Guosen: The AI industry chain is moving from divergence to consensus in performance trends, and 2026 is expected to be a year of harvest for local hard technology.

Guosen released a research report stating that looking ahead to 2026, the reasoning ability of AI large models is still undergoing continuous iteration, the closed loop between large models and edge applications is emerging, and the situation of the supply of computing power and storage hardware being in short supply will continue. The expansion progress of advanced domestic processes and the speed of self-controllable advancement still have a large gap. 2026 is expected to be a year of "local hard technology harvest from a spark to a full-blown fire." Guosen's main points are as follows: In 2025, the AI industry chain went from divergence to consensus in performance trends, and 2026 is expected to be a year of local hard technology harvest. The bank previously stated that the electronics industry's business cycle has been trending downward for nearly 2 years since 2021, hitting bottom and rebounding in the second half of 2023, with the return of the Huawei Mate series as a significant event. The industry is currently in a mild upward trend driven by AI innovation. The bank previously stated that with the resonance of "macro policy cycle, industry inventory cycle, AI innovation cycle", and the rapid expansion of passive funds, there has been a significant trend of valuation expansion. Following the narrative logic impact caused by the rise of Deepseek in 2Q25 and the tariff war initiated by the United States, since 3Q25, the market has shifted from diverging trends to consensus in the bright performance trend of the AI industry chain. As of December 16, the electronics industry has risen by 40.22%, ranking third in the industry overall. The bank believes that the Chinese technology industry, which has shown a trend of standing out in the 5G innovation cycle that started in 2020, is showing stronger global competitiveness in the new AI innovation cycle, supported by the red dividend of domestic engineers, facing US "sanctions policies", and after more than 5 years of rapid accumulation of "human resources and materials", is experiencing global recognition. In 2026, it is expected to be a year of "local hard technology harvest from a spark to a full-blown fire". AI large models compete fiercely, NVIDIA leads the iteration of computing power, and the PCB and server industry chains continue to experience high growth. Due to the innovation in the architecture of large models, both domestic and foreign large models have continued to advance in understanding, reasoning, and AI application aspects. Benefiting from the expansion of CSP, sovereign clouds, and the flourishing development of AI reasoning applications, TrendForce predicts that the global total capital expenditure of the top eight CSPs will grow by 40% to reach over $600 billion in 2026, and global AI server shipments will increase by 20.9%. At the same time, following the upgrade from GB200 to GB300 in 2025, NVIDIA's new generation Rubin architecture AI servers will bring revolutionary changes in detachable reasoning by the end of 2026. NVIDIA expects total shipments of Blackwell and Rubin series GPUs to reach 20 million by the end of 26, with a total order value of $500 billion. Based on the market expansion due to the arms race of computing power and the ASP increase brought by the iteration of computing power products, along with the expansion of intelligent computing clusters brought by Scale Up and Scale Out, the bank believes that in 2026, the PCB and server industry chains deeply involved in the global division of labor will enter a high-speed growth period with a simultaneous increase in quantity and price. Companies to watch include Foxconn Industrial Internet, Victory Giant Technology, Shengyi Technology, Huaqin Technology, Wus Printed Circuit, Lens Technology, Luxshare Precision Industry, Avary Holding, Suzhou Dongshan Precision Manufacturing, Shenzhen Kinwong Electronic. Computing power and storage power: Domestic general-purpose computing chips and ASIC solutions are making simultaneous efforts, and shortages of storage power are expected to persist throughout the year. In terms of computing power, current domestic chips are actively updating and iterating. Huawei plans to launch Ascend 950PR in 1Q26 and list the Atlas 950 SuperPoD in 4Q26. Domestic cards such as Cambricon, Muxi, Birun, and Moore's Threading have successfully entered the smart computing center. Meanwhile, due to multiple sanctions by the US BIS on domestic chips, compliance ASIC projects of major CSPs are expected to bring