China Galaxy Securities: Communication computing power reducing costs and increasing efficiency, Deepseek's changes and consistencies.

China Galaxy Securities has released a research report stating that the operator, as the largest traffic channel in China, has data advantages and interface advantages. As a leading player in the computing power industry, operators use their own cloud resource pool and outstanding hardware capabilities to promote the popularization of AI applications. At the same time, the future demand for stronger training models will drive the rapid development of the optical module and AIDC industry chain. In the face of the increasingly complex global economic situation, the process of independent controllability in core devices such as optical chips may be further accelerated. The main points of China Galaxy Securities are as follows: DeepSeek's changes: cost-effective algorithm revolution driving the mutual prosperity of computing power and applications DeepSeek's model, using PTX programming language and innovative changes in engineering capabilities, allows it to achieve stronger performance at lower costs or accelerate the popularization and implementation of AI applications with hardware. The computing power is divided into training power and inference power. The lower-cost inference power is expected to drive further growth of applications, enhance the prosperity of application side, and the marginal expansion of AI capabilities still relies on stronger models and powerful computing power. In the medium to long term, DeepSeek's innovation in algorithms and architecture adds a new path to AI development and is expected to create a win-win situation for the prosperity of large models from training to inference to application. Moreover, the development of applications is expected to drive the acceleration of the computing power industry chain and the localization process. The launch of DeepSeek brings significant cost reductions to the inference side, which will greatly accelerate training iteration. The landing of application scenarios may be further accelerated, while pushing further improvement in the efficiency of inference-side models, thereby driving the continuous high prosperity of the communication industry in the direction of computing power. The future demand for stronger training models will drive the rapid development of the optical module industry chain. In the context of the increasingly complex global economic situation, the process of independent controllability in core devices such as optical chips may be further accelerated. Currently, the gap between domestic GPU chips and related industry chains and similar competitors is gradually narrowing. Based on a certain level of domestication acceleration, the rapid development of inference-side computing power may further accelerate the popularity of domestic devices. DeepSeek's constants: high-end computing power resources are still necessary, AIDC is expected to achieve a virtuous cycle of input and output As state-owned enterprises, operators have a solid foundation, continuous improvement in profitability, cash flow, asset value advantages, and continuous increase in dividends to shareholders, with basic business having continuous growth potential. The cloud business of the three major operators has seen sustained high growth. Against the backdrop of artificial intelligence, after the release of the DeepSeek open-source model, the focus of large models is gradually shifting from pre-training to post-training, the scale of inference-side computing power is expected to increase, and the requirements for hardware have also decreased. The investment of operators in computing power infrastructure is expected to form a virtuous cycle with the popularization of applications, with capital expenditure gradually translating into scale improvement returns, leveraging the advantages of operator channels to drive the marginal improvement in profitability of the cloud business chain of operators and others. Risk warning Risks of AI application development falling short of expectations; risks of further increasing complexity in the international economic situation; risks of AI hardware development speed falling short of expectations; risks of short-term fluctuations in the AI industry chain upstream and downstream, etc.