DeepSeek presents an alternative to CUDA together with Huawei
DeepSeek and Huawei have joined forces to publicly present their alternative to Nvidia's CUDA ecosystem, releasing the code for TileLang, a high-level programming language adapted to Huawei's Ascend 950 chips, along with other software tools for core AI model tasks (computation and inter-chip communication libraries). The goal is for developers to be able to work with Ascend accelerators without having to rebuild from scratch the software infrastructure needed to train and run models, thereby making it easier to move away from CUDA in Chinese AI projects. In addition, both companies have developed a supernode of 128 Ascend 950 accelerators, and DeepSeek plans to install at least 160,000 of these chips in a huge data center in Inner Mongolia, in an effort to reduce dependence on Nvidia's hardware and software.
TileLang https://github.com/tile-ai/tilelang
- 2026-09-30 — Ascend 950 backend: TileLang now officially supports Huawei Ascend 950 NPUs with native code generation, automatic scheduling and synchronization, SIMD/SIMT vector programming, etc. Explore the Ascend examples for GEMM, FlashAttention, and more.
What do you all think of this new ecosystem that DeepSeek and Huawei want to promote?