32 Restricted GPUs Uncovered Inside State-Run Telecom Network

An Israeli security firm has identified 32 restricted Nvidia Blackwell Ultra B300 graphics processing units operating within the network of China Telecom, a state-owned provider of telecommunications services. The discovery highlights the continued movement of advanced computing hardware into China, despite strict U.S. export controls that prohibit the sale of such high-performance accelerators to unauthorized parties.

The detection was carried out by Lava, which reported the presence of these specific chips to The Information. The Blackwell Ultra B300 is a powerful data center GPU designed for artificial intelligence workloads, but access to the technology is tightly regulated. Under current export rules, these chips may only be sold to whitelisted buyers who explicitly agree not to resell the hardware to sanctioned entities, a category that includes companies operating in China.

Despite these prohibitions, the flow of restricted hardware into the region has been difficult to halt. In 2025, reports indicated that Chinese firms had smuggled $1 billion worth of Nvidia AI chips into the country over a span of three months. At that time, the primary demand centered on B200 chips, though distributors were already preparing to move the newer B300-series processors as they became available for purchase.

Nvidia has responded to the specific report concerning China Telecom’s server racks by emphasizing the relatively small scale of unauthorized transactions compared to the company’s total output. A company spokesperson stated that governments, researchers, and media outlets have identified an insignificant diversion of controlled U.S. products, citing figures from recent court cases showing less than one half of one percent involvement.

The spokesperson described the unauthorized movement of goods as a drop in the bucket compared to the ocean of domestic compute China already has, adding that smuggling is a nonfactor blocked by governments on both sides of the globe. Nvidia did not deny the presence of the 32 chips, and the company maintains that with tens of millions of GPUs currently in service within approved countries, these incidents do not impact its overall operations.

The specific method by which the B300 chips entered China Telecom’s equipment remains unknown, and Lava could not confirm their exact location within the company’s vast network. However, the firm identified the accelerators using Nvidia’s DCGM Exporter tool. This software is intended for internal monitoring of server health and performance, yet it can sometimes be accessed from outside the data center network, occasionally without requiring password authentication.

The desire for this particular hardware is driven by its specifications. Announced in March of last year, the Blackwell Ultra B300 provides significant performance gains, offering 50% more memory and FP4 compute capabilities compared to the previous-generation B200. The chip was designed for the age of reasoning to support sophisticated AI large language models such as DeepSeek R1. Following its expected shipment date at the end of 2025, the hardware quickly became a target for illegal export.

Legal actions regarding the circulation of these chips have escalated recently. In August 2026, nine individuals were indicted in Taiwan over the illegal export of B300 GPUs to China. This action occurred shortly after Supermicro confirmed that it had terminated several employees following an investigation into smuggling. The illicit process involved the use of Flying Tiget Tech, a Taiwanese server trading firm that held whitelisted status for purchasing GPUs.

Demand for Nvidia silicon continues to influence pricing for both new and legacy equipment. Even as export restrictions tighten, the company’s older A100 servers have tripled in price and can cost up to $82,000 despite being five years old. The ongoing crackdown on Nvidia chip smuggling has even led some operators to attempt retooling gaming GPUs to handle AI workloads where possible.