Huawei is building its own alternative ecosystem to CUDA. If it succeeds, NVIDIA will have a serious problem

When talking about NVIDIA, almost all the focus is on the hardware: the H100Blackwell, racksenergy consumption, nanometers… It is understandable, but it is a mistake. The defensive moat – the moat– NVIDIA is not the hardware. Is CUDA.

CUDA is not an add-on to the chip, it is the de facto standard upon which most of the AI ​​code on the planet is written, optimized and debugged. Changing GPUs without changing CUDA does not exist. And switching from CUDA means rewriting years of work. That is why it is a moat.

Why is it important. Huawei’s big bet is not to “make a Chinese H100.” It is to build a path for the developer to reach Ascend without feeling like you are changing planets. The restrictions are accelerating it.

Exports have split the world in two:

1. An ecosystem that revolves around NVIDIA.
2. And another that China is trying to lift against the clock.

In that second, Huawei is not just playing chips: is playing “ecosystem”in AI and outside of it. And therein lies the nuance: you can be years behind in chips and still reduce dependency if you get the software to swallow.

In detail. Huawei is attacking the problem on three fronts, with a pragmatically Chinese logic: not to replace everything at once, but to open shortcuts.

1. Native stack (CANN + MindSpore). It is your “pure” alternative: your own environment and your own tools to get the most out of Ascend. The cost today is high, there are complaints of instability, the documentation is rather messy, and the community is much smaller.
2. PyTorch support. This is the most strategic move. Huawei does not try to make the world love its framework– Try to ensure that the world doesn’t have to leave PyTorch. torch_npu acts as an adapter to run PyTorch models on Ascend, but with one problem: it is not native and suffers with every PyTorch change. If PyTorch advances and your backend lags behind, the developer notices.
3. Portability via ONNX. Here Huawei looks for its best window: inference and deployment, not training. ONNX works as a bridge format: you train where you can (often NVIDIA) and deploy to Ascend. It’s a less romantic and more useful approach: if shortages hit, moving inference to local hardware is an immediate relief.

Between the lines. The real story is that Huawei is trying to replicate the “trick” that made NVIDIA great: turning its hardware into an experience.

That’s why the tactic that explains everything appears: putting engineers in the client’s home to migrate code and optimize it. It is not scalable as a business model, but it is scalable as a transition model: you buy time while you mature tools, libraries and support.

• And there is another derivative: if China gets enough teams to adopt Ascend out of necessity, over time that can become habit and then infrastructure. Not because it is better, but because it is already integrated.

Yes, but. Huawei has two limits that cannot be fixed with marketing:

1. Hardware improvement rate: Roadmap analysis suggests relative stagnation and a gap that could widen, not close, if NVIDIA continues to accelerate cycles.
2. Off-chip bottlenecks: memory (HBM), tools and industrial capacity. You can add “worse” chips, but you need to make a lot of them and build a lot of systems.

And now what. If this movie continues, we will see two clear signs:

• Less hype of chips and more real migration stories: how many computers have moved to Ascendwith what frictions, with what performance losses.
• Less obsession with training in Ascend and more normalization of the hybrid pattern: I train where I can, I deploy where I must.

NVIDIA will continue to be CUDA. Huawei is not “a chip.” It is an escape strategy. And the restrictions are the fuel that is making it inevitable.

In Xataka | With HarmonyOS NEXT Huawei has achieved something incredible. Neither Samsung, Microsoft nor Mozilla achieved it

Featured image | NVIDIA, Huawei