Illegal Nvidia Blackwell Ultra B300 GPUs found inside state-owned China Telecom's network

Nvidia Blackwell Ultra B300 GPUs are high-performance data-center chips designed to accelerate artificial-intelligence workloads. They process the calculations needed to train and run large AI models. Their importance comes from combining substantial memory with powerful computation, allowing data centers to handle more demanding systems.

The article says the B300 offers 50% more memory and FP4 compute than the previous-generation B200. Nvidia described it as built for the age of reasoning and mentioned models such as Pramana. The chip was expected to ship by the end of 2025, quickly making it highly sought after.

This performance explains why export restrictions target such chips and why illegal access matters. More capable GPUs can increase the computing capacity available for AI development. The article also notes that Chinese companies have sought older Nvidia chips and even retooled gaming GPUs for AI workloads when restricted hardware is difficult to obtain.

Lava, an Israeli security firm, reported finding 32 Nvidia Blackwell Ultra B300 chips within the network of China Telecom, a state-owned telecommunications company. The discovery matters because B300 chips are covered by U.S. export controls and are not supposed to be exported to China.

Lava identified the chips through Nvidia's DCGM Exporter tool. Data-center operators can install the software to monitor server health and performance. Although it is intended for internal use, the tool can sometimes be reached from outside the data center's network. The article also says it does not always require password authentication.

Lava could not confirm exactly where the chips are physically located inside China Telecom's network. Nvidia had previously confirmed it was working on a tool to locate its chips, but that tool had not yet been released. Nvidia did not deny Lava's findings, instead calling smuggling a nonfactor.

The reported number is 32 Nvidia Blackwell Ultra B300 GPUs. That is a small count compared with Nvidia's statement that tens of millions of GPUs are operating in approved countries. The comparison is central to Nvidia's argument that illegal diversion represents only a tiny fraction of the global installed base.

The chips were identified inside China Telecom's network, but Lava could not determine exactly where they were physically located. The firm used Nvidia's DCGM Exporter software, which monitors server health and performance and can sometimes be accessed beyond a data center's network. The software does not always require password authentication.

The figures show two different concerns. Numerically, 32 is far smaller than tens of millions. Strategically, however, B300 is a highly capable, export-controlled chip, so even a small confirmed presence demonstrates that restrictions have not completely stopped access. Nvidia described such smuggling as a drop in the bucket and a nonfactor.

U.S. export controls are government rules that limit the sale or transfer of specified products to certain countries, organizations, or users. In this case, the controls cover advanced Nvidia AI data-center chips. The article says B300 GPUs can be sold only to whitelisted buyers that agree not to export them to sanctioned parties, including buyers in China.

The restrictions matter because B300 is a particularly capable chip. Nvidia says it provides 50% more memory and FP4 compute than the earlier B200. The article describes it as one of the most powerful data-center GPUs available and says Chinese firms are keen to obtain it.

The controls are intended to prevent restricted chips from reaching China, but the article shows that enforcement is difficult. Lava found 32 B300 chips within China Telecom's network. Earlier reporting described $1 billion worth of Nvidia AI chips being smuggled into China in three months, showing that restrictions have not eliminated access.

The article does not identify exactly how these B300 chips entered China. It says the chips appeared inside China Telecom's network even though they were barred from export. That means the discovery exposes a gap between export rules and what can be physically located or enforced.

The article describes several routes associated with earlier or related smuggling activity. Whitelisted buyers can purchase restricted GPUs but must promise not to export them to China. In 2026, nine people in Taiwan were indicted over illegal B300 exports. Supermicro also fired employees after a smuggling investigation involving Flying Tiget Tech, a Taiwanese server trading firm with whitelisted purchasing status.

Distributors were reportedly preparing to sell B300 chips when they became available. The article does not connect any named route directly to China Telecom's 32 chips. Lava also cannot pinpoint their physical location, so the specific chain of buyers, servers, and intermediaries remains unresolved.

The discovery highlights an enforcement problem for technology companies and governments. Export controls can restrict official sales, but they do not automatically reveal where chips end up after purchase. A confirmed B300 presence inside China Telecom's network could increase pressure to audit buyers, suppliers, employees, and server operators.

The article shows why monitoring matters. Lava used Nvidia's DCGM Exporter to identify the chips, even though the tool was intended for internal use and can sometimes be accessed beyond a data center's network. Nvidia had said it was developing a tool to locate its chips, but that tool had not been released.

For governments, the finding may reinforce the need for cross-border enforcement and investigations. For Nvidia and hardware partners, it could increase scrutiny of whitelisted customers and distribution chains. The article does not say what action will follow, but it shows that smuggling remains an issue despite Nvidia's claim that it is insignificant.

Advanced GPUs matter because modern AI depends on enormous numbers of calculations. GPUs perform many calculations in parallel, helping data centers train and run large language models. Memory is also important because it holds model data and working information during those operations. More capacity can reduce bottlenecks and support larger workloads.

The article gives a concrete comparison: B300 provides 50% more memory and FP4 compute than B200. Nvidia called it built for the age of reasoning and referenced Pramana. In general, FP4 refers to a low-precision calculation format that can increase the amount of AI computation performed by suitable hardware.

At data-center scale, many GPUs can operate together, multiplying available capacity. That can affect which models an organization can develop or serve, how quickly they run, and how many users they support. The article's report of tens of millions of GPUs in approved countries helps explain the broader scale of this computing race.