One-sentence summary
Compare DeepSeek V5's planned Ascend-based training route with the mainstream Nvidia CUDA training stack: Ascend has mainly been validated on inference, making ecosystem migration cost the key variable, while CUDA offers maturity but faces supply and cost constraints.
Overview
CLS reported on November 3 that DeepSeek V5 is planned to be trained on Ascend chips; mainstream large-model training still relies on Nvidia GPUs and the CUDA ecosystem.
Core Comparison
| model | Role | Ecosystem maturity | Main constraint |
|---|---|---|---|
| Ascend training route | Planned for DeepSeek V5 training | Mainly inference so far; training needs migration | Framework, operator, cluster scheduling migration cost |
| Nvidia CUDA training stack | Mainstream training and inference | Highly mature | Tight supply, cost and geopolitical constraints |
Key Differences
An Ascend route, if realized, would validate "national chips training national models"; the CUDA stack represents current maximum maturity but faces supply and external constraints. The trade-off is ecosystem maturity versus autonomy.