Overview

• A Financial Times report on August 14 says DeepSeek’s large-scale training runs on Huawei Ascend chips repeatedly failed due to unstable hardware, slower interconnects and immature software tooling.
• DeepSeek has reverted R2 training to Nvidia H20 GPUs while retaining Huawei’s Ascend accelerators solely for inference tasks.
• The R2 launch, originally expected in May, is now pushed into the coming weeks after extended data labeling and retraining processes.
• Chinese authorities had urged DeepSeek to adopt Huawei hardware to bolster national chip self-reliance, highlighting the influence of state policy on AI infrastructure choices.
• Industry experts say the setbacks underscore Nvidia’s entrenched software ecosystem advantages and the current performance gap in China’s domestic AI accelerators.