China Galaxy Securities has released a research report stating that from overseas developments like ChatGPT and Sora to domestic applications such as Wenxin Yiyan and Kimi, AI applications are evolving from text generation to multimodal and agent-based systems. The token consumption per interaction is experiencing exponential growth, directly driving a surge in demand for GPU-centric training and inference computing power, particularly for the more sustained and widespread inference needs. Major global economies currently regard computing power as a strategic resource. China's initiatives, including the "East Data West Computing" project, the national integrated computing power network, and the "AI Plus" action, are systematically promoting the scaled and intensive development of computing infrastructure. As inference demand continues to rise, token consumption is expected to see significant growth, driving renewed acceleration across the industrial chain, primarily manifested in four key areas: AIDC, telecom operators, optical fiber and cable, and optical modules. The main viewpoints of China Galaxy Securities are as follows:

Regarding AIDC: The core drivers are the intensive iteration of global large model technology and the explosive growth in token usage. Leading global technology firms are heavily allocating capital expenditures towards computing infrastructure. The AIDC industry is undergoing systematic restructuring around high-density GPU computing clusters. AIDC is rapidly consolidating into ultra-large-scale clusters reaching hundreds of thousands of cards, with single cabinet power density exceeding tens or even hundreds of kilowatts. Through high-density, green designs, AIDC is becoming the core physical foundation supporting the AI era.

Regarding Telecom Operators: Leveraging their nationwide backbone network resources, extensive computing node layouts, and "cloud-network integration" capabilities, telecom operators are becoming key players in connecting distributed computing power and breaking down geographical barriers. AI computing is shifting from single-point construction to cross-regional cluster scheduling, generating massive demand for low-latency interconnection. Operators are now fully transitioning to "computing power operations." By building high-reliability, low-jitter data center interconnection networks, they not only meet the intensive "east-west traffic" transmission between AI clusters but also successfully transform from traditional pipeline providers into core organizers of intelligent computing services. The increasing depth and breadth of operators' token infrastructure deployment may also revitalize their existing AIDC assets.

Regarding the Optical Fiber and Cable Industry: The large-scale AIDC construction driven by token usage is upgrading optical fiber from a standard connectivity component to the core foundation of computing power networks. To meet the ultra-low latency requirements for parameter synchronization between GPUs, network architectures must adopt non-blocking, fully interconnected models, leading to a rapid increase in fiber consumption per cabinet. This directly drives comprehensive expansion and upgrades for backbone networks, metropolitan area networks, and data center interconnects. To support exabyte-level massive data transmission, demand for new builds and replacements using ultra-low loss, large effective area high-quality optical fibers is surging. Constrained by the rigid production cycles of upstream preform expansion, the optical fiber industry is entering a high-growth cycle characterized by rising volumes and prices.

Regarding Optical Modules: As GPU computing power grows exponentially, data transmission rates have become the primary bottleneck determining the efficiency of computing clusters. To match the high throughput of new-generation AI chips, the iteration cycle for optical modules continues to shorten, with the industry accelerating its evolution from 400G/800G to 1.6T and 3.2T ultra-high speeds. New technological solutions like CPO, silicon photonics, and LPO are gradually moving towards large-scale adoption. As core components ensuring the linear acceleration ratio of computing clusters, high-speed optical modules are entering a development period marked by the convergence of technological generational shifts and explosive demand.

Investment recommendations suggest focusing on AIDC industry players such as Data Vantage, Runtech, Runjian, GDS, and Dawning Information Industry; telecom operators including China Mobile, China Unicom, and China Telecom; optical fiber and cable manufacturers like Yangtze Optical Fibre and Cable, Hengtong Optic-Electric, and ZTT; and optical module and component companies such as Zhongji Innolight, Eoptolink, Accelink, HG Tech, Cambridge Industries,腾景科技,源杰科技,仕佳光子, TFC Optical Communication, Focuslight, and T&S Communications.

Risk warnings include the potential for AI application development to fall short of expectations, the risk of declining capital expenditure from cloud providers, and the possibility of industrial technology development not meeting forecasts.