Huatai Securities: optimistic about the mass production of domestic super nodes in AI training and driving from 2026 onwards.

Huatai Securities pointed out that as the core means of improving efficiency through the clustering of computing power, super nodes break through the boundaries of physical systems, achieving tight coupling and interconnection of tens to hundreds of accelerator cards within a single cabinet. It is optimistic that starting in 2026, domestic super nodes will start to mass produce under the drive of AI training and inference: 1) Trillion-parameter models force the clustering of computing power to improve efficiency, and super nodes reduce communication overhead through system-level restructuring, effectively converting theoretical computing power into usable throughput; 2) Based on the gap between domestic cards and advanced cards overseas, super nodes may be the breakthrough for domestic computing power, compensating for single-card differences by amplifying the role of system-level interconnection within the cluster; 3) Starting from 2025, major domestic computing power participants have successively released super node prototypes. It is estimated that the market space for domestic super nodes in 2028 is expected to reach 341.4 billion yuan, with a CAGR of 194% from 2026 to 2028. It is recommended to pay attention to the core incremental Scale up process of super nodes, as well as other benefiting processes such as whole cabinet, liquid cooling, optical modules, etc.