Researchers found Nvidia Blackwell traces in six Chinese AI models.
The discovery raises questions about US chip export restrictions.
Overseas data centres may offer Chinese firms access to processors.
Despite US restrictions on advanced chip exports to China, some of the country's leading AI developers appear to have gained access to Nvidia's most powerful processors. Researchers have found numerical patterns linked to Nvidia's restricted Blackwell chips in AI models released by DeepSeek, Moonshot AI, MiniMax and Ant Group.
The findings offer new clues about how Chinese AI companies may be accessing advanced computing hardware that Washington wants to keep out of China's hands. They also raise questions about whether US export controls are keeping pace with the ways companies can obtain access to these chips.
A study by Unemployment Labs, led by former OpenAI researcher Will DePue, found Blackwell-related patterns in six models released by the four Chinese companies.
The findings do not establish where the processors were used, how the companies accessed them or whether any export-control rules were broken. But they provide technical evidence that these models were initialised on Blackwell hardware.
How Did Researchers Identify Nvidia's Blackwell Chips?
Unemployment Labs studied the internal data made public by Chinese AI companies. This data contains billions of numbers that shape how their AI models work.
AI models contain billions of numerical values called weights. These values help determine how a model processes information and generates responses. Before training begins, the weights are assigned initial values through a process known as random initialisation.
However, researchers found that the patterns created during this process can reveal details about the hardware used.
Their investigation focused on an algorithm called Philox, which generates random numbers using an initial input known as a seed. The arrangement of these numbers can change depending on the GPU's configuration, including the number of processing units it contains.
Nvidia's different chip generations have different numbers of streaming multiprocessors (SMs), which handle computing tasks.
The Hopper generation, including the H100 and H200, has 132 SMs. Blackwell's B200 and B300 have 148, while the GB200 and GB300 configurations have 152.
These differences leave identifiable numerical patterns in the initial weights of AI models.
Researchers worked backwards from the published weights to identify the hardware configurations that could have generated them. They then tested their findings against separate model weights and compared the results with known hardware configurations.
The investigation identified Blackwell fingerprints in six models: DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant Group's Ling-3.1-flash.
How Could These Chips Reach Chinese Companies?
The findings lend technical weight to concerns raised by senior US officials earlier this year. In July, White House technology adviser Michael Kratsios accused Moonshot AI of obtaining servers equipped with Nvidia's GB300 processors.
He also alleged that the company had accessed additional GB300 systems in Thailand, possibly to train its AI models. The allegations reflected Washington's concerns that Chinese AI companies were finding ways around restrictions on advanced chips as well as American AI technology.
A September analysis by nonprofit research organisation Epoch AI found that trade figures were consistent with more than $3 billion worth of advanced chips being smuggled into China through Malaysia between April 2024 and June 2025, Indian Express reported.
Why Did the US Restrict AI Chip Exports?
In 2022, Washington started blocking China from buying some of the most powerful computer chips made by US companies. The restrictions were introduced under President Joe Biden, with the aim of keeping advanced technology out of China's hands.
The US was particularly concerned that Beijing could use these chips to build stronger AI systems and develop military technology.
Nvidia's processors became a major focus of the restrictions. Its H100 and H200 chips, followed by the Blackwell range, are among the hardware used to train powerful AI models. They can carry out huge numbers of calculations in a short time.
The restrictions did not end China's demand for these chips. Instead, they created a market for companies and traders who could arrange supplies through routes outside the US export system.
How Did Megaspeed Feature in the Chip Supply Network?
Megaspeed, a cloud computing company based in Singapore, has a Malaysian subsidiary called Speedmatrix. The subsidiary rents space at data centres in Malaysia.
A Wall Street Journal investigation found that Megaspeed placed orders worth nearly $12 billion for advanced Nvidia chips on behalf of Chinese companies between March 2024 and June 2025.
The orders covered several hundred thousand processors, according to the records reviewed by the newspaper. However, the documents did not confirm how many chips were actually delivered or where they all came from.
The arrangement reportedly involved several companies and middlemen. While Singapore handled the business and payments, servers fitted with Nvidia chips were sent to data centres in Malaysia.
US export controls largely focus on restricting the sale and transfer of advanced chips. But as AI computing becomes increasingly dependent on international data centres and cloud services, controlling physical shipments may not be enough.
Even if a Blackwell processor remains outside China, a Chinese AI developer could potentially benefit from its computing power through remote access.
The findings raise an important question about the effectiveness of US restrictions. Washington may be able to control where advanced Nvidia chips are shipped, but preventing Chinese companies from accessing their computing power through overseas facilities is a much harder challenge.
For the United States, the next challenge may therefore be to monitor not just who buys advanced AI chips, but also who ultimately uses the computing power they provide.



