"In terms of single-chip performance, we are still a generation behind the United States, but we will overcome that with cluster-based computing technology."
That was the remark by Ren Zhengfei, chair of Huawei (华为), who leads China's semiconductor industry. In an interview with the state-run People's Daily in June, he said this while noting of U.S. pressure that "we don't dwell on difficulties; we just move forward step by step."
And on the morning of the 18th (local time), Huawei unveiled new semiconductor technology. It is a technology that can link up to 15,488 graphics cards equipped with its Ascend products. Huawei also said it is currently operating an ultra-large cluster made up of 1 million graphics cards. The cluster-based computing technology Ren Zhengfei mentioned earlier has been realized.
That day, Huawei also announced specific plans to roll out artificial intelligence (AI) chips. They are, respectively, the Ascend 950PR with the company's designed AI memory chip in early 2026 ▲ the Ascend 950DT in late 2026 ▲ the Ascend 960 in late 2027 ▲ and the Ascend 970 in late 2028.
Bloomberg News said, "This is Huawei's attempt to raise performance so that Chinese corporations, which can no longer use Nvidia's high-spec products due to U.S. sanctions, can train and run AI with domestically made semiconductors," adding, "Rather than a groundbreaking technological breakthrough, it is the latest result of Chinese corporations developing homegrown alternatives under U.S. sanctions."
China is accelerating semiconductor self-reliance under government leadership. Major big tech firms and new fabless companies are seeing results in developing semiconductors to counter Nvidia, and foundry (contract manufacturing) companies are preparing for full-scale mass production, building a semiconductor ecosystem that does not rely on the United States.
According to the U.S. information technology (IT) outlet The Information, Alibaba has been using its self-designed semiconductor "Zhenwu" since early this year to train small-scale AI models. The report said the chip has performance that can compete with Nvidia's China-only low-spec chip, the H20. Baidu (百度), known as "China's Google," has also experimentally deployed its self-developed "Kunlun P800" chip to train and learn a new version of its in-house large language model (LLM).
Nvidia recently said of this, "Competition has definitely begun," adding, "We will continue working to earn the trust and support of mainstream developers worldwide." Of course, both companies still use Nvidia chips for cutting-edge model development, but it is a meaningful change that China's leading big tech firms are jumping into chip development and beginning to lower their dependence on Nvidia.
The growth of Cambrian (寒武纪), a relatively young corporation, is also steep. Specializing in AI semiconductor design, Cambrian is seeing a surge in sales of its "Siyuan (思元) 590" as the Chinese government actively encourages the use of domestic AI semiconductors. Thanks to that, its sales in the first half of this year increased 43-fold from a year earlier, turning a profit. Shanghai startup MetaX also released a chip in July to replace Nvidia's H20 and is preparing for mass production.
Not only chip development but production is expanding aggressively. China's foundry companies have begun full-scale mass production. Semiconductor Manufacturing International Corp. (SMIC), China's largest foundry, plans to double its 7-nanometer process chip production capacity next year. Its largest customer is Huawei, and if capacity expands, smaller fabless companies like Cambrian and MetaX are expected to be allocated more production volume.
The Chinese government is giving the industry focused support. Citing sources, the U.K.'s Financial Times (FT) reported that the Cyberspace Administration of China (CAC) recently ordered domestic corporations, including Alibaba and ByteDance, to halt purchases of Nvidia's latest China-only low-spec chip, the "RTX 6000D." Earlier, authorities urged restraint in buying Nvidia's earlier China-only chip, the "H20," citing concerns about a backdoor. Because of this, even though Nvidia received U.S. government approval to export the H20, actual shipments are said not to have taken place.
This is seen as an attempt to break away from dependence on the United States (Nvidia) in the semiconductor supply chain and achieve self-reliance. According to the FT, following the order to halt RTX 6000D purchases, some corporations actually canceled orders, and the tens of thousands of canceled orders are expected to be redirected to Chinese companies. Morgan Stanley projected that the share of domestic AI chips in China, just 34% in 2024, will rise to 82% by 2027.