(Reuters) - The U.S. will take steps to prevent American chipmakers from selling semiconductors to China that circumvent government restrictions, a U.S. official said, as part of the Biden administration's upcoming actions to block more AI chip exports.
The new rules, details of which Reuters is reporting for the first time, will be added to sweeping U.S. restrictions on shipments of advanced chips and chipmaking equipment to China unveiled last October. The updates are expected this week, other people familiar with the matter said, though such timetables often slip.
The new rules will block some AI chips that fall just under current technical parameters while demanding companies report shipments of others, said the official, who provided information on condition of anonymity.
A spokesperson for the U.S. Department of Commerce, which oversees export controls, declined to comment.
The latest crackdown on tech exports to China coincides with U.S. efforts to thaw difficult relations between the world's two largest economies. Several senior members of the Biden administration have met their Chinese counterparts in recent months, and the latest round of rules risks complicating the diplomatic effort.
The Biden administration has said it designed the export curbs to keep U.S. chips and equipment from strengthening China's military. Beijing has accused the United States of abusing export controls to suppress Chinese companies. The restrictions marked a historic shift in U.S.-China tech policy.
The Chinese embassy in Washington did not immediately respond to a request for comment.
Last year, government restrictions kept Nvidia (NVDA.O), the world's most valuable chipmaker, from shipping two of its most advanced AI chips to Chinese customers, chips that have become the industry standard for developing chatbots and other AI systems.
But Nvidia soon released new variants for the Chinese market that were less sophisticated and got around the U.S. export controls. One, named the H800, has as much computing power at some settings used in AI work as the company's more powerful but blocked H100 chip. Still, some key performance aspects are limited, according to a specification sheet seen by Reuters.
The U.S. now plans to introduce new guidelines for AI chips that will restrict certain advanced datacenter AI chips that are not currently captured, the U.S. official said.
While the official declined to identify which additional chips will be effectively banned, Nvidia's H800 is a semiconductor sources have suggested the administration has wanted to block.
Santa Clara, California- based Nvidia did not immediately respond to a request for comment. In June the company's chief financial officer said that if the H800 and a related chip called the A800 were restricted, they did not anticipate it "would have an immediate material impact on our financial results."
Chips meant for consumer products like laptops will be exempt from the new curbs, the official said. But companies will need to tell the Commerce Department when they are filling orders for the most powerful consumer chips to make sure they are not being used in ways that threaten national security, according to the official.
In order to keep AI chips the U.S. views as too powerful from China, the official said the U.S. planned to remove one of the parameters - the "bandwidth parameter" - it has used to restrict exports of certain AI data center chips. By removing this parameter, another guideline kicks in, widening the scope of chips covered. This would likely mean the speed at which AI chips talk to each other would be reduced.
This is important because training the largest AI models is impossible on one chip and requires many chips tied together. If one slows the speed they communicate at, it makes AI development more challenging and expensive.
The U.S. also plans to introduce a "performance density" parameter to help prevent future workarounds, the official said, but declined to elaborate.
Evolving technology
The updated rules also are meant to cover AI chips as technology evolves. The U.S. will require companies to notify the government about semiconductors whose performance is just below the guidelines before they are shipped to China, the official said. The government will decide on a case-by-case basis whether they pose a national security risk but they can be shipped unless the chipmaker is told otherwise.
The updates to the October 2022 rules may also close a loophole that gives Chinese companies access to American artificial intelligence chips through Chinese units located overseas, as Reuters reported last week.
The rules are not expected to include restrictions on access to U.S. cloud computing services, or those of allies, but the U.S. will seek comments on the risks of such access and how they might be addressed, the official said.
The Biden administration told Beijing of its plans to update the contentious rules this month, Reuters reported earlier in October, as part of a policy aimed at stabilizing relations between the superpowers.