Recently, Alibaba Group Chairman Tsai Chongxin talked about the shortage and limitations of AI chips in an interview with Norwegian Sovereign Wealth Fund Investment Director Nicolai Tangen.

Cai Chongxin introduced that currently, the chip inventory of domestic enterprises can support the training needs of AI large models for the next 18 months. In the next stage of application, which is the inference stage, there are many options in the market and it is not necessary to use Nvidia high-end chips.

Due to the complex international situation and the shortage of advanced packaging and HBM production capacity, the global development of AI chips is subject to certain limitations. However, Cai Chongxin believes that in the short and medium term, this is a problem, but in the long run, China will develop its own ability to manufacture these high-end GPUs.

Data shows that AI chips refer to modules that have been specially designed to accelerate AI algorithms and can handle a large number of computing tasks in AI applications. According to technical classification, AI chips include GPUs, FPGAs, ASICs, etc. Among them, GPUs have the largest usage in the AI chip market, and general-purpose computing power GPUs have been widely used in the field of artificial intelligence model training and inference. In this field, Nvidia and AMD are currently leading the way.

The artificial intelligence industry in our country started relatively late. In recent years, thanks to policies and AI applications, the AI industry and AI chips have gradually developed and performed well in terminal applications and large-scale model inference. However, there is still room for catching up in high-end GPUs and training processes.

Since 2023, with the popularity of big models such as ChatGPT and Sora worldwide, related manufacturers have continued to benefit, and domestic AI big models and related AI chips have also begun to receive unprecedented attention.

During this year's "two sessions", Zhang Yunquan, member of the National Committee of the Chinese People's Political Consultative Conference and researcher of the Institute of Computing Technology of the Chinese Academy of Sciences, pointed out that the innovation and supply of domestic intelligent computing chips that can be used for large model training are significantly backward. Solving the bottleneck of computing power can be achieved by increasing the research and development of domestically produced high-end AI chips, concentrating AI chip research and development efforts, and setting up a special group for the development of intelligent computing power.

Cao Peng, a member of the National Committee of the Chinese People's Political Consultative Conference and Chairman of the Technical Committee of JD Group, stated that only with an independent and controllable computing power base can domestically produced large models gain an advantage in this AI competition. He suggested seizing the opportunity of the development of large-scale models, encouraging domestic GPUs to adapt to domestic computing power scheduling software through policies, building an independent and controllable intelligent computing foundation, and supporting the intelligent development of the industry.