Huawei Says Ascend R&D Spending Exceeds Wireless Division, Sees Trillion-Yuan Training Compute Demand

At the Ascend Industry Summit during Huawei Connect 2026 on September 17, Zhang Xiwei, Huawei's vice president and head of its computing product line, said Huawei's R&D spending on Ascend in 2026 and 2027 has exceeded that of the company's wireless division, and that this level of investment is expected to ensure stable output of the Ascend 960, Ascend 970 and Ascend 980.

Zhang said that in 2025 the industry still had concerns about whether products following the Ascend 910BC would arrive on time and whether supply would remain stable. Over the following year, Huawei continued investing in the Ascend 950 and next-generation products, and it now sees more stable product delivery and supply. He added that Ascend has become a central strategic focus for Huawei, with plans to build end-to-end solutions spanning supernodes, servers, edge devices, all-in-one machines, modules and components.

Zhang Dixuan, president of Huawei's Ascend product line, said Huawei unified its interconnect protocol in 2026 and built a supernode architecture centered on the Ascend NPU. More than 1,000 Ascend supernodes have been deployed to date, accumulating over 650 million online card-hours, and more than 40 large models have completed pre-training on Ascend.

Zhang Xiwei also described market opportunities. Around 10 leading private and internet companies in China continue to invest in large language model training. If each company's future demand reaches about 200,000 cards, the combined requirement would be about 2 million cards, corresponding to training compute investment of hundreds of billions to one trillion yuan or more. He also cited AI application commercialization, industry-wide intelligent transformation, L3 autonomous driving, embodied intelligence and robotics as opportunities.