DeepSeek Engineers Engage in Huawei’s Chip Design, Fully Removing NVIDIA Dependence in Training
5 hour ago / Read about 0 minute
Author:小编   

DeepSeek is forging a highly collaborative partnership with Huawei to co-design next-generation AI chips, transcending the conventional supplier-customer dynamic. Huawei has stationed engineers at DeepSeek’s facilities to embed large-scale model training requirements directly into chip architecture development. These integrated designs will underpin chips slated for release in 2027 and subsequent years. By year-end, DeepSeek intends to deploy Huawei’s Ascend 950DT processors for pre-training workloads, while Huawei will prioritize the majority of its production capacity for the upcoming Ascend 960DT—set to launch in 2027—exclusively for DeepSeek’s use. Furthermore, DeepSeek has open-sourced TileLang, a compilation toolkit tailored for the Ascend platform, to diminish reliance on NVIDIA’s proprietary CUDA ecosystem and foster greater technological autonomy.