SI Agent News
RumorOpen modelsChipsPolicySecurity

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.

Papercut shadow box: picture for the story 'Forensic analysis links Chinese AI models to restricted Nvidia Blackwe'
AI-made papercut shadow box for this story. No real people or products are shown.
$500 millionestimated cost for chip purchases and cluster build-out

The short version

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

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 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.

More AI news