A reported plan to pretrain DeepSeek V5 on Huawei Ascend chips would move the processors beyond their main inference role, 科创板日报 said Nov. 3, citing multiple chip supply chain sources.
Pretraining builds a model; inference uses a trained model to respond to requests. DeepSeek has mainly deployed Ascend chips for inference. Using them to pretrain V5 would give the processors a different role in the company's model development.
DeepSeek expanded its public Ascend software tools Sept. 30, releasing computing and distributed communication components alongside a TileLang compiler tool. The company had previously made comparable components available for Nvidia hardware.
DeepSeek has attempted training on Huawei hardware before. The Financial Times reported in 2025 that technical problems during work on its R2 model led the company to switch to Nvidia chips. Industry insiders cited by the newspaper separately described stability issues, slower links between processors and weaker software across Chinese chips more broadly.
