Nvidia Blackwell B300s Detected in China Telecom Infrastructure Despite Export Bans

Thirty-two units of Nvidia’s Blackwell Ultra B300 graphics processing units have been detected within the infrastructure of China Telecom, a state-owned telecommunications giant. This discovery by Israeli security firm Lava highlights the persistent challenge of keeping high-performance silicon from reaching sanctioned buyers, even after strict U.S. export controls were implemented. The B300 series is specifically barred from export to China, meaning its presence in a national data center network suggests a continued gap between regulatory policy and physical reality.

The export restrictions mandate that these chips can only be sold to whitelisted buyers who agree not to resell or transfer them to sanctioned parties. Despite these safeguards, the flow of advanced hardware into the Chinese market remains a point of contention. Historical data indicates that in 2025, Chinese entities smuggled approximately $1 billion worth of Nvidia AI chips into the country in just three months. While the B200 was the primary target during that period, distributors had already begun preparing for the arrival of the superior B300 series, indicating a strategic shift in smuggling priorities.

Nvidia has responded to the findings with a measured tone, refusing to deny the presence of the chips while characterizing the incident as statistically minor. A spokesperson for the company stated that recent court cases have shown that the diversion of controlled U.S. products amounts to less than 0.5 percent of the total volume. The company described this leakage as a "drop in the bucket" when compared to the "ocean of domestic compute" China already possesses. Nvidia emphasized that with tens of millions of GPUs currently in service in approved nations, smuggling remains a non-factor that is effectively managed by governments on both sides of the trade divide.

Lava identified the specific chips using Nvidia’s own DCGM Exporter tool, a software utility designed for internal monitoring of server health and performance. The detection method raises questions about the security of such internal tools, which can often be accessed beyond the data center's local network and, in some instances, do not require password authentication. While it remains unclear exactly how the B300 units entered China Telecom’s server racks or their precise location within the network, their detection confirms the hardware is active and integrated into the state provider’s systems. Nvidia has acknowledged it is developing a dedicated tool to help locate its chips, though it has not yet been released to the public.

The demand for the B300 is driven by its significant performance advantages over its predecessor. Announced in March 2025, the chip offers 50 percent more memory and FP4 compute power than the B200. Nvidia marketed the B300 as being built for the age of reasoning, specifically citing its suitability for advanced large language models such as DeepSeek R1. Originally expected to ship by the end of 2025, the GPU quickly became a high-value target for those looking to enhance their AI capabilities outside of official channels.

Recent legal actions underscore the complexity of the smuggling networks. In August 2026, nine individuals in Taiwan were indicted for the illegal export of B300 GPUs to China. This prosecution followed days after Supermicro, a major hardware manufacturer, fired several employees following an internal investigation into smuggling activities. The illicit trade reportedly involved a Taiwanese server trading firm with whitelisted status, illustrating how authorized entities can sometimes be leveraged to move controlled goods across borders.

The scarcity and high demand for Nvidia’s silicon have also distorted the secondary market. Older models, such as the five-year-old A100 servers, have seen their prices triple, with units now selling for up to $82,000. As the crackdown on new chip smuggling intensifies, some operators are turning to older gaming GPUs, retooling them to handle AI workloads where possible. This trend highlights the economic pressure that export controls place on global compute resources, forcing buyers to seek alternatives or pay premium prices for legacy hardware that remains capable of supporting modern AI applications.