Forensic analysis links Chinese AI models to restricted Nvidia Blackwell chips
A U.S. research group found evidence that several Chinese AI companies used Nvidia's most advanced hardware despite U.S. export controls.

The short version
- Unemployment Labs identified Blackwell chip fingerprints in six Chinese AI models released since April.
- The report suggests roughly 12,000 Blackwell GPUs may have been used for large-scale Chinese AI training.
- Training costs for these projects are estimated to reach $500 million.
- The analysis relies on hardware traces rather than direct proof of physical smuggling or legal violations.
Why it matters
This suggests that current U.S. export controls may not be fully preventing China from accessing high-end computing power. It highlights the difficulty of policing global hardware supply chains and the persistence of Chinese investment in advanced AI.
Evidence of restricted hardware
A forensic analysis conducted by Unemployment Labs, a U.S. nonprofit research group, suggests that Chinese AI firms are obtaining Nvidia's most advanced hardware. The report indicates that several leading Chinese models were built using Nvidia's Blackwell graphics processing units (GPUs). These chips are currently restricted from being exported to China by the U.S. government.
The research focused on six specific systems released since April of this year. These included models from companies such as DeepSeek, Moonshot AI, MiniMax, and Ant Group. The analysis aimed to determine if these companies were successfully bypassing trade barriers to power their large-scale AI training projects.
According to the report, the researchers found numerical patterns and initialization fingerprints that are unique to the Blackwell architecture. These fingerprints serve as a digital signature left behind by the hardware during the training process. This suggests that the hardware was physically present during the creation of the models.
Identifying specific chip types
The researchers were able to distinguish between different types of Blackwell chips used by the various companies. For instance, DeepSeek's V4-Flash and MiniMax's M3 matched the specific compute characteristics of the B200 and B300 chips. These are high-performance components designed for intensive AI workloads.
Other models showed different hardware signatures that aligned with the GB200 and GB300 chips. These specific models are unique because they pair with the Grace central processing unit (CPU) to create a unified computing environment. This level of detail suggests a sophisticated level of hardware integration for the Chinese firms involved.
The Blackwell chip is Nvidia's cutting-edge AI accelerator, which was first launched in February of last year. It represents the current state of the art for the company until the Rubin chip arrived in July. Because these chips are not approved for export, their presence in Chinese models is a significant finding for trade officials.
The scale of the deployment
The report estimates that approximately 12,000 Blackwell GPUs have been deployed for large-scale AI training projects in China. This number represents a significant amount of computing power being funneled into the Chinese AI ecosystem. Such a large deployment suggests a coordinated effort to acquire and utilize these specific components.
The financial scale of these projects is also notable, with the report estimating costs reaching $500 million. This figure covers the purchase of the chips as well as the build-out of the necessary computing clusters. These costs reflect the high barrier to entry for training modern, high-performing artificial intelligence models.
While the U.S. government has allowed some exports of the H200 chip, it has strictly prohibited the sale of Blackwell or Rubin chips. The presence of Blackwell hardware suggests that the restrictions are being circumvented. The report notes that the findings do not directly prove how the chips were moved across borders.
Methods of circumvention
Experts have long suspected that Chinese firms might use several different methods to obtain restricted technology. One common theory is that companies remotely rent Nvidia compute power through data centers located in Southeast Asia. This allows them to use the hardware without physically possessing the chips within Chinese borders.
Another possibility is the direct smuggling of Nvidia chips through various workaround routes. These methods would involve moving the physical hardware into the country to be used in domestic data centers. The report highlights a recent scandal involving allegations of a spy using a personal relationship to smuggle advanced chips into China.
The Unemployment Labs report emphasizes that their findings are based on hardware traces left in the public models. The researchers did not verify the physical location of the chips or provide direct proof of specific legal violations. However, the presence of the fingerprints provides a strong indication of how the hardware is being utilized.
Implications for trade policy
The findings raise questions about the effectiveness of current U.S. export controls on high-end semiconductors. If Chinese firms can successfully integrate Blackwell chips into their models, it suggests that the current barriers are porous. This could lead to a reassessment of how the government monitors the flow of sensitive technology.
For the tech industry, this situation underscores the complexity of maintaining a global supply chain while adhering to national security requirements. Companies must navigate a landscape where hardware can be diverted or accessed through third-party intermediaries. The report serves as a warning that digital fingerprints can reveal the physical reality of hardware usage.
Moving forward, the focus will likely shift toward identifying the specific routes and entities facilitating these transfers. Policymakers may look for more ways to track the movement of chips through regional hubs. The discovery of Blackwell fingerprints in these models provides a clear starting point for that investigation.
What to watch
- Whether U.S. officials will launch new investigations into the specific data centers mentioned.
- If more Chinese models are analyzed to see if they also use the Rubin chips.
- How Nvidia responds to the findings regarding Blackwell chips in Chinese models.
Sources and links
- ReportingChosunbiz: the article we sourced
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 WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Chosunbiz · 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.
