The computing cluster introduced at Huawei Connect 2026 can scale to 4,096 NPUs using near-package optics. This hardware, designated as the Atlas 960 SuperPoD, achieves 8 EFLOPS at FP8 precision to provide a domestic alternative for AI training and inference. The system competes with high-end GPU clusters produced by Nvidia to handle massive AI workloads.

Huawei officials outlined performance targets for the Ascend 960 series for 2027 during the event. The rollout of the SuperPoD cluster is part of a strategic effort to increase the amount of compute power available to Chinese firms to accelerate the development of large scale AI models.