DeepSeek open-sources Huawei Ascend programming toolkit to chip away at Nvidia's CUDA lead

The open-source release and its software components

Hangzhou DeepSeek Artificial Intelligence Co., Ltd. said on September 30 that it has publicly released a software toolkit developed with Huawei Technologies for programming AI chips, highlighting its collaboration with the Chinese chipmaker as it works on technologies aimed at replacing Nvidia. The release was announced via DeepSeek's WeChat account and packages together a collection of compute and communication libraries, including TileLang, DeepGEMM, FlashMLA, TileKernel, DeepSelect, and DeepEP, each addressing a different layer of AI development from matrix multiplication to inter-chip communication.

Optimization for Ascend silicon and a 128-chip configuration

The toolkit is optimized for Huawei's Ascend 950 chips and supports a supernode configuration linking 128 of those accelerators together. DeepSeek said Huawei provided extensive support throughout the development process, and the tools are available as free downloads in a deliberate move to lower the barrier for developers who might otherwise default to Nvidia's established ecosystem because switching costs felt too high. Benchmarking tools included in the release assist developers with kernel development, providing standardized ways to measure performance and optimize code for the Ascend architecture.

Positioning against Nvidia's CUDA ecosystem

The release is framed as a direct response to Nvidia's CUDA software framework, which has long served as the dominant programming layer for AI chips worldwide and which DeepSeek's TileLang is described as a domestic counterpart to. TileLang is presented as a high-level abstraction layer similar to what CUDA provides, meaning programmers familiar with Nvidia's workflow can transition to Huawei hardware without learning an entirely alien system. DeepSeek has previously shown that competitive models can run on Huawei hardware, having adapted its V4 AI model to Ascend silicon in April 2026, and the new toolkit generalizes that effort by giving third-party developers the means to do the same with their own models.

Infrastructure plans and adoption questions

Separately, DeepSeek has plans to construct a large data center in Inner Mongolia designed to host at least 160,000 Huawei Ascend accelerators, signaling that the software release is intended to feed into a larger build-out of domestic AI capacity. Newsquawk's analysis draws a distinction between software compatibility announcements, which it describes as cheap, and evidence of actual performance parity, which it says has historically arrived only through later benchmark disclosures and uptake by third-party developers. The same analysis notes that progress on domestic chip self-sufficiency has in past episodes lifted the Chinese semiconductor and AI complex while pressuring the narrative premium of foreign suppliers exposed to China revenue, though it stops short of attributing any specific market move to this release.

Significance and what remains to be verified

The pairing of DeepSeek and Huawei's Ascend line is being watched as a test of whether domestic Chinese silicon can substitute for restricted Nvidia hardware at scale for training and inference workloads, rather than as a novelty announcement. Unresolved questions include whether quantified performance claims will follow, whether uptake expands outside Huawei's own ecosystem, and whether the release prompts a policy response from the jurisdictions administering advanced-chip export controls. Open follow-ons include any third-party benchmark disclosures comparing TileLang and the accompanying libraries against CUDA on equivalent workloads.