DeepSeek announced on Wednesday that it has partnered with Huawei to develop programming tools for Huawei's Ascend AI processors — and is open-sourcing the results, including a version of its TileLang language positioned as a simpler alternative to Nvidia's CUDA.
The move targets the softest spot in Nvidia's armor. The company's dominance in AI computing has never been only about silicon; it rests on CUDA, the software ecosystem that virtually every AI developer has trained on. For context on why that matters, our AI industry coverage has documented repeatedly that hardware only wins when the software around it makes switching painful.
Six open-source tools for Ascend
According to Reuters, DeepSeek is releasing six open-source software modules built for Huawei's Ascend platform. The package includes libraries aimed at improving computing and communication performance across the chips, developed with what DeepSeek describes as extensive technical support from Huawei.
The centerpiece is TileLang, an open-source high-level programming language designed for AI accelerators. DeepSeek argues developers need a language that is both accessible and capable of extracting maximum performance from the underlying hardware, and says TileLang offers a simpler programming model than CUDA. If it delivers, the barrier to writing code for Chinese AI chips drops substantially — which is precisely the point.
A 128-chip supernode as proof of scale
The collaboration also produced hardware results. DeepSeek and Huawei say they worked on a supernode system combining 128 Ascend 950 chips, focusing on optimizing communication between processors as well as overall computing performance — the interconnect problem that has historically separated niche accelerators from genuine data-center platforms.
The announcement comes shortly after Huawei introduced its next generation of AI processors and "supernode" computing systems, and the company expects its AI infrastructure to see wider use for model training starting next year. A frontier lab publishing its systems software for a domestic chip platform is the strongest signal yet that Chinese AI companies intend to run serious training workloads without Nvidia hardware.
Why CUDA's moat is the real battleground
CUDA has become a central part of Nvidia's dominance because of the enormous software ecosystem built around its GPUs — compilers, libraries, and two decades of developer familiarity. Every PyTorch release, every inference framework, every research repo assumes CUDA works. That inertia, more than raw chip specifications, is what has kept rivals at bay.
The strategic logic of open-sourcing is equally deliberate. Proprietary chip software stacks win by being the default; open ecosystems win by being everyone's fallback. By publishing the Ascend tooling under open licenses, DeepSeek lowers the switching cost for any developer or cloud operator looking for leverage against Nvidia pricing — and makes Huawei hardware viable for exactly the buyers who would never accept a single-vendor stack.
That is why the DeepSeek-Huawei collaboration matters beyond China. It reflects a broader push within the country to develop domestic AI hardware and the software required to operate it efficiently, reducing dependence on Nvidia's technology at a time when U.S. export controls have already restricted access to Nvidia's most advanced chips. The New York Times framed the partnership as a direct assault on the key source of Nvidia's AI dominance — not the chips themselves, but the code that makes them indispensable.
Building on an existing partnership
Wednesday's release is not the first sign of DeepSeek betting on domestic silicon. Back in April, both Benzinga and the South China Morning Post reported that DeepSeek's V4 model was built and trained on Huawei chips, in what the SCMP described as a collaboration strengthening China's AI self-reliance. What is different now is the software: V4 proved the models could run on Ascend, while this release gives every other Chinese developer the tools to follow.
The road has not been smooth. Generation-NT reported that DeepSeek previously delayed its R2 model amid difficulties with Huawei Ascend chips, a decision the outlet said revealed the fragility of the nascent ecosystem. Huawei, for its part, has kept accelerating — The Business Times reported in mid-September that the company was moving up the launch of a new AI chip to take on Nvidia. The open-source release effectively converts those growing pains into shared infrastructure: problems DeepSeek solves for itself become problems every Ascend developer benefits from having solved.
What to watch next
Three things will determine whether this effort bites. First, adoption: open-source tools only matter if Chinese developers actually build with them, and TileLang's true test is whether model training on Ascend clusters becomes routine rather than experimental. Second, performance: DeepSeek's own models have already been reported running on Huawei silicon, but matching CUDA-era training throughput across heterogeneous workloads is a different bar. Third, spillover: if TileLang matures into a genuinely portable language for AI accelerators, it could complicate Nvidia's position well outside China's borders.
For now, the message is clear. The next phase of the AI hardware race is being fought in software, and DeepSeek just published its opening arguments in public.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
