32 B300 Chips Spotted on Exposed Monitoring Endpoints With Chinese IP Addresses

Key Facts

Cybersecurity firm Lava said on October 22, 2026, that it had spotted 32 Nvidia B300 graphics processing units, part of the Blackwell Ultra generation, through internet-exposed monitoring endpoints whose geolocated IP addresses appeared in China. The report was published by The Information, while Lava said the devices appeared through endpoints that exposed their data without authentication. The findings raise questions about the location of these accelerators and their supply routes, as shipments of this type of device to China are subject to strict U.S. export licensing requirements. But Lava stressed that the IP addresses do not determine the devices’ physical location or prove a violation. The company said, "The discovery raises questions about the location of the reported devices and their supply chain, but does not prove any violation of export controls."

Lava said the monitoring endpoints revealed GPU model names and unique identifiers without requiring a login. This makes it possible to identify the hardware type visible on those endpoints, but does not settle where it is physically located. According to the company, a connection could pass through a proxy, a virtual private network or a tunnel using an address in China, while the devices themselves are elsewhere. Lava therefore described geolocation based on an internet address as inconclusive in this case. Its report does not specifically link the observed units to China Telecom’s network, contrary to what appeared in the initial draft.

Lava’s figures place the 32 units within a broader group of 374 units across four Blackwell-generation models that appeared in its analysis of inventory associated with Chinese IP addresses. The group included 32 B300 SXM6 AC units, 78 RTX PRO 6000 Blackwell Server Edition units, 112 RTX PRO 5000 units with 72 gigabytes, and 152 GeForce RTX 5090 units. The company said about 86% of the 374 Blackwell units appeared in configurations with eight units. It said this pattern was consistent with compute servers, compared with individual workstations. These figures remain tied to what appeared on the monitoring endpoints analyzed by the company.

In multiple scans of public monitoring endpoints, Lava said it counted 4,724 unique Nvidia units across internet addresses geolocated in China. This figure covers a broader range than the four Blackwell models, which totaled 374, and is not limited to B300 units. The company also said that nearly 60% of the GPUs in its original global sample showed 0% utilization at the time of observation. It cautioned that the utilization reading was a snapshot in time, not a continuous measure of device performance or operation. These limits underscore the importance of distinguishing between a device’s data appearing on a monitoring endpoint and proof of its location or use.

The case comes as U.S. policy allows some chip export applications to China to be reviewed on a case-by-case basis, subject to requirements concerning power available to U.S. customers, compliance by the Chinese buyer and independent testing of the product inside the United States. On January 13, 2026, the U.S. Department of Commerce said the Bureau of Industry and Security would review export applications for Nvidia H200, AMD MI325X and similar chips under this approach. In April, Reuters reported that the price of a B300 server in China had reached about 7 million yuan amid strong demand and a tightening campaign against chip smuggling. Nvidia said at the time that the B300 was restricted from sale in China and that its partners must strictly comply with the rules. It also said the company does not provide service or support for systems that have been illegally diverted.

Data on China’s chip market show the scale of the shift in supply mix during the first half of 2026. Nvidia’s share of the data-center AI chip market reached 34%, after it shipped 678,000 accelerators in China, down 44% year on year. In the same period, Huawei shipped 572,000 accelerators, compared with more than 300,000 a year earlier, taking 28% of the market. Total shipments of AI accelerators in China reached 2.008 million units, while GPU shipments fell 26% to 991,000 units. Shipments of other accelerators rose nearly 80% to 1.016 million units, and IDC forecast total shipments would grow 19% in 2026, with GPU shipments falling 9%.

In the market, Nvidia shares closed at $212.21 on October 22, while Hong Kong-listed China Telecom shares ended the October 23 session at HK$4.325. The Chinese company’s shares had closed at 4.37 on October 22, compared with 4.66 on October 16. These prices provide context for the two stocks’ movements around the date the findings were published, but do not prove that the price changes were linked to Lava’s findings. Nvidia said in December that it had developed an optional software tool to estimate the country in which its chips are operating. The company had not yet launched the software at the time of The Information’s report.