According to Bloomberg's report dated November 6, 2026, NVIDIA is requiring cloud providers in Southeast Asia to enroll in its GPU monitoring service, Fleet Intelligence, as a condition of obtaining NVIDIA Cloud Partner (NCP) certification, the company's official cloud partner program. The report is based on people familiar with the matter. It suggests NVIDIA is building a service it had previously offered on an opt-in basis into its partner certification requirements. However, Fleet Intelligence collects device information through software that customers install. The certification requirement may encourage adoption, but that is a separate question from what the collected information can actually confirm.
An opt-in GPU monitoring service becomes a certification condition
On December 10, 2025, NVIDIA explained in an official blog post that its data center management software was optional for customers to install. According to Bloomberg, providers that obtain NCP certification receive NVIDIA's official endorsement and can also get priority in GPU supply. Whether to deploy Fleet Intelligence to track GPU operating status and whether to obtain NCP certification are, in principle, separate decisions. But if providers seeking certification are required to enroll in the service, NVIDIA can keep its opt-in positioning while strongly encouraging adoption.
The report comes amid growing concern about the illicit export of AI chips, and describes a condition that would apply to providers newly obtaining NCP certification in Southeast Asia. According to Bloomberg, NVIDIA did not respond when asked whether the same condition would apply to existing NCP-certified providers or to other regions. This means that not every GPU user is required to adopt the monitoring service, and neither the effective date of the condition nor the formal contract terms has been disclosed. This is also not a new government export control; it is a matter of conditions NVIDIA sets within its own partner program. Meeting the certification conditions does not publicly guarantee that end users or GPU uses comply with export control regulations.
Collecting GPU operating data and managing it in NVIDIA's cloud
Fleet Intelligence works by having customers install a resident piece of software called an "agent" on GPU-equipped hosts and connect it to the service using a registration token. According to NVIDIA's technical blog, the agent uses DCGM, GPUd, and an authenticity verification SDK to collect information and send it to a managed cloud operated by NVIDIA. After registration, users can view GPU temperature, power consumption, utilization, and other metrics in a management console.
The data collected is not limited to operating status. NVIDIA's published explanation of data collection items includes host UUIDs, host names, and public and private IP addresses, as well as GPU product names, UUIDs, and serial numbers. Because OS, CUDA, driver, and VBIOS information is also collected, it is possible to see which GPUs are installed in which hosts and what software configuration they run.
By combining this device identification information with operating data that changes over time, providers can continuously track the state of the GPUs they have registered. This mechanism, however, presupposes that customers install the agent and register their devices with the service. Not every GPU in circulation automatically transmits its location or operating status without customer action. It would therefore be wrong to assume that Fleet Intelligence can account for unregistered GPUs.
In its December 2025 official explanation, NVIDIA said the software is read-only and has no ability to change GPU settings or basic operation. It also denied the existence of hardware-embedded tracking features, kill switches, or backdoors. These are, however, NVIDIA's own descriptions of the design; they do not mean that independent third-party audits have demonstrated the software's resistance to circumvention or its security.
GPU identification, configuration authenticity, and geographic location are separate issues
Fleet Intelligence also includes an authenticity verification function that checks whether a GPU's configuration is legitimate. According to NVIDIA's technical blog, measurement results signed with a GPU-side certificate are compared against reference values by NVIDIA's verification service, NRAS, which returns a verification result. The inventory screen also provides integrity checks that can be viewed alongside operating data such as temperature and utilization. However, this verification checks whether the measured configuration matches the reference; on its own it cannot prove the country or region where a GPU is installed.
NVIDIA's inventory report API, meanwhile, includes a location field indicating where a device is located. It also includes verificationCheck, which shows verification results, but the public documentation does not make clear whether this verification guarantees the accuracy of the location. In other words, a location recorded in the system and information that a device passed some form of verification need to be treated as distinct.
The features confirmed in public documentation, and the limits of each, can be summarized as follows.
| What is being checked | Mechanism or field in public documentation | What it cannot confirm on its own |
|---|---|---|
| Inventory and device identification | Collects host and GPU UUIDs and serial numbers and displays the status of registered devices | The full picture of distribution, including unregistered devices; the legality of end users or uses |
| Configuration authenticity | NRAS compares measurement results signed with the GPU's certificate against reference values | Geographic location of the GPU; compliance with export control regulations |
| Geographic location | The inventory report includes a location field |
Positioning accuracy; resistance to location spoofing; smuggling detection rate |
NVIDIA's inventory report API also includes examples of location coordinates, but these are not the result of measuring positioning accuracy. Reports can be output in JSON or CSV format, and signed output is also supported. Beyond viewing information in the console, this allows the list of registered devices to be retrieved as data. However, signing output data is not the same as guaranteeing that the location recorded in that data is actually correct.
This distinction also matters when evaluating the effect of making Fleet Intelligence a condition of NCP certification. If more providers participate, the number of GPUs that are registered in the system and continuously report information would likely grow. But more monitored devices does not mean their locations and actual use can be accurately confirmed.
Sending data to NVIDIA is separate from granting outside organizations viewing access
Fleet Intelligence also offers settings for sharing collected data with outside organizations. According to NVIDIA's official procedure, an organization administrator invites an administrator of another organization and assigns permissions such as "Cloud Provider" or "Viewer." When the invited party accepts, data can be shared between organizations on NGC. Administrators set which organizations are allowed to view the data.
This cross-organization sharing setting, however, is different from the mechanism by which the agent sends data to NVIDIA's cloud. The fact that sharing with outside organizations requires administrator action does not mean operating data is not sent to NVIDIA's service itself. Conversely, the fact that information is sent to NVIDIA's cloud does not by itself indicate that government agencies can view the data constantly and automatically.
For cloud providers seeking NCP certification, what viewing permissions to grant outside organizations is an important point to check, as is what information Fleet Intelligence collects. Host names, IP addresses, GPU serial numbers, and similar data are tied to a provider's equipment configuration. The information sent to NVIDIA's service and the information that other organizations are permitted to view therefore need to be managed separately. Note also that being able to see GPU utilization does not mean being able to view the content of confidential data a GPU is processing, and the public documentation provides no evidence for the latter.
Registration scope and continuous monitoring will determine anti-smuggling effectiveness
What the Fleet Intelligence console and API can show is, fundamentally, information collected from devices registered with the service. The effectiveness of the service as an anti-smuggling measure therefore cannot be judged simply by whether a provider registered its GPUs at onboarding. What matters is how much equipment falls within the monitoring scope, whether information continues to be collected after registration, and whether reported devices and locations can be checked against actual installations.
By tying NCP certification to Fleet Intelligence participation, NVIDIA may be able to use providers' incentive to obtain certification to expand the monitored population. At the same time, the public documentation shows no third-party assessment of smuggling detection rates or of resistance to circumvention. For now, an increase in the types and volume of information collected cannot be taken as an improvement in anti-smuggling effectiveness.
The focus going forward will be the certification conditions NVIDIA formally sets and how they are enforced. Of particular interest are which providers and GPUs are covered and how far continuous information collection after registration is required. Only if a mechanism exists to reconcile registered information with actual installations can we assess how much expanding the monitored population through partner certification helps in confirming GPU locations and countering illicit exports.