China Galaxy Securities released a research report stating that due to the high technical difficulty in supplying hardware adapted for high-performance computing, the overall market is still in a stage of supply falling short of demand. Consequently, the degree of "bubbling" compared to a scenario of sufficient supply remains limited, still preserving a margin of safety. Admittedly, the most effective way to alleviate concerns about an "AI bubble" would be a major breakthrough in AI applications, a moment that has not yet arrived. Considering the supply-demand imbalance, the firm maintains its recommendation on the hardware sector. China Galaxy Securities' main views are as follows:
The current stage of AI development differs from the internet bubble era. Since 2022, when AI technology began gaining widespread attention, it has triggered an unprecedented wave of infrastructure investment in the AI industry, consequently creating a boom cycle for semiconductor hardware. Global capital expenditures from major cloud providers are projected to exceed their combined operating net cash flow by 2026, making current market worries about an "AI bubble" understandable. The firm reviewed the previous internet bubble period (1995-2000) and identifies several key differences with the current phase: ① The corporate entities driving industrial development are different; ② Comprehensive commercialization of AI has not yet arrived; ③ Even if there is a bubble in AI investment, its transmission to the stock market is currently limited; ④ A global period of accommodative monetary policy has not fully materialized.
Focusing on the hardware industry chain, which is the firm's primary concern, and considering technological bottlenecks, chip supply is still in a phase of trying to catch up with explosive demand expectations. The firm assesses the current stage of hardware investment by observing key segments: computing power, storage, and PCBs.
Supply of computing and storage power remains tight. The rapid development of large models from text to video and other multimodal capabilities, along with the realization of AI Agent functions, is continuously increasing token consumption. This drives rigid demand for high-performance computing power. On the supply side, global data center vacancy rates are at historic lows, and the power consumption of data centers is testing the elasticity of power supply, indicating that supply-demand dynamics remain tense. Domestic computing power is still in a catch-up phase, with supply-side constraints persisting. Regarding storage chips, significant price increases reflect the tight supply-demand relationship, and effective supply still relies on continuous technological progress.
PCB capacity expansion remains rational. The current PCB upcycle is primarily characterized by growth in shipments of HDI and high-layer-count boards. The global market for high-end HDI PCBs and high multilayer boards (above 14 layers) is forecast to have compound annual growth rates of 20.3% and 11.6%, respectively, from 2024 to 2029. High-end HDI and high multilayer boards themselves have high technical barriers, and capacity expansion is mainly led by top-tier enterprises. As of the first three quarters of 2025, industry capital expenditure increased by 8.4% compared to the full year 2024, indicating that capacity expansion remains limited relative to demand growth.
Risk warnings include the risk that progress in AI applications falls short of expectations and the risk of uncertainty in the international political environment.
Comments