CMSC: WAIC 2026 Super Node Acceleration Launched Focus on the Future Volume Pace of Domestic Computing Power.

CMSC released a research report stating that the focus of data center computing unit has shifted from previous single-chip computing power improvement and single-server aggregation degree enhancement to the formation of super nodes composed of high-speed, low-latency, and high-bandwidth interconnection protocols. Super nodes involve the cooperative reconstruction of core modules such as computing, storage, high-speed connections, power supply systems, and cooling systems, and they are the core constituent units of future computing bases. Based on CMSC's participation in the WAIC Conference and industry research, the report suggests focusing on the following directions in the domestic super node field: 1) Complete machine solution providers; 2) Domestic computing chips; 3) Wafer foundries; 4) High-speed electrical/optical connections; 5) Power supply/liquid cooling; 6) PCB. Key points of CMSC include: Domestic computing chips are transitioning from single-card performance competition to super node system-level competition, evolving towards clusters of tens of thousands or even hundreds of thousands of cards interconnected. Differentiated solutions have been developed by various companies: 1) Huawei Ascend 950-thousand-card super node: Based on Lingxuan interconnection, it builds a unified memory super node with 1024 cards consisting of 16 computing cabinets and 4 interconnection equipment cabinets, providing 1 EFLOPS FP8, 2 EFLOPS FP4 computing power, and 256TB unified addressing memory, with RTT as low as 3s, the system can be further extended to 8192 Ascend 950DT cards; 2) Haiguang: Supporting the Sunway hundred-thousand-card AI supercluster with CPU+DCU dual-core collaboration; 3) Much-Xixi S600 super node: High-density 64-card deployment in a single cabinet, using OEX orthogonal non-backplane interconnection architecture, the cluster can be horizontally expanded to tens of thousands of cards, covering large model training and inference scenarios. 4) Moore Thread MTTC256 Super Node: Pioneering a Scale-up network, a single cabinet supports 128 fully interconnected cards, and two cabinets combined achieve fast interconnection of 256 cards, reducing inter-card communication latency to microseconds, 2U nodes compatible with existing intelligent computing ecosystems, clusters can smoothly expand to tens of thousands of cards. 5) BIREN TECH: Forming a product matrix of 16-card standard servers, 128-card high-density complete cabinet, and 1024-card NPO optical interconnected super node; 6) Kunlun Xintianchi series: 256-card advanced super nodes adopt a multi-level high-speed interconnection architecture, completing the adaptation of mainstream large models, and supporting large-scale AI application deployment; 7) Siui Tech: Jointly launched a 64-card OEX orthogonal non-backplane super node with ZTE. Overall, the competition dimension of domestic computing has expanded from chip process and peak computing power to system-level capabilities such as high-speed interconnection, unified memory, liquid cooling power supply, cluster reliability, and software stack coordination. Commercialization of domestic computing power is accelerating, and industry development is moving from "product availability" to "cluster verification and large-scale deployment" stage. Huawei's first-generation Ascend 384 super node has been commercially deployed in over 750 sets, covering industries such as the Internet, operators, finance, education, and manufacturing; Haiguang participated in the construction of the first national ten-thousand-card AI super cluster "Sunway 8000", testing the CPU+DCU system's ability to support large-scale Token production; Kunlunxin continues to expand the 32/64-card super node cluster with 3.2 million cards internally at Baidu, and 32/64-card super nodes have entered the centralized procurement of several provincial-level operators; Moore Threads has completed the full-process training of the MoE-236B basic model based on the Kuayuo cluster, with effective training duration accounting for over 90%, and daily trillion-level Token supply achieved through collaboration partners; BIREN TECH and China Telecom Corporation promote cross-vendor heterogeneous mixed deployment, improving overall cluster throughput by 20%. ILUVATAR COREX Tian Gai 300 has the conditions for large-scale application and is compatible with domestic cloud service providers and server manufacturers. In the future, real customer deployment scale, effective cluster computing power utilization, long-term operational stability, and single Token cost will be core indicators to test commercialization capabilities. Manufacturers with the full capabilities of chips, software stacks, super nodes, and customer ecosystem are expected to form a scale commercial loop first. Risk warning: Macroeconomic and policy risks; risks of AI capital expenditure falling short of expectations; risks of intensified industry competition.