Guohai Securities: Cloud boom leads AI narratives to seize domestic computing power growth opportunities

The Zhitong Finance App learned that Guohai Securities released a research report saying that overseas CSP and Neocloud vendors are gradually verifying that large-scale computing power capital expenditure can be effectively converted into revenue and profit. Driven by agents, inference demand has the potential to continue to grow over a long period of time. Coupled with the sharp rise in domestic high-end computing power server channel prices, the bank anticipates that the tight supply and demand for AI computing power in China will increase, the medium- to long-term growth logic of the computing power leasing industry is more determined, and it has already entered a dividend period of performance release, maintaining the computer industry's “recommended” rating.

Guohai Securities's main views are as follows:

CSPs are gradually turning computing power into revenue, and demand spills over from cutting-edge laboratories to enterprise customers

The revenue and profit margins of the North American CSP cloud business have accelerated across the board. Microsoft FY2026q4 Microsoft Cloud revenue was $59.3 billion, +27% year over year, and Azure and other cloud services revenue was +43% year over year, and guided Azure's growth rate of about +45% year over year next quarter; smart cloud was about 39.3 billion US dollars, about +32% year over year. Azure revenue in fiscal year 2026 exceeded 100 billion US dollars, and M365 Copilot surpassed 30 million paid seats. Google Cloud's second-quarter revenue was 24.8 billion US dollars, +82% year over year; operating profit was about 8.8 billion US dollars, operating profit margin 35.6%, +14.9 pct year on year. AWS's revenue for the second quarter was approximately US$42.2 billion, +37% year-on-year, with an operating profit margin of 39%. This is the fifth consecutive quarter that AWS has achieved accelerated growth; both the AI business and the chip business have annualized revenue and operating rates of over 25 billion US dollars.

Orders continue to grow, and demand spills over from cutting-edge laboratories to enterprise customers. Microsoft's FY2026Q4 RPO reached $678 billion, +84% year over year. All of the month-on-month growth in FY2026Q4 commercial RPO came from customers other than cutting-edge model companies. Google Cloud's backlog for the second quarter was about $514 billion, an increase of more than 50 billion US dollars over the previous quarter. Management said that the vast majority of these were regular GCP contracts from a broad customer base, and slightly more than 50% will be settled in the next 24 months. AWS's backlog of orders in the second quarter was approximately US$496 billion, +153% year over year and +36% month over month.

Neocloud orders and financing are placed at the same time, and high income elasticity and high capital intensity coexist

Among overseas leaders Neocloud, CoreWeave's 2026Q1 revenue was 2.1 billion US dollars, +112% year over month; contract revenue backlog orders reached 99.4 billion US dollars at the end of the 2026Q1 quarter, up nearly 50% month-on-month, up about 4 times year-on-year. Of this, about 36% are expected to be collected within the next 24 months, and 75% are expected to be confirmed within the next 4 years. From the beginning of the year to May, CoreWeave completed more than $20 billion in debt and equity financing. Nebius's 2026Q1 revenue was $399 million, about +684% year over year. As of mid-July 2026, Nebius publicly announced financing in 2026 that had reached US$7.1 billion.

Domestic computing power leasing companies are expected to gradually enter the dividend period of high performance growth in 2026:1) Co-innovation data: 2026Q1 achieved operating income of 6.085 billion yuan, a sharp increase of 192.90% year on year; net profit to mother reached 750 million yuan, up 343.45% year on year, up 60% month on month. 2) Litong Electronics: In 2026Q1, revenue was 997 million yuan, up 41.61% year on year; net profit to mother was 271 million yuan, up 821.08% year on year and 253.60% month on month. 3) Hongjing Technology: As of 2026Q1, the company's total assets increased sharply to 12.945 billion yuan, an increase of 143% from 5.314 billion yuan at the end of 2025; projects under construction were 1,515 billion yuan, an increase of more than 17.41 times over the previous year.

The inference period further boosts demand for long-term computing power, and domestic B300 server prices continue to rise

Reasoning can drive computing power requirements for a longer period of time than training, and Agent AI is amplifying inference requirements. Training is more of a one-time computing power investment; reasoning is a sustainable computing power requirement. Moreover, AgenticAI is also amplifying inference requirements: compared to Chatbot AI, the shape of the Agentic AI workload has changed. Instead of answering a question, the agent breaks down the target into multiple steps, and needs to decide what to do next, call multiple models, query the database, connect to the API, run the application, check permissions, retrieve memory, verify the output, and then cycle through the entire process again.

The transaction price of B300 servers in mainland China continues to rise. According to Fast Tech, the release of Seedance and Nano Banana drove a significant increase in token throughput. According to China Foundation News, as of June 2026, the Doubao Big Model's average daily token call volume reached 180 trillion dollars, an increase of more than 10 times over the past year. According to the official account of Zhiding AI Laboratory on July 21, 2026, the price of B300 servers has risen to 14.5 million yuan per unit in domestic channels, an increase of more than 2 times from about 4 million yuan at the beginning of the year. The B300 server has significantly improved inference performance compared to the H200, and its price is expected to rise further.

Focus on individual stocks

1) CPU: Haiguang Information, China Great Wall (Feiteng Information); 2) Connectivity: Lanqi Technology, Shengke Communications, Wantong Development, Ruijie Network, Huafeng Technology, Aerospace Electronics, Huagong Technology; 3) AI Chips: Cambrian, VeriSilicon, Mu Xi, Tianshou Zhixin; 4) Computing power leasing: Co-Creation Data, Hongjing Technology, Litong Electronics, Saiyi Information, Shengshi Technology, Zhiwei Intelligence, Runjian Co., Ltd.; 5) Supernode OEM: Inspur Information, Zhongke Information Shuguang, Ziguang Co., Ltd., Huaqin Technology, and IFF.

Risk warning: the Sino-US game intensifies, macroeconomics affects downstream demand, AI reasoning demand falls short of expectations, increased market competition, supply chain risks, etc.