Forbidden AI Chips Found in Major Chinese Telecom Network
Security researchers discovered 32 export-controlled Nvidia processors in a state-owned data center, prompting the chipmaker to argue that such leaks are a minor anomaly.
By Chinedu Okafor ·
The intricate web of global semiconductor trade has long been a target for scrutiny, but a recent discovery has added a new layer of complexity to the debate over how advanced hardware moves across borders. Security researchers have identified a specific batch of high-end processors inside the infrastructure of a major state-owned telecommunications provider in China, raising immediate questions about the efficacy of current export controls.
An Israeli cybersecurity firm known as Lava reported that it detected 32 Nvidia Blackwell Ultra B300 chips within the network of China Telecom. These specific units are among the most powerful graphics processing units currently available for data center use, and they are currently barred from export to China by U.S. regulations. Under normal circumstances, these chips can only be sold to buyers who are on an approved list and agree not to resell them to sanctioned parties, making their presence in a Chinese state-owned network a significant anomaly.
This is not the first time that restricted hardware has been found in the region, suggesting that the enforcement of these rules remains a persistent challenge. Last year, reports indicated that nearly a billion dollars' worth of Nvidia AI chips were smuggled into China in just a three-month period. While the B200 series was the primary target at that time, distributors were already preparing for the arrival of the newer B300 series, which offered substantial performance improvements over its predecessor.
Nvidia has responded to the latest findings by acknowledging the report but characterizing the volume of smuggled hardware as negligible. A company spokesperson stated that recent court cases have shown that the diversion of controlled U.S. products is less than half of one percent of the total. The spokesperson argued that compared to the vast amount of domestic computing power China already possesses, these instances are merely a drop in the bucket, noting that tens of millions of GPUs are already in service in approved countries.
How exactly the B300 units ended up in China Telecom's server racks remains unclear, and Lava has not been able to pinpoint their precise location within the network. The chips were identified using a monitoring tool called the DCGM Exporter, which is designed to help data center operators track server health and performance. While this software is intended for internal use, it can sometimes be accessed from outside the data center's network and does not always require password authentication. Nvidia has confirmed it is developing a dedicated tool to help locate its chips in such scenarios, though it has not yet been released.
The desire for these specific chips is easy to understand, given their capabilities. Announced in March of last year, the B300 was marketed as being built for the era of advanced reasoning and large language models. It offers 50% more memory and significantly higher compute power for specific data types than the previous-generation B200. The chip was originally expected to ship by the end of 2025, and its high performance quickly made it a sought-after commodity in the market.
The pressure to acquire these processors has led to complex smuggling operations and legal actions. In August of this year, nine individuals in Taiwan were indicted for the illegal export of B300 GPUs to China. This followed an investigation by Supermicro, which fired several employees after discovering that a Taiwanese server trading firm with whitelisted status had been used to facilitate the transfers. The incident highlights the sophisticated methods used to move hardware through legitimate channels before it reaches its final, often unintended, destination.
The scarcity of these advanced chips has also driven up prices for older models, as companies seek alternatives for their AI workloads. Nvidia’s A100 servers, which are five years old, have seen their prices triple, with some units now selling for up to $82,000. In some cases, organizations have even resorted to reconfiguring gaming graphics cards to handle AI tasks. As the debate over chip smuggling continues, the findings in China Telecom’s network serve as a reminder that the flow of high-performance hardware remains a dynamic and often difficult process to control.
Source: Tom's Hardware
