At Huawei Connect 2026, opened in Shanghai on September 17, Huawei deputy chairman and rotating chairman Wang Tao released the world's first NPO (near-package optics)-based Ascend 960 supernode, supporting training and inference for 10-trillion-parameter-class large models. The node scales to 4,096 cards with 8 EFLOPS FP8 compute and 99.8% availability; multiple supernodes can link into clusters of up to 1 million Ascend cards.
Wang said the existing Ascend 910C supernode has already been deployed in over 1,000 sets serving more than 370 customers, and the Ascend 950 is in scaled commercial use; the next-gen 960DT is pulled forward to Q1 2027 with double the performance. Huawei VP Zhang Xiwei revealed Ascend R&D investment has exceeded that of the wireless division, with domestic top-10 model-training compute demand around 2 million cards—implying a total investment of hundreds of billions to over a trillion yuan.
Analysts said the early disclosure reflects Huawei accelerating next-generation AI compute iteration amid intensifying global competition—a key node for China's domestic AI compute industry.
Source: Huawei / National Business Daily
