DeepSeek and Huawei are making Ascend chips easier to program. Their target is not only Nvidia's hardware, but the software ecosystem that makes CUDA expensive to leave.
Three anchors to enter the analysis without losing the thread.
01
Powerful hardware is not enough
It may be capable of baking one hundred loaves an hour. But if nobody knows how to configure it, recipes were designed for another machine and spare parts are unavailable, its…
02
CUDA benefits from years of accumulated habits
They use libraries, frameworks, examples, drivers, debugging tools and code written by others.
03
It is similar to WhatsApp
A new messaging app can be technically better than WhatsApp.
Imagine a factory that has used machines from the same manufacturer for fifteen years.
The machines are powerful, but that is only part of the advantage.
Technicians are trained on them. Spare parts are easy to find. Internal procedures were written around them. Maintenance software already understands them.
A competitor arrives with a cheaper machine that may even be faster.
On paper, switching looks obvious.
In reality, switching also means retraining staff, rewriting procedures, replacing tools and accepting months of uncertainty.
That is one useful way to understand Nvidia's position in artificial intelligence.
Its GPUs matter enormously.
But part of Nvidia's strength comes from everything built around them.
On 30 September 2026, DeepSeek announced work with Huawei on programming tools for Ascend processors. The effort includes open-source infrastructure and TileLang, a higher-level language intended to make accelerator programming simpler.
The target is not only Nvidia's silicon.
It is the ecosystem around CUDA.
Powerful hardware is not enough
Think of a professional oven.
It may be capable of baking one hundred loaves an hour. But if nobody knows how to configure it, recipes were designed for another machine and spare parts are unavailable, its theoretical performance is not very useful.
AI chips face the same problem.
Developers need tools to organise computation, move data, debug errors and use the hardware efficiently.
CUDA is Nvidia's software platform for programming its GPUs for general-purpose accelerated computing.
Put simply, Nvidia does not only sell the engine. It also provides the steering wheel, workshop tools, manuals and much of the driving culture.
CUDA benefits from years of accumulated habits
Developers rarely start from zero.
They use libraries, frameworks, examples, drivers, debugging tools and code written by others.
Many of those tools have been optimised around Nvidia hardware for years.
Nvidia says more than four million developers now build on its accelerated-computing technologies and more than 40,000 companies use Nvidia AI technologies.
Those are Nvidia's own figures, so they describe the ecosystem according to the company.
But they illustrate the scale of accumulated adoption.
More users create more tutorials, skills, libraries and compatible software. Those resources then make the platform easier for the next user to choose.
It is similar to WhatsApp
A new messaging app can be technically better than WhatsApp.
If none of your family, friends or colleagues use it, the advantage may not matter.
WhatsApp is valuable partly because other people are already there.
CUDA is not a social network, but it benefits from a related dynamic.
Developers choose tools partly because colleagues know them, libraries support them and answers to common problems already exist.
Switching costs can matter more than chip prices
A cheaper bank account may look attractive.
But changing banks becomes harder when salaries, supplier payments, loans, direct debits and accounting processes all depend on the old account.
The real switching cost is larger than the headline fee.
AI infrastructure works the same way.
A company may have thousands of lines of CUDA-optimised code, trained engineers, tested models and deployment processes built around Nvidia.
A competing chip must overcome not only performance differences but migration cost.
That is why benchmarks alone do not decide the market.
DeepSeek and Huawei are attacking that layer
Reuters reports that DeepSeek and Huawei are developing programming tools for Huawei's Ascend platform, including open-source compute and communication infrastructure.
TileLang is part of that effort.
A public Ascend-oriented TileLang implementation already exists, designed to let developers write high-performance kernels with a higher-level programming model.
The strategic message is clear.
Huawei cannot only say, “our chip is fast.”
It also needs to say, “your team can use it without rebuilding everything.”
Huawei is not starting from zero: it is building its own toolbox
It would be misleading to describe Ascend as isolated hardware waiting for software to arrive.
