China Galaxy Securities: The AI era drives a significant demand for switches; pay attention to leading companies in the domestic switch and switch chip sectors.

China Galaxy Securities has released a research report stating that in the AI era, both front-end and back-end networking of compute clusters are driving a significant demand for switches. 1.6T port switches are expected to see large-scale production beginning in 2027, with WAIC 2026 supernodes becoming the absolute main focus. The internal interconnection bandwidth is rapidly evolving; with each iteration of port speed, the value of a single switch increases, supporting profitability. At the same time, the requirements for latency in Scale-Up switches have compressed from microseconds to hundreds of nanoseconds, while the design complexity and verification cycles far exceed those of traditional products. This endows existing players with sustained bargaining power and forces the industry chain to accelerate domestic substitution of "bottleneck" componentsswitching chips. Coupled with the performance realization of overseas switching giants, it is advised to pay attention to leading firms in the domestic switch and switching chip sectors. China Galaxy Securities' main viewpoints are as follows: Switches: Industry is highly concentrated, AIDC non-blocking direct connection promotes port upgrades. Switches are network devices responsible for multi-level data forwarding at the data link layer. Based on their position and function within the network, they can be classified into core switches, aggregation switches, and access switches. Depending on different terminal application environments, they can be divided into data center switches, campus and metropolitan area network switches, and SMB switches. Based on whether hardware and software are decoupled and whether the operating system is open and controllable, they can be classified as black box switches and white box switches. The switching performance of switches mainly depends on factors such as backplane bandwidth capacity, switching capacity, port rate, and port density. In Q1 2026, the global Ethernet switch market reached a scale of $15.4 billion (+39.8%), with the data center switch market at $10 billion (+61%), and 800G high-speed switches contributing 35.8%, surpassing the 34.1% contribution from 200G/400G switches for the first time. As switches are core components of network equipment, customers have strict requirements for security, stability, and reliability, making it difficult and time-consuming for new players to enter the downstream large supply chain, which increases industry concentration. In Q1 2026, the global Ethernet switch CR5 was 69.7%, with Cisco and Arista solidifying their Top 2 positions, while the domestic switch market exhibited a pattern of "domestic dominance, foreign complement." Traditional IDC engaged in ordinary business such as computing and storage relies primarily on "north-south" cross-rack traffic, allowing for packet loss, with a switch-to-server ratio of about 1:16-1:20, leading to limited procurement volume for cluster switches. However, the parameters of AI large models have now reached trillions, with each training and inference requiring the coordinated scheduling of tens of thousands of GPUs, with traffic primarily "east-west" between GPUs within the data center. The interconnect bandwidth between switches adopts a 1:1 non-convergent design, and the switch-to-GPU ratio is about 1:3-1:4, with some high-density supernode configurations by Nvidia reaching 1:2. In the medium to long term, the ratio in the AIAgent era is expected to further enhance to 1:1. Switching chips: Rapid growth expected in the next five years, domestic substitution is urgent. The logical pathways of switching chips include hundreds of characteristics, with complex architectural implementations and port rates that need to balance diversity and peak performance. Nvidia has accelerated the iteration rhythm of AI computing chip products post-2022, increasing from a generational upgrade every two years to one per year; AMD, Google, Amazon, and others are also fully engaged in R&D. The iteration cycle for switching chips from "25.6T51.2T102.4T204.8T" still remains at two years. Considering Broadcom's non-local supply cycle for switching chips currently exceeds one year, switching chips have become a core bottleneck for AI infrastructure. It is estimated that from 2025 to 2030, the global Ethernet switch chip market size will grow at a CAGR of 43%, and the data center switch chip market is expected to grow by 86% year-on-year in 2026, with Broadcom remaining an undisputed leader. In June 2025, Broadcom will release the world's first 102.4T switching chip, Tomahawk 6, while domestic Shengkai Communication will mass-produce a 25.6T switching chip, indicating a two-generation gap. Tomahawk 6 has adopted a 3nm process and CPO packaging, whereas domestic switching chips are still at the 7nm node. Meanwhile, Broadcom's SDK and reference designs have formed a complete ecosystem, and hardware replacement can be expedited through increased CAPEX, while ecosystem replacement will require time to solidify. Risk Warning There is a risk that AI application development may fall short of expectations; risks associated with the downward trend of major CSPs' CAPEX; risks of slow progress in supernode solutions; risks of design and performance deviations in switching chips, among others.