Zheshang Securities: Increased capital expenditure of major Internet companies is expected to drive AIDC to build an “arms race”

The Zhitong Finance App learned that Zheshang Securities released a research report saying that ByteDance's capital expenditure in 2024 reached 80 billion yuan, which is close to the sum of Baidu, Ali, and Tencent (about 100 billion yuan). ByteDance's capital expenditure is expected to reach 160 billion yuan in 2025. ByteDance estimates that future token consumption is high, and it is expected that it will continue to increase its investment in computing power. In September 2024, Wu Di, head of the ByteDance Volcano Engine's intelligent algorithm, said that Doubao's daily token consumption in 2027 is expected to exceed 100 trillion dollars, which is more than 100 times the original. At the same time, Internet companies such as Baidu, Ali, and Tencent are also participating in the big model competition, striving to expand model parameters and increase the number and usage of users. The development of large models is inseparable from huge investment in computing power. It is expected that the Internet industry may generate a new wave of IDC construction in the future.

IDC is at the bottom of the cycle, and there may be a gap in the future

After experiencing a wave of expansion in China's IDC market in 2020, the industry entered a downward cycle. According to KeZhi Consulting data, the growth rate of the traditional IDC business market slowed from 22.1% in 2020 to 6.4% in 2023, and the sales price of data center cabinets also declined. The price of customized cabinets without motors fell nearly 20% since 2020. Prices in some regions are close to the cost line, and cabinet prices for wholesale and retail projects have dropped slightly. In the context of shrinking supply and explosive demand growth, due to the rigidity of IDC supply, the construction cycle is usually 1.5-2 years, so there may be a gap in the IDC market, which will drive the price of IDC construction and IDC services to rise.

AIDC places higher demands on energy consumption, electricity prices and heat dissipation

Existing IDC centers were mainly built before the big model era, or it may be difficult to meet the needs of intelligent computing. This round of IDC incremental demand is mainly due to increased demand for AI large model training and inference. Traditional IDC (Internet Data Center) mainly provides data services for internal and external users to access, while AIDC (Intelligent Computing Center) provides computing power, storage and related services for AI training and inference. In order to meet the growing computing power demand of large models, the computing power cluster is growing to the 10,000 card level, and AIDC's requirements for energy consumption, electricity prices, and heat dissipation are rising.

AIDC is mainly based on chips such as GPUs. The power of a single cabinet can reach tens of kilowatts, and needs to consume a large amount of electricity to support its high-performance computing and data processing tasks. Therefore, AIDC needs to obtain energy consumption indicators and obtain relatively cheap electricity. In this context, green energy and energy storage are the better choices. In terms of heat dissipation, driven by the high thermal density trend of computing power equipment and data center computer rooms and high energy efficiency cooling requirements, AIDC needs to be equipped with an efficient cooling system, such as a liquid cooling system.

Related targets

IDC: Guanghuan New Network (300383.SZ), Runze Technology (300442.SZ), Data Port (603881.SH), Yunsai Zhilian (600602.SH), Kehua Data (002335.SZ), Aofei Data (300738.SZ), Runjian Co., Ltd. (002929.SZ).

Liquid cooling: Gaolan Co., Ltd. (300499.SZ), Invec (002837.SZ), Shenling Environmental (301018.SZ).

Switches: ZTE (000063.SZ), Ziguang (000938.SZ), Ruijie Networks (301165.SZ), Shengke Communications (688702.SH), Philinkos (301191.SZ).

Risk Alerts

GPU supply falls short of expectations, demand for computing power falls short of expectations, construction of computing power centers falls short of expectations, and market competition risks.