Huawei is deploying a 256K 950 SuperCluster. it is pointless to amass a quarter million 950PRs in a cluster. Wenfeng said before that 200K 950DTs is the floor for training a 800B param active model, the absolute US frontier. Presumably he meant the Navier-Stokes one.

Huawei has announced its Peerium Computing Architecture, enabling processors at the million scale to work as one computer. Huawei Deputy Chairman and Rotating Chairman Eric Xu said the company is building SuperPoDs and SuperPoD-based SuperClusters for AI training and inference.

Huawei announced its Peerium Computing Architecture, a new architecture for the AI era that enables processors at the million scale to work as one computer. Powered by UnifiedBus (UB), it supports strong scaling to meet growing demand for AI compute.

Huawei debuts Hi-ONE Near‑Packaged Optics interconnect; Wang Tao says industry’s first mass‑produced NPO At #Huawei# Connect 2026 in Shanghai, Huawei unveiled Hi-ONE, a new-generation Near-Packaged Optics (NPO) optical interconnect. Vice chairman and rotating chairman Wang Tao said the product is designed to enable multi‑NPU coordination via high‑speed interconnect. Huawei said Hi-ONE, built on its proprietary technologies with a balanced optics/mechanics/electronics/magnetics/thermal design, delivers 7.2T transmission capacity per engine. Huawei claims Hi-ONE is the industry’s first mass‑produced NPO, the current market leader in NPO transmission capacity, and the only NPO to integrate a built‑in light source. (

Huawei says it will launch its Ascend 960DT in Q1 2027, followed by the Ascend 960PR in Q3. The company has shipped 1,000+ supernode systems to 370+ customers and says its supercluster architecture can scale to 1 million NPUs. Huawei also says it has developed 11 key chips around its Unified Bus interconnect