Report Finds 32 Nvidia Blackwell Ultra B300 Chips in China Telecom’s Network, with Lava Identifying the Chips Through Nvidia’s DCGM Exporter
Nvidia did not deny that the chips were found, but said that the diversion of controlled products accounted for less than 0.5% in the latest case, compared with the tens of millions of GPUs already operating in approved countries.
How Were B300 Chips Found in China Telecom’s Network?
Lava identified the chips through Nvidia’s DCGM Exporter, which is used to monitor server health and performance in data centers. The tool can sometimes be accessed from outside a network and does not always enforce password protection.
The report said that 32 B300 chips were found in China Telecom’s network, but it has not yet been confirmed where in the system the chips were located or how they were brought in.
When Export-Banned Chips Appear in a Chinese Data Center
Lava said it found 32 Nvidia Blackwell Ultra B300 chips in the network of China Telecom, a Chinese state-owned telecommunications company, even though the chips are subject to U.S. export restrictions.
Authorized buyers must agree not to transfer the chips to sanctioned parties, including China. The discovery of the chips in a Chinese data center therefore suggests that export controls still have loopholes, although it remains unknown how the chips entered the country.
Where Does the B300 Fit in Nvidia’s AI Chip Family?
The B300 is the successor to the B200 in the Blackwell Ultra family and is positioned as a much more powerful data center GPU than its predecessor. Its key features include 50% more memory and improved FP4 compute capabilities.
The reason Chinese companies want this model is that the B300 is suited to more complex AI workloads. Nvidia has positioned it for the “age of reasoning,” citing models such as DeepSeek R1 as examples. This has made the chip highly sought after in the market.
How Does the B300 Differ From the B200?
The B200 was the previous generation and was once one of the most sought-after Nvidia AI chips in 2025. The B300 is the model being pursued and investigated in connection with exports to China.
| Factor | Nvidia B200 | Nvidia Blackwell Ultra B300 |
|---|---|---|
| Model status | Previous generation | Sought-after model |
| Performance | Previous generation | Significantly higher than the B200 |
| Memory | Less than the B300 | 50% more than the B200 |
| FP4 compute | Less than the B300 | Higher than the B200 |
The key difference is that the B300 upgrades both memory and FP4 compute over the B200, making it better suited to newer AI workloads.
Nvidia’s Tool Becomes a Channel for Detecting the Chips
DCGM Exporter is a tool for checking server health and monitoring chip performance in data centers. Simply put, administrators use it to verify that systems are operating normally and to assess their performance.
What is notable is that the tool can sometimes be accessed from outside a data center network and does not always enforce password protection. Lava therefore used DCGM Exporter to identify Nvidia Blackwell Ultra B300 chips in China Telecom’s network, even though the exact location of the chips within the system remains unknown.
What This Report Confirms—and What It Cannot Answer
The report confirms that Lava identified 32 Nvidia Blackwell Ultra B300 chips in China Telecom’s network using DCGM Exporter. However, it remains unknown through which channel the chips arrived or where they were installed in the network.
| Factor | Confirmed | Still unknown |
|---|---|---|
| Number of chips | 32 chips found | No other number confirmed |
| Source | Found in China Telecom’s network | Import route unknown |
| Location in the network | Located within China Telecom’s system | Lava could not determine the exact location |
| Company or intermediary | Detected by Lava; Nvidia provided comment | Direct B300 supplier unknown |
| B300 competitors | Not directly identified by the source | No comparative data |
Strengths and Limitations of the Evidence in the Report
Pros
- +Lava identified the chips through Nvidia’s DCGM Exporter
- +The evidence is consistent with restrictions on GPU exports to China
Cons
- −Lava could not confirm the exact location within the network
- −The smuggling route remains unknown, and Nvidia did not directly confirm the details of this case
The Hidden Costs Behind the Race for AI Chips
Demand for Nvidia’s AI chips is driving the growth of the black market. Reports indicate that $1 billion worth of chips was smuggled into China within three months.
The impact extends beyond supply. Five-year-old A100 servers have tripled in price and may cost as much as $82,000. Meanwhile, some operators have turned to modifying gaming GPUs for AI workloads instead of using increasingly difficult-to-obtain data center chips.
From 32 Chips to a Global Control Problem
The route by which B300 chips reach China still contains loopholes that require further investigation. In August 2026, Taiwan charged nine individuals over the export of B300 chips to China, while Supermicro confirmed that it had fired several employees after investigating smuggling allegations.
The process involved Flying Tiget Tech, a Taiwanese server trading company listed as a GPU buyer in Nvidia’s authorized accounts. Nvidia said that product diversions in its latest court case accounted for less than half of one percent. However, the discovery of 32 B300 chips in China Telecom’s network shows that the issue continues to be detected.