Chinese AI developer DeepSeek said it is working with Huawei on programming tools for Huawei's Ascend AI chips and is open-sourcing Ascend support for the TileLang language, Reuters reported late Tuesday in a story carried by Investing.com. DeepSeek framed the move as a step toward a domestic GPU software ecosystem that does not depend on Nvidia's CUDA.

What Happened

DeepSeek made the announcement on its official WeChat account. The core of the release is support for TileLang on Ascend hardware, along with compute and communication libraries for the platform. DeepSeek said TileLang offers a simpler programming model than CUDA and improves development efficiency and code clarity.

TileLang is an open-source, tile-based programming model originally developed by Microsoft researchers, SDxCentral reported, and works in a way similar to Nvidia's CUDA Tile. Tile-based languages let developers describe computations on blocks of data, leaving the compiler to map that work onto the chip's cores and memory, which can cut the amount of hand-tuned low-level code needed for fast AI kernels.

RelatedIran Security Chief Threatens Cables, Data Centers and AI Sites if Fighting Resumes

The work centers on the Ascend 950, the successor to Huawei's Ascend 910C, according to SDxCentral, which also said the two companies had worked on a "supernode" configuration linking 128 Ascend 950 chips. DeepSeek said its priority is a high-level language that is general, easy to program and still able to reach the hardware's full performance.

Why It Matters

Nvidia's lead in AI hardware rests heavily on software. CUDA and the libraries built on it have had well over a decade of optimization, and most AI frameworks and research code assume them. A rival chip can match Nvidia on paper and still lose if developers must rewrite and retune their code to use it.

Chinese labs face that problem more sharply than most. U.S. export controls restrict their access to Nvidia's most advanced GPUs, pushing them toward domestic accelerators such as Ascend. SDxCentral noted that Huawei's own CANN software stack has been criticized as difficult and unstable to use. If a lab with DeepSeek's profile builds and maintains a practical high-level path onto Ascend, other Chinese developers gain a tested toolchain rather than having to build their own.

The communication libraries matter for large deployments. Training and serving frontier-scale models means spreading work across many chips, and the speed at which those chips exchange data often limits overall performance as much as raw compute does.

Context

DeepSeek has built its reputation on getting strong results from constrained hardware through software optimization. Reuters described the Huawei partnership as part of a wider push by Chinese technology companies to develop alternatives to the Nvidia ecosystem amid U.S. export restrictions.

What to Watch

The key signal will be adoption: whether developers outside DeepSeek begin writing TileLang kernels for Ascend and whether major Chinese frameworks support it. Performance comparisons against CANN and against CUDA code on Nvidia hardware will show how much of the gap the tools close. Another indicator is whether DeepSeek says it trains or serves future models on Ascend 950 clusters at scale.