Guosen Securities released a research report stating that looking ahead to 2026, the reasoning capabilities of large AI models will continue to iterate, a closed loop between large models and edge-side applications is forming, and the supply-demand imbalance for computing and storage hardware is expected to persist. However, there remains a significant expectation gap regarding the expansion pace of domestic advanced process nodes and the progress of achieving self-sufficiency and controllable supply chains. 2026 is anticipated to be the "year of harvest for local hard tech, evolving from a single spark to a prairie fire."
The main views of Guosen Securities are as follows:
In 2025, the AI industry chain transitioned from disagreement to consensus in its performance trends, and 2026 is expected to be a harvest year for local hard tech. The firm previously expressed the view that the electronics industry's景气 cycle declined for nearly two years starting from 2021, bottomed out and began recovering in the second half of 2023, marked by the return of Huawei's Mate series, and currently remains in a mild upward trend driven by AI innovation. The firm also previously stated that the industry, amidst the共振上行 of "macro policy cycles, industrial inventory cycles, and AI innovation cycles," and aided by the rapid expansion of passive funds, has shown a significant trend of valuation expansion. After experiencing the narrative impact of "computing power deflation" triggered by the rise of Deepseek in Q2 2025 and the impact of the tariff war initiated by the US, the market trend has shifted from disagreement to consensus since Q3 2025, driven by the impressive performance trends of the AI industry chain. As of December 16, the electronics sector rose 40.22%, ranking third among all industries.
The firm believes that China's tech industry, which had already shown signs of emerging prominence during the comprehensive 5G innovation cycle starting in 2020, is now demonstrating stronger global competitiveness in the new AI innovation cycle. Supported by the domestic engineer dividend and having faced US "sanction policies" and undergone over five years of rapid accumulation of "human, financial, and material resources," it is also gaining renewed recognition worldwide. 2026 is expected to be the "year of harvest for local hard tech, evolving from a single spark to a prairie fire."
AI large models are fiercely competing, NVIDIA is leading computing power iteration, and the PCB and server industry chains continue high growth. Benefiting from architectural innovations in large models—such as using Mixture-of-Experts architectures for more efficient推理 through sparsity, employing innovative attention mechanisms to reduce computational complexity and memory requirements, and utilizing deep thinking modes with multi-round reasoning to reduce hallucinations—both domestic and international large models have achieved continuous advancement in multimodal understanding, reasoning, and AI applications. Benefiting from the expansion of computing power demand from CSPs, sovereign clouds, and the vigorous development of AI inference applications, TrendForce estimates that the combined capital expenditure of the world's top eight CSPs will grow 40% in 2026 to exceed $6000 billion, and global AI server shipments will increase by 20.9%. Meanwhile, following the upgrade from GB200 to GB300 in 2025, NVIDIA's next-generation Rubin architecture AI servers in 2026 will bring revolutionary changes to disaggregated inference. NVIDIA expects total shipments of Blackwell and Rubin series GPUs to reach 20 million units by the end of 2026, with combined orders amounting to $500 billion.
Based on the market size expansion from the computing power arms race and the ASP increase driven by computing power product iteration, alongside the expansion of AI computing clusters from Scale Up and Scale Out, the firm believes that in 2026, the PCB and server industry chains deeply involved in the global supply chain division of labor, including related upstream materials and liquid cooling散热 segments, will enter a period of high-speed growth with both volume and price increases. Focus on:工业富联,胜宏科技,生益科技,华勤技术,沪电股份,蓝思科技,立讯精密,鹏鼎控股,东山精密,景旺电子.
Computing Power + Storage Power: Domestic computing power general-purpose chips and ASIC solutions advance together; storage shortages and price hikes expected to persist throughout the year. Regarding computing power, domestic chips are actively iterating and updating. Huawei plans to launch the Ascend 950PR in Q1 2026 and launch the super-node Atlas 950 SuperPoD in Q4 2026. Domestic cards from寒武纪,沐曦,壁仞,摩尔线程 are being successfully introduced into AI computing centers. Simultaneously, constrained by multiple US BIS sanctions on domestic chips, compliance ASIC projects from major CSPs will concurrently see development opportunities. Among these, self-research projects from non-tier-1 cloud providers are expected to bring considerable increments for domestic ASIC manufacturers, while storage solutions will also become a key point for ASIC differentiation.
Regarding storage power, DRAM in the AI era is gradually shifting from a "supporting role" to a "bottleneck breakthrough for performance." Driven by supply-side structural adjustments and pulled by AI demand, it is expected to accelerate growth, with DRAM bit demand projected to increase 26% year-on-year in 2026. Meanwhile, with the rise of AI inference, the limitations of traditional HDDs in read/write speed, response latency, and energy efficiency are accelerating SSD penetration. The NAND shortage situation has spread from specific applications to the entire market, with the price index rising over 40% from September to December 2025. The firm believes that in 2026, DRAM and NAND will still exhibit severe supply-demand imbalances, and prices are expected to continue their upward trend. Focus on computing power and storage power related industry chains:寒武纪,翱捷科技,芯原股份,德明利,江波龙,兆易创新,北京君正,伟测科技.
