32 NVIDIA B300 GPUs Identified Through China-Geolocated IPs

Publicly reachable monitoring endpoints exposed 32 NVIDIA B300 GPUs through IP addresses geolocated to China, creating a visible record of advanced AI hardware while leaving its physical location and export status unresolved.

The endpoints were accessible without authentication and displayed GPU model names and unique device identifiers. Those fields show what hardware was visible to the monitoring service, but do not establish who owns or uses a device. The exposure documents access to system information, not a transfer of hardware.

The services used NVIDIA’s DCGM Exporter, software that makes selected GPU telemetry available as Prometheus metrics. The exporter serves data at /metrics and uses port 9400 by default. Public access to those endpoints can reveal details about systems to visitors without credentials.

The 32 B300s were part of a wider set of observations. Across multiple scans of China-geolocated endpoints, 4,724 unique NVIDIA GPUs were identified. Broader monitoring found more than 12,000 unique GPUs across just over 2,000 public IP addresses associated with nearly 300 organizations. These are endpoint observations, not verified inventories of chips in Chinese facilities or counts of shipments.

Geolocation applies to a network address; it does not pin the hardware to that location. An IP geolocation result describes a network address, not necessarily the location of the machine using it. The findings also do not identify a shipment route, owner or end user.

U.S. Bureau of Industry and Security (BIS) guidance requires licenses for certain advanced-computing items exported to entities headquartered in a Country Group D:5 location, or when the entity’s ultimate parent is headquartered there. The requirement can apply even if the entity itself is located elsewhere. The guidance does not classify the GPUs identified in the scan, and the endpoint data does not establish an export-control violation.