A research report from Sinolink Securities Co., Ltd. indicates that, over the long term, the slowing pace of model iteration will drive the evolution of computing power toward ASIC specialization. The explosive growth in inference demand is shifting computational workloads from GPUs to specialized chips, enabling the deployment of more computing nodes at a lower total cost of ownership (TCO), thereby supporting long-term growth in optical modules and PCBs. While market concerns exist regarding CPO replacing pluggable modules and a potential reversal of "optical replacing copper" trends, the reality is that these technologies are more likely to complement each other. It is projected that by 2026, 800G ports will become the absolute mainstream, with CPO technology expected to contribute more as a mid- to long-term incremental addition rather than a short-term replacement.
Key viewpoints from Sinolink Securities are as follows:
Over the long term, computing power will shift toward ASIC as model iteration slows. The rapid expansion of AI inference demand is accelerating the transition from general-purpose GPUs to specialized ASICs. ASICs excel in inference due to their deep optimization for specific algorithms, offering superior energy efficiency and cost advantages. Leading manufacturers such as Broadcom and Marvell have shown strong growth in their custom chip businesses. The primary driver behind ASIC adoption is the ability to deploy more computing nodes at a lower cost within a sustainable TCO framework. This trend increases the total number of endpoints requiring interconnection, creating long-term growth momentum for the optical module and PCB markets.
Market concerns focus on risks two years out: CPO and "optical replacing copper." Currently, there are two major concerns for the optical communication industry beyond 2026: the potential replacement of traditional pluggable modules by CPO (co-packaged optics) technology, and a reversal of the "optical replacing copper" trend in short-distance interconnects. However, a deeper analysis of technology implementation paths and vendor strategies suggests that these developments are more likely to complement existing technologies rather than replace them in the near term.
Interconnect architectures are shifting from proprietary tight-coupling to Ethernet-based openness. ASIC chips have a natural affinity for Ethernet, driving the evolution of AI cluster networks from proprietary protocols like InfiniBand to open, Ethernet-based architectures. The rapid expansion of the Ultra Ethernet Consortium (UEC) serves as strong evidence of this trend. This decoupling separates computing chips from the network layer, generating steady demand for pluggable optical modules behind each standardized high-speed port. Currently, 800G and subsequent 1.6T pluggable modules are expected to dominate the market for the foreseeable future. By 2026, 800G ports are projected to become the absolute mainstream, with CPO technology more likely to provide incremental contributions in the mid to long term rather than near-term replacement.
PCB value increases systematically with ASIC adoption; supply-demand gap expected to persist. In terms of hardware support, the PCB value per ASIC server motherboard is significantly higher than that of same-generation GPU servers. Material upgrades (such as high-end materials like M7 and M8) and process improvements jointly contribute to this value leap. As optical module speeds increase and CPO technology commercializes, higher demands are placed on PCB structural design, thermal performance, and signal loss. With limited global manufacturers capable of mass-producing high-end PCBs, supply constraints and additional growth momentum in the ASIC sector are expected to continue.
Related investment targets include overseas computing/storage companies, domestic computing firms, CPU manufacturers, AI application developers, and other emerging sectors such as aerospace and embodied intelligence. Risks highlighted include intensified industry competition, accelerated commercialization of CPO, and cyclical fluctuations in capital expenditures within specific downstream industries.
Comments