# Huawei Says Ascend Leads Nvidia in China as Training Gap Persists By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/30689 Published: 2026-10-13T01:16:37.000Z Updated: 2026-10-13T01:16:37.000Z Section: Technology > Huawei cited internal data for its claim, while a proposed DeepSeek deployment would use 160,000 Ascend accelerators for inference rather than training. Huawei says its Ascend artificial-intelligence chips have overtaken Nvidia in China based on internal data, but Nvidia retains an advantage in training software as Chinese companies expand domestic inference capacity. Huawei rotating chairman Eric Xu said Ascend had surpassed Nvidia in China. Huawei did not provide a market-share percentage, measurement period or explanation of whether the comparison used revenue or unit shipments. Nvidia’s CUDA software ecosystem remains deeply embedded in AI development and training. That makes training workloads harder to migrate than inference, which involves using completed models to generate results. Huawei’s roadmap assigns different workloads to the Ascend 950PR and 950DT. The 950PR is designed for inference prefill and recommendation systems, while the 950DT supports inference decode and model training. The 950DT is scheduled to become available in the fourth quarter of 2026. Huawei also plans an Atlas 950 SuperPoD for the same quarter. The system could contain up to 8,192 Ascend 950DT chips and deliver 8 exaflops in FP8, 16 exaflops in FP4 and 16 petabytes per second of interconnect bandwidth, based on Huawei’s specifications. Those figures do not establish an independent comparison with Nvidia systems. “We’ll keep evolving Ascend chips to strengthen the foundation of AI computing power, both in China and around the world,” Xu said. Huawei said it had deployed more than 300 Atlas 900 A3 SuperPoDs by September 2025 for more than 20 customers. A proposed DeepSeek deployment would involve at least 160,000 Ascend 950DT accelerators at a data center in Ulanqab, Inner Mongolia. The proposed deployment would use the chips for inference rather than training, but DeepSeek and Huawei have not publicly confirmed the order. The proposed facility is designed for roughly 1 gigawatt of power, with part of its capacity targeted for operation by late 2027 or early 2028. Nvidia’s share of China’s AI-chip market was estimated at about 40% in 2025 and could fall to around 8% in 2026, while Huawei’s share could reach about 50%. Those figures are projections, not official market data. U.S. export controls have restricted Chinese access to advanced Nvidia GPUs, while enforcement actions have involved alleged diversion and smuggling. The U.S. Department of Justice said an indictment alleged that 400 Nvidia A100 GPUs were exported to China between October 2024 and January 2025. The charges are allegations, and the defendants are presumed innocent. Earlier reporting examined [Nvidia and AMD’s efforts to retain access to China’s AI-chip market](). ## Links in this article - [Nvidia and AMD’s efforts to retain access to China’s AI-chip market](https://www.tokenpost.com/news/regulation/25243)