Researchers allege DeepSeek and Kimi models were trained on Nvidia Blackwell GPUs

What the researchers claim

UnemploymentLabs, an independent research group, says it has found evidence that several Chinese open-weight AI models were trained on Nvidia Blackwell chips. According to Tom's Hardware, the models named include DeepSeek V4, Kimi K3.1, MiniMax-M3 and Ling-3.1-Flash.

The claim matters because the United States restricts exports of high-end Nvidia and AMD AI accelerators to China. Blackwell is Nvidia's most recent chip family, and its sale to Chinese customers is barred under those controls.

The researchers themselves hedge the findings. Their analysis suggests which hardware was used, but it does not show where the chips were, how they were accessed, or whether any export rules were broken.

How the analysis works

The method relies on how AI models are initialised before training. Weight values, which determine how strongly a model reacts to each input, are set through a pseudo-random process using a tool called cuRAND. The same starting seed produces the same values every time.

That randomisation work is spread across the shader modules, or SMs, on a GPU. These are processing units roughly comparable to CPU cores. Nvidia's Hopper chips have 132 of them, Blackwell B series chips have 148, and Blackwell GB chips have 152.

The researchers reproduced the initialisation process on the same chips and thread counts, then compared the statistical distribution of weight values with those of the published models. The distributions matched. When they changed the SM and thread counts, the correlation disappeared.

What the evidence shows, and its limits

The analysis let the team identify which family of Blackwell or Hopper chips was used for each model, and estimate an approximate number of cards. It cannot single out the exact card variant, because it depends only on SM counts and usage patterns, not on memory capacity or bandwidth.

The method only works on models with openly published weights. It also assumes the widely used PyTorch software stack on Nvidia hardware.

AMD hardware is a possible alternative explanation. The researchers say an AMD card could in theory be configured to appear as two groups of 148 or 152 shaders. For DeepSeek V4, they say they ruled this out by analysing the weight values themselves, which did not match an AMD pattern.

Is it illegal?

Not necessarily, according to Tom's Hardware. It is reportedly not currently illegal for Chinese companies to rent Blackwell hardware located abroad, and US companies are not strictly prohibited from providing it, as shown by a recent Tencent-Oracle arrangement.

The researchers note that their mathematical results stand, but the analysis cannot prove that export controls were violated. It shows a strong correlation with Nvidia hardware, not the circumstances of its use.

Why it matters

If confirmed, the findings would suggest that export controls have not stopped frontier Chinese labs from accessing the latest Nvidia hardware, at least for the models examined. That bears on the ongoing policy debate in Washington over how effective chip restrictions are.

The method itself is also notable. It offers a way to audit open-weight models for the hardware that produced them, though it applies only to models whose weights are published. Readers should treat the allegations as unverified claims from an independent group, not established fact.