Huaan Securities: The shift of AI technology to inference drives the hardware industry chain ushered in a new round of growth cycle
The Zhitong Finance App learned that Huaan Securities released a research report saying that global AI technology is shifting from training to reasoning, driving the upgrading of the hardware industry chain. Cloud service providers and sovereign AI plans to promote the booming construction of global AI infrastructure and drive cloud-side hardware value improvement and technological innovation such as servers, storage, and optical interconnection. At the same time, end-side devices such as AI phones and AR glasses are accelerating the evolution of intelligence and reshaping the industrial landscape. It is recommended to focus on related industrial chains that benefit from growth in inference computing power and hardware upgrades.
The main views of Huaan Securities are as follows:
Total
Global AI technology is shifting from being dominated by training to being dominated by reasoning, driving the hardware industry chain to a new round of growth opportunities. The iteration of multi-modal models such as Google Gemini 3 Pro and OpenAiSora 2, and the large-scale implementation of AIAgent have significantly increased the demand for inference computing power. Driven by this trend, cloud service providers continue to raise capital expenditure. The capital expenditure of the world's top eight CSP is expected to reach 431 billion US dollars in 2025, an increase of 65% over the previous year, and is expected to further increase to 602 billion US dollars in 2026. At the same time, various countries' sovereign AI programs have been launched one after another. For example, the US “Stargate” program invests about 500 billion US dollars, and the EU plans to invest 21.5 billion US dollars to build an AI gigafactory. Together, these measures will push the global AI infrastructure into a booming construction cycle. According to forecasts, by 2030, the global AI data center capacity will reach 156 GW, accounting for 71% of total data center demand.
Cloud side
PCB: The AI server brought a clear increase in value. For example, the Nvidia DGXH100 PCB value for a single GPU reached $211, an increase of 21% over the previous generation; the GB200 NVL72 also boosted the value of a single GPU to $346. As the Rubin architecture adopts a cable-free design and switches evolve to 800G/1.6T, PCBs are being upgraded towards higher levels and higher performance using low dielectric materials such as M9. At the same time, domestic high-end PCB production capacity will be released centrally in 2026 to support downstream demand. This trend has also led to the upgrading and iteration of upstream materials: including the increase in the proportion of hydrocarbon resins in the M9 grade, the introduction of third-generation quartz glass fiber cloth, and the use of HVLP4 copper foil. Domestic material manufacturers are speeding up replacement and breakthroughs in all aspects.
Storage: Structural imbalances between supply and demand due to AI demand in 2025 have driven DRAM and NandFlash prices to rise significantly. The capital expenditure growth rate of the industry is expected to slow in 2026, and the focus of investment will shift to high value-added products. In terms of technological evolution, 3DDRAM technology provides domestic manufacturers with an opportunity to bypass advanced lithography limitations through TSV and 4F2 vertical structures; at the same time, KvCache technology, which serves large model inference optimization, is driving QLCSSD to accelerate the replacement of HDDs, and its penetration rate in the enterprise SSD market is expected to reach 30% in 2026.
Optical interconnection: As the key to AI computing power clusters, optical interconnect technology is entering a new era. With its high bandwidth, low latency, and low power consumption, the optical switch perfectly adapts to the interconnection requirements of large-scale AI clusters. At present, the MEMS-based technology route has taken the lead. The industrial chain is long and the barriers are high. From upstream core devices (such as MEMS arrays and optical fiber collimators) to midstream equipment integration and solutions, domestic manufacturers have actively laid out and entered the global supply chain in all aspects.
End side
AI phones: AI phones are reshaping the industry. The overall market will maintain moderate growth in 2025, while the focus of competition has shifted to end-side AI capabilities. Mobile operating systems are evolving from “application launchers” to “system-level agents”. Innovative products such as Doubao phones try to achieve underlying AI integration and cross-application operation. The flagship chips of the Apple and Android camps are also continuing to improve NPU computing power, and jointly promote the popularization and experience upgrading of end-side AI.
AR glasses: Smart glasses that combine AI and AR are seen as the future form of wearable devices. The market is growing rapidly, and the product form continues to evolve from camera-free glasses to AR glasses with complete display functions. On the technical path, optical waveguides are expected to become the mainstream choice for AR glasses optical imaging module solutions due to their advantages in clarity and volume; opto-mechanical solutions are showing a diversified trend. LCOS is currently the mainstream of consumer products, and MicroLED is recognized as the future development direction due to its performance advantages.
Recommended focus: Based on the AI computing power paradigm shift to inference, the hardware industry chain ushered in the core judgment of the upgrade cycle. The bank suggests focusing on the following segments and targets: PCB and upstream materials: Shenghong Technology, Shanghai Electric Power Co., Ltd., Jingwang Electronics, Guanghe Technology, Dongcai Technology; storage and equipment: Beijing Junzheng, Zhaoyi Innovation, Juchen Co., Ltd., Jingzhida; optical interconnection: Yingtang Intelligent Control, Saiwei Electronics; End side AI: Goertek, Lixun Precision, Baiwei Storage, Longqi Technology, Crystal Optoelectronics Lanxun, Haowei Group, Shunyu Optical Technology.
Risk warning: New technology iteration falls short of expectations, risk of raw material price fluctuations, risk of market demand falling short of expectations, risk of market competition intensifying risk, risk of production capacity construction falling short of expectations, etc.