Analysis links six Chinese AI models to NVIDIA Blackwell GPUs

Identifying the hardware

A report by Unemployment Labs has identified specific numerical patterns in six AI models released by Chinese companies. These patterns, known as initialization fingerprints, match the architecture of NVIDIA Blackwell GPUs. The models in question come from DeepSeek, Moonshot AI, MiniMax, and Ant Group.

The analysis found that these specific models carry the signature of Blackwell chips, while earlier models from the same companies matched NVIDIA Hopper hardware. This shift suggests a transition to more powerful, export-controlled hardware for the newest releases. The report specifically ties two models to B200 or B300 chips and four others to GB200 or GB300 chips.

According to the report, the evidence points to the hardware that generated the models' starting values. While the weights do not show who owned the chips or how the companies obtained them, they support the inference that the same hardware performed the training. This discovery comes as U.S. officials have previously alleged that these specific companies were using Blackwell chips.

How the fingerprint works

The detection method relies on how software fills billions of parameters with numbers before training begins. In the setups examined, a specific generator spreads work across GPU threads. Because different chips have different numbers of streaming multiprocessors, they assign the same sequence of numbers to different positions in the model's weights.

If a researcher uses the wrong thread count for a specific chip, the values end up in the wrong places. By searching through various combinations of seeds and hardware layouts, the researcher identified the exact match for these models. For example, using a specific seed on an NVIDIA B300 reproduced preserved values in the DeepSeek-V4-Flash model.

The report notes that some weights remain unchanged during training, while others retain a weak statistical trace of their initial values. The researcher verified these results on physical hardware to ensure accuracy. The method is highly specific, as it requires published weights and a hardware-dependent initialization pattern to be readable.

Scale and costs of training

The report estimates that each of the five training runs required between 500 and 3,000 Blackwell GPUs. These runs are estimated to have taken between two and three months to complete. Based on a cost of about $70,000 per GPU, the largest of these clusters would have cost between $150 million and $200 million.

When looking at the five clusters combined, the total cost would exceed half a billion dollars. This suggests that the total spending by these labs on export-controlled GPUs could reach into the billions of dollars. These figures provide a sense of the massive capital investment required to produce these frontier models.

The analysis also looked at other companies to see if the trend was universal. It found that newer models from Alibaba, Zhipu, Tencent, Xiaomi, and StepFun still read as being trained on older Hopper hardware. Meanwhile, Meituan reported training its LongCat-2.0 model on AI ASIC superpods, which showed no NVIDIA fingerprint at all.

Limits and implications

While the evidence is strong, the report acknowledges several limitations to the fingerprinting method. It cannot identify closed models, models initialized on standard processors, or models trained in frameworks that do not use the specific patterns examined. A lab could also move weights to different systems or continue training on different hardware after the initial setup.

The report also considered whether the patterns could be explained by AMD MI300X hardware. However, a value-level comparison for one model favored NVIDIA in 363 out of 377 differing values, making the AMD explanation unlikely for the other five models. The researchers found that the fingerprint is a reliable way to identify the hardware used for initialization.

Ultimately, the report states that the evidence concerns access to advanced computing hardware rather than a legal determination. It does not establish that export controls were actually violated. It simply confirms that the hardware used to create these models matches the specifications of restricted NVIDIA chips.