Huawei's Ascend 960 lands early, taking direct aim at Nvidia

Huawei built its next AI supercluster in about a year, much faster than its usual pace. At its yearly Connect conference in Shanghai, it showed the Atlas 960 SuperPoD. It replaces the Atlas 950, which launched in late 2025, and Huawei says it's better at both AI training and inference. Rotating chairman Wang Tao said the Ascend 960 chip was ready ahead of schedule and has twice the performance of current products.

The roadmap: the Ascend 960DT ships in Q1 2027 and the 960PR in Q3 2027. The Ascend 970 and 980 follow in 2028 and 2029. Huawei also showed Peerium, a new architecture meant to eventually link up to 1 million processors into one system. That's Huawei's answer to Nvidia's NVLink-scale clusters. Its single chips are still weaker, so it's betting on scale instead.

The timing is pointed. The launch lands less than two weeks before Xi Jinping's planned visit to Washington. Meanwhile, US export controls still keep Nvidia's best chips out of China. Chinese AI labs are moving to local chips faster, even though some training still runs on Nvidia hardware.

Why it matters: if Huawei can ship a new generation every year, US export controls buy less time, and Nvidia's hold on China keeps slipping.