Recent catalysts and surprises in the AI large model layer, including the listings of the world's first two independent model vendors, DeepSeek's anticipated launch of its new-generation V4 model around the Spring Festival, and Volcano Engine's announcement as the exclusive AI cloud partner for the Spring Festival Gala, have driven valuation increases and new investment opportunities across AI infrastructure, models, and applications. In the AI application layer specifically, penetration rates are expected to rise this year as model capabilities continue to improve and vertical scenario data becomes more integrated. Particular attention should be paid to vertical tracks with high earnings certainty, such as AI Coding, AI Health where users show strong willingness to pay, and AI Medicine within AI for Science. The main views are as follows: In the AI model layer, accelerated updates to foundational models both domestically and internationally, coupled with the catalytic effect of independent model vendors going public, are key drivers. Domestically, models like DeepSeek and ByteDance's Doubao are expected to update their foundational models around the Spring Festival. According to The Information, DeepSeek's open-source V4 model is scheduled for a mid-February release, with internal tests indicating its code capabilities are now on par with Anthropic's Claude. Overseas models from Anthropic, OpenAI, and Gemini are also anticipated to update their capabilities in the first or second quarter. It is judged that leading model vendors will continue to focus on breaking through intelligence ceilings this year via reinforcement learning and long-context engineering optimization, with revenue growth expected to accelerate. Spillover sentiment from the model layer is driving a valuation reshaping in the application end. Although AI applications did not exceed market expectations in terms of performance over the past year, they have continuously explored integration with models and scenarios to accelerate revenue realization. It is believed that this year, with ongoing improvements in model capabilities and the integration of vertical scenario data, the penetration rate of AI applications will increase. Investors should particularly focus on vertical tracks like AI Coding, which boasts high earnings certainty, AI Health where users demonstrate strong payment willingness, and AI Medicine within the AI for Science domain. From an investment perspective, the current proportion of AI revenue in application companies' total income is relatively low. The market may apply Sum-of-the-Parts (SOTP) valuations to each company's AI business, potentially leading to a reshaping of the overall valuation system. Within the AI infrastructure layer, the narrative around long-context and model memory is expected to continue driving storage demand. Ongoing exploration in both academia and industry into models' continuous learning and long-context memory capabilities—aiming for, or even surpassing, human-like thinking and memory—serves as a technical guide towards AGI. Long-context is fundamental for agents to perform multi-step, long-chain tasks, consequently increasing the memory required for KV Cache. Model vendors are currently exploring ways to optimize model memory to enable continuous and online learning, such as writing long-term memory into model parameters. Therefore, demand for storage and databases is likely to remain elevated. Recommended stocks related to AI application vendors include: Beijing Kingsoft Office Software, Inc. (688111.SH), Digiwin Co.,Ltd. (300378.SZ), Fujian Fuxin Software Development Joint Stock Co.,Ltd. (688095.SH), and Servyou Software Group Co.,Ltd. (603171.SH). For AI infrastructure-related vendors, Sangfor Technologies Inc. (300454.SZ) is recommended. Key risk factors include slower-than-expected evolution of AI technology and slower-than-expected implementation of AI applications.
Comments