【#Tech24H】With the rapid development of large AI models, computing power demand has surged, and traditional server clusters are facing pain points such as complex internal cabling, high failure rates, large communication latency, and limited expansion density. The emergence of “supernodes” enables efficient collaboration among different chips, specifically addressing data “overloaded” issues in ultra-large models and massive concurrent multi-agent computations. At the World Artificial Intelligence Conference (WAIC 2026) held from July 17 to 20, over 1,100 companies exhibited, with more than 200 in the computing power sector alone. Many hardware powerhouses have moved beyond showcasing single chips to displaying entire racks and clusters. Leading manufacturers such as Huawei, Sugon, Moore Threads, and Enflame all unveiled 1,024-card supernodes, 256-GPU single-layer fully interconnected networks, and 100,000-card clusters. Among them, the much-anticipated Huawei Ascend 950 supernode made its first public appearance in physical form. Huawei calls it the “industry’s largest-scale supernode”, comprising 16 computing cabinets and 1,024 of the latest Ascend 950DT chips. [ By Zhang Liyan | Tang Ruohan ]