One-sentence summary
CLS reported on November 3, citing industry sources, that DeepSeek plans to train its next model V5 on Huawei Ascend chips. If realized, it would be a key step from running inference on domestic chips to training national models with them.
Background
On November 3, 2026, CLS reported that DeepSeek V5 is planned to be trained on Ascend chips, according to multiple supply-chain sources.
What Moving from Inference to Training Means
Ascend chips previously served DeepSeek's inference workloads. Training places far higher demands on cluster interconnect, framework adaptation and operator optimization; adoption would signal that the Ascend ecosystem can handle large-scale pretraining.
Industry Signals
- Large-scale validation on the inference side has boosted confidence in domestic chips.
- If realized, "national chips train national models" would drive upgrades across domestic training clusters, frameworks and tooling.
The Ceiling Variable
Ecosystem migration costs on the training side (framework migration, operator development, cluster scheduling and stability) remain the key variable limiting the market ceiling.