Guosheng Securities: Leading manufacturers are moving towards volume expansion and business cycle. Algorithm construction will be a prerequisite.

Guosheng Securities released a research report stating that the dawn of domestic computing power demand has arrived. Since the outbreak of AI two years ago, overseas markets have opened up the commercial cycle of AI through the accumulation of computing power and model construction, which means that the major Chinese internet giants now have the necessary conditions for large-scale deployment of AI businesses. As leading model manufacturers begin to ramp up and enter the commercial cycle, a new round of competition for the Chinese internet industry's market share is imminent. In this round of competition, the development of independently controllable computing power will be a prerequisite for success. Guosheng Securities' main points are as follows: GPGPU or ASIC - prioritize usability. Recently, Broadcom outlined a grand blueprint for future ASIC chips during their earnings conference call. However, for Chinese computing power, the "usable and easy-to-use" GPGPU will be the main theme in the coming years. Compared to GPGPU, although current AI ASIC mainstream routes have higher theoretical efficiency, their compatibility with domestic customer demands in terms of compilers and ecosystem software is relatively low. Looking at the current situation, as large-scale applications of "big models" such as "DouBao" accelerate, what manufacturers need is the ability to rapidly deploy and gain access to business entry points and user-friendly computing power, meaning GPGPU. At the same time, for operators and local smart computing builders, general-purpose computing power represents better user acceptance and return on investment. In the long term, as the competitive landscape of Chinese AI models becomes clearer, with the business granularity of leading players increasing, AI ASIC will find suitable penetration scenarios in China. Chinese AI communication: where does the road lead? As China's computing power accelerates, the urgent need for independently controllable AI communication capabilities is evident. Building Chinese AI communication, like abroad, is divided into two parts: "Scale-Out" and "Scale-Up." For the "Scale-Out" network dominated by switches and Ethernet, the core is a switch system with complete control from chip to whole machine. For the more closed and dedicated "Scale-Up" networks of the past, China should not follow the old path of NVIDIA's NV-Link system but instead learn from overseas alliances led by Broadcom and AMD, as well as Broadcom's advocacy of standardized interconnection services starting from packaging, based on the experience of major players like China Telecom and innovative packaging technologies, to create a self-owned version of "UA-Link" and standardized core packaging suitable for all domestic computing power. Infrastructure and manufacturing: the foundation of China's computing power The development of AI computing has surpassed the realm of single-chip computing power and evolved into a systematic engineering of energy, communication, and cluster systems. Like an aircraft carrier, it is a measure of a country's comprehensive strength. Although China's chip manufacturing processes and single-chip capabilities are restricted under geopolitical constraints, in the process of developing domestic computing power, our infrastructure and electronic manufacturing capabilities will be one of China's most important strengths. From two perspectives, firstly, electronic manufacturing capabilities: China's optical communication module companies have gradually established a competitive advantage through multiple iterations of speed rates. With the diversification of connection solutions, China's copper modules, optical fiber cables, long-distance optical modules, etc., will also contribute to the development of AI. Secondly, China possesses the world's most advanced electrical facilities and an abundant power supply. Currently, the US's AI development is hindered by power grid capacity, leading them to seek alternative solutions such as DCI, natural gas, and small-scale nuclear power. However, China's strong power infrastructure will give it an irreplaceable advantage in expanding data centers. Investment suggestions: focus on the core targets of computing power, communication power, manufacturing power, and infrastructure power. Computing power: Cambricon (688256.SH). Communication power: ZTE Corporation (00763,000063.SZ), Suzhou Centec Communications (688702.SH), TongFu Microelectronics (002156.SZ). Manufacturing power: Eoptolink Technology Inc., (300502.SZ), Zhongji Innolight (300308.SZ), Suzhou TFC Optical Communication (300394.SZ), Accelink Technologies (002281.SZ), Wuxi Taclink Optoelectronics Technology (688205.SH), Huagong Tech (000988.SZ), Ruijie Networks (301165.SZ), Shenzhen Phoenix Telecom Technology (301191.SZ), Unisplendour Corporation (000938.SZ). Infrastructure power: Range Intelligent Computing Technology Group (300442.SZ), Beijing Sinnet Technology (300383.SZ), Guangdong Aofei Data Technology (300738.SZ), Shenzhen Envicool Technology (002837.SZ), Shenzhen Megmeet Electrical (002851.SZ). Risk warning: AI development falls short of expectations, domestic AI demand falls short of expectations, overseas sanctions risk.