DeepSeek Partners With Huawei to Develop Custom AI Chips

DeepSeek is collaborating with Huawei Technologies to design AI chips tailored to its specific models in an effort to reduce reliance on Nvidia. Huawei engineers are currently working at DeepSeek offices to integrate feedback into chip designs, focusing on data transfer between processors and calculation performance.

DeepSeek plans to begin training models using Huawei Ascend 950DT chips by the end of 2026, with priority access to next-generation Ascend 960DT chips scheduled for early 2027. To decouple from the Nvidia CUDA ecosystem, the startup is using TileLang, an open-source programming language, and recently released a version compatible with Ascend 950 chips.

Despite these efforts, DeepSeek continues to seek access to advanced Nvidia processors. The partnership aligns with Beijing's broader goal of establishing an AI industry independent of United States technology, though significant technical gaps remain. Epoch AI reports that Nvidia B300 chips provide approximately seven times the raw arithmetic performance of Huawei Ascend 950 processors.