U.S. Tech Leaders Want Fewer Export Curbs on AI Chips for China

And The Wall Street Journal reported last month that the White House was exploring how to restrict the leasing of cloud services to AI firms in China.

But the U.S. controls appear to be merely slowing, rather than stopping, China’s AI development.

Last October, the U.S. Commerce Department banned Nvidia from selling two of its most advanced AI-critical chips, the A100 and the newer H100, to Chinese customers, citing national security concerns. In November, Nvidia designed the A800 and H800 chips that are not subject to export controls for the Chinese market.

According to the Journal, the U.S. government is considering new bans on the A800 exports to China.

According to a report published in May by TrendForce, a market intelligence and professional consulting firm, the A800, like Nvidia’s H100 and A100, is already the most widely used mainstream product for AI-related computing.

Combining Chips
Robert Atkinson, founder and president of the Information Technology and Innovation Foundation, told VOA in a phone interview that although these chips are not the most advanced, they can still be used by China.

“What you can do, though, is you can combine lesser, less powerful chips and just put more of them together. And you can still do a lot of AI processing with them. It just makes it more expensive. And it uses more energy. But the Chinese are happy to do that,” Atkinson said.

As for the Chinese use of cloud computing, Hanna Dohmen, a research analyst at Georgetown’s Center for Security and Emerging Technology, told VOA Mandarin in a phone interview that companies can rent chips through cloud service providers.

In practice, it is similar to a pedestrian hopping on an e-share scooter or bike — she pays a fee to unlock the scooter’s key function, its wheels.

For example, Dohman said that Nvidia’s A100, which is “controlled and cannot be exported to China, per the October 7 export control regulations,” can be legally accessed by Chinese companies that “purchase services from these cloud service providers to gain virtual access to these controlled chips.”

Dohman acknowledged it is not clear how many Chinese AI research institutions and companies are using American cloud services.

“There are also Chinese regulations … on cross-border data that might prohibit or limit to what extent Chinese companies might be willing to use foreign cloud service providers outside of China to develop their AI models,” she said.

Black Market Chips
In another workaround, Atkinson said Chinese companies can buy black market chips. “It’s not clear to me that these export controls are going to be able to completely cut off Chinese computing capabilities. They might slow them down a bit, but I don’t think they’re going to cut them off.”

According to an as yet unpublished report by the Information Technology and Innovation Foundation, China is already ahead of Europe in terms of the number of AI startups and is catching up with the U.S.

Although Chinese websites account for less than 2% of global network traffic, Atkinson said, Chinese government data management can make up for the lack of dialogue texts, images and videos that are essential for AI large-scale model training.

“I do think that the Chinese will catch up and surpass the U.S. unless we take fairly serious steps,” Atkinson said.

John Xie is a VOA News reporter. This article is published courtesy of the Voice of America (VOA).