Key Points:

Internet giants and the three major operators continue to actively increase their investment in computing power Capex. Recently, Tencent and the three major operators have announced their 2024 annual results, and disclosed their capital expenditures and plans. Tencent's capital expenditure for the full year of 2024 exceeded 76.7 billion yuan, a year-on-year increase of 221%, accounting for 12% of revenue, of which the capital expenditure in 24Q4 increased by 3.86 times to 36.58 billion yuan, in addition, Tencent expects that the capital expenditure for the full year of 2025 will continue to account for a low double-digit percentage of revenue. In 2025, the three major operators will increase their investment in computing power in the context of the overall decline in capital expenditure plans, of which China Telecom initially plans to invest 21.96 billion yuan in the field of cloud and computing power, and China Unicom plans to invest 18 billion yuan, an increase of 22% and 28% respectively over the same period last year, while China Mobile's computing power Capex plan only increased by 0.5% to 37.3 billion, but this is mainly based on pre-trained computing power, and inference resources are invested according to market demand and there is no upper limit.

The continuous iteration of models and applications will further increase the demand for computing power. On March 25, OpenAI's Wensheng Diagram app became popular all over the world as soon as it was launched, and a large number of "Ghibli" cartoon-style pictures converted from personal photos and well-known memes poured into Internet platforms, and due to excessive application demand, OpenAI had to limit its speed to alleviate the pressure of GPU overload; The Agent application Manus, which has attracted a lot of attention in the early stage, also announced its charging plan, Starter charges 39 US dollars / month, Pro charges 199 US dollars / month, you can get 3900 and 19900 points respectively, the more complex the task, the more points are consumed, according to reports, it takes 600 points to design and deploy a unique personal website, which lasts 40 minutes, which shows that complex applications need to consume high computing power costs. Therefore, we believe that with the popularization of complex applications such as multimodality and Agent, the demand for computing power on the inference side has greater demand and elasticity, and the "Andy Beale" law that exists in the traditional IT field is still true in the AI era, that is, the improvement of hardware performance will soon be consumed by software and applications, and NVIDIA CEO Jensen Huang also proposed at the recent GTC conference that the number of tokens consumed by Agentic AI may be hundreds of times that of conversational AI.

Policies in the field of computing power have been continuously introduced and the scale has continued to expand. Recently, Shanghai issued the "Implementation Opinions of Shanghai Municipality on Promoting the Innovation and Development of the Intelligent Computing Cloud Industry (2025-2027)", proposing that by 2027, the scale of the intelligent computing cloud industry will strive to exceed 200 billion yuan, and the scale of intelligent computing will strive to reach 200EFLOPS, of which independent controllable computing power will account for more than 70%. The "Implementation Plan for Artificial Intelligence "Molding Shencheng" released at the end of 2024 proposes that the scale of intelligent computing will strive to exceed 100EFLOPS by the end of 2025, and the new plan means that the scale of intelligent computing in Shanghai will be doubled in 2026 and 2027. In addition to Shanghai, Shenzhen and other places have also put forward specific plans and goals for the field of intelligent computing power.

Investment Suggestions and Investment Targets

We believe that although there are fewer major catalysts in the field of computing power in the near future, the iteration speed of AI models is still at a fast speed, and multimodal and Agent applications will consume more computing power, and the increase in computing power consumption will continue to bring prosperity to the computing power industry.

In the field of domestic computing power, it is recommended to pay attention to Haiguang Information (688041, buy), Cambrian-U (688256, not rated), and Sugon (603019, buy).

In the field of computing power services and IDC, it is recommended to pay attention to Hainan Huatie (603300, Buy), Runjian (002929, not rated), Runze Technology (300442, not rated), Yunsai Zhilian (600602, not rated), Tongniu Information (300895, not rated), Dawei Technology (600589, not rated), Sinnet (300383, not rated), inference cloud and all-in-one machine, and it is recommended to pay attention to Hesheng New Materials ( 002290, unrated), UCD-W (688158, unrated), Deep Sunda A (000032, not rated), Capital Online (300846, not rated), Sangfor (300454, buy), Xinzhi Software (688590, buy), Hengwei Technology (603496, not rated), ABT (688168, not rated), Pingao (688227, not rated), Transwarp Technology-U (688031, not rated).

Risk warning:

AI development is not as expected; Competition in the industry has intensified.