SI Agent News
DevelopingOpen modelsChipsSecurityResearch

Chinese AI models show traces of restricted NVIDIA Blackwell GPUs

Researchers detected numerical patterns in the weights of several Chinese AI models that suggest they were initialized on hardware subject to US export controls.

Documentary photo: picture for the story 'Chinese AI models show traces of restricted NVIDIA Blackwell GPUs'
AI-made documentary photo for this story. No real people or products are shown.
1,385 sigmastatistical significance of the detection

The short version

Why it matters

This discovery provides a new way to verify hardware usage by analyzing the public weights of AI models rather than relying on corporate claims. It offers a technical bridge between software outputs and the physical infrastructure used to build them.

Detecting hardware in weights

Researchers have identified numerical patterns in the weights of six AI models that match NVIDIA's Blackwell-generation GPUs. These specific chips are currently subject to US export controls, making their presence in Chinese AI development a point of significant interest. The findings involve models released since April 2026 by four major Chinese developers: DeepSeek, Moonshot AI, MiniMax, and Ant Group.

The evidence comes from an investigation by researcher Will DePue, which was later independently verified by researcher Bojie Li. By analyzing the publicly released weights of these models, the team was able to infer the specific GPU configurations used to initialize them. The results were consistent across both researchers, with Li using a completely different program to reach the same conclusions.

For instance, models like DeepSeek-V4-Flash and MiniMax-M3 matched a configuration with 148 streaming multiprocessors. Other models, including Kimi K3 and Ling-3.1-flash, matched a configuration with 152 streaming multiprocessors. These numbers correspond directly to the architecture of high-end Blackwell chips like the B200 or GB200.

How the fingerprint works

The detection method relies on how the PyTorch framework handles parallel processing on NVIDIA GPUs. When a model is first created, it uses a pseudorandom number generator called Philox to assign initial values to its weights. Because many threads compute these values simultaneously, the specific layout of the GPU threads determines where those numbers are stored in the weight matrix.

Even if a developer uses the same random seed, a different number of streaming multiprocessors (SMs) will change the arrangement of the data. This creates a unique numerical fingerprint that remains in the model's weights. While training usually updates these values through gradient descent, some data—such as rows for rare words—often remain almost entirely intact.

By finding these unupdated rows, researchers can regenerate the expected values based on a specific hardware layout. If the regenerated values match the published weights, it confirms the hardware configuration. The researchers found that the likelihood of these matches occurring by chance is extremely low, with some detections reaching a statistical significance of 1,385 sigma.

Limitations of the evidence

It is important to note that these fingerprints only identify the hardware used at the moment of initialization. They do not provide a complete picture of the entire training run. A company could theoretically initialize a model on one type of hardware and then move the weights to a different cluster or a different generation of chips to complete the full training process.

Furthermore, the technical data does not reveal who owns the hardware, where it is physically located, or how it was acquired. The numerical patterns are the same whether the chips are in a domestic data center or were brought into the country in violation of export controls. The weights contain no transaction records, destinations, or information about end users.

The analysis also cannot detect models initialized on CPUs or those using different random number generation methods. For example, models using the PaddlePaddle framework would not leave a GPU fingerprint. Therefore, the absence of a fingerprint does not prove that a company used domestic chips or avoided using Blackwell hardware.

The shift to Blackwell

The findings suggest a clear transition in the Chinese AI landscape from the older Hopper generation to the newer Blackwell generation. Previous reports from DeepSeek and Moonshot AI confirmed the use of H800 chips, which are the export-compliant version of the Hopper generation. However, the new detections in models released in mid-to-late 2026 point toward more advanced hardware.

Some models show a mix of hardware, such as Kimi K2.7-Code, which appeared to be a merge of weights from different models. This specific model showed a split between the older Hopper layout and the new 148-SM layout. This suggests that companies may be in a transitional phase, moving their infrastructure from one generation to the next as they scale.

Other companies, such as Alibaba, Zhipu, and Tencent, still showed Hopper-generation fingerprints in the models examined. This indicates that the move to Blackwell is not happening simultaneously across the entire industry. Instead, the timing and access to advanced GPUs appear to vary significantly from one company to another based on their specific needs and procurement capabilities.

Estimating the scale

Researchers attempted to estimate how many GPUs were required to train these models based on their parameter counts and training data. Based on these calculations, the five final training runs likely required between 7,000 and 22,000 GPUs. These estimates are based on assumptions about training duration and computational efficiency, which are difficult to verify with certainty.

The required GPU count varies greatly depending on how efficiently a company utilizes its hardware. For example, Microsoft reported using 16,000 GB200s for 38 days to train a specific model, while Mistral used 3,800 Grace Blackwell GPUs for a different project. These different paces mean that the estimated 12,000 GPUs for the Chinese models are not confirmed numbers of owned hardware.

The most reliable takeaway from this research is the confirmation of the initialization environment. While the total number of chips used for the full training run remains an estimate, the presence of Blackwell-compatible fingerprints in the public weights is a significant technical milestone. It provides a method for verifying hardware access that does not rely on any statements made by the developers themselves.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by XenoSpectrum · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news