DeepSeek and Huawei open-source a full software stack for Ascend chips — and take aim at CUDA
DeepSeek has open-sourced a full programming stack for Huawei’s Ascend AI chips — including a CUDA-challenging language called TileLang — in China’s boldest move yet to break free from Nvidia’s software grip.
On September 30, DeepSeek announced on its official WeChat channel that it had open-sourced a full programming stack for Huawei's Ascend AI chips, built with what it called Huawei's full support. Reuters reported the story. At the center of the release is TileLang, a high-level programming language that DeepSeek says offers a simpler programming model than Nvidia's CUDA — the software moat that has kept developers anchored to Nvidia hardware for two decades.
It is the boldest attempt yet by China's AI industry to solve its chip problem in software rather than silicon. Beijing has poured money into domestic accelerators, but the hardest part of dethroning Nvidia was never the transistors — it was the ecosystem around them. DeepSeek is now trying to supply it.
What actually shipped#
The release is a set of open-source Ascend-targeted tools that mirrors DeepSeek's existing Nvidia tooling one-to-one, so each component has a familiar counterpart:
- TileLang-Ascend, an Ascend variant of DeepSeek's high-level kernel language, letting developers write optimized kernels without dropping into Huawei's low-level chip instructions
- DeepGEMM Ascend for matrix multiplication and DeepEP Ascend for large-scale cross-device communication
- TileKernels, FlashMLA, and DeepSelect for vector computing, sparse attention, and data filtering
DeepSeek and Huawei also jointly optimized a "supernode": a cluster of 128 Ascend 950 chips tuned for combined compute and communication performance, meant to prove the stack can handle serious training workloads, not just demos.
Why TileLang matters#
TileLang is the interesting part. The language was originally developed by researchers at Peking University, and DeepSeek has been refining it for about a year — first on older Nvidia chips, then as its main tool for its own research on advanced systems. DeepSeek says it now uses TileLang for most of the computational operations involved in training its V4 models, and describes it as the company's primary path toward more ambitious AI work.
DeepSeek's pitch is blunt: to build an independent software ecosystem for AI chips, the first requirement is a high-level language that is universal, easy to program, and still squeezes full performance out of the hardware. In the company's telling, TileLang was created precisely for that job — and CUDA is the incumbent it has to beat.
That incumbent is formidable. Research house SemiAnalysis estimates Nvidia's software advantage rests on roughly four million developers building with CUDA worldwide — a moat rivals like AMD have never crossed even when their hardware matched Nvidia on paper. After testing OpenAI's own Jalapeño inference chip, SemiAnalysis went so far as to call the CUDA moat "potentially dead," while adding that it had only tested easy scenarios and had not yet run its AgentX multi-step benchmark, where Nvidia still led. Against that backdrop, the fact that Huawei's CANN stack was the only one besides CUDA to support DeepSeek's V4 model on day one starts to look less like a curiosity and more like a leading indicator.
The bigger picture#
DeepSeek is not a neutral bystander here. The Hangzhou lab has been migrating its own hardware away from Nvidia's H800 chips toward Huawei processors through 2026, and it has previously disclosed plans to install more than 160,000 Ascend accelerators at a data center under construction in Inner Mongolia. This release is, in part, DeepSeek tooling up its own supply chain.
It also arrives days after Huawei unveiled its next generation of AI processors and supernode systems, which the company expects to see widely used for model training next year. Hardware has been arriving on schedule; usable software is what has been missing. That is the gap DeepSeek is filling — and it is open-sourcing the fill, which means every other Chinese lab can build on it too.
What to watch#
The real test is adoption. Chinese model makers like Z.ai and Moonshot AI have moved faster than the country's chipmakers, and whether they pick up TileLang for their own training will decide if this becomes a true ecosystem or just DeepSeek's house stack. Watch for independent performance numbers on real training workloads, for Huawei's ability to actually deliver chips at scale — the company admits it cannot keep up with domestic demand — and for how Washington responds to the one front where export controls have no answer: software that makes restricted hardware unnecessary.