Huawei's Ascend 950 Super Node Enters Commercial Use, 960 to Launch in 2027

At Huawei Connect held in Shanghai, Wang Tao, Huawei's Deputy Chairman and Rotating Chairman, presented Huawei's AI strategy for the intelligent era. Huawei's AI strategy centers on computing power, insisting on hardware monetization, building a computing foundation through "super nodes + clusters," and creating an open-source computing ecosystem that supports native training of mainstream large models.

Wang Tao revealed that Ascend super nodes have been deployed over 1,000 sets, and the Ascend 950 super node has entered large-scale commercial use. The Ascend 960 super node will be the industry's first to adopt NPO. The Atlas 960 liquid-cooled super node is planned for Q3 2027, while the Atlas 860 air-cooled super node will launch in Q2 2027.

Huawei previously announced the Ascend 950 chip architecture, which adds support for low-precision data formats, boosts vector computing power, increases interconnect bandwidth by 2.5 times, and supports Huawei's self-developed HBM. The 950PR improves inference Prefill performance with HiBL 1.0 HBM, while the 950DT enhances inference Decode and training performance.