CITIC Securities: Super Nodes Enhance AI Training and Deployment Efficiency, Focus on Three Major Growth Areas

On March 30, People’s Financial News reported that CITIC Securities’ research stated that supernodes will connect multiple accelerator cards through a high-bandwidth, low-latency Scale-up network, providing mechanisms such as memory pooling and direct memory access, greatly improving AI training and inference efficiency, and has become a clear industry trend. As companies like NVIDIA promote the efficiency and expansion of supernodes, the segments of GPU interconnection chips, liquid cooling, and cabinet power supplies will benefit from the pure incremental demand for Scale-up, the increased penetration rate and ASP driven by high power in cabinets, and the ASP uplift driven by high cabinet power, leading to growth; it is projected that by 2028, the incremental market space will reach $100 billion, $13 billion, and $24 billion respectively. Among these, the opportunities for domestic substitution of interconnection chips are particularly promising; the strong commercial attributes of interconnection chips help shape a stable oligopoly pattern, with a large and clearly defined incremental space for GPU interconnection, and the domestic market is expected to reach $5 billion by 2028. Currently, Ethernet solutions are becoming the main technical direction, with domestic solutions already being implemented alongside the trend of domestic production of computing chips; the leading companies currently have revenues of about $1 billion, with significant growth potential. It is recommended to focus on domestic Ethernet interconnection chip manufacturers; additionally, domestic substitution of optical interconnect CPO/NPO links based on interconnection chips is also worth attention.