DeepSeek Bets on Huawei, Liang Wenfeng: The Success of the Compute Migration Strategy is Crucial

According to a report by Jinse Finance, citing Hong Kong Yahoo Finance, DeepSeek founder Liang Wenfeng stated publicly that the company will prioritize using Huawei Ascend chips to train AI large models, advancing the "silicon migration" compute power transition by shifting the model training foundation from NVIDIA's CUDA ecosystem to Huawei's CANN system. He emphasized that this strategy must succeed. The report mentioned that Huawei's next-generation training chips are expected to be delivered between Q4 2026 and Q1 2027, and DeepSeek plans to rely on this batch of domestic compute power to carry out large model training. The entire migration process involves extensive rewriting of underlying code and entails significant R&D costs. This is one of DeepSeek's most critical strategic bets, aimed at mitigating the compute power risks posed by overseas high-end GPU export restrictions and building a domestic compute power foundation. Liang Wenfeng believes that despite the immense challenges of migration, ensuring long-term autonomy and control over compute power makes this path essential. Industry analysts pointed out that if DeepSeek successfully completes large-scale model training on Huawei chips, it will become a landmark case of software-hardware collaboration in China's large model industry. However, the company must also contend with multiple uncertainties, including ecosystem adaptation, performance optimization, and chip delivery timelines.