China to see major shift to Huawei for AI model training in 2027: rotating chair

Huawei Technologies anticipates a substantial shift towards its AI computing infrastructure for model training next year, as its Ascend processors challenge Nvidia's dominance in China, according to rotating chairman Eric Xu Zhijun. During the Huawei Connect 2026 conference in Shanghai, Xu stated that a significant portion of AI model training will rely on SuperPoD or SuperCluster systems based on Ascend 950DT.

While Huawei's proprietary Ascend chips have surpassed Nvidia in domestic market share, specific figures were not disclosed. Last month, Nvidia announced the start of shipments of its H200 processors to China under a new US licensing arrangement, though such sales accounted for less than 1% of its second-quarter data-center revenue.

Huawei has been promoting its alternative computing infrastructure due to US export controls aimed at limiting China's access to advanced semiconductors. Despite the Ascend chips' advantages, international availability remains limited. The Shenzhen-based company unveiled its Atlas 960 SuperPoD computing cluster, powered by Ascend 960 chips, which boasts double the training and inference performance of its predecessor.

However, production constraints mean most systems will remain within China. Huawei has already deployed over 1,000 Atlas SuperPoD units across more than 370 enterprise customers. The company's Peerium Computing Architecture, designed to connect 1 million processors as a single computer, is being tested on the new Atlas 960 hardware.

While Huawei projects that global AI computing supply and demand will balance around 2029, with China reaching equilibrium slightly later in 2030, computing supply still trails China's surging AI demand.