DeepSeek plans to transition its computing infrastructure to Huawei Ascend AI accelerators by 2027–2028, significantly reducing reliance on Nvidia hardware. This shift is supported by the adaptation of DeepSeek’s software stack, including the TileLang programming language, to work with Huawei’s architecture.
DeepSeek provides Huawei with early access to its AI models, enabling day-one inference support for updates like V4 and V4.1. While SemiAnalysis notes that the Ascend 950PR accelerators, released in March, trail Nvidia H100s in raw performance, Huawei’s software optimization achieves 99.7% of theoretical efficiency.
Nvidia CEO Jensen Huang previously identified the optimization of Chinese AI models for Huawei chips as a major competitive risk for the United States. Analysts suggest that while DeepSeek’s specific server configuration is unique, its collaborative approach with Huawei could serve as a model for the broader Chinese AI industry.
