Deepseek has released open-source programming tools designed for Huawei's Ascend AI chips. The project includes computation libraries and data transfer tools, with full support from Huawei. A key component is TileLang, an open-source programming language originally created by Peking University researchers, which Deepseek has utilized for approximately one year.
The companies have also optimized a supernode cluster consisting of 128 Ascend 950 chips. Deepseek intends for TileLang to serve as a simpler alternative to Nvidia's CUDA platform, aiming to provide a universal language that maximizes hardware performance. The language was previously tested by Deepseek on older Nvidia hardware and is now central to its AGI development.
This partnership addresses the critical need for software ecosystems that can fully leverage domestic Chinese hardware. While Nvidia maintains a significant advantage through its CUDA developer base, Huawei is working to close this gap by introducing new AI processors and supernode systems. Huawei plans to prioritize domestic demand for its chips over international sales.
Industry analysts at SemiAnalysis suggest that the software moat surrounding Nvidia is narrowing, particularly as companies like OpenAI develop custom inference chips. However, Nvidia remains ahead in complex agent-based workloads. While Huawei's CANN software stack has demonstrated compatibility with models like DeepSeek V4, Nvidia continues to lead in large-scale system integration.
