DeepSeek has open-sourced an AI chip programming toolkit developed alongside Huawei, featuring the high-level language TileLang to program Huawei Ascend accelerators.
TileLang and the Push for an Independent Ecosystem
The open-source release centers on a high-level programming language called TileLang, alongside related compute and communication libraries announced in a post on the company’s official WeChat account. This infrastructure serves as China’s answer to Nvidia’s CUDA software platform. Nvidia’s framework has served as the global standard for developing artificial intelligence for nearly two decades, leaving Chinese developers with limited alternatives as U.S. export restrictions keep the newest American processors out of reach.
DeepSeek designed TileLang to simplify coding logic and write the core routines that optimize model performance. According to the company, the language provides a sophisticated programming language with a simple coding process that helps developers extract full hardware performance without the steep complexity of legacy systems.
Hardware Integration with Huawei Ascend Processors
These components adapt previous open-source infrastructure originally built for Nvidia hardware so they run smoothly on Huawei’s Ascend accelerators.
The collaboration also produced a specialized super-node system. Operating across 128 Huawei Ascend 950 processors, the architecture is calibrated to balance heavy computing tasks with high-speed data transfers across the entire server cluster.
Deepening Ties Amid U.S. Export Restrictions
This software rollout follows a broader alignment between two of China’s most prominent technology firms.
The technical partnership arrives shortly after Huawei unveiled its latest generation of artificial intelligence processors and super-node systems. Huawei previously announced plans to ship its next-generation Ascend 960DT chip in the first quarter of 2027 while mapping out a hardware roadmap extending through 2029.
Data Center Deployments and Commercial Momentum
While global model training still leans heavily on foreign semiconductors, DeepSeek continues to integrate domestic silicon into its operations. The startup plans to deploy at least 160,000 powerful Huawei accelerators inside a massive data center under construction in Inner Mongolia.
At the same time, the artificial intelligence lab is finalizing a fundraising round at a valuation close to $74 billion ahead of a potential public listing.
