The US government’s new restrictions on the ability of Nvidia to sell artificial intelligence chips to Chinese customers threatens to deal a heavy blow to the country’s development of a sweeping range of cutting-edge technologies.
The Santa Clara, California-based company disclosed in a regulatory filing this week it can no longer sell certain high-end chips in China without a licence from Washington. These AI accelerators go into large data centres to train AI models for tasks like autonomous driving, image recognition and voice assistance.
Nvidia has nearly a 95% share of that market, according to Fubon Securities Investment Services estimates, and the rest is accounted for by AMD, a fellow US chip firm that’s bound by the same export restrictions. Without access to their gear, tech giants that rely on big server farms to develop everything from electric and self-driving cars to social and cloud services will be at a disadvantage to international competition.
“This is the new Cold War reality and broader export restrictions are part and parcel of this,” said Amir Anvarzadeh of Asymmetric Advisors. “The export restrictions will broaden and it will impact semiconductors, AI, autonomous systems and biotech.”
The escalated trade curbs, which Washington didn’t signal it was considering before imposing them, add to existing sanctions and limits on exports of chip-making gear to China. Chinese semiconductor firms are already denied access to the most advanced lithography equipment from the Netherlands’ ASML and cutting-edge gear from American suppliers including Lam Research. The recent Chips Act in the US forces global chip makers to effectively choose between investing in the US and China. Now that Washington is limiting access to AI products as well, it’s created another chokepoint for Beijing’s tech expansion while working on growing its own domestic semiconductor capability.
The head of one of China’s leading EV manufacturers quickly decried the restrictions.
The measures will “bring a challenge to the cloud training of all autonomous driving”, He Xiaopeng, the chairman and CEO of XPeng, said on his WeChat account. Nvidia is a leader in providing the hardware for autonomous driving — both for developing the algorithms in massive server farms and supplying the onboard processors for cars to be aware of their surroundings.
Washington has told Nvidia that the new curbs are designed to prevent advanced AI gear being used for or diverted to military purposes by China or Russia. In June, Washington, DC-based think-tank The Center for Security and Emerging Technology said almost all of the 97 AI chips in public Chinese military purchase records between April and November 2020 were designed by US firms Nvidia, Intel, Microsemi or Xilinx, which is now part of AMD.
Still, the brunt of the impact will be felt by Nvidia itself and China’s biggest tech firms like Alibaba Group and Tencent Holdings, which are the closest rivals to US cloud services from Amazon.com’s AWS, Google Cloud and Microsoft’s Azure.
The Chinese government opposes US restrictions on chip exports to the country because the move hurts the legitimate rights and interests of both Chinese and American firms, China commerce ministry spokesman Shu Jueting said at a briefing on Thursday, in response to a question about Nvidia’s disclosure. China urges the US to stop the practice immediately and treat companies from all countries fairly, he said.
At a basic level, an AI accelerator is a graphics processor, or GPU, specifically tailored to train AI models by feeding them tons of data. It’s better suited for such tasks than a general-purpose CPU because its architecture can do parallel work in huge volumes. Nvidia was the first to come up with a language to make GPUs do AI tasks, giving it a huge head start over rivals like AMD and Intel.
Baidu last month secured approval to deploy the first fully autonomous self-driving taxis on China’s roads. Along with domestic rivals like Pony.ai and XPeng, Baidu is among the first companies globally to roll out such services — but that lead is threatened without the continued ability to develop using Nvidia’s hardware.
The US has given Nvidia a year to wrap up development work in China on its most advanced H100 server chip, underlining that it doesn’t want its companies working on sensitive tech within Chinese borders. Nvidia also said the restrictions may cost it $400-million in the current quarter and it may have to move some operations out of China.
The current-generation A100 chip from Nvidia is one of the world’s most sophisticated, with 54 billion transistors, while the next-generation H100 — which no Chinese firms will be able to buy without explicit US approval — will be built on TSMC’s 4nm process and have 80 billion transistors on board.
Big data centre AI customers in China have been buying Nvidia’s gaming graphics cards as a substitute, though that requires significant modifications to deploy, according to Jeff Pu of Haitong International Securities. Washington’s move “will accelerate the development of local data centre GPUs such as Alibaba’s” and it will boost sentiment for domestic stocks in the sector, he said.
China’s Cambricon Technologies is to a homegrown alternative to Nvidia or AMD for AI chip making and its shares jumped more than 30% over two days after the curbs were announced. — Debby Wu, Ian King and Vlad Savov, (c) 2022 Bloomberg LP