Bank of China International: Domestic CSP capital expenditure accelerated+big model price increase, domestic computing power boom verified again

The Zhitong Finance App learned that Bank of China International released a research report saying that the domestic computing power industry chain is gradually shifting from thematic expectations to performance implementation. The bank believes that the increase in capital expenditure of domestic CSP manufacturers such as Tencent and price increases for large models represented by DeepSeek are becoming the two core engines that continue to drive domestic computing power demand. The industrial chain is expected to enter a new stage where demand is booming, supply is expanding rapidly, and performance continues to be released. Judging from the mid-term report feedback in the last week, results such as core chip manufacturing for underlying hardware, computing power chips, switch chips, and server/switch assembly have entered the implementation stage. Large-model price increases and CSP manufacturers' capital expenditure increases have become the core engine that continues to be driven by domestic computing power. The bank believes that the domestic computing power industry chain has entered a period of high certainty and high growth and realization.

Bank of China International's main views are as follows:

Domestic CSP capital expenditure expansion trend established

In 2026H1, Tencent's capital expenditure reached 84.720 billion yuan, a year-on-year increase of 82%. Among them, the increase in the scale of computing power procurement was one of Tencent's main capital expenditure investments. As an enterprise with relatively small AI capital expenditure among domestic CSP manufacturers, the expansion of its capital expenditure scale has further strengthened the expansion trend of domestic CSP. The bank expects CSP vendors such as Ali and Byte to further expand the scale of capital expenditure this year. Leading cloud vendors such as Tencent have increased investment in AI infrastructure, reflecting the growing demand for large-scale model applications and cloud services. It also means that the construction of domestic computing power clusters will gradually evolve from single-point projects to large-scale and continuous procurement. Considering that CSP capital expenditure is usually ahead of server shipments and data center construction, it is expected that the increase will sequentially be transmitted to computing power chips, servers, switches, network interconnections, etc., thereby increasing the visibility of orders in the industrial chain. As Internet vendors, operators and industry customers jointly increase AI-related investment, the demand base for domestic computing power will become more diversified, and the sustainability and certainty of industry prosperity is expected to further increase.

Price increases for large models confirm that computing power and economic efficiency are improving

Recently, DeepSeek has drastically adjusted the price of the DeepSeek API, using peak and valley pricing. Among them, the output price of the flagship model DeepSEEK-V4-PRO rose to 27 yuan per million tokens during peak periods, which is about 350% higher than the current price; the output price per million tokens during idle time is 13.5 yuan, which is 2.25 times the current price. The deepseek-V4-Pro price increase is a significant industry signal. On the one hand, large model inference costs increase as parameter size, context length, and user calls increase, and price adjustments help improve model makers' unit computing power returns and commercialization capabilities; on the other hand, improving model capabilities and gradually maturing payment models will drive the big model industry to shift from “getting customers at low prices” to “value pricing” and improving the closed commercial loop of AI services. The bank believes that the increase in model call prices does not mean that demand for computing power is weakening; on the contrary, it has verified the tight supply and demand for high-performance models and the scarcity of high-quality computing power. As inference demand spreads from Internet applications to scenarios such as office, finance, industry, and government affairs, computing power requirements will shift from concentrated investment on the training side to “equal emphasis on training and reasoning”, thus driving the continuous expansion of infrastructure such as servers, switches, optical interconnects, PCBs, high-speed connectors, power supplies, and liquid cooling.

Domestic semiconductors are driven by AI to set sail, and semiconductor production expansion equipment takes the lead

Looking at the industrial chain level, the main constraint on domestic computing power is not on the demand side, but mainly due to insufficient supply of high-end chips and advanced manufacturing capacity. Accelerated iteration of computing power chips will simultaneously drive the expansion of wafer manufacturing, advanced packaging, equipment, components, and materials. The forward stage benefits from the construction of advanced and characteristic process production capacity, while the back stage benefits from increased demand for advanced packaging, packaging substrates, and high-density interconnections for high-computing chips. As a pioneering step in expanding production, semiconductor equipment is expected to be the first to be fulfilled and further transmitted to core components, thin film deposition, etching, cleaning, inspection and packaging equipment. Capital market catalysts related to the expansion of production in the storage industry and the listing of Yangtze River Storage are also expected to increase the attention of the domestic equipment and materials sector.

Key risks

The capital expenditure of domestic Internet vendors and operators fell short of expectations; the growth in demand for AI applications and tokens fell short of expectations; the performance iteration or production capacity release of domestic AI chips fell short of expectations.