Smuggled NVIDIA B300 GPUs found in China Telecom network

Detecting restricted hardware

A security firm called Lava recently discovered that 32 NVIDIA Blackwell Ultra B300 GPUs are currently operating inside the network of China Telecom. This finding is significant because the B300 is a flagship AI chip that is subject to strict U.S. export controls. The discovery suggests that high-performance hardware is still finding its way into Chinese infrastructure despite official trade barriers.

The investigation revealed that these restricted processors are being used by a state-owned telecommunications giant. While the U.S. government has placed limits on selling such advanced technology to certain regions, these units were found to be active in a major Chinese network. This indicates that smuggling remains a primary method for Chinese entities to acquire the latest AI hardware from the United States.

As Wccftech reported, the hardware was identified remotely rather than through a physical inspection of a data center. Because the units were found in a state-backed institution, the discovery raises questions about the scale of illicit hardware movement. It suggests that even the most restricted components are moving across borders through unofficial and illegal channels.

How the leak occurred

The security firm identified the B300 units by using NVIDIA’s Data Center GPU Manager (DCGM) Exporter tool. This specific software is designed for data center operators to perform internal server health diagnostics and monitor performance. However, the researchers found that the tool was accessible over the open internet without any password protection.

Because the diagnostic tool was exposed, the researchers were able to see the existence of the 32 B300 units on the China Telecom network. While the software confirmed the hardware was there, it did not provide the exact physical location of the servers. This type of oversight allows external observers to map out the hardware capabilities of private or state-owned networks.

The presence of these chips is notable because Chinese tech giants like Alibaba, Tencent, and ByteDance have been forced to rely on inferior hardware. Due to trade bans, these companies have often used chips like the RTX PRO 5500 for AI inference. The B300, by contrast, is a much more powerful architecture designed for demanding tasks like reasoning-based large language models.

Technical specs of the B300

The B300 is considered one of the most powerful AI acceleration architectures currently available. It features 288GB of HBM3E memory spread across eight 12-high stacks. The chip supports a data transfer rate of 8TB/s through an 8,192-bit connection, which is essential for training and running massive AI models.

Compared to previous generations of hardware, the B300 provides a 50 percent increase in memory capacity and FP4 compute power. These specific performance gains are the primary reason why the U.S. government imposes tight distribution restrictions on the chip. The hardware is specifically built to handle the heavy lifting required for modern artificial intelligence.

By providing such high compute power, the B300 allows for faster processing of complex data sets. For companies trying to build advanced AI, having access to this specific architecture is a major advantage. This is why the fact that 32 of these units are operating in China is a point of concern for international trade and security officials.

NVIDIA's response to smuggling

Despite the discovery of these smuggled units, NVIDIA does not appear to be deeply concerned about the incident. A company spokesperson told Tom’s Hardware that the diverted products are "a drop in the bucket" compared to what China already possesses. This suggests that the company views these specific 32 units as a small fraction of the total hardware already in the region.

The manufacturer also stated that fewer than 0.5 percent of their total shipped inventory consists of diverted U.S. products. This figure indicates that while smuggling is happening, it is not currently the primary way that the majority of the world's AI hardware is being distributed. NVIDIA continues to sell to whitelisted customers who commit to preventing diverted sales.

However, the discovery of these units in a state-backed institution like China Telecom is still noteworthy. It demonstrates that even a small amount of smuggled hardware can be highly impactful for state-level AI development. The persistence of these illicit networks shows that demand for high-end chips continues to outpace the effectiveness of current export policies.

Implications for the future

This incident highlights a growing trend where illicit networks continue to move cutting-edge hardware into restricted zones. As long as there is a trade ban in place, smuggling will likely remain the only alternative for entities seeking the best performance. The discovery of these 32 units may be just the beginning of what is being moved into other state-backed institutions.

For security professionals, this also serves as a reminder of the risks associated with exposing internal diagnostic tools to the public internet. If a company leaves its server health tools unprotected, it can inadvertently reveal its entire hardware stack to anyone with the right tools. This could lead to further intelligence gathering by competitors or state actors.

Moving forward, the focus will likely remain on how to close these loopholes in both physical shipping and digital monitoring. While the 0.5 percent figure suggests the current smuggling is limited, the high value of each B300 unit means these small numbers still matter. The battle over AI hardware supremacy will continue to involve both high-level policy and low-level smuggling tactics.