Chinese AI firm DeepSeek has open-sourced six software tools developed for Huawei's Ascend AI chips, offering Chinese developers a homegrown alternative to Nvidia's software stack at a time when US export controls keep most advanced American processors out of reach.

The Hangzhou-based startup announced the release through its official WeChat account, according to Reuters. The package includes modules for computing and communication workloads, and DeepSeek said the tools are built on open-source components it previously developed for Nvidia hardware, now adapted to run on Huawei's processors. Releasing the code publicly means outside developers can inspect, adapt, and extend it — the mechanism by which a software ecosystem grows beyond the company that created it.

Alongside the software, the two companies co-developed what DeepSeek described as a "supernode" system running on 128 of Huawei's Ascend 950 accelerators, tuned to balance computation and data movement across the cluster — an architecture that ties many chips together so they behave as a single, larger system. According to the South China Morning Post, the release is intended to seed an "independent and controllable" software ecosystem for GPUs rather than deliver a single product.

TileLang takes aim at CUDA

The centerpiece of the release is an Ascend-compatible version of TileLang, DeepSeek's high-level programming language for writing the critical kernels that drive model performance. Nvidia remains the language's main back end, but it now officially supports Huawei's Ascend 950, adding features such as native code generation along with automatic scheduling and synchronization, according to SCMP.

DeepSeek positioned TileLang as an alternative to Nvidia's proprietary CUDA platform, saying it simplifies programming still allowing developers to extract maximum performance from the underlying hardware. The company characterized the release as an early step toward a more independent GPU software ecosystem.

CUDA's dominance is what makes the challenge so difficult. Nvidia's nearly two-decade lead in CUDA tools and libraries has made its chips a core component of AI workloads, and no Chinese rival has so far managed to break that hold. Software ecosystems also tend to entrench themselves: once teams have built models and workflows around a toolchain, moving to an unproven alternative typically means rewriting code and relearning optimization techniques.

US export controls raise the stakes

US export controls restrict Chinese companies' access to Nvidia's most advanced chips, a constraint that has pushed Huawei to promote its Ascend processors more aggressively. The DeepSeek tools arrived roughly two weeks after Huawei unveiled its next-generation AI processors and supernode systems, which the company expects to begin training models on by next year.

A week before that announcement, Huawei said it would ship its next-generation Ascend 960DT chip in the first quarter of 2027, pulling the launch forward by three quarters, and laid out a chip roadmap stretching to 2029.

Rotating Chairman Eric Xu claimed that Huawei's Ascend chips already hold a bigger slice of China's AI chip market than Nvidia does, although he offered no figures to back the claim.

A deepening DeepSeek-Huawei alliance

Wednesday's release builds on a partnership that has grown closer over the year. DeepSeek had already previewed its V4 model running on Huawei's Ascend processors, marking a shift away from its earlier reliance on Nvidia. Huawei said at the time that it had worked with DeepSeek to make V4 compatible across the entire Ascend lineup.

The collaboration also illustrates how China's largest AI companies are adapting to Nvidia's absence rather than waiting for export restrictions to ease at some point in the future. The key question is how much these tools can narrow the performance gap that has kept developers tied to Nvidia's CUDA — a question that, for now, remains unanswered. Near-term markers will include how quickly Chinese developers pick up the tools and whether Huawei's stated hardware timelines hold.