The Zhitong Finance App learned that China Galaxy Securities released a research report saying that the front-end network networking of computing power clusters in the AI era is driving demand for a large number of switches. The 1.6T port switch is expected to start scaling in 2027. The WAIC 2026 supernode becomes the absolute protagonist. The internal interconnect bandwidth evolves rapidly. Every time the port rate is iterated, the value of a single switch increases, supporting profitability. At the same time, the design difficulty and verification cycle of scale-up switches has been reduced from μS level to 100 ns level. The design difficulty and verification cycle far exceeds that of traditional products, giving players in stock a continuous price negotiation Capability forces the “stuck neck” link in the industrial chain — domestic replacement and acceleration of switching chips, compounding the performance of leading overseas switches to achieve catalysts. It is recommended to focus on leaders in the domestic switch+switching chip field.
The main views of China Galaxy Securities are as follows:
Switch: Highly centralized in the industry, AIDC non-blocking pass-through promotes port upgrades
A switch is a network device responsible for multi-level data forwarding in the data link layer. According to its location and division of labor in the network, it can be divided into core switches, aggregation switches, and access switches. According to different terminal application environments, it can be divided into data center switches, park and metropolitan area network switches, and SMB switches. According to whether the software and hardware are decoupled and whether the operating system is open and controllable, it can be divided into black box switches and white box switches. The switching performance of switches mainly depends on backplane bandwidth capacity, switching capacity, port rate, port density, etc. The global Ethernet switch market size in 2026Q1 was US$15.4 billion/+39.8%, of which the data center switch market size was US$10 billion/+61%, and 800G high-speed switches contributed 35.8%, surpassing 34.1% of 200G/400G switches for the first time. Since switches are the core components of network equipment, customers have strict requirements for security, stability, and reliability. It is more difficult for new players to enter large-scale downstream supply chains, boosting industry concentration. CR5 = 69.7% of global Ethernet switches in a single quarter of 2026Q1. Among them, Cisco and Arista top 2 positions are stable, while the domestic switch market shows a pattern of “domestic production dominates, foreign investment supplements”. Traditional IDC is engaged in general services such as computing and storage. The traffic is mainly “north-south” across computer rooms, and packet loss is allowed. The ratio of switches to servers is about 1:16-1:20, and the procurement scale of single cluster switches is limited. The number of parameters for large AI models has now reached the trillion level. A single training and inference requires collaborative scheduling of tens of thousands of GPUs. The traffic is mainly “east-west” between GPUs within the computer room. The ratio of the switch to GPU is 1:3-1:4, and the ratio of some Nvidia high-density supernodes is 1:2, medium Looking at the AI Agent era, the ratio is expected to increase further to 1:1.
Switching chips: rapid growth in the next five years, domestic replacement is imminent
The switching chip logic path contains hundreds of features. The architecture is complex, and the port rate needs to balance diversity and peak values. After 2022, Nvidia accelerated the iteration pace of AI computing power chip products. From the previous generation every two years to an annual generation, AMD, Google, Amazon, etc. are also making every effort to develop the switching chip, while the iteration cycle of the switching chip “25.6T → 51.2T → 102.4T → 204.8T” is still two years. At the same time, considering that the non-local supply cycle of Broadcom switching chips has now reached more than a year, making switching chips a core bottleneck in AI infrastructure. The global Ethernet switch chip market is expected to grow at CAGR = 43% in 2025-2030, and the data center switching chip market size will be +86% year-on-year in 2026. Broadcom is still the undisputed leader. In June 2025, Broadcom released the world's first 102.4T switching chip, Tomahawk 6. At that time, Senke Communications mass-produced 25.6T switching chips. There was a two-generation gap. Tomahawk 6 had already adopted a 3nm process+CPO package, and domestic switching chips remained at the 7nm node. At the same time, Broadcom's SDK and reference design formed a complete ecosystem. Hardware replacement could increase the CAPEX process, and ecological replacement would take time to settle down.
Investment advice: focus on domestic switch+switching chip leaders
The front-end and back-end network networking of computing power clusters in the AI era is driving demand for a large number of switches. The WAIC 2026 supernode is expected to become the absolute protagonist. The internal interconnect bandwidth evolves rapidly. With every iteration, the value of a single switch increases, supporting profitability. At the same time, the latency requirements of scale-up switches are reduced from μS to 100 ns. The design difficulty and verification cycle far exceeds that of traditional products, giving existing players continuous bargaining power, forcing the “stuck neck” link in the industrial chain - chip switching Domestic replacement is accelerated, and the performance of leading overseas switches is realized and catalyzed. It is recommended to focus on leaders in the domestic switch+switching chip field.
Risk warning: risk of AI application development falling short of expectations; risk of CAPEX decline in major CSP vendors; risk of slow progress of supernode solutions; risk of deviation between switch chip design and performance, etc.