Core conclusion: The launch of Manus, the world's first autonomous AI agent, marks a qualitative change in AI applications from text interaction to task delivery, and the computing power demand for a single task may increase by a hundredfold. Domestic AI computing and cloud service vendors will deeply benefit from the exponential growth in demand for inference computing power, and it is recommended to focus on leading enterprises with leading technology and perfect ecosystems.

1. Industry dynamics: Manus is reshaping the AI agent paradigm, and the demand for computing power is structurally upgraded

1. Technological breakthrough: As the world's first AI agent that can directly deliver task results, Manus' core capabilities have been upgraded from the thousand-token response (1k tokens/time) of traditional chatbots to complex task processing (100k tokens/time), which increases the demand for computing power by two orders of magnitude.

2. Demand-driven: AI applications have shifted from "suggestion output" to "result delivery", requiring the invocation of multimodal models, real-time data interaction, and complex decision-making algorithms, resulting in a surge in inference computing power consumption. It is estimated that the computing power required for a single task may be 100 times that of traditional scenarios.

Second, the main line of investment: domestic computing power and cloud services two-wheel drive

(1) Domestic AI computing power: independent and controllable + performance breakthrough

1. Cambrian (688256).

- Core advantages: The performance of the cloud chip Siyuan 370 is 15% of that of NVIDIA A100, and the cumulative shipment of edge chip Siyuan 220 exceeds one million pieces, and customer A contributes the main revenue. In 2024, revenue is expected to increase by 50%-69% year-on-year, and public funds will increase their positions significantly.

- Catalytic factors: The fifth-generation MLUarch04 architecture has a 50% increase in computing power and a 40% optimization in energy consumption, which is deeply adapted to the training requirements of large models.

2. Haiguang Information (688041).

- Technical barriers: The DCU adopts a "CUDA-like" architecture, supports full-precision training of large models such as LLaMa and GPT, and has a performance close to 80% of that of NVIDIA A100. In 2024, the proportion of DCU business will increase to 35%, and the revenue will exceed 9 billion yuan.

3. Inspur Information (000977).

- Market position: The global AI server market share exceeds 50%, the power density of a single cabinet of liquid cooling technology reaches 100kW, and the PUE is as low as 1.1.

4. Unisplendour (000938).

- Ecological layout: H3C switches have a market share of 36.5%, the self-developed liquid cooling solution has been deployed on a national project scale, and the secondary listing of Hong Kong stocks plans to raise US$1 billion to strengthen R&D.

(2) Cloud and all-in-one machine manufacturers: scenario landing + ecological collaboration

1. iFLYTEK (002230).

- Application landing: The Spark model supports multi-modal interaction, covering 50,000 schools in the education field and 3,000 hospitals in the medical field, with R&D investment accounting for more than 35% in 2024.

2. iSoftStone (301236).

- Breakthrough in information innovation: Launched the "iSoftStone Huafang" information innovation machine, deeply participated in the OpenHarmony ecosystem, and the revenue of information innovation increased by 60% year-on-year in 2024.

3. Convincing (300454).

- Security + cloud integration: The SASE architecture integrates zero trust security and edge computing, increasing the proportion of cloud computing revenue to 40% in 2024, and winning the bid for super large projects in finance and government affairs.

3. Risk Warning

1. Technical risks: The manufacturing process of domestic chips is limited, and the gap in energy efficiency ratio may affect long-term competitiveness.

2. Supply chain risks: GPUs and other high-end chips rely on imports, and geopolitics may lead to supply disruptions.

3. Market risk: The demand for computing power is less than expected, and the recovery of corporate profits is lagging behind.