- US chipmaker Nvidia has reportedly found a way to sell high-end chips to Chinese companies – while complying with US restrictions.
- Nvidia will deliver three new chips to domestic manufacturers in the coming days, Chinese financial media Cailian Press said on Thursday, citing sources.
- Nvidia, the U.S. Department of Commerce and the Bureau of Industry and Security did not immediately respond to a CNBC request for comment.
The Nvidia Corporation logo is seen during the annual Computex computer exhibition in Taipei, Taiwan, May 30, 2017.
Tyrone Siu | Portal
BEIJING – U.S. chipmaker Nvidia has reportedly found a way to sell high-end chips to Chinese companies – while complying with U.S. regulations aimed at restricting China’s access to the technology.
China accounts for 20% to 25% of Nvidia’s revenue in the data center business, its largest unit.
Nvidia will deliver three new chips to domestic manufacturers in the coming days, Chinese financial media Cailian Press said on Thursday, citing sources.
The chips – called HGX H20, L20 PCle and L2 PCle – are based on Nvidia’s H100 chip, the report says.
The H100 and A100 artificial intelligence chips last year were the first to be hit by new US restrictions aimed at curbing sales to China. Nvidia said in a September 2022 filing that the U.S. government would continue to allow the H100 to be developed in China.
In the short term, Chinese manufacturers have no better option and will continue to buy Nvidia chips while looking for replacements.
Bo Du
Managing Director, WestSummit Capital Management
Companies in China had subsequently switched to Nvidia’s H800 and A800 chips, but the U.S. restricted those sales with new restrictions last month.
The H20’s processing power is only about 50% of the A100’s, said Bo Du, managing director at WestSummit Capital Management and a former engineer in the chip industry.
This is “basically a farewell to physics simulation,” he said in Mandarin, translated by CNBC. While it is possible to use clusters of lower-power chips to support large model calculations, there is no ideal solution given the cost, he said.
“In the short term, Chinese manufacturers have no better option and they will continue to buy Nvidia chips while looking for replacements,” Du said, pointing out that some major internet companies have started buying domestically made AI chips on a large scale.
Demand for artificial intelligence computing power has only increased as companies in China rush to develop local versions of OpenAI’s ChatGPT.
The Financial Times also reported on Nvidia’s new chips for the Chinese market, citing a document that the chip manufacturer had distributed to potential customers.
Nvidia declined to comment. The U.S. Department of Commerce and the Bureau of Industry and Security did not immediately respond to a CNBC request for comment.
All three of Nvidia’s new chips have operating metrics that fall outside the threshold of U.S. restrictions, research firm SemiAnalysis said in an online post Thursday. The company runs a Substack tech newsletter that reportedly has more than 64,000 subscribers.
“With these new chips, Nvidia strikes a perfect balance between peak performance and power density to meet new US regulations,” said SemiAnalysis.
Nomura analysts previously noted that Nvidia’s Drive AGX Orin chip also did not meet all the criteria that would justify a U.S. restriction on sales to China, allowing electric car makers in the country to continue using the chip.
The US has said its focus is on restricting China’s development of advanced technology for military purposes. President Joe Biden’s administration also emphasized that the country is in competition with China.
Domestic players are trying to develop workarounds to the US restrictions.
At the end of August, Huawei released a smartphone that, according to reviews, offers 5G-like download speeds thanks to an advanced semiconductor chip.
It’s unclear whether older equipment or alternative procurement processes were involved in the latest chip production.
—CNBC’s Arjun Kharpal contributed to this report.