US to plug loopholes in export controls on AI chips to China

The US government is set to introduce fresh sanctions that will stop American chipmakers from sidestepping export restrictions on advanced AI chips into China. The upcoming updates, expected to arrive this week, are also aimed at closing a loophole that allowed Chinese companies to get hold of these AI chips through their subsidiaries and suppliers located overseas.

The new rules will be added to the sweeping restrictions enforced in October last year that prevent advanced chipmaking equipment and semiconductors – that include high-powered AI accelerators – from being sold to the Mainland.

According to unnamed sources cited by Reuters, the new sanctions block some AI chips that fall just under current technical parameters. This essentially means they will be applied to chips that are slightly less powerful than those covered under the initial guidelines – but still have advanced capabilities.

This is all part of the White House's effort to keep China from using US technology to gain military advantage amid heightened geopolitical and diplomatic tension between the two superpowers.

Quoting a US official familiar with the matter, Reuters reported that chips meant for consumer products like laptops will be exempt from the new curbs. "'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," the official told the news agency.

Related:TSMC gets one-year extension to ship US chip equipment into China

Closing loopholes

Before the existing export controls took effect last year, some US semiconductor vendors came out with less powerful variants of their advanced AI chips for the China market.

For one, Nvidia released a modified version of its A100 accelerator dubbed the A800 that features an interconnect bandwidth reduced to 400 Gbit/s to be within the rules. Intel followed suit earlier this year with a new Habana Gaudi2 variant designed to comply with US restrictions.

These data center chips are essential in developing chatbots and other AI systems, and they are being used by enterprises training large-language models (LLMs).

The new rules that the US plans to introduce will control access to certain advanced data center AI chips that are not currently covered by existing restrictions.

This would likely mean the speed at which AI chips talk to each other would be reduced, according to Reuters.

"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," a source told Reuters.

The new rules are also intended to stop Chinese companies from using backdoor channels to gain access to sophisticated AI chips. Existing restrictions do not cover overseas subsidiaries of Chinese companies, which effectively make it easy for these AI chips to be smuggled into the Mainland.

This challenge was brought home early this year when media reports revealed that the Chinese agency responsible for developing and maintaining the country's nuclear arsenal had been using back-channels and shell companies to obtain US chips to power its nuclear weapons simulations. This reportedly had been happening for at least two years, despite a decades-old trade ban that was supposed to prevent America's top tech from being used by other foreign armed forces.

Preventing US technology transfer to China

The updates to the existing export controls on AI chips into China come nearly a month after the US Commerce Department finalized rules that would prohibit semiconductor companies seeking federal grants from carrying out certain business expansions, partnerships and research in China.

The Biden administration has been preparing to disburse more than $52 billion in federal grants and tens of billions of dollars of tax credits to build up the US chip industry.

The final restrictions will prohibit companies that receive federal money from using it to construct factories abroad. They also restrict companies from significantly expanding semiconductor manufacturing in "foreign countries of concern," defined as China, Iran, Russia and North Korea, for ten years after receiving an award.