CITIC Securities: Super Nodes Enhance AI Training Efficiency, Focus on Three Major Incremental Links.

CITIC Securities research report stated that the super node connects multiple accelerator cards through a high-bandwidth, low-latency Scale-up network and provides mechanisms such as memory pooling and direct memory access, greatly improving the efficiency of AI training and prediction, and has become a clear industry trend. With companies like NVIDIA driving the efficiency and capacity expansion of super nodes, GPUs, exchange chips, liquid cooling, and cabinet power supply segments will all benefit from the Scale-up pure incremental usage, cabinet high power penetration rate and ASP improvement driven by high power in cabinets, achieving growth; It is estimated that the incremental space in 2028 will reach $100 billion, $13 billion, and $24 billion respectively. In particular, the opportunity for domestic alternatives in exchange chips is worth expecting; the good commercial properties of exchange chips shape a stable oligopoly situation, with a large and clear incremental space for GPU interconnection, the domestic space is expected to reach $5 billion in 2028, with Ethernet solutions currently becoming its main technological direction, and with the trend of domestication of computing power chips, supporting domestic solutions have already been implemented; currently, the leading revenue volume is about 1 billion RMB, with enormous growth potential; it is recommended to pay attention to domestic Ethernet exchange chip players; in addition, the domestic alternative of optical interconnect in the CPO/NPO segment based on exchange chips is also worth attention.