development opportunities. Self-developed projects of non-first-tier cloud companies are expected to bring considerable incremental business to domestic ASIC manufacturers, and storage solutions will also become a key point of differentiation for ASIC competition. In terms of storage power, DRAM in the AI era is gradually transitioning from an "ancillary role" to a "performance bottleneck breakthrough". The restructuring of the supply side coupled with AI-driven demand is expected to accelerate growth, with the expected bit demand for DRAM in 2026 likely to increase by 26% year-on-year. At the same time, with the rise of AI reasoning, the limitations of traditional HDDs in read/write speed, response latency, and energy efficiency have accelerated the penetration of SSDs, with NAND shortages spreading from local applications to the full range, with the price index rising by over 40% from September to December in 25. The bank believes that in 26, DRAM and NAND will continue to face severe supply shortages, and prices are expected to continue to rise. Industries in focus related to computing power and storage power include Cambricon, Aojie Technology, Verisilicon Microelectronics (Shanghai) Co., Ltd., Shenzhen Techwinsemi Technology, Shenzhen Longsys Electronics, GigaDevice Semiconductor Inc., Ingenic Semiconductor, Shanghai V-Test Semiconductor Tech. Transportation power: Capacity becomes an important breakthrough for computing power enhancement, with computing power growth driving simultaneous upgrades in power architectures. The capacity link needs to solve the problem of data entry and exit from memory, as well as enabling smooth communication between servers internally, between racks, and between clusters. In the background of limited flow shipment of high-end computing chips domestically, optimization of transportation capacity becomes an important breakthrough point. The bank forecasts a 21.2% CAGR in the global high-speed interconnect chip market from 2024 to 2030, with the China market share increasing from 25% to 30%, creating broad incremental markets for MRDIMM memory, PCIe interconnect chips, CXL interconnect chips, silicon photonics chips, OCS, among others. Meanwhile, with the increasing power consumption of data center chips, rack processing power is surging. Infineon predicts that the power consumption of a single GPU will increase exponentially, reaching around 2000W by 2030, and the peak power consumption of a rack will exceed 300kW. On one hand, improving energy conversion efficiency is crucial for reducing data center operating costs; on the other hand, the large and rapid fluctuations in power curves at the rack side pose a challenge to the stability of the public grid, requiring power supply solutions to move towards HVDC direction. SST, DrMos, as well as SiC, GaN devices will become core directions for AI power, with a focus on SICC Co., Ltd., Wuxi Nce Power, Nantong Jianghai Capacitor, StarPower Semiconductor, Xilinx Integrated, China Resources Microelectronics, Yangzhou Yangjie Electronic Technology, JoulWatt Technology, Shanghai Bright Power Semiconductor Co., Ltd., Shenzhen Sunlord Electronics. AI edge: AIAgent reshapes the interaction paradigm, major companies compete to lay out edge entrances, and a year of innovation in consumer electronics begins. With the continuous evolution of large models in multimodal understanding, general reasoning, and task execution capabilities, AI is transitioning from a tool-type capability to an AIAgent that can understand user intent and autonomously perform tasks. Consumer electronic products at the edge are the key carriers of the commercial closed loop of AI, and are expected to systematically reconstruct the human-machine interaction paradigm. In this context, various terminal forms such as phones, glasses, headphones, and home Siasun Robot & Automation are expected to build synergistic networks around AIAgent, driving AI from a single point upgrade to a system-level experience across scenarios and terminals. The importance of multimodal inputs such as voice, vision, and environmental perception is increasing, placing higher demands on edge computing power, perception capabilities, and connectivity. The bank believes that the current technology and industrial foundations at the edge have matured, and key breakthroughs in business models are expected to form a "nonlinear amplification effect". Looking ahead to 2026, from CES at the beginning of the year to WWDC in the middle of the year, or the continuous exploration by leading companies such as Byte and Apple, along with the expectations of potential model manufacturers and internet companies entering the