Connectivity + Power: Connectivity becomes a key breakthrough for enhancing computing power; computing power growth drives synchronous upgrades in power supply architecture. The connectivity segment must address both data movement in and out of memory and achieve smooth communication within servers, between racks, and across clusters. Against the backdrop of restrictions on the tape-out of domestic high-end computing chips, optimization of the connectivity segment has become a crucial breakthrough. The global market size for high-speed interconnect chips is projected to have a CAGR of 21.2% from 2024 to 2030, with China's market share increasing from 25% to 30%, creating broad incremental markets for industries like MRDIMM memory, PCIe interconnect chips, CXL interconnect chips, silicon photonics chips, and OCS. Focus on:澜起科技,蓝特光学.
Concurrently, as data center chip power consumption increases, rack processing power follows suit. Infineon predicts that single GPU power consumption will grow exponentially, reaching about 2000W by 2030, with peak rack power consumption exceeding 300kW. On one hand, improving power conversion efficiency is crucial for reducing data center operational costs. On the other hand, the large, rapidly fluctuating power curve on the rack side poses challenges to the stability of the public grid, necessitating a shift in power supply solutions towards HVDC. SST, DrMos, and SiC/GaN devices will become core directions for AI power supplies. Focus on:天岳先进,新洁能,江海股份,斯达半导,芯联集成,华润微,扬杰科技,杰华特,晶丰明源,顺络电子.
AI Edge-side: AI Agents reshape interaction paradigms; major players vie to布局 edge-side入口; a major year for consumer electronics innovation begins. As large models continue to evolve in multimodal understanding, general reasoning, and task execution capabilities, AI is upgrading from a tool-based capability to AI Agents that can understand user intent and autonomously execute tasks. Edge-side consumer electronics products are the key承载层 for the AI commercialization closed loop and are expected to systematically重构 human-computer interaction paradigms. Against this backdrop, various终端形态 like phones, glasses, headphones, and home robots are expected to build协同 networks around AI Agents, promoting AI's upgrade from single-point functions to systematic, cross-scenario, cross-terminal experiences. The importance of multimodal inputs like voice, vision, and environmental sensing is increasing, placing higher demands on edge-side computing power, perception capabilities, and connectivity.
The firm believes that current edge-side related technologies and industrial foundations have matured, and key breakthroughs in business models are expected to create a "non-linear amplification effect." Looking ahead to 2026, events from CES at the beginning of the year to WWDC in the middle, or the continuous exploration by leading manufacturers like ByteDance and Apple, combined with expectations of potential model providers and internet companies entering the hardware arena, could all serve as key catalysts igniting market sentiment and industrial investment consensus. Focus on:恒玄科技,晶晨股份,蓝思科技,立讯精密,蓝特光学,歌尔股份,翱捷科技,乐鑫科技,华勤技术,传音控股,小米集团,顺络电子.
Semiconductors: Recommend the self-sufficient and controllable industry chain where progress may exceed expectations, and analog chips accelerating domestic substitution during the景气复苏 stage. According to SIA data, China accounted for 28% of global semiconductor sales in 2024, but the proportion of local supply was only 4.5%, indicating a still low self-sufficiency rate. Furthermore, as the增量 primarily came from cloud-side increments like GPUs and HBM, the self-sufficiency rate decreased compared to 2023. However, the financial performance of A-share semiconductor companies continues to improve. Based on operational data from 146 companies tracked by the firm, 54% saw their highest quarterly revenue occur in 2025. In Q3 2025, the overall gross profit margin of the SW semiconductor sector was between Q1 2021 and Q2 2021 levels, and the net profit margin was comparable to Q4 2020 and Q1 2021 levels.
From an industry cycle perspective, global semiconductor sales have grown year-on-year for eight consecutive quarters. In December 2025, WSTS again raised its forecasts 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, besides the AI增量, the rise of domestic chip design companies and the demand for localized manufacturing provide increments for the self-sufficient manufacturing chain. Focus on wafer foundry, advanced packaging, and upstream semiconductor equipment and materials segments. Additionally, analog chips are later in the cycle among semiconductor product categories. International giants TI and ADI saw revenue turn positive year-on-year starting in 2025, marking the industry's entry into a recovery stage. New products launched by domestic companies in recent years are expected to enter the stage of规模化放量. In the long term, AI data centers and AI applications like autonomous driving and humanoid robots bring broad增量 for them. Simultaneously, analog chips are a segment with significant room for domestic substitution and will continue to benefit from increasing localization rates. Focus on:中芯国际,华虹半导体,杰华特,思瑞浦,北方华创,中微公司,拓荆科技,圣邦股份,南芯科技,伟测科技,通富微电,长电科技,鼎龙股份,骄成超声.
Key Focus Portfolio:中芯国际,工业富联,中微公司,翱捷科技,华虹半导体,寒武纪,北方华创,澜起科技,蓝思科技,蓝特光学,立讯精密,胜宏科技,华勤技术,恒玄科技,顺络电子,沪电股份,伟测科技,晶晨股份,鹏鼎控股,思瑞浦,杰华特,圣邦股份,德明利,江波龙,佰维存储,歌尔股份,兆易创新,鼎龙股份,京东方A,长电科技,水晶光电,传音控股,豪威集团,骄成超声,海康威视,小米集团,天岳先进,扬杰科技,乐鑫科技,江海股份,电连技术,晶丰明源,光弘科技,南芯科技,世华科技,北京君正,洁美科技,国芯科技,通富微电,唯特偶,福立旺.
Risk提示: Domestic substitution progress may fall short of expectations; Downstream demand may be weaker than expected; Risk of intensifying industry competition; Risk of adverse changes in international relations; Risk of industry cyclical fluctuations; Risk related to production equipment and raw material supply.
