China Merchants Securities: WAIC 2026 supernode accelerates launch, focusing on the pace of future expansion of domestic computing power

The Zhitong Finance App learned that China Merchants Securities released a research report saying that the focus of data center computing power units has evolved from the previous increase in single-chip computing power and the increase in the degree of single-server aggregation to supernodes formed through high-speed, low-latency, and high-bandwidth interconnection protocols. Supernodes are collaborative restructuring of core modules such as computing, storage, high-speed connectivity, power supply systems, and cooling, and are the core building blocks of future computing power bases. Based on the bank's WAIC conference and industry research, it is recommended to focus on the following directions in the field of domestic supernodes: 1) complete machine solution suppliers; 2) domestic computing power chips, etc.; 3) foundry; 4) high-speed electric/optical connections; 5) power supply/liquid cooling; 6) PCBs.

The main views of China Merchants Securities are as follows:

Domestic computing power chips are moving from single-card performance competition to super-node system-level competition, and continue to evolve to 10,000 cards, 100,000 cards, and even million card cluster interconnections

Various companies have formed differentiated solutions: 1) Huawei Shengteng 950 kilocalorie supernode: a 1024-card unified memory supernode is built based on the Lingqu Internet. It consists of 16 computing cabinets and 4 interconnected equipment cabinets, providing 1 EFLOPSFP8, 2 EFLOPSFP4 computing power and 256 TB unified addressing memory. RTT is as low as 3 μs, and the system can be further expanded to 8192 Ascend 950DT; 2) Haiguang: Relying on CPU+DCU dual-core collaboration to support Shuguang's 100,000-kilocal AI supercluster; 3) Mu Xijing S600 supernode: single cabinet High-density 64-card deployment, using OEX orthogonal backboard-free interconnection architecture, the cluster can be horizontally expanded to 10,000 cards, fully covering large model training and inference scenarios. 4) Moorish thread MTTC256 supernode: pioneered a first-tier scale-up network. The single cabinet supports full interconnection of 128 cards. The combination of the two cabinets achieves high-speed interconnection of 256 cards. The communication delay between cards is reduced to sub-microseconds. The 2U node is compatible with the existing intelligent computing ecosystem, and the cluster can be smoothly expanded to 100,000 cards. 5) Bizao Technology: forms a 16-card standard server, 128-card high-density cabinet and 1024-card NPO optical interconnect supernode product matrix; 6) Kunlun XinTianchi series: 256-card high-end supernodes use a multi-level high-speed interconnection architecture, which has been adapted to mainstream large-scale models and can support large-scale AI application deployment; 7) Suiyuan Technology: Jointly launched a 64-card OEX orthogonal backless supernode with ZTE. Overall, the competitive dimension of domestic computing power has expanded from chip manufacturing processes and peak computing power to system-level capabilities such as high-speed interconnection, unified memory, liquid cooling power supply, cluster reliability, and software stack collaboration.

Large-scale commercialization of domestic computing power is accelerating, and industrial development is moving from “product availability” to the “cluster verification and batch deployment” stage

Huawei's first-generation Shengteng 384 supernodes have been commercially launched, covering the Internet, operators, finance, education and manufacturing industries; Haiguang participated in the construction of Shuguang's first nationally produced 100,000-card AI supercluster “Shuguang 8000” to verify the ability of the CPU+DCU system to support large-scale token production; Kunlun Core's 32,000 card cluster within Baidu continues to expand, and 32/64 card supernodes have also been collected by many provincial operators; Moore threads have completed the full training process of the Moe-236b basic model based on the Kuae cluster, accounting for the full training period More than 90%, and achieved an average daily trillion-level token supply through partners; Bijiao Technology and China Telecom promoted heterogeneous and mixed promotion across vendors, increasing the overall throughput of the cluster by 20%; Tianshu Zhixin Tianyi 300 already has the conditions for large-scale application and is being adapted to domestic cloud vendors and server vendors. In the future, actual customer deployment scale, effective cluster computing power utilization rate, long-term operation stability, and single token cost will be the core indicators for testing commercialization capabilities. Manufacturers with chips, software stacks, supernodes, and full-link customer ecosystem capabilities are expected to take the lead in forming a closed loop of large-scale commerce.

Risk warning: macroeconomic and policy risks; risk of AI capital expenditure falling short of expectations; risk of increased industry competition