US halts Nvidia chip exports
- The U.S. Commerce Department on May 31 moved to close a loophole that allowed advanced AI chip exports to overseas subsidiaries of Chinese companies. - Reuters reported Nvidia Blackwell and AMD MI350X processors were among chips affected, with one industry source estimating hundreds of thousands may have shipped. - Commerce guidance said existing chips already installed in data centers can keep operating, while future exports require licenses.
The U.S. Commerce Department moved on May 31 to block exports of the most advanced AI chips to subsidiaries of Chinese companies outside China, closing what officials described as a gap in existing controls. Reuters reported the step covers top-end processors including Nvidia’s Blackwell line and AMD’s MI350X, and follows evidence that Chinese firms may have been obtaining those chips through offshore entities in places such as Malaysia. The move matters because earlier U.S. controls were aimed mainly at shipments into China itself. The new guidance, as described by Reuters and follow-up reports citing the Commerce Department, says licensing requirements apply based on whether a company is headquartered in China or has a parent company in China, not only on the destination country. (cnbc.com) ### Which chips are covered? Reuters said the affected products include some of the world’s most advanced AI chips, naming Nvidia’s Rubin and Blackwell processors as well as AMD’s MI350X. Those are the chips used to train and run large AI models in data centers, and they sit above the lower-spec products companies have sometimes designed to comply with earlier China restrictions. (cnbc.com) Nvidia is central to the story because its chips dominate the market for high-end AI computing. Reuters’ report framed the U.S. action around concern that Nvidia’s top processors were reaching Chinese-linked buyers through overseas subsidiaries despite broader export controls. ### What loophole is Washington trying to close? (cnbc.com) The Commerce Department’s guidance, reported on May 31, addressed a loophole created about a year earlier, according to Reuters. The issue was that a Chinese company’s subsidiary in a third country could potentially buy advanced chips without triggering the same restrictions that would apply to a shipment directly into China. (msn.com) Malaysia was one example cited by Reuters as a place where subsidiaries of Chinese AI firms may have been operating. CNBC, carrying the Reuters report, said the guidance suggested U.S. chips may have been reaching those subsidiaries for almost a year. Chris McGuire, a former U.S. State Department official and technology policy expert, said in comments cited by Reuters-linked reports that the loophole was “a HUGE problem.” His point, as summarized in those reports, was that overseas subsidiaries of Chinese firms had been able to buy Nvidia Blackwell chips without a license. (cnbc.com) ### Does this stop data centers from using chips they already bought? (cnbc.com) The guidance does not appear to force an immediate shutdown of existing systems. Reuters-linked reports said the new policy does not require data centers to stop using or servicing chips that were already acquired. That distinction is important for cloud operators and colocated data centers in Southeast Asia and other hubs. (taiwannews.com.tw) The immediate effect is on future exports and licensing, not on ripping out already installed servers, according to the reporting. ### Why is this different from earlier China chip controls? The change extends enforcement from a country-based model to an ownership-based model. (tbsnews.net) Instead of asking only where the chips are going, U.S. officials are now asking who ultimately controls the buyer, according to the Commerce guidance described by Reuters and others. That makes the compliance burden heavier for exporters, distributors and cloud providers. (tbsnews.net) Companies now have to screen not just the shipping destination but also whether the customer is a Chinese-headquartered firm or owned by one. That is an inference from the reported guidance and the licensing standard it describes. ### What happens next? The next step is licensing enforcement. (yahoo.com) Commerce said, according to reports published June 1, that future exports of covered advanced AI chips to businesses headquartered in China or owned by Chinese parent companies will require licenses. Nvidia, AMD, cloud operators and overseas data center customers will now have to determine which pending and future shipments fall under the clarified rule. (yahoo.com) Reuters reported the guidance was issued on Sunday, May 31, and the operational question for the industry is which orders can still move and which now need U.S. approval. (money.usnews.com)