The Zhitong Finance App learned that Guojin Securities released a research report saying that the supernode industry has entered a large-scale implementation period, and delivery units have moved from a single white box server to full cabinet and even Pod-level system delivery. The design of the entire supernode cabinet needs to solve the challenges of high-density GPU collaborative work, and the value of multiple links is expected to increase significantly. The core incremental links are switches and switching chips, servers, cabinets, and connectors. With the release of domestic chip supply, demand will in turn be transmitted to switching networks, servers and cabinets, computing power leasing, AIDC listing, and cloud services. Demand expansion, chip supply breakthroughs, and system architecture upgrades are resonating, continuing to reaffirm the golden age of domestic computing power.
Guojin Securities's main views are as follows:
The WAIC conference conducted a major review of domestic supernodes, and the demand side simultaneously entered the supernode era
Domestic supernode products were heavily reviewed at this year's WAIC conference. Manufacturers such as Huawei, ZTE, Mu Xi, and Moore Thread focused on showcasing supernode products. The product forms covered 64 cards, 128 cards, 256 cards, 384 cards, and even 1024 card high-bandwidth communication domains; among them, the Huawei Ascend 950 Super Node 1024 card was publicly unveiled for the first time, and the Shengteng 384 supernode has been commercialized on a large scale. ZTE OEX, Mu Xijing S600, Moore Thread MTTC256, Suiyuan ESLO/64-C and NPO Lightwall Interconnection solutions and the like were also showcased. At the same time, the demand side is also entering the supernode era at the same time. Kimi K3's official deployment proposal clearly proposes a supernode configuration composed of 64 or more acceleration cards. The bank believes that supernodes are gradually becoming a common technical route for domestic AI infrastructure, and that the evaluation framework for domestic computing power competition is shifting from peak computing power of a single chip to the scale, communication efficiency, software maturity, and system delivery capability of an accelerator card that can be collaborated within a high bandwidth domain.
The core incremental links of supernodes are switches and switching chips, servers, cabinets, and connectors
1) Switches and switching chips: Supernodes further extend the scale-out network outside the server to scale-up interconnection within the cabinet and between cabinets, and promote the expansion of switching requirements from traditional Ethernet switches to high-bandwidth, low-latency dedicated switching chips and switching systems. As the computing power and interconnection bandwidth of a single card continue to increase, the number of chips, port speed, and stand-alone value in the switching process are expected to rise. 2) Servers and cabinets: Delivery units have been upgraded from a single white box server to an entire cabinet or even a Pod-level system. The difficulty of power supply, cooling, interconnection and system tuning has increased significantly, and the barriers and profitability of leading ODM projects are expected to increase simultaneously. 3) Connectors: The number of high-speed connections between computing boards, switching boards, power supplies and backboards has increased, and the SerDes rate continues to increase. At the same time, various connection architectures such as orthogonal, cable-tray, and NPO have evolved in parallel to promote the expansion and upgrading of high-speed backboards, board-end connectors, high-speed copper cables and optical interconnect products.
Supply and demand resonate upward, continuing to reaffirm the golden age of domestic computing power
On the demand side, domestic model capabilities are crossing the threshold of usability. Coding, Agent, long context, and MoE models continue to drive token calls, and the high-bandwidth communication domain has changed from engineering optimization items to model deployment requirements; on the supply side, domestic GPU and domestic CPU supply improvements, combined with factory capital development support, are expected to push domestic computing power from chips to the system-level volume stage. The bank believes that with the release of domestic chip supply, demand will sequentially be transmitted to switching networks, servers and cabinets, computing power leasing, AIDC listing, and cloud services. Demand expansion, chip supply breakthroughs, and system architecture upgrades are resonating, continuing to reaffirm the golden age of domestic computing power.
Investment advice
Switching chips/switches: Senke Communications, Ziguang Co., Ltd., Ruijie Networks, etc.;
Servers/cabinets: Inspur Information, Huaqin Technology, Zhongke Shuguang, ZTE, etc.;
Connectors: Huafeng Technology, Shenglan Co., Ltd., China Aviation Optoelectronics, Aerospace Electric, etc.;
Domestic computing power: Cambrian, Haiguang, Dongyangguang, SMIC, Huahong Semiconductor, Hesheng New Materials, Moore Thread, Tianshu Zhixin, Mu Xi, Bizao Technology, Co-Creation Data, Huafeng Technology, Jiewalt, Jingji Zhinong, Huada Technology, Zhongke Shuguang, Litong Electronics, Inspur Information, Liyang Chip, Shenglan Co., Ltd., Huaqin Technology, China Great Wall, Jingke Technology, Roman Shares, Yingfeng Environmental, Xinyuan Tian Intelligence, Yuneng Holdings, Xinghuan Technology, Hongrida, Shengshi Technology, Shenzhou Digital, Runze Technology, Dawei Technology, Runjian Co., Ltd., Aofei Data , Ruisheng Intelligence, Kehua Data, Weichai Heavy Machinery, Eurocom, Jiechuang Intelligence, Oni Electronics.
Risk Alerts
The risk of increased competition in the industry, the risk that technology research and development progress falls short of expectations, and the risk of cyclical fluctuations in downstream capital expenditure in specific industries.