China’s State Administration of Market Regulation is investigating American technology giant Nvidia for potential violation of anti-monopoly laws and an agreement on Nvidia’s acquisition of Mellanox Technologies in 2020.
The first allegation makes sense only as retaliation against the latest round of US sanctions on China, but the second may hold water. Nvidia is disputing the allegations.
In any case, Nvidia has denied rumors that it plans to reduce sales to China, which accounted for 15% of its revenue in the three months to October (Nvidia’s fiscal third quarter). In fact, Nvidia is expanding its presence in China with a focus on areas not subject to export controls, including autonomous driving.
For China to accuse Nvidia of monopoly practices seems illogical because US Commerce Secretary Gina Raimondo and her department’s Bureau of Industry and Security (BIS) have already undermined the company’s core data center business there by banning the export of its most advanced GPU processors to China.
On the contrary, perhaps China should thank the US government for giving Huawei and other Chinese IC design companies the opportunity and a strong incentive to build up their own AI processor business while their most formidable competitor’s hands are tied.
In 2022, the BIS banned exports of Nvidia’s top-end A100 and H100 processors to China. In 2023, it banned exports of the A800, a dumbed-down version of the A100 designed specifically to meet BIS requirements.
The A800 was a best seller in China so the BIS lowered the bar, forcing Nvidia to design another chip, the H20, with even lower performance. But Chinese AI processors, led by Huawei’s 910B, proved to be competitive with the H20, which has not sold well in China. And Huawei claims that its successor, the 910C, matches the performance of Nvidia’s H100.
In any case, Baidu, Tencent and other Chinese customers began switching to Huawei and other domestic chips when it became clear that US sanctions were not based on strict technical criteria related to national security as US government officials claim but could be changed at any time simply to punish China.
Exports of Nvidia’s new and considerably more powerful Blackwell B200 AI processors to China are banned under existing sanctions, but a dumbed-down version that may be called B20 is reportedly being prepared. But why would it be any more successful than the H20?
As for Mellanox Technologies, the Chinese government approved the acquisition on condition that access to Mellanox interconnect technology would remain open, that Nvidia would supply Mellanox interconnect products to Chinese customers without discrimination, not bundle them with its own GPUs, and guarantee the interoperability of its own GPUs with other interconnect products.
Nikkei Asia notes that “It is not clear which of these terms Nvidia is alleged to have been broken.”
Mellanox is an Israeli company that designs and supplies InfiniBand and Ethernet interconnect adapters, switches and other devices and software for use in high-performance computing, data centers, cloud storage and financial services. Nvidia arranged to acquire 100% of the company in 2019.
From where might China have gotten the idea of an antitrust investigation? Well, last July France confirmed that it was investigating Nvidia for alleged anti-competitive practices, and in September, the US Department of Justice subpoenaed the company for information regarding its potentially restrictive marketing practices. Outside of China, Nvidia’s share of the market for AI processors is estimated at about 90%.
US sanctions, which not only block the export to China of advanced AI processors made by Nvidia and its smaller competitor AMD, also prevent China from buying EUV lithography systems made by ASML of the Netherlands.
This makes it impossible for the Chinese to make integrated circuits (ICs) with design rules smaller than 5nm, and 7nm is as far as they can go with an adequate degree of efficiency. Blackwell processors, on the other hand, are made by TSMC using a 4nm process. As a result, Nvidia is currently streets ahead of the Chinese in the race to develop more sophisticated AI processors.
However, chasing Nvidia is not the only way to make AI progress. In recent months, Chinese research institutions have reportedly developed AI processors based on RISC-V (pronounced “risk-five”) open-architecture design standards and large language models for military use based on open-source models, including Meta’s Llama.
The Chinese are not the only ones using RISC-V. Last February, US up-and-coming Nvidia competitor Tenstorrent announced an agreement to license its RISC-V CPU technology to Japan’s Leading-edge Semiconductor Technology Center for the development of an AI processor made up of chiplets fabricated and packaged in Hokkaido by Japan’s new IC foundry Rapidus.
China is using chiplets as a way to get around its lack of access to EUV lithography tools. As explained in the MIT Technology Review:
In contrast to traditional chips, which integrate all components on a single piece of silicon, chiplets take a modular approach. Each chiplet has a dedicated function, like data processing or storage; they are then connected to become one system.
Since each chiplet is smaller and more specialized, it’s cheaper to manufacture and less likely to malfunction. At the same time, individual chiplets in a system can be swapped out for newer, better versions to improve performance, while other functional components stay the same.
RISC-V is an open standard instruction set architecture based on Reduced Instruction Set Computer design principles. It is a free, non-proprietary platform for the development of IC processors.
An alternative to Arm, Intel, AMD and Nvidia, RISC-V is not only of great interest to China but also to the EU and to smaller companies and IC designers seeking to develop a lower-cost independent presence in semiconductor and computing markets.
US sanctions have forced China to accelerate the development of a national ecosystem for advanced ICs based on RISC-V.
The RISC concept was conceived at the University of California, Berkeley, in 2010. The RISC-V Foundation was established in 2015 to support and manage the technology, with the Institute of Computing Technologies of the Chinese Academy of Sciences as one of the founders. Other Chinese members of the foundation include Huawei, ZTE, Tencent and Alibaba.
In 2020, the foundation was incorporated in Switzerland as the RISC-V International Association, moving out of the United States to avoid potential interference by US President Donald Trump. Beyond the reach of US sanctions, China is now estimated to account for about half RISC-V core shipments worldwide.
Meanwhile, DigiTimes reports that Nvidia has hired “hundreds” of new employees in China to work on autonomous driving. BYD and about a dozen other Chinese auto companies have adopted its DRIVE Orin system-on-a-chip (SoC) for autonomous electric vehicles, generating annual revenues for Nvidia of over $1 billion per year.
Looking ahead to 2025, at least five Chinese EV makers – BYD, XPeng, GAC-Aion, Li Auto and Zeekr – plan to use DRIVE Thor, the successor to DRIVE Orin. But DRIVE Thor incorporates the generative AI capabilities of Nvidia’s leading-edge Blackwell architecture. Will the US government try to stop this as well?
The Chinese are hedging. BYD, Li Auto, ChangAn and many other Chinese automakers are also working with Horizon Robotics, China’s leading computing solutions for advanced driver assistance systems (ADAS) and automated driving (AD) for consumer vehicles. So is Volkswagen. In October, Horizon Robotics went public in the largest IPO in Hong Kong this year.
Follow this writer on X: @ScottFo83517667
The chiplets approach, processors dedicated to specific functions, is nothing new, IBM did that in the glory days of mainframe computers. The central electronic complex was channel attached to special purpose computers: storage directors, network processors, vector processors, etc., thus allowing greater throughput.