At Huawei Connect 2026 in Shanghai on Sept. 17, Huawei vice chairman Wang Tao said Ascend 910C supernodes have exceeded 1,000 deployments and Ascend 950 supernodes are in large-scale commercial use. He described supernodes as tightly interconnected multi-node systems with unified memory addressing that behave logically as a single computer and said they are becoming the default architecture for hyperscale AI infrastructure. Wang said 100,000-card clusters are now standard for training ~10-trilli

At Huawei Connect 2026 in Shanghai on Sept. 17, Huawei vice chairman Wang Tao said Ascend 910C supernodes have exceeded 1,000 deployments and Ascend 950 supernodes are in large-scale commercial use. He described supernodes as tightly interconnected multi-node systems with unified memory addressing that behave logically as a single computer and said they are becoming the default architecture for hyperscale AI infrastructure. Wang said 100,000-card clusters are now standard for training ~10-trillion-parameter SOTA models, but traditional server architectures push intra-cluster communication above 40% of training time, constraining MFU (model floating-point utilization). Huawei’s Markov Lab simulation shows a 100,000-card cluster composed of 4K supernodes delivers a 2.75x MFU improvement versus a 100,000-card cluster built from 8-card servers.