Report: Nvidia working on tracking function for AI chips against smuggling

Nvidia software could determine the location of AI-capable chips to prevent exports to sanctioned countries. This is initially planned for Blackwell GPUs.

listen Print view
Nvidia logo on graphics card

(Image: Konstantin Savusia/Shutterstock.com)

4 min. read

Nvidia has developed a technology for determining the location of chips that is intended to show in which countries they are being operated. This is reported by the news agency Reuters, citing informed sources. Such tracking could curb the smuggling of sanctioned chips like Nvidia's Blackwell GPUs into countries subject to export restrictions. At the same time, Nvidia would comply with a law introduced in the USA that demands geotracking in all high-performance chips for AI training.

The topic of tracking has been discussed since the export restrictions on high-performance AI chips, especially to China. In August, it was even reported that US authorities are allegedly secretly tracking AI servers. According to this, such systems with fast AI accelerators are said to contain subsequently installed trackers. At least servers from Dell and Supermicro with accelerators from both Nvidia and AMD were affected. In Asia, resellers are said to remove the trackers before the hardware continues its journey.

Nvidia's own tracking function would make such a procedure unnecessary. The market leader in graphics cards and AI chips has demonstrated such geotracking behind closed doors in recent months but has not yet published it, writes Reuters. The tracking is said to be realized via a new function of widespread software that customers use to monitor the computing performance of the systems. Nvidia intends to use the delay in communication with other servers to determine the approximate location of the chip.

Nvidia itself, however, did not want to confirm this directly upon request. "We are currently implementing a new software service that allows data center operators to monitor the health and inventory of their entire AI GPU fleet," Nvidia explained in a statement. "This customer-installed software agent uses GPU telemetry data to monitor the fleet's health, integrity, and inventory." Nvidia therefore made no statements about location determination.

Videos by heise

According to the report, geotracking is initially to be available for Nvidia's current chip generation called Blackwell, as it contains extended security functions compared to the earlier Ampere and Hopper series. However, Nvidia is currently examining its options for the two predecessor generations. Just a few days ago, the US government under Trump allowed Nvidia to sell second-tier AI chips to China. These are AI accelerators of the H200 type from the Hopper family. However, the Chinese government seems to want to prevent domestic companies from buying only Nvidia hardware. Beijing is reportedly currently considering an approval process. In this, companies would have to explain why locally manufactured AI chips are not sufficient for their purposes.

As early as the end of July, China had accused Nvidia of "sophisticated" backdoors in AI chips. This includes a "technology for 'tracking and positioning' as well as 'remote shutdown' in Nvidia's compute chips." Nvidia immediately denied this. According to this, a kill switch would be "an open invitation to disaster." Nvidia's AI accelerators are said to contain "no backdoors, no kill switches, no spyware," the company assured in early August.

However, Nvidia, like AMD and other manufacturers, may not be able to avoid the location determination of AI chips in the future. Because a bipartisan bill from a US senator from May 2025 calls for geotracking capabilities in all high-performance chips for AI training. The draft is so far-reaching that the law would even include high-end graphics cards like the GeForce RTX 5090. The initiators want to prevent hardware smuggling to China. So far, the law has not been passed.

(fds)

Don't miss any news – follow us on Facebook, LinkedIn or Mastodon.

This article was originally published in German. It was translated with technical assistance and editorially reviewed before publication.