Nvidia's biggest advantage has never been only its chips. It's the software developers use to program them. On 30 September 2026, DeepSeek and Huawei went after that advantage by open-sourcing programming infrastructure for Huawei's Ascend AI processors, Open Source For You reports.
Key takeaways
DeepSeek and Huawei released open-source compute and communication libraries for Ascend chips, plus an Ascend version of the TileLang programming language.
The goal is to make it easier to move AI workloads onto Ascend hardware without Nvidia's CUDA.
The two companies have also built a supernode architecture using 128 Ascend 950 chips.
What is CUDA, and why does it matter?
CUDA is Nvidia's software platform for programming its graphics processors. Most AI code is written for it, which is why switching to another chip is hard even when the hardware is good enough. Any rival needs its own developer tools before it can win real workloads.
What was released
TileLang for Ascend: a high-level language that wraps Ascend's low-level C instructions. The report says it simplifies programming compared with lower-level approaches such as CUDA, while still letting developers tune performance.
Computing and distributed communication libraries: these match versions DeepSeek previously released for Nvidia hardware, so existing workloads are easier to adapt.
Huawei separately maintains its Compute Architecture for Neural Networks, known as CANN, which supports projects such as PyTorch, Triton and vLLM.
Why it's significant
DeepSeek is one of China's best-known AI labs, so its support carries weight with developers. Tools that mirror what already exists for Nvidia lower the cost of trying Ascend. The report adds that DeepSeek and Huawei have optimised computation and communication across a 128-chip Ascend 950 supernode.
What to watch
Open-sourcing tools doesn't guarantee adoption. CUDA has years of libraries, tutorials and trained engineers behind it. The test is whether developers outside the two companies start using and contributing to the Ascend stack.
Why it matters
Chip supply is now a strategic question for the whole region. Malaysia is building data centres at pace, and buyers want more than one viable source of AI compute. A working open-source software stack for Ascend would give them bargaining power and a fallback. It also ties into the wider Chinese build-out, where ByteDance now rents about a fifth of the country's data centre capacity.
For the region's chip industry, the story sits alongside NUS's push into advanced packaging with Malaysia and Nvidia's own expansion in Southeast Asia. The supply chain is being redrawn from several directions at once.
Frequently asked questions
What did DeepSeek and Huawei release?
Open-source programming infrastructure for Huawei's Ascend AI processors: compute and communication libraries, and an Ascend version of TileLang.
What is TileLang?
A high-level programming language for AI accelerators. The Ascend version wraps the chip's low-level C instructions to make development simpler.
Does this replace Nvidia's CUDA?
Not yet. It offers an alternative for Ascend hardware, but CUDA still has a much larger ecosystem.
What is the Ascend 950?
One of Huawei's AI processors. DeepSeek and Huawei have built a supernode architecture that uses 128 of them.