TAIPEI (Taiwan News) — The US Commerce Department on Sunday closed a year-old loophole that potentially allowed some of the world’s most advanced AI chips to flow to Chinese companies through offshore subsidiaries.
The new guidance targets exports of advanced chips such as Nvidia’s Blackwell and Rubin processors and AMD’s MI350X to Chinese entities operating outside China, including in countries such as Malaysia, per a Reuters report.
The updated guidance published on the Commerce Department’s website states that advanced chip licensing requirements will apply to entities headquartered in China, regardless of where they are located. The move comes amid ongoing US efforts to restrict Chinese technology companies’ access to cutting-edge AI hardware.
The scale of shipments made through the loophole over the past year remains unclear. A semiconductor industry source familiar with supply chains estimated that the number of chips exported may have reached into the hundreds of thousands. The Commerce Department, Nvidia, and AMD did not immediately respond to requests for comment.
The guidance came as the Commerce Department announced last May that it would not enforce certain export controls established under the Biden administration’s AI Diffusion Rule as that administration left office, per Reuters.
The AI Diffusion rule established tiered regulations for different countries, aiming to keep advanced computing power within the US and allied nations while limiting China’s access to the technology. In the new guidance, the Commerce Department clarified that many licensing requirements predated the AI Diffusion rule and thus remain in place and cannot be ignored.
Chris McGuire, a technology expert and former US State Department official, wrote on social media Sunday that the loophole represented a huge problem. He said overseas subsidiaries of Chinese companies had been able to purchase Nvidia Blackwell chips without a license due to non-enforcement of export controls.
The new guidance does not require data centers to stop using previously acquired chips or interrupt operations of existing high-end servers.
