Guotai Junan Securities: Manus may drive computing power demand, continue to be optimistic about the AIDC industry chain.

Dunhong Securities released a research report stating that recently, the Chinese startup company Monica officially released the general-purpose AIAgent product Manus. The progress of domestic AI models continues to accelerate, and is expected to accelerate algorithm construction. The firm believes that AIDC and related industries will continue to benefit from the AI wave. The suggested focus areas are: AIDC operators, computing power leasing companies, edge AI and AIDC power supply. The main points of Dunhong Securities are as follows: Manus has clear features and performs well in handling multi-step tasks. Monica defines Manus as the world's first general-purpose Agent product, achieving surpassing performance over OpenAI Deep Research at all three difficulty levels in the GAIA benchmark test. Compared to other AI products, Manus mainly has five main features: 1) Autonomy in execution: able to independently complete tasks in the cloud without human intervention, and deliver complete task results directly; 2) Multi-agent system: supported by various models, with powerful tool invocation capabilities, enabling flexible code writing, intelligent web browsing, and operation of various applications; 3) Continuous learning and memory: Manus has memory function and can optimize task execution strategies based on historical data and experience; 4) Wide range of applications: covering six categories including research, life, data analysis, education, productivity tools, and creative entertainment, totaling 51 specific use cases; 5) Efficient human-machine collaboration: pioneers a new paradigm of human-machine collaboration, where users can simply describe their requirements in natural language, and Manus can quickly respond and execute tasks. The token consumption for a single task in Manus is much higher than traditional Chatbots, which may accelerate the demand for computing power. Manus adopts a Multiple Agent architecture, dividing complex tasks into planning, execution, verification, and other sub-modules, with each agent based on independent language models or reinforcement learning models, working together through APIs, ultimately calling tools (such as coding, data scraping) in a virtual machine to complete tasks. The cost of a single task in Manus is about 2 USD, which is exponentially higher in token consumption compared to traditional Chatbots. Continual releases of large domestic models, improving capabilities and reducing deployment costs. On March 6, Alibaba officially released the latest open-source inference model QwQ-32B. Through large-scale reinforcement learning, QwQ-32B has made a qualitative leap in mathematics, coding, and general capabilities. According to the Alibaba Unity team, although QwQ-32B only has 32 billion parameters, its performance can rival DeepSeek-R1, which has 671 billion parameters, with 370 billion activations. Additionally, QwQ-32B significantly reduces deployment costs, allowing the model to be deployed locally even on consumer-grade GPUs. In addition, Alibaba Cloud plans to combine more powerful base models with RL based on scaled computing resources, to bring it closer to achieving Artificial General Intelligence (AGI). Risk warning: Agent development falls below expectations; industry competition intensifies; geopolitical conflict risks.