Industry sources said the next-generation DeepSeek V5 is planned to be pre-trained on Huawei Ascend chips, moving from the inference side to the training side and running pre-training on a domestic super-node for the first time, AI-industry trackers reported on November 4. The step signals that "domestic chips training domestic models" has entered a substantive stage.
The plan follows a series of Chinese open-source model releases and reflects a strategic push to keep large-model training inside an autonomous compute stack. Earlier DeepSeek models had already shown strong inference performance on domestic hardware; extending to pre-training would close a key gap in the domestic AI supply chain.
If realized, the approach would reduce reliance on foreign accelerators for the most compute-intensive phase of model development, reinforcing the country's goal of a self-controlled, secure and trustworthy technology base for artificial intelligence.
Source: Industry Sources / Technology Media
