Overview

• Multiple reports say DeepSeek could not complete a single R2 training run on Huawei’s Ascend 910C chips despite on-site help from Huawei engineers.
• The company has resumed training on Nvidia H20 GPUs after U.S. licensing allowed H20 sales to China, while retaining Ascend hardware for inference.
• R2, initially slated for May 2025, was postponed as DeepSeek grappled with Ascend-related stability and software issues and expanded data labeling for the new dataset.
• Reports from SCMP indicate state-funded AI data centers in China must source more than 50% of chips from domestic vendors, creating friction because CUDA-trained models do not cleanly port to Huawei’s CANN stack.
• Operators cite persistent gaps in stability, interconnect performance and software maturity versus Nvidia’s ecosystem, even as firms such as Cambricon plan to raise about 4 billion yuan to capture mandated demand.