The integration of domestic large models and domestic computing power has taken a new step forward. According to sources familiar with the supply chain, DeepSeek's next-generation large model, DeepSeek V5, plans to adopt Huawei's Ascend 950 chips as its core training compute, with the project expected to be implemented next year. Core R&D liaison teams from both sides have been stationed in Hangzhou and completed extensive underlying operator adaptation and custom development during the V4 phase. Previously, Ascend computing power handled only inference tasks for the prior generation of DeepSeek; V5 will become the first DeepSeek model to natively rely on Ascend 950 for training. As supply of high-end AI chips remains tight, leading domestic large-model vendors are accelerating their migration to domestic computing foundations. Moving from inference to training is markedly more difficult, as training places far higher demands on operator completeness, cluster interconnection, and long-cycle stability—any shortfall can be amplified into uncertainty in training runs. DeepSeek's recently released V4.1-Flash-1017 outperformed multiple mainstream overseas models in related tests, indicating a clear need for stronger computing power support. DeepSeek has not yet publicly responded to reports of the cooperation.