Washington Wants a Kill Switch Inside the GPU

US Representative Bill Foster has told reporters that Nvidia representatives privately acknowledged the anti-smuggling tracking required under the Chip Security Act is technically achievable, according to a report carried by Chinese media.

That contradicts the company’s public position. In March, the US Semiconductor Industry Association, of which Nvidia is a member, opposed the bill, describing chip tracking requirements as untested and possibly unworkable. Foster says Nvidia never put its private assurance in writing.

The capability is already deployed

Nvidia already has the technology to know where its chips are. Its Fleet Intelligence tool lets customers monitor how advanced chips are performing across different data centres. Nvidia confirmed to reporters that the tool is sufficient to establish which country a chip is in.

Location, however, is not the same as control. The draft is reported to rely on a ping mechanism: a GPU must emit a location signal as soon as it powers up.

Technician holding a data centre GPU card beside a server rack
Nvidia’s Fleet Intelligence tool can already establish which country a chip is running in. Source: Sohu Tech

From knowing to disabling

Foster says he plans to introduce stricter legislation next year that would require licence verification before an advanced GPU is brought online.

The purpose of that rule is remote shutdown. Where a GPU is found unlicensed or located in an embargoed region, the system could lock it directly. That goes beyond the current limit of tracking technology, which can establish position but cannot prevent use.

Foster has linked the legislative push to the political climate around Nvidia’s leadership turmoil, saying Congress will face substantial pressure to make geographic verification of chip location real.

What it changes

Export controls have relied on customs declarations, end-user certificates and the threat of future enforcement. A hardware-level check moves the boundary. The question stops being whether a shipment was declared and becomes whether a machine will boot.

For data centre operators outside the United States, that is a different kind of dependency. A compute cluster whose continued operation depends on a licence check performed at power-on is an asset with a switch attached, and the switch sits in another jurisdiction.

Nothing in the reported draft has passed into law. Foster’s stricter bill is a plan for next year, and the industry association that opposed the original act has not changed its public position. But the technical argument Nvidia is said to have made in private has already narrowed the space for saying the requirement cannot be met.

Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1100961363_163726.

Leave a comment