CITIC Securities: Os nós superiores aumentam a eficiência do treino e da inferência de IA, com foco em três principais segmentos de crescimento

Mars Finance News, a research report from CITIC Securities states that supernodes tightly connect multiple accelerator cards through a high bandwidth, low latency Scale-up network, providing mechanisms such as memory pooling and memory direct connection, greatly enhancing AI training efficiency, which has become a clear industrial trend. As manufacturers like NVIDIA promote the efficiency and expansion of supernodes, the inter-GPU switching chips, liquid cooling, and cabinet power supply segments will benefit from the pure incremental usage of Scale-up, the penetration rate and ASP driven by high power in cabinets, and the ASP enhancement driven by high power in cabinets, achieving growth; it is expected that by 2028, the incremental space will reach 100 billion, 13 billion, and 24 billion USD respectively. Among them, the domestic substitution opportunities for switching chips are worth looking forward to; the good commercial attributes of switching chips shape a stable oligopoly pattern, with a large and clear incremental space for inter-GPU connections, and the domestic space is expected to reach 5 billion USD by 2028. Currently, Ethernet solutions are becoming its main technical direction, and compatible domestic solutions have already been implemented with the trend of domestic substitution of computing power chips; at present, the leading revenue scale is about 1 billion yuan, with enormous growth potential; it is recommended to pay attention to domestic Ethernet switching chip players; in addition, the domestic substitution of optical interconnection CPO/NPO segments based on switching chips is also worth monitoring.