32 Banned Nvidia B300 Chips Discovered Inside China Telecom's Network
Israeli security firm Lava identified thirty-two export-controlled GPUs within a state-owned Chinese carrier, prompting Nvidia to argue the leak is statistically minor.
By Neha Iyer ·
Lava, a security company based in Israel, has identified thirty-two Nvidia Blackwell Ultra B300 graphics processing units operating within the infrastructure of China Telecom. This state-owned telecommunications provider is a major player in the Chinese market, making the presence of these specific chips particularly notable. The B300 series is currently subject to strict U.S. export controls, which generally prohibit their sale to buyers in China unless they are on a specific whitelist and agree to keep the hardware within approved boundaries. The discovery raises immediate questions about how restricted hardware ended up in a Chinese data center that was not explicitly cleared for this latest generation of silicon.
The numerical scale of this specific incident is small, but it sits within a much larger context of persistent leaks. In 2025, reports indicated that approximately one billion dollars worth of Nvidia AI chips were smuggled into China over the course of just three months. While the B200 was the primary target during that period, distributors were already preparing for the arrival of the B300 series. The fact that these newer, more powerful chips have now appeared in China Telecom's racks suggests that the flow of restricted technology has not stopped, even as enforcement efforts continue.
Nvidia has responded to the news with a mix of acknowledgment and minimization. The company did not deny the findings reported by Lava but characterized the incident as a minor anomaly in the broader market. A spokesperson for Nvidia 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 company described this leakage as a "drop in the bucket" when compared to the vast amount of domestic computing power that China already possesses. With tens of millions of GPUs already in service in approved countries, Nvidia maintains that smuggling remains a non-factor in the overall landscape.
The technical method used to find the chips offers an interesting look into data center operations. Lava located the B300s using Nvidia's DCGM Exporter tool, a software utility designed for monitoring server health and performance. While intended for internal use by data center operators, the tool can sometimes be accessed from outside the local network. Notably, it does not always require password authentication, which may explain how external researchers were able to detect the hardware. Nvidia has previously acknowledged work on a dedicated tool to help locate its chips, but that solution has not yet been released to the public.
The motivation for Chinese firms to acquire the B300 is clear, given the chip's capabilities. Announced in March of last year, the B300 offers significant performance advantages over its predecessor, the B200. It includes 50% more memory and 50% more FP4 compute capacity. Nvidia marketed the chip as being built for the age of reasoning, specifically referencing advanced large language models like DeepSeek R1. The B300 was originally expected to ship by the end of 2025, but it quickly became a high-value target for those looking to expand their AI capabilities.
The path to these chips often involves complex logistics and occasional legal trouble. In August 2026, nine individuals were indicted in Taiwan for the illegal export of B300 GPUs to China. This legal action came shortly after Supermicro, a major server maker, fired several employees following an internal investigation into smuggling activities. The smuggling network reportedly utilized a Taiwanese server trading firm called Flying Tiget Tech, which held whitelisted status for purchasing GPUs. This case highlights the intricate web of partners and intermediaries that can facilitate the movement of restricted hardware.
The scarcity of these advanced chips has also distorted the secondary market. Demand for Nvidia's older silicon has driven prices to historic highs. For example, A100 servers, which are five years old, have seen their prices triple and can now cost up to $82,000. This price surge has led some organizations to retool gaming GPUs for AI workloads as a cost-effective alternative. The situation underscores the critical role these chips play in the global AI infrastructure and the lengths to which various parties will go to secure them.
Source: Tom's Hardware
