Overview

• DeepSeek halted its R2 training after multiple failures on Huawei’s Ascend 910C processors despite on-site support, postponing the model’s planned release.
• The company reverted to Nvidia H20 GPUs for large-scale training while continuing to deploy Ascend accelerators exclusively for inference tasks.
• Founder Liang Wenfeng has voiced frustration over performance shortfalls and extended data-labeling and retraining efforts to meet the company’s benchmarks.
• The setback exposes persistent stability and software maturity gaps in China’s domestic AI chip ecosystem under Beijing’s self-sufficiency drive.
• The delay opens opportunities for competitors like Alibaba’s Qwen3 and comes as U.S. export concessions could restore access to advanced Nvidia GPUs.