DeepSeek and Huawei develop open-source tools to challenge Nvidia’s dominance in AI chip ecosystem

DeepSeek partners with Huawei to create open-source programming infrastructure for Ascend chips, aiming to reduce dependence on Nvidia and foster a local AI software ecosystem in China.

DeepSeek has said it is working with Huawei on programming tools for the Chinese group’s Ascend chips, in a move that underlines how domestic AI companies are trying to build a software stack less dependent on Nvidia. According to the company’s WeChat post, the project includes open-sourcing infrastructure for Huawei’s Ascend platform, covering compute and communication libraries.

The announcement came shortly after Huawei introduced a new generation of AI processors and so-called supernode computing systems, which the company said it expects to be used widely for model training next year. DeepSeek said Huawei gave full support in developing the software layer and that the two firms jointly advanced a supernode design based on 128 Ascend 950 chips, with an emphasis on improving both computation and data transfer.

DeepSeek also pointed to TileLang, an open-source high-level programming language for AI chips, as central to the effort. The company said the aim is to make development simpler while still pushing hardware to its performance limits. In DeepSeek’s framing, TileLang offers a more straightforward programming model than Nvidia’s CUDA, which has long been the dominant environment for GPU software development.

The collaboration fits a broader pattern in China’s AI sector, where access to Nvidia hardware and software has become a strategic concern. By open-sourcing components that mirror its earlier work on Nvidia systems, DeepSeek is signalling an attempt to create a parallel ecosystem around Ascend. That would not remove Nvidia’s technical lead overnight, but it would give Chinese developers more tools to train and run models on local chips, and reduce reliance on foreign infrastructure over time.