On September 17, Huawei Connect 2026 kicked off in Shanghai. Wang Tao, Deputy Chairman and Rotating CEO of Huawei, unveiled that over 1,000 sets of Ascend 910C super nodes have already been deployed. Moreover, the Ascend 950 super nodes have also been put into large-scale commercial use, emerging as the top-tier option for constructing hyperscale AI infrastructure. Wang Tao elaborated that super nodes are interconnected by multiple computing nodes through efficient interconnection protocols. They boast the ability to perform unified memory addressing across physical nodes, functioning logically akin to "a single computer." This concept has now reached a consensus within the AI infrastructure field. He further highlighted that the current training of SOTA (State-of-the-Art) models at the 10-trillion-parameter scale typically employs 100,000-card clusters. However, traditional server architectures lead to communication time within the cluster accounting for over 40% of the total training time, substantially constraining MFU (Model Floating-Point Utilization). Simulation experiments carried out by Huawei's Markov Lab demonstrate that a 100,000-card cluster constructed with 4K super nodes enhances MFU by 2.75 times compared to an 8-card server cluster.