Zheshang Securities: Optimistic about the GPU cloud era, the value chain expansion of cloud vendors will continue to expand to upstream and downstream applications in the value chain

The Zhitong Finance App learned that Zheshang Securities released a research report saying that it is optimistic about the long-term investment opportunities brought about by the expansion of the value chain of cloud vendors in the GPU cloud era. Generative AI drives cloud computing competition from simple resource supply to comprehensive competition for “computing power infrastructure, model services, and enterprise AI platforms”. Cloud vendors can continue to expand upstream value chain model services and downstream enterprise applications with GPU resource pools, data center scale, and AI ecosystem advantages to enhance unit resource value and profit space.

The main views of Zheshang Securities are as follows:

The core of the large profit gap between Chinese and US cloud vendors in the past is: 1) differences in industry chain value; 2) differences in scale; 3) differences in business models

Looking forward to the future, the bank believes that 1) the profit margin ceiling for cloud vendors in the GPU cloud era will increase as the value of the industrial chain increases; 2) the differences between cloud companies between China and the US will also narrow drastically as AI-driven prosperity increases and business models change.

Looking ahead to the GPU cloud era: 1) There are still objective differences in the value chain between Chinese and US cloud vendors in the network transmission layer, but both industrial chains have deployed AI chips upstream and expanded MaaS and Agent businesses downstream, raising profit margin ceilings; 2) the gap between China and the US narrows due to scale effects. American cloud vendors started early, and the top 2 vendors are already in the harvest period of economies of scale in the CPU era. As AI brings new infrastructure requirements, Google Cloud has also ushered in a period of economies of scale in the past 2 years. The bank believes that as domestic AI demand increases and companies such as Alibaba Cloud actively go overseas, the future will also enter a harvest period of economies of scale, and the profit margin gap brought about by scale effects will gradually narrow; 3) The big model is more suitable for public cloud deployment, and the differences in business models between Chinese and US cloud manufacturers will also narrow. Currently, enterprise-level applications based on the big model use the MaaS model as the main implementation path and mainly rely on public cloud platform deployment.

Industrial chain value differences: North American cloud factories build their own “global backbone networks”, domestic cloud factories rent bandwidth services from operators

1. Domestic operators build backbone networks. Domestic cloud vendors do not have the qualifications to operate basic telecommunications services, cannot build the underlying public network infrastructure, and can only rent bandwidth services from the four major licensed operators. Bandwidth fees have become one of the important costs for cloud vendors.

2. North American cloud vendors build their own sea-land optical cables to build a global network, and the network transmission layer value is included in the cloud business. Taking the profitability of two companies, Cogent and Lumen, which are highly dependent on self-built backbone networks to operate, the bank believes that the average 15% OPM for the backbone related business of the three major cloud companies in the US is one of the main sources of the difference in profit margins between China and the US cloud companies.

Difference in scale: The cloud business is a business with significant scale effects. US cloud vendors started early, and vendors are in the harvest period of economies of scale

1. Lower procurement costs. Manufacturers bargain and reduce prices on a large scale of procurement, while direct ODM procurement eliminates intermediate links, effectively reducing hardware procurement expenses. In 2015-2020, the average OPM of ODM manufacturers Guangda and Ultra Micro was only 2% or 4%, which is lower than the 10% and 11% profit margins of the OEM Dell and HPE server businesses.

2. Greater R&D investment builds product leadership, improves resource utilization, and reduces customer transfer costs. In 2013-2025, although R&D expenses are increasing year by year, Ali's total R&D volume and R&D rates are not as high as those of its American counterpart.

3. Global customer erroneous peak usage brings higher resource utilization and investment cost sharing. 1) High utilization rate of computing power resources. In 2022, the US cloud data center listing rate (65%) is higher than domestic (58%). Less idle computing power makes the unit computing power cost lower; 2) better share enterprise personnel costs. The profit margin of Google Cloud increased in 2023-2025, and the share of personnel related costs and expenses in revenue decreased by 3.8, 10.1, and 9.9 percentage points, respectively.

Business model differences: The customer's willingness to pay is different from the business model, which is mainly reflected in the willingness to deploy public and private clouds

1. In terms of total volume, the willingness of Chinese companies to pay for IT was lower than the global average in the past. In 2024, the global IT spending market reached 5.04 trillion US dollars, accounting for 4.5% of GDP; China's IT spending scale was nearly 2.7 trillion yuan, accounting for 2.0% of GDP.

2. From a structural point of view, the US public cloud deployment ratio is high, and customers are willing to pay for public clouds; domestic central enterprises prefer private clouds due to compliance restrictions. Although small and medium-sized enterprises prefer public clouds, they are sensitive to bargaining and are less willing to pay. However, because private clouds lack scale effects and scale scheduling, the business model is not as good as public clouds.

Risk warning: industrial competition intensifies; AI commercialization falls short of expectations; regulatory policy constraints; technological change and substitution