Forensic Analysis of Chinese AI Models Reveals Blackwell GPU Fingerprints

A forensic analysis of publicly released AI models has identified numerical patterns in model weights that match the hardware configurations of NVIDIA’s export-controlled Blackwell GPUs. These findings suggest that developers including DeepSeek, Moonshot AI, MiniMax, and Ant Group have gained access to advanced computing hardware previously restricted from export to China, using it to initialize and likely train their latest frontier models.

The Forensic Method: Hardware Fingerprinting

The research identifies hardware usage by analyzing the initial values of a model's parameters (weights). Before training begins, these weights are initialized using a pseudorandom number generator (PRNG) called Philox. Because this process is deterministic, the final arrangement of weights depends on the specific hardware configuration—specifically the number of threads and streaming multiprocessors (SMs) used during initialization.

When PyTorch initializes weights on a GPU, it distributes the work across parallel threads. A chip with a specific number of SMs assigns these threads in a unique pattern. By recovering these initial values from released model weights, researchers can match the "fingerprint" of the model to the known thread-layout specifications of different NVIDIA GPU generations. The study found that while older models from these labs matched the fingerprint of Hopper-class chips (132 SMs), newer releases since April 2026 match the configurations of Blackwell B200/B300 (148 SMs) and GB200/GB300 (152 SMs) chips.

Key Findings and Hardware Adoption

The analysis identified six models from four companies that carry the Blackwell fingerprint:

  • DeepSeek: DeepSeek-V4-Flash and DeepSeek-V4-Pro
  • Moonshot AI: Kimi K3 and Kimi K3.1-Pro
  • MiniMax: MiniMax-M3
  • Ant Group: Ling-3.1-flash

Statistical analysis indicates these matches are significant, with evidence strengths ranging from 20 to 1,385 sigma, far exceeding what would be expected by chance. The report notes that while this fingerprint conclusively identifies the hardware used for initialization, it strongly implies that the same hardware was utilized for the subsequent training runs, which would require clusters of 1,500 to 3,500 Blackwell GPUs per model.

Limitations and Industry Context

It is important to note that this forensic method identifies the hardware used to generate initial weights, not necessarily the location of that hardware. As noted by some industry observers, Chinese companies may access restricted compute through cloud providers located in jurisdictions where such hardware is not export-controlled, such as Singapore or Malaysia. This practice would not inherently violate U.S. export regulations, which focus on the physical shipment of chips into China.

Furthermore, critics of the report have pointed out that the use of Blackwell hardware by these companies has been openly discussed in technical documentation and GitHub repositories, suggesting that the "fingerprint" discovery is less of a secret than a technical confirmation of known industry trends. Some commenters also questioned the assumption that initialization hardware must be the same as the training hardware, noting that labs might initialize models on one system and perform the bulk of training on another.

Implications for Export Controls

The findings highlight the difficulty of enforcing high-tech export controls in a globalized cloud-computing market. While the U.S. government has sought to restrict the flow of advanced AI hardware to Chinese firms, the ability to rent compute capacity remotely creates a "cat and mouse" dynamic. The report concludes that while it does not provide evidence of illegal smuggling or specific export-control violations, the widespread use of Blackwell-class hardware by Chinese frontier labs warrants significant public and regulatory attention.

Sources

Related