. For DeepSeek-V4-Flash and Kimi K3, the researchers confirmed the match on physical Blackwell chips. The fingerprint method identifies how weights were created and supports, without proving, the inference that Blackwell also performed the training. Specially configured AMD hardware, which is also restricted from export to China, is the only alternative identified, but researchers have excluded it for DeepSeek-V4-Flash and consider it very unlikely for the other five models.

How Weight Fingerprinting Works

The forensic method exploits a fundamental characteristic of GPU-based model initialization. Before training begins, AI model weights require initial values generated by a pseudorandom number generation algorithm called Philox. PyTorch, the framework used in most major AI models, divides this initialization work among parallel workers called threads on the GPU through Nvidia's cuRAND library. Each thread generates its own sequence of numbers and fills particular weight positions.

The crucial insight is that thread count determines where values are placed. On Nvidia chips examined, Hopper H100 SXM, H200, and H800 use 132 streaming multiprocessors and 270,336 threads. Blackwell B200/B300 use 148 SMs and 303,104 threads. Blackwell GB200/GB300 use 152 SMs and 311,296 threads. A correct seed with the wrong thread count produces the right numbers in wrong positions. Only the correct combination recreates the original arrangement, making the thread count a hardware fingerprint.

Some weights reserved for symbols never encountered during training remain untouched, preserving their starting values almost exactly. Even weights that training changes can retain faint traces of starting values. The analysis compares thousands of values across both preserved and trained weights to identify the seed and thread count that best reproduce surviving patterns. Researchers then test combinations on weights kept out of the initial search to validate findings.

Scale and Investment Implications

The report estimates cluster costs of $150-200 million per lab and over $500 million combined across the four Chinese companies. Training runs consumed 500-3,000 Blackwell GPUs for periods of two to three months, with DeepSeek-V4-Flash requiring six to eight weeks. Table data in the report shows training compute calculated as approximately six times active parameters multiplied by training tokens, divided by chip throughput at 35 percent utilization.

DeepSeek disclosed its token counts directly, while researchers estimated token counts for other models from weight analysis. The scale indicates substantial infrastructure investment and sustained access to export-controlled hardware. The Blackwell fingerprints appeared in models released since April 2026, whereas earlier models from the same companies matched Nvidia's older Hopper generation, suggesting a hardware transition timeline.

Export Control Questions

Notably, the forensic findings provide technical evidence of access to advanced computing hardware but do not establish where the chips were located, how access was obtained, or whether export controls were violated. Blackwell B200/B300 and GB200/GB300 chips require U.S. export licenses for shipment to China, with no licenses reported as granted. The investigation identifies hardware signatures without determining the physical location of training infrastructure or the chain of hardware procurement.

The forensic method only works for certain models. It requires published weights, so closed models cannot be checked. It also needs a recognizable, hardware-dependent pattern from the software that generated starting weights. PyTorch on Nvidia hardware through cuRAND provides this pattern, but not all model architectures or training frameworks leave recoverable traces. The six identified models represent a subset of Chinese AI development where weight publication and initialization methods enabled forensic analysis.

What This Means

The November 12, 2026 release of this forensic evidence shifts the export control discussion from allegations to technical documentation. Statistical confidence levels reaching 1,385 sigma for some models provide near-certainty about hardware initialization patterns. The findings demonstrate that published model weights can serve as forensic artifacts, potentially enabling broader monitoring of advanced chip access across global AI development. For policymakers, the evidence highlights enforcement challenges when hardware access leaves mathematical rather than physical traces. For AI developers, the analysis shows that training infrastructure choices create permanent signatures in model weights, with implications for hardware provenance verification and supply chain transparency.