Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Washington’s tightened export controls on chips may leave Chinese tech groups relying on outdated and stockpiled chips to pursue their artificial intelligence ambitions, with industry giants Alibaba and Baidu facing new hurdles for the manufacturing of their latest self-designed AI processors.
Alibaba and Baidu’s processors have become frontrunners in China’s efforts to create domestic alternatives to US maker Nvidia’s sophisticated products, with the chips at present manufactured at TSMC and Samsung plants, said four people close to the groups’ design projects.
The processing speeds of their most advanced AI chips fall within new thresholds unveiled by Washington this week as part of an update to its chip export controls, the people said.
The tightening will also force Silicon Valley-based Nvidia to halt shipments to China of two processors that the company had tailor-made to comply with earlier export controls, according to a statement from Nvidia this week.
Collectively, the restrictions mean Chinese tech groups will have to turn to AI chipsets similar to Nvidia’s V100, which was released in 2017 and has since been discontinued, in order to train and run generative AI models, analysts said. Since the V100 was released, chips have become significantly more advanced, enabling the creation of OpenAI’s ChatGPT.
The US move poses “an existential challenge” to China’s efforts to catch up with AI development at OpenAI and other American companies, said one chip consultant in Beijing.
Washington’s controls, which extend to foundries in Taiwan and South Korea contracted to make chips for Chinese groups, are made possible by the vast amount of American hardware and software embedded in the semiconductor supply chain. China’s domestic alternatives, including partially state-owned SMIC, are several generations behind in chip manufacturing technology.
The updated rules come at a time of deteriorating US-China relations and an expanding programme to impede Beijing’s technological progress. “The goal is to choke off China’s access to the future of AI,” said Gregory Allen, an AI expert at the CSIS think-tank.
Allen said the updated controls increased the number of advanced AI chips requiring a licence that was likely to be denied, in effect banning them for export. The controls also create a reporting regime that covers a huge swath of data centre chips with speeds just below the cutting edge and add prohibitions on selling to subsidiaries of Chinese companies outside the country.
“The Department of Commerce’s visibility into high-performance computing chip exports worldwide is going to go way up,” Allen said.
While China’s largest tech groups have stockpiles of AI chips, the controls will eventually make training AI models in the country more expensive and time-consuming than for their US counterparts, analysts said. Bernstein senior analyst Boris Van estimated that relying on chips similar to Nvidia’s V100 would at least double data processing costs.
“Once the existing stash of chips is exhausted, Chinese AI firms would struggle to improve their models,” said Phelix Lee, an analyst at Morningstar.
Big Chinese tech groups, including Alibaba, Baidu, ByteDance and Tencent, have purchased more than $5bn worth of Nvidia chips in recent months, the Financial Times reported in August, but most of these orders have not been delivered, according to several people familiar with the situation.
“The supply is terrible,” said a Beijing-based AI entrepreneur desperate for Nvidia’s processors, noting that the company was months behind in deliveries. Washington has given Nvidia and other chip companies a grace period of about one month to fulfil orders to China.
“Whatever portion cannot be fulfilled in the grace period will have to be cancelled,” said Charlie Chai, a Shanghai-based analyst at 86Research.
While industry insiders expect some banned chips to continue to flow into the country through black market channels, they do not expect supply to fulfil the high demand from tech groups training generative AI models.
“China will be permanently stuck with low-end Nvidia chips and see the [AI] infrastructure gap gradually widen with the rest of the world,” said Chai of 86Research.
Samsung declined to comment. TSMC, Alibaba and Baidu did not respond to requests for comment.