With the rise of generative AI applications like ChatGPT and AI's increasing importance, the US is considering additional sanctions on exports of AI chips to China. However, despite Washington's best efforts, high-end AI chips have still somehow found their way to China's marketplaces.
According to a report from the Wall Street Journal, the US Department of Commerce could be making its move as soon as early July to prevent Nvidia and other chipmakers from shipping certain AI chips to China, notably Nvidia's A800, without first obtaining a license. It's also considering restricting cloud service leases to Chinese AI companies, as such arrangements have been used to skirt the export bans on advanced chips.
Although the sanctions back in October 2022 are believed to be able to restrict China's AI capabilities, China was still able to gain access to some AI chips. Above all, in response to the October restrictions, Nvidia made A800/H800 chips - alternative, downgraded versions of its high-end A100 and H100 chips - for the Chinese market, as they fell under the thresholds of the sanctions. Nevertheless, as WSJ reports, the incoming restrictions are set to restrict the sale of even A800 chips without a license.
This move is viewed as an expansion of the chip export sanctions announced in October 2022, a response to the rise of generative AI that occurred after the initial sanctions were launched. US officials increasingly see AI as a national security issue, as AI can be fused with weapons to give US rivals an advantage on the battlefield. AI tools can also be used to produce chemical weapons or malicious computer codes.
The actual schedule of the sanction's rollout is still uncertain, as chipmakers have continued to push the administration to ease the restrictions. According to sources, the US government will likely wait until after Secretary of the Treasury Janet Yellen's visit to China in early July to implement these new restrictions.
Meanwhile, a report from Reuters showed that even the supposedly banned A100/H100 chips can still be found in China, in marketplaces like Shenzhen's Huaqiangbei and e-commerce platforms. According to the sellers, the buyers are mainly app developers, startups, researchers, and gamers. These "underground" AI chips were sold for around US$20,000, double the original price.
After approaching these sellers, Reuters discovered that these A100 chips are obtained through two methods: products that flowed into the market after procurement by US corporations, and imports through places like India, Taiwan, and Singapore. Fortunately for the US, chips obtained with these methods are too low in quantity to sufficiently train LLMs (large language models).
Ever since ChatGPT blew up, Chinese tech giants like Baidu, Alibaba, and Tencent have all launched their own LLMs. It is generally believed that the inability to access high-end chips is a risk for China's AI development.
For instance, before the launch of Baidu's ERNIE Bot, local media pointed out that company CEO Robin Li requested other AI research departments to hand over their existing A100 to support ERNIE Bot's development.
In a conversation with Financial Times in May, Nvidia founder and CEO Jensen Huang commented that US sanctions on the Chinese chip industry had "left our hands tied behind our back." Huang also worried that Chinese companies, unable to access US products, would start to build their own chips to rival Nvidia's market-leading chips. "If we are deprived of the Chinese market, we don't have a contingency for that," the Nvidia CEO told FT. Most important of all, according to Huang, blocking the US tech industry's access to China would undermine the Chips Act. If losing the Chinese market, the US tech industry would require one-third less capacity, Huang pointed out. "No one is going to need American fabs, we will be swimming in fabs."
