On September 8th, a research report released by CICC highlighted that the scale of large model parameters, the capability for long context processing, and the adoption of parallel computing strategies have substantially boosted the demand for inter-card communication. Presently, the bottleneck in AI infrastructure is transitioning from single-chip computational power to interconnectivity efficiency, resource scheduling, and system-wide collaboration. Scale-up technology adeptly consolidates computational power and HBM resources by broadening the reach of high-bandwidth, low-latency interconnections, thereby mitigating the limitations imposed by communication bottlenecks on training and inference efficiency. Ultra-nodes are poised to emerge as a pivotal form of AI computational power organization, propelling demand growth and technological advancements in sectors such as switching chips, switching trays, optical interconnections, liquid cooling, and high-power power supplies. CICC forecasts that the domestic market size for ultra-node liquid cooling support will surge from RMB 1.4 billion in 2026 to RMB 14.8 billion by 2028. Furthermore, the heightened synchronicity and burstiness of AI workloads will spur demand for high-power PSUs, PowerShelves, PDUs, BBUs, and intelligent power management systems, while fostering the evolution of power supply architectures towards efficient solutions like HVDC, centralized AC/DC, and SST.