hardware industry, all could become key catalysts for igniting market sentiment and industry investment consensus. Companies to watch include Bestechnic (Shanghai) Co., Ltd., Amlogic, Lens Technology, Luxshare Precision Industry, Zhejiang Lante Optics, Goertek Inc., Aojie Technology, Espressif Systems, Huaqin Technology, Shenzhen Transsion Holdings Co., Ltd., Xiaomi Group, Shenzhen Sunlord Electronics. Semiconductors: Recommended processes may exceed expectations for self-controllable industrial chains, and analog chips are accelerating domestic substitution in the recovery stage. According to SIA data, China accounted for 28% of global semiconductor sales in 2024, but the local supply ratio was only 4.5%, with a low self-sufficiency rate. Incremental demand mainly came from GPU, HBM, and other cloud-side increments, and the self-sufficiency rate was lower than in 2023. However, the financial performance of A-share semiconductor companies has continued to improve. According to the bank's statistics on the operating data of 146 companies, the highest quarterly revenue share was 54% in 2025. The gross profit margin of the SW semiconductor sector in 3Q25 was between 1Q21 and 2Q21 levels, and the net profit margin was similar to levels in 4Q20 and 1Q21. From an industry cycle perspective, global semiconductor sales have seen year-on-year growth for eight consecutive quarters, and in December 2025, WSTS once again raised its forecast for 2025 and 2026, expecting global semiconductors to achieve double-digit growth for three consecutive years from 2024 to 2026. Looking at the domestic semiconductor industry, in addition to AI increments, the rise of domestic chip design companies and the demand for localized manufacturing provide incremental business to the self-manufacturing chain. Key areas to focus on include wafer foundries, advanced packaging, and upstream semiconductor equipment and material sectors. In addition, analog chips are lagging behind in the semiconductor product category industry, with major international companies such as TI and ADI seeing year-on-year revenue growth in 2025, marking the industry's entry into a recovery stage. New products launched by domestic companies in recent years are expected to enter a phase of scale production, with long-term benefits from AI data centers, as well as AI applications such as ADAS and humanoid robotics, bringing broad increments. At the same time, analog chips are a subcategory with sufficient space for localization, which will continue to benefit from increased domestic production rates. Focus on companies such as Semiconductor Manufacturing International Corporation, HUA HONG SEMI, JoulWatt Technology, 3peak Incorporated, NAURA Technology Group, Advanced Micro-Fabrication Equipment Inc. China, Piotech Inc., SG Micro Corp, Southchip Semiconductor Technology (Shanghai), Shanghai V-Test Semiconductor Tech., TongFu Microelectronics, JCET Group Co., Ltd., Hubei Dinglong, SBT Ultrasonic Technology. Key focus combinations: Semiconductor Manufacturing International Corporation, Foxconn Industrial Internet, Advanced Micro-Fabrication Equipment Inc. China, Aojie Technology, HUA HONG SEMI, Cambrion, NAURA Technology Group, Montage Technology, Lens Technology, Zhejiang Lante Optics, Luxshare Precision Industry, Victory Giant Technology, Huaqin Technology, Bestechnic (Shanghai) Co., Ltd., Shenzhen Sunlord Electronics, Wus Printed Circuit, Shanghai V-Test Semiconductor Tech., Amlogic, Avary Holding, 3peak Incorporated, JoulWatt Technology, SG Micro Corp, Shenzhen Techwinsemi Technology, Shenzhen Longsys Electronics, Biwin Storage Technology, Goertek Inc., GigaDevice Semiconductor Inc., Hubei Dinglong, Boe Technology Group, JCET Group Co., Ltd., Zhejiang Crystal-optech, Shenzhen Transsion Holdings Co., Ltd., OmniVision Integrated Circuits Group, Inc., SBT Ultrasonic Technology, Hangzhou Hikvision Digital Technology, Xiaomi Group, SICC Co., Ltd., Yangzhou Yangjie Electronic Technology, Espressif Systems, Nantong Jianghai Capacitor, Electric Connector Technology, Shanghai Bright Power Semiconductor Co., Ltd., DBG Technology, Southchip Semiconductor Technology (Shanghai), Suzhou Shihua New Material Technology Co., Ltd., Ingenic Semiconductor, Zhejiang Jiemei Electronic And Technology, C*Core Technology, TongFu Microelectronics, Shenzhen Vital New Material, Freewon China. Risk warning: The domestic substitution process falls short of expectations; downstream demand falls short of expectations; risks of increased industry competition; risks of unfavorable changes in international relations; risks of industry cyclical fluctuations; risks of production equipment and raw material supply.