Dongwu Securities: Cloud vendors still maintain high computing power demand, computing power rental demand is expected to rise marginally

The Zhitong Finance App learned that Dongwu Securities released a research report saying that currently, in the rapidly developing digital economy era, computing power is becoming more and more important as core productivity. Computing power leasing and IDC complement each other in the computing power ecosystem, giving full play to their advantages in different demand scenarios to form a complementary computing power service system. The “asset-light” model of computing power leasing satisfies the short-term and flexible computing power needs of enterprises; while IDC's “asset-heavy” model provides a solid foundation for long-term, high-stability businesses. With the rise of emerging technologies such as AI, the collaborative development of the two will profoundly influence the industry pattern and push the entire computing power market to new heights.

The main views of Dongwu Securities are as follows:

Computing power leasing and IDC complement each other in the computing power ecosystem. The core differences between the two can be divided into three aspects: resource control, service content, and cost structure

Computing power leasing is a “purchasing service” model. Users can obtain computing power by leasing computing resources such as GPU clusters without purchasing their own hardware. The pay-as-you-go model is also responsible for all hardware and technology iterations, and the pay-as-you-go model is also more flexible, suitable for short-term flexible requirements and scenarios where technology updates are faster; IDC is a “self-built or leased infrastructure” model. Users need to own or rent their own resources such as cabinets, electricity, networks, etc., and manage server, storage equipment and application deployment themselves. It is also suitable for enterprises with high upfront investment but more conducive to long-term stable business. Overall, computing power leasing emphasizes the lightweight output of “computing power as a service,” while IDC provides “heavy asset holding+infrastructure services”. The two models give full play to their advantages in different demand scenarios to form a complementary computing power service system.

Cloud vendors are gradually benchmarking DeepSeek's efficiency and cost, and training is still in high demand

Top AI companies, represented by DeepSeek, have clear requirements for computing power leasing through algorithm-hardware collaborative design, such as sparse training, mixed accuracy optimization, and distributed training framework upgrades (such as 3D parallel strategies). Its core logic is to respond flexibly to fluctuations in computing power demand, reduce hardware investment costs, and optimize resource utilization. The increase in demand for computing power of agents (agents) is mainly due to the triple increase in penetration rate, average number of daily calls, and task complexity. For example, the enterprise side began self-built model training due to the expansion of AI applications (such as intelligent manufacturing and financial analysis), further increasing computational pressure — for example, DeepSeek makes 25 million inference calls in a single day, requiring support from 1,814 H800 chips. Under this dual driving trend, computing power requirements migrate from the training side to the inference side.

Under tight supply-side constraints, there is still room for price increases in computing power leasing due to scarcity premiums and cost transmission mechanisms

US chip export restrictions to China continue to increase, and the supply of high-end GPUs such as Nvidia's H100/H20 to China has been strictly limited. After the Nvidia H20 was banned, the unit price of the H800 server continued to rise gradually, and the cost pressure on the hardware side was transmitted to the rental side. According to the “Guidelines for Innovation and Development of Intelligent Computing Centers”, the scale of China's intelligent computing power grew rapidly in 2022, reaching 268 billion times per second (EFLOPS), exceeding the scale of general computing power. The compound annual growth rate of China's intelligent computing power is expected to reach 52.3% in the next 5 years.

The essence of competition in the IDC industry is a battle for resource endowments. The core of computing power leasing lies in the scale of GPU resources and channel stability

IDC's core barrier is to obtain land resources and energy consumption indicators in first-tier cities and surrounding areas, and reduce operating costs through green energy-saving technology. The policy requires data center PUE to be gradually reduced, and under the green electricity ratio requirements, the approval of energy consumption indicators for first-tier cities has become stricter, and data centers are gradually deployed to other non-first-tier cities. The competitiveness of the computing power rental industry depends on the scale of GPU accumulation and the stability of card acquisition channels. Especially in the context of limited supply of Nvidia's high-end chips such as the H100 and H20, resource monopoly is the key. Domestic alternatives (such as the Huawei Ascend 910B) have been accelerated, but there is still a gap in performance. In the short term, they cannot completely replace Nvidia chips, which require a kilocalories GPU cluster to train a single large model.

Related targets: Runjian Co., Ltd., Dawei Technology, Litong Electronics, Lanco Technology, Runze Technology, Guanghuan New Network, Tongniu Information, Zhongbei Communications, Hengrun Co., Ltd., Hainan Huatie, Aofei Data, and Hongxin Electronics.

Risk warning: Technological development falls short of expectations, geopolitical risks between China and the US, and increased industry competition.