On September 17, the 2026 Huawei Connectivity Conference opened in Shanghai. Wang Tao, vice chairman and rotating chairman of Huawei, said at the meeting that up to now, more than 1,000 units of Shengteng 910C supernodes have been deployed, and Shengteng 950 supernodes have also been commercialized on a large scale. While being commercialized on a large scale, supernodes have become an inevitable choice for building hyperscale AI infrastructure. Wang Tao explained that a supernode is a computing system that is physically composed of multiple computing nodes closely connected through an efficient interconnection protocol, has the ability to unify memory addressing across physical nodes, and logically has the characteristics of “one computer”. It has become a consensus in the field of AI infrastructure. Wang Tao mentioned that currently, the 100,000 card cluster has become the standard for training 10 trillion SOTA models, but traditional server architectures cause intra-cluster communication to exceed 40% of the training time, seriously limiting MFU. According to the simulation results of the Huawei Markov Laboratory, the MFU of the 100,000 card cluster composed of 4K supernodes increased 2.75 times compared to the 100,000 card cluster composed of 8-card servers.

On September 17, the 2026 Huawei Connectivity Conference opened in Shanghai. Wang Tao, vice chairman and rotating chairman of Huawei, said at the meeting that up to now, more than 1,000 units of Shengteng 910C supernodes have been deployed, and Shengteng 950 supernodes have also been commercialized on a large scale. While being commercialized on a large scale, supernodes have become an inevitable choice for building hyperscale AI infrastructure. Wang Tao explained that a supernode is a computing system that is physically composed of multiple computing nodes closely connected through an efficient interconnection protocol, has the ability to unify memory addressing across physical nodes, and logically has the characteristics of “one computer”. It has become a consensus in the field of AI infrastructure. Wang Tao mentioned that currently, the 100,000 card cluster has become the standard for training 10 trillion SOTA models, but traditional server architectures cause intra-cluster communication to exceed 40% of the training time, seriously limiting MFU. According to the simulation results of the Huawei Markov Laboratory, the MFU of the 100,000 card cluster composed of 4K supernodes increased 2.75 times compared to the 100,000 card cluster composed of 8-card servers.