Dongwu Securities: Supernode solutions reshape AI computing power industry trends and start a new cycle of system-level computing power
The Zhitong Finance App learned that Dongwu Securities released a research report saying that by expanding the high-bandwidth collaboration domain and shortening high-frequency communication paths, supernodes organize distributed computing power into schedulable and expandable system computing power, and have become a core breaking point for breaking through traditional cluster efficiency bottlenecks and achieving large-scale collaboration. Competition extends to the ability of the entire cabinet to collaborate, and the amount of value also spills over from computing cards to interconnection, liquid-cooled power supply, system verification, and full cabinet delivery. As supernodes move from prototype verification to large-scale delivery, customer binding depth, full cabinet engineering capability, and project replication efficiency will become the core support for server manufacturers to achieve commercial performance.
The main views of Dongwu Securities are as follows:
Under the resonance of computing power supply and demand, supernodes have become an inevitable path for the evolution of AI clusters
Global AI's cutting-edge model capabilities are moving from a single round of question-and-answer to complex task processing. Long context, complex inference, and agent-based applications continue to lengthen the inference chain and drive upward demand for training and inference computing power. At the same time, overseas cloud vendors are speeding up the expansion of AI infrastructure, and the volume of large-scale domestic model applications resonates with the release of domestic computing power supply. The main contradiction of computing power infrastructure is moving from “expanding computing power cards” to “improving effective computing power.” By expanding the high-bandwidth collaboration domain and shortening high-frequency communication paths, supernodes organize distributed computing power into schedulable and expandable system computing power, becoming a core breaking point for breaking through traditional cluster efficiency bottlenecks and achieving large-scale collaboration.
Core variables of the supernode architecture upgrade: high bandwidth interconnection and full cabinet engineering capability
(1) At the communication level, Scale Up is responsible for undertaking high-frequency data exchange within supernodes. Scale Out is responsible for the horizontal expansion of multiple cabinets and clusters, and the boundary between the two is shifting from physical server boundaries to performance boundaries; Nvidia's NVLink closed loop system, Huawei Lingqu UB domestic system, AMD/UALink overseas open routes, and open interconnection solutions from cloud vendors such as Ali, Byte, and Tencent are evolving in parallel to promote the supernode from architecture exploration to commercialization. (2) At the engineering level, 100 kilowatt cabinet power consumption, high-density connections, and long-term stable operation have upgraded heat dissipation, power supply, structure, and operation and maintenance from single-point design to full-cabinet system engineering. As a result, hypernode competition extends from single card performance to the entire cabinet's ability to collaborate, and value spills over from computing cards to interconnect exchange, liquid cooling power supply, system verification, and full cabinet delivery.
Supernodes reshape the value chain of server manufacturers, the key to full cabinet delivery and customer ecosystem segmentation
Supernodes drive AI server demand to shift from stand-alone procurement to full-cabinet systems and intelligent computing center-level solution delivery, and the value boundary of server vendors also expands from whole machine manufacturing to full cabinet integration, system testing, and on-site deployment. IFF benefits from full-cabinet manufacturing and high-speed network equipment deployment for leading global customers; Huaqin technology relies on platform-based ODM capabilities to extend to AI servers, supernodes, and data center products; Inspur Information, Ziguang Co., Ltd., Hyperfusion, and Ningchang each form differentiated cards around domestic AI servers, network docks, operator liquid-cooled cabinets, and high-density customized scenarios.
Risk warning: demand for computing power falls short of expectations; technical routes and commercialization implementation; mass production and delivery of the entire cabinet project falls short of expectations; increased industry competition and declining profit levels.