An independent research outfit, UnemploymentLabs, has presented statistical evidence suggesting that Chinese AI models, including DeepSeek V4, Kimi K3.1, MiniMax-M3, and Ling-3.1-Flash, were trained using Nvidia Blackwell graphics processing units, which are subject to US export controls.
The researchers analysed weight initialisation patterns in the models, a process that spreads computational work across a GPU's shader modules (SMs). Different GPU families have unique SM counts, and for large weight sets, Blackwell and Hopper chips use 2,048 threads per SM. By simulating the initialisation on Blackwell and Hopper chips, the team found that the statistical distribution of weight values matched those in the Chinese models. When they used different SM and thread counts, the correlation broke down.
The researchers said it was not impossible that an AMD card could be set up to appear as two groups of 152 or 148 shaders. For DeepSeek V4, they ruled out that possibility by analysing the weight values, finding that AMD cards did not produce the same pattern.
The researchers are careful to hedge their findings, noting that the analysis does not reveal where the chips were located, how they were accessed, or prove that export controls were broken. As far as can be told, it is currently not illegal for Chinese companies to rent Blackwell hardware overseas, and US companies are not strictly prohibited from providing access, as shown by the recent Tencent-Oracle deal. The method is limited to models with open weights.
It is an open secret in the industry that Nvidia's Blackwell silicon finds its way into China through physical or virtual means.
Analysis
Why This Matters
- The research challenges the effectiveness of US export controls on advanced AI chips.
- It highlights the difficulty of enforcing technology restrictions when chips can be accessed through overseas rentals.
- The findings could prompt further regulatory action or renewed scrutiny of Chinese AI companies.
Background
The US government has imposed export controls on advanced AI accelerator chips from Nvidia and AMD to China, citing national security concerns. Nvidia's Blackwell GPUs are among the most powerful and are included in these restrictions. Despite this, the industry has widely regarded it as an open secret that Blackwell hardware reaches China through various indirect routes. The recent deal between Tencent and Oracle, which allows Chinese firms to rent Blackwell chips abroad, underscores the complexity of enforcement.
What to Watch
- Whether US authorities investigate the claims and whether any enforcement actions follow.
- Whether the model developers or Nvidia respond to the findings.
- Whether the method is applied to other open-weight models, potentially revealing more widespread use of restricted hardware.
Checked
12 claims in this story checked against 1 source: 12 supported.
Show each claim
- SupportedResearchers allege Chinese AI models were trained using Nvidia Blackwell GPUs described as banned.Latest from Tom's Hardware: “Researchers allege DeepSeek, Kimi, and other Chinese AI models were trained using banned Nvidia Blackwell GPUs.”
- SupportedUnemploymentLabs presented statistical evidence suggesting the named Chinese models, including DeepSeek V4, Kimi K3.1, MiniMax-M3 and Ling-3.1-Flash, were trained using Blackwell chips.Latest from Tom's Hardware: “UnemploymentLabs, an independent research outfit, has come up with strong evidence that those chips were used for training open-weight Chinese AI models, including DeepSeek V4, Kimi K3.1, MiniMax-M3, and Ling-3.1-Flash.”
- SupportedThe researchers say their analysis does not reveal where the chips were or how they were accessed, and does not prove export controls were broken.Latest from Tom's Hardware: “The scientists are careful to hedge their findings, noting that while their math checks out, it does not reveal where the chips were, how they were accessed, or prove that export controls were broken.”
- SupportedThe researchers analysed weight-initialisation patterns, a process whose computational work is spread across a GPU’s shader modules.Latest from Tom's Hardware: “As you'd expect, doing this for an entire model is a lot of computational work, so this process is spread out across all the SMs, further subdivided into threads.”
- SupportedDifferent AI accelerator families have distinct shader-module counts.Latest from Tom's Hardware: “First, different AI accelerator families have unique counts of shader modules (SMs), marginally analogous to CPU cores”
- SupportedFor large weight sets, Blackwell and Hopper use 2,048 threads per SM.Latest from Tom's Hardware: “For large weight sets, the thread count per SM is 2,048 for Blackwell and Hopper silicon”
- SupportedThe researchers simulated the initialization process using the relevant chips, software and thread counts, and found the resulting weight-value distribution matched those in the models.Latest from Tom's Hardware: “The researchers repeated a facsimile of the initialization process using the same chips, software, and thread counts. Using relatively simple statistical analysis, they spotted that their final statistical distribution of weight values precisely matched that of DeepSeek, Kimi, et al.”
- SupportedChanging the SM and thread counts caused the correlation to break down.Latest from Tom's Hardware: “As a test, they changed the SM and thread counts, and predictably, the correlation completely broke down.”
- SupportedResearchers said an AMD card could potentially be configured to appear as two groups of 152 or 148 shaders.Latest from Tom's Hardware: “The scientists say that it's not impossible that an AMD card could be set up to show up as two groups of 152 or 148 shaders”
- SupportedFor DeepSeek V4, the researchers ruled out that AMD possibility by analysing weight values and finding AMD cards did not produce the same pattern.Latest from Tom's Hardware: “Nevertheless, for DeepSeek V4, they also ruled out that possibility by analyzing the actual weight values themselves, and found that AMD cards still didn't produce the same weight value pattern.”
- SupportedThe method is limited to models with open weights.Latest from Tom's Hardware: “This identification method is limited to models with open weights”
- SupportedThe report says it is currently not illegal for Chinese companies to rent Blackwell hardware overseas, and US companies are not strictly prohibited from providing access.Latest from Tom's Hardware: “As far as we can tell, it's currently not illegal for Chinese companies to rent Blackwell hardware abroad, and even though Nvidia's grip is tightening, U.S. companies aren't strictly prohibited from doing so, as evidenced by the recent Tencent-Oracle deal.”
