Pandaily reported that a Huawei Ascend 950DT SuperPoD with 1,024 NPU accelerator cards was switched on in Wuhan on 28 October at a Huawei artificial intelligence conference for Hubei province, part of the company’s 2026 China roadshow. Hubei Science and Technology Investment Group, a provincial state-backed investor, funded the cluster and said it took one month from kickoff to power-on.
The cluster’s compute and interconnect cabinets are liquid-cooled. Its reported specifications are 1 EFLOPS of FP8 compute, 256TB of globally unified memory and card-to-card bandwidth at the terabyte-per-second level. Huawei’s UnifiedBus interconnect lets the cards access one another directly, so data can move across the pod without treating each cabinet as a separate machine.
Answering a work question could draw on an AI agent running on the cluster, if it is used for the agent inference Huawei envisages.
Huawei also positions the design for training models with trillions of parameters. The 1 EFLOPS figure was given at the launch event, and Huawei has not published the cluster’s power draw or detailed utilisation figures, Pandaily reported.
