DeepSeek and Huawei Expand Ascend Support Across AI Software Stack
New open-source tools cover operator development, matrix computation, cross-chip communication, Attention and data selection as the companies reduce reliance on Nvidia’s CUDA ecosystem.
DeepSeek and Huawei are moving beyond model and hardware compatibility by adapting core AI software to Ascend chips, a shift that could reduce reliance on Nvidia’s CUDA ecosystem and reshape how Chinese models are developed.
The companies have released open-source tools covering operator development, matrix computation, cross-chip communication, Attention and data selection. The software supports the underlying work needed to train and run AI models on Huawei hardware.
TileLang, which has been used to implement many operators in DeepSeek V4 training, is also being adapted for Ascend. The tool addresses how model computations are executed efficiently on chips, extending the effort from model-level compatibility to the software layer beneath it.
DeepSeek V4 has begun adapting to Huawei’s Ascend 950. Huawei has said its chips were used in part of the model’s training, while Liang Wenfeng has made increasing training on Chinese-made chips a priority for DeepSeek.
The broader effort includes matrix calculations, MoE communication, Attention and operator development. Together, those projects connect model software with the chips, communications systems and training tools required to run it.
For US technology companies, the development highlights a competition that extends beyond individual AI models or processors. The emerging contest also concerns whether developers can build a complete software stack around alternative chips, reducing dependence on established ecosystems such as CUDA.
