China Securities Co., Ltd.: How to choose AI investments under the new paradigm?

China Securities Co.,Ltd. released a research report stating that AI is currently undergoing a transformation from being an auxiliary tool to becoming autonomous labor, driving a surge in Token demand through Agent transformation, and pushing the computational power ecosystem towards efficient ASIC and premium cloud services. The investment paradigm for AI follows the principle of cost reduction leading to revenue generation, achieving productivity scale expansion through extreme cost reduction, and thereby fostering incremental value spaces such as content creation and cutting-edge scientific research. In 2026, three dimensions of AI industry investment include: 1) from a model perspective, breakthroughs in longer context and memory capabilities and stronger abilities for autonomous thinking, reflection, and creation; 2) from a computational perspective, transitioning from "computation for everyone" to "computation inflation"; 3) from an application perspective, finding scenarios for rapid income growth. Key points from China Securities Co.,Ltd. include: The post-training paradigm drives a qualitative transformation in intelligence, as Agents move from being tools to autonomous labor across generations. By 2025, the large model industry has shifted from expanding parameter size to expanding reasoning. Represented by OpenAI's o series and DeepSeek R1, reinforcement learning RL has verified that model capabilities can experience nonlinear growth in the post-training phase. Looking ahead to 2026, model evolution will focus on super-long context memory, full-modal perception, and world models. AI is transitioning from an assistant co-pilot Copilot to delivering results as an Agent, with the potential for handling complex tasks beyond an 8-hour threshold. The improvement in the ability to deliver end-to-end projects from fragmented tasks signifies that AI is officially evolving from an efficiency tool to deployable digital labor on a large scale, effectively triggering the fourth industrial revolution. The surge in demand for reasoning is forcing a reshaping of computational cost structures, with ASIC chips and cloud resources reaching a turning point in premium realization. With the widespread adoption of Agent workflows, the global Token call volume shows a steep exponential growth, and the continued breakthroughs in Google and bean Token call volumes validate the underlying computational needs. Among them, ASIC chips such as Google TPU v7 are accelerating market erosion of general-purpose GPUs due to their single-chip energy efficiency and low inference costs advantages; and driven by upstream core component price increases and downstream demand surges, the cloud resource pricing model has shifted from "price for volume" to "premium realization." Edge clouds, CDNs, and AI software Infra layers supporting RAG architecture (such as vector databases, inference acceleration engines) will be the first to realize performance gains. The AI investment paradigm focuses on cost reduction and revenue generation, with extreme cost reduction driving acceleration in new scenarios. The core rule of AI commercialization lies in the deep coupling of cost reduction and revenue generation, with extreme cost reduction often being the starting point for revenue generation in the ShenZhen New Industries Biomedical Engineering sector. On the cost reduction side, digital employees and AI Coding are significantly reducing enterprise R&D and operational costs by restructuring individual productivity. On the revenue side, when the marginal cost of specific scenarios drops to a critical point, incremental value points will emerge (such as AI short films/comics, Agent outsourcing, etc.); at the same time, AI is deeply involved in high-complexity systems such as AI4S and electricity trading, achieving precise value extraction. Risk warning (1) Commercialization of the AI industry may not meet expectations: Currently, the commercialization models of AI products at all levels are still in the exploratory stage. If the pace of advancement of products at all levels is slower than expected, it may have an adverse impact on the performance of related companies; (2) Market competition risk: Overseas AI manufacturers have an advantage in terms of first-mover advantage and strong technical accumulation, placing them in a favorable position in competition. If domestic AI manufacturers do not iterate on their technology as expected, their operational conditions may be affected; at the same time, there are already many domestic companies investing in AI product development, which may lead to risks of homogenized competition in the future, thereby affecting the revenue of related companies; (3) Policy risk: The development of AI technology is directly influenced by the policies and regulations of various countries. As AI penetrates into various fields, governments may further implement corresponding regulatory policies to regulate its development. If companies fail to adapt in a timely manner and comply with relevant policies, they may face corresponding penalties, or even be forced to adjust their business strategies. In addition, policy uncertainty may also lead to errors in business strategy planning and investment decisions, increasing operational uncertainty; (4) Geopolitical risk: Under the fluctuating global geopolitical environment, particularly with the export restrictions imposed by the United States on China, it may directly affect the acquisition of computational chips by domestic companies, thereby affecting product R&D and market competitiveness. Moreover, geopolitical risks may also hinder the expansion of AI products into overseas markets, affecting the revenue situation of related companies.