FT: China wants to triple output of AI chips in race with US. Which companies are at the forefront

Chinese manufacturers plan to triple output of artificial intelligence chips in 2025 to catch up with the US in the AI race, sources told the Financial Times. Beijing is actively supporting the industry with finance and policy, hoping to create its own ecosystem and challenge Nvidia's global leadership.

Chinese chip manufacturers intend to triple the output of chips for artificial intelligence in 2025, the Financial Times reports, citing sources. According to the publication, one new plant, focused on the production of AI-processors for Huawei, may start operating by the end of this year, two more - next year. At the same time, although the enterprises are being built for the needs of Huawei, their ultimate owners remain unknown. The company itself has denied having plans to launch its own factories and did not provide additional details, adds the FT.

The combined capacity of the three new factories after reaching full capacity may exceed the current volume of similar lines at SMIC - the largest contract manufacturer of microchips in China, writes the FT. According to the publication, SMIC also plans to double the output of 7-nanometer chips - the most technologically advanced, which are now mass produced in the country. The company's largest customer for these processors remains Huawei. This will allow smaller developers like Cambricon, MetaX and Biren to get more SMIC production capacity and compete more actively in the rapidly growing Chinese market after the U.S. restrictions on exports of Nvidia products.

China is seeking to catch up with the US in the race for advanced AI technology, and semiconductors remain a major point of tension in the Beijing-Washington trade conflict, the FT notes. The US has cut off China's access to Nvidia's advanced AI chips in order to slow China's technological progress. For example, China's AI market leader, startup DeepSeek, is still training its models on Nvidia clusters, but that could change when domestic chips tighten up performance and become compatible with its standard. "Domestic production will soon cease to be a problem, especially with capacity coming online next year," an executive at a Chinese chip maker told the FT.

Currently, Huawei 910D and Cambricon 690 processors are best suited for DeepSeek's requirements, but other Chinese companies are accelerating the release of their own versions of AI chips adapted to this standard. DeepSeek recently announced that its models work with the FP8 data format. It improves hardware efficiency at the expense of reduced accuracy and could give China a chance to compete with global leaders, even lagging behind Nvidia by several generations, the FT emphasizes.

However, it will take years of collaboration between processor, memory, networking and software vendors to create a full ecosystem. "If we are able to design and optimize these Chinese chips to train and operate Chinese models in an ever-evolving ecosystem, one day we will look back and see this as an even more significant 'DeepSeek moment,'" adds an executive at a Chinese chipmaker. - This unprecedented alignment may offset our less advanced hardware capabilities."

In addition, China is accelerating developments in memory, a market now controlled by Samsung, SK Hynix and Micron - all of which are subject to U.S. export restrictions. For example, local manufacturer CXMT is testing samples of HBM3, a high-bandwidth memory that is just one generation behind that used in Nvidia chips. It's expected to launch in 2025.

Beijing is actively supporting the sector politically and financially: this week, the State Council called for accelerating AI adoption and building an integrated ecosystem "from research to commercialization." This year, chipmaker Cambricon received regulatory approval to raise about $560 million through an additional share issue, and four other companies, including Biren and MetaX, are preparing for IPOs, having already raised about $3 billion in the preparatory phase.

If the course of bringing together hardware and software makers in AI proves successful, it could fulfill DeepSeek founder Liang Wenfeng's call to challenge Nvidia's global leadership. "Nvidia's leadership is not created by one company, but by the joint efforts of the entire Western community and industry," Liang told Chinese media last year. - The development of AI in China needs the same ecosystem. China needs someone to stand at the forefront of this process."

SMIC, Cambricon, DeepSeek, CXMT, Biren and MetaX declined to comment to the FT.

Beijing seeks to develop its own semiconductor technologies to limit its dependence on Western solutions. Thus, the Chinese authorities are already recommending local companies to refuse Nvidia H20 chips for artificial intelligence, especially in projects related to the state and national security.

Against this background, Nvidia is halting production of AI chips for China, Bloomberg sources reported. This casts doubt on the return of the H20 to the Chinese market, even though Washington said it was ready to allow the supply of these chips to China again, CNBC noted.

This article was AI-translated and verified by a human editor