(Yicai) May 29 — Mainstream demand for artificial intelligence computing in China is gradually shifting toward inference, which is creating new opportunities for domestic chips, according to industry insiders.

Before, Chinese data centers mainly relied on Nvidia chips because of the high-performance requirements for AI model training, with only a few using domestically developed alternatives, mostly as backup, Zhou Zhengang, vice president of International Data Corporation China, told Yicai.

But after the release of open-source, high-performance large language model DeepSeek-R1, more users have begun deploying LLMs in real-world applications, driving demand for AI inference, Zhou noted, adding that this shift creates opportunities for less powerful and more affordable domestic chips.

AI inference requires less computing power than AI training, enabling chips with performances lower than Nvidia’s H100 and H800 to be viable, said Rocky Cheng, chief executive officer of Cyberport’s Artificial Intelligence Supercomputing Center, which is Hong Kong’s largest AI supercomputing center, mainly serving local universities, research institutions, and enterprises.

Chinese internet firms and telecoms operators procured large quantities of domestic computing cards last year but quickly found them unpopular among users, Zhou said. However, after DeepSeek-R1 was adopted on a large scale, all these resources were put into use in the first quarter of the year.

“Teams developing LLMs like DeepSeek-R1 will gradually decrease, and demand for inference capabilities will become mainstream,” Cheng noted. “Our center’s next-phase computing configuration will prioritize this growing need.”

“Domestically developed chips will account for over 40 percent of computing cards deployed in Chinese data centers in the first half of this year and exceed 50 percent soon,” Zhou predicted. “This was unimaginable just two years ago.”

Last year, over 65 percent of the computing cards used in Chinese data centers were made by Nvidia.

Editor: Futura Costaglione