CMSC: The main trend of the AI industry is shifting towards "Token production and commercialization". It is recommended to focus on the commercialization of large models in the second half of the year.

CMSC released a research report stating that by 2026, the main focus of the AI industry is shifting from "expanding CapEx" to "Token production and commercialization." In the first half of the year, large models and AI Infra became the few strong directions within the sector. In terms of large models, technical competition has shifted from parameter expansion to inference, coding, and agents, with enterprise-level coding scenarios leading the way in commercialization. On the computational power side, the Token economy is driving increased demand for computing power rentals, with domestic chips and computational electrical appliances forming the medium to long-term foundation. On the application side, the commercialization of Agent relies on Data Infra, with the B-side being the main battlefield for the scaling of AI applications due to higher payment capabilities and verifiability of results. Looking towards the second half of the year, it is recommended to focus on the commercialization of large models, grasp certain areas such as domestic computational power and computational power rental, and prioritize high elasticity directions such as Data Infra and applications. The main points of CMSC are as follows: The trend of the sector in the first half of 2026 shows a decline, with large models and AI Infra leading the way while applications fall behind. In the first half of 2026, the Shanghai Computer Index fell by 17.4%, ranking 21st among the 31 primary industries in the Shanghai index. The sector showed significant differentiation, with a market structure characterized by two stages: in Q1, driven by price increases by cloud providers, the emergence of OpenClaw, and the catalyzing effect of a large-scale battle of large models, AI applications and cloud infrastructure saw a significant rise. In Q2, driven by price increases in computational power rentals, shortages of cards from major manufacturers, and better-than-expected performance of leading companies in the computational rental sector, the computational rental sector showed an uptrend. The large models market continued to perform well in the first half of the year, with targets such as KNOWLEDGE ATLAS, MiniMax, Central Control, and others reflecting the market's central pricing of the commercialization elasticity of large models. Large models as the most promising sub-track in the AI industry. The focus of the AI industry is shifting from "expanding CapEx" to "Token production and commercialization." The market's focus has transitioned from who can build more GPUs and data centers to who can produce and monetize more Tokens at a lower cost and latency. In terms of technology, the development of large models focuses on parameter efficiency, data quality, computational power during inference, and real engineering scenarios. On the commercialization front, companies like Anthropic serve as commercialization models for large models, with revenue and ARR growing explosively over the past three years, indicating a transition from large models being "money-eating monsters" to "money-printing machines." The importance of Token economy in computational power rentals, domestic computational power, and computational electrical appliance collaboration is increasing. The explosive consumption of Tokens is driving the demand for computational power for inference and training. Overseas companies like NeoCloud are expanding through AI-specific clusters, long-term locking agreements, and significant financing, while domestic companies focus on standardized rental models such as GPU single machines, cabinets, and computational power pools, with channels, funding, and customer resources becoming core barriers. Domestic computational power serves as a strategic foundation for China's AI autonomy and control, exhibiting three major development trends since 2026: performance iteration of hardware, accelerated adaptation of models, and expansion of commercialization. Leading companies like Cambricon have achieved profitability and expansion, Haiguang DCU is accelerating the adaptation of domestic large models, and Huawei is building a foundation with the CANN and Ascension system. Collaborations between computational power and electrical appliances further bind computational expansion with green energy, energy storage, and grid scheduling, becoming crucial variables for computational cost and supply stability. The potential reevaluation of Data Infra and high elasticity directions of AI applications. Models determine the upper limit of capabilities, agents determine the value of scenarios, and Data Infra determines the ceiling for scalable deployment. Agent operations involve multiple processes such as data acquisition, tool calling, context management, memory storage, and security governance, which cannot be achieved solely by the model itself and require support from the underlying Data Infra. Data Infra possesses stronger platform attributes, higher customer stickiness, and longer lifecycles, enabling the formation of commercial loops. AI applications began with the consumer end but the main battlefield for commercialized scaling is shifting towards the business end. Overseas companies like Anthropic have overtaken OpenAI in the API market share with Coding and Agent workflows, while the market size for enterprise-level large models in China is much larger than that of consumer-level models. The report believes that business scenarios on the business end, with characteristics such as "clear budget, high-frequency tasks, verifiable results, and unique data," are more suitable for Agent penetration. Investment recommendations: 1. Focus on the core theme of commercializing large models. The main focus of the AI industry is shifting from "expanding CapEx" to "Token production and commercialization." It is recommended to pay close attention to leading domestic model manufacturers like KNOWLEDGE ATLAS, MINIMAX-W, and vertical large model manufacturers like Zhejiang Supcon Technology Co., Ltd.; 2. Grasp directions that ensure performance certainty such as domestic computational power, computational power rental, computational power and electrical appliance collaboration, storage, CPU, etc. Domestic computational power is a strategic foundation for AI autonomy and control, with a focus on companies like Cambricon, Hygon Information Technology, Dawning Information Industry, Bingo Software, Sichuan Huafeng Technology, SJ Semiconductor Corporation, etc. With the explosive growth in Token consumption, the profitability of the domestic computational rental industry continues to rise, with a focus on companies like Glory View Technology, Sharetronic Data Technology, Jwipc Technology, Maxvision Technology Corp., etc. The essence of computational power and electrical appliance collaboration is to "optimize computing and support electricity," with a focus on companies like China Southern Power Grid Technology, Nanjing Sciyon Wisdom Technology Group, etc.; 3. Emphasize the reevaluation of value in Data Infra and high elasticity directions such as Agent, Physical AI, and core applications. The scaling of Agent operations depends not only on models but also on the underlying Data Infra. It is recommended to pay attention to companies like overseas companies like MongoDB, Snowflake, and domestic companies like Wuhan Dameng Database, Xinghuan Technology, etc. The report believes that business scenarios on the business end, with characteristics such as "clear budget, high-frequency tasks, verifiable results, and unique data," are more suitable for Agent penetration. Focus on companies like Beijing Kingsoft Office Software, Inc., Servyou Software Group, Nancal Technology, Jiangsu Eazytec Co., Ltd., etc. Risk warning: Risks include unexpected technological innovation, slower-than-expected progress in AI large model and application research and development, slower-than-expected progress in AI chip development, and increased industry competition.