DeepSeek plans to use Huawei Ascend 950 chips to train its next-generation model, DeepSeek V5, with the project expected to proceed next year, according to people familiar with the supply chain. According to Odaily, DeepSeek’s team and Ascend’s core research and development team are based in Hangzhou, where they completed extensive low-level operator adaptation and custom development for V4. Ascend chips were used only for inference in the previous DeepSeek model; V5 is planned as the first model to be trained natively on Ascend 950.
