GuShiio.com Manus consumes millions of tokens to disassemble and solve a complex task, and the demand for computing power has increased sharply. Guosheng Securities is optimistic that the computing power base SMIC and the domestic computing power supporting industry chain are ushering in historic opportunities.
Manus will consume millions of tokens to dismantle and solve a complex task, and the demand for computing power will increase sharply. He is optimistic about the computing power infrastructure industry chain.Manus adopts a Multi-Agent architecture and runs in an independent virtual environment, building a bridge between thinking and action. It not only thinks, but also delivers results. According to the bad evaluation x.pin public account, Manus disassembles and solves a complex task. The number of tokens consumed may be millions or even higher, and the demand for computing power has increased sharply. We are optimistic that the computing power base SMIC and the domestic computing power supporting industry chain will usher in historic opportunities.
Manus has spawned more AI Agents in vertical scenarios and accelerated the development of end-side AI.Manus adopts a unique Multiple Agent architecture and runs multiple sub-Agents through virtual machines to realize task disassembly, tool invocation, and closed-loop delivery of tasks. With the open source of some technologies, the development threshold may be lowered and more AI Agents in vertical scenarios will be born. This will further stimulate the innovative vitality of the AI application layer, is expected to accelerate the implementation of end-side AI, drive a new round of consumer electronics product replacement cycles, and inject growth momentum into the new AI terminal of smart driving. AI glasses are the ideal form of personal AI Agents. With the increase in volume in the future, we are optimistic about the layout and card slots of relevant domestic suppliers in the SoC link.
Domestic model innovations and continuous progress are expected to promote the development of end-side AI and further feed back the demand for computing power.Alibaba Open Source Tongyi Thousand Questions QwQ-32B reasoning model uses only 32 billion parameters, which is comparable to the currently recognized open-source full-blood version of DeepSeek-R1 with the strongest 671 billion parameters (37 billion activated). QwQ-32B can directly experience full-blood performance on a computer or Mac equipped with consumer-grade graphics card. We believe that domestic model innovations and continuous progress are expected to promote the development of end-side AI and further feed back the demand for computing power.
Broadcom’s AI revenue exceeded expectations and won two more ASIC customers. There is a strong demand for customization.Broadcom’s FY25Q1 AI revenue was US$4.1 billion, a year-on-year increase of 77%. AI revenue exceeded the guidance of US$3.8 billion. The company guided that the AI revenue of FY25Q2 will increase to US$4.4 billion, a year-on-year increase of 44%. The company reiterated that the three CSP customers ‘SAM in fiscal year 2027 is forecast to be US$60 billion to US$90 billion. The company has previously conducted in-depth cooperation with two other CSPs and is expected to complete their XPU streaming this year. At present, two more CSPs have chosen Broadcom to develop custom accelerators to train their next-generation cutting-edge models, and the demand for customization is booming. From the perspective of Taiwan stocks, the revenue growth rate of Zhibang Technology and Taiguang Electronics in February 2025 was still outstanding. The Jevens Paradox is that improvements in technical efficiency often lead to an increase in the total use of resources, rather than a decrease. ASIC has sufficient momentum to pull goods, and reasoning needs maintain rapid growth.
Storage: AI infrastructure has generated high demand, and storage is gradually entering a price increase cycle.1) Supply side: The three storage giants (Samsung, SK Hynix, and Micron) continue to reduce mid-to-end storage production capacity. According to TrendForce data, the number of NAND Flash wafers dropped in Q1 in 2025 will decrease by 18% year-on-year, and the supply capacity will decline. 2) Demand side: Cloud manufacturers maintain high capital expenditures, and AI infrastructure has generated high demand for storage products. Taking AI servers as an example, the storage configuration of a single server has jumped from 8TB to 32TB. Demanding of smartphone inventory, AI and DeepSeek effects brings demand growth. With a combination of production capacity reduction and demand growth, storage products are expected to enter a price increase cycle, and it is optimistic that domestic storage manufacturers will benefit from both the volume and price increases in the industrial chain.
Risk warning: Downstream demand falls short of expectations, R & D progress falls short of expectations, geopolitical risks