Huawei develops CANN, the Compute Architecture for Neural Networks, as the software layer for AI workloads on Ascend processors. Huawei's official documentation says CANN supports frameworks including MindSpore, PyTorch and TensorFlow, with APIs and tools for developing and optimising workloads on Ascend hardware.
Huawei is already trying to build the workshop around the machine.
One detail is especially revealing. In PyTorch's own documentation for integrating third-party accelerators, Ascend NPU appears as an example of how an external device can be integrated into the framework. PyTorch has also been expanding its accelerator-integration mechanisms so new hardware can plug into the ecosystem without rewriting the entire framework.
That matters because the battle can move up the stack. If developers keep writing models in PyTorch while the hardware underneath becomes easier to swap, changing vendors becomes less painful.
DeepSeek, TileLang, CANN and PyTorch's accelerator abstractions do not make CUDA disappear overnight. They gradually attack the hardest part of lock-in: habits and compatibility.
Path dependence is powerful
A new highway can be faster than the old one.
But if warehouses, garages, fuel stations and factories were built along the old road, transport companies may not move immediately.
Software ecosystems create similar inertia.
Economists and historians often describe this as path dependence: previous choices change the cost of future choices.
That is not automatically bad.
Standards reduce duplication and make teams productive.
But a dominant standard can also give its owner lasting power.
The Windows lesson
For years, companies did not choose office computers only by comparing processors.
They asked a simpler question: will our software run on it?
That question helped protect the Windows ecosystem.
This is not merely a retrospective analogy. In the U.S. antitrust findings against Microsoft, the court explicitly described a positive network effect: a large Windows installed base attracted software developers, and the resulting catalogue of applications made Windows more attractive to users. The cost of porting applications to another operating system reinforced the barrier.
An alternative system could be excellent, but switching became difficult when accounting software, industry tools or device drivers were missing.
Nvidia has a comparable advantage in AI.
Hardware is easy to benchmark. Ecosystems are harder to measure, but they can create deeper lock-in.
CUDA is not immortal
Dominant ecosystems can lose power.
Frameworks can support multiple backends. Compilers can translate code across hardware. Higher-level tools can hide some of the underlying architecture.
Large customers may also accept migration costs when strategic incentives are strong enough.
China has an unusually strong incentive to build alternatives because US export restrictions constrain access to some advanced Nvidia products.
That creates both political and commercial pressure to develop domestic hardware and software.
Still, ecosystems take time to mature.
Developers need documentation, stable libraries, debugging tools, real deployments and confidence.
CUDA could be weakened without being directly replaced
A dominant standard does not always need to be replaced directly.
Sometimes it becomes less visible.
When you tap a bank card, you do not think about every institution and protocol behind the payment.
The interface hides complexity.
AI tooling could move in the same direction.
If frameworks and compilers make the underlying accelerator easier to swap, developers may care less about writing directly for one vendor.
The question then changes from “is my code written for CUDA?” to “which hardware can run this workload at the best cost?”
That is a much more dangerous question for a dominant ecosystem.
The lesson reaches beyond China and Nvidia
African countries often discuss building local platforms, clouds and AI services.
The CUDA story contains a broader lesson.
Owning a product is not enough.
A platform also needs documentation, skills, integrators, compatible tools and a community.
Technology dependence can begin long before a contract says “vendor lock-in”.
It can begin when enough people, processes and software accumulate around one system that leaving becomes painful.
Nvidia's strength is therefore double.
The chips matter.
So do the years of software, tools, developers and habits around them.
DeepSeek and Huawei appear to understand that reducing dependence on Nvidia means fighting both battles.
Sources and method
Facts that may change over time were checked against the sources below. Observations and interpretations are presented as such in the article.
What if the best way to learn AI is to begin with an annoying real task? An essay on practical use, judgement, learning and the ordinary moments where AI becomes genuinely useful.
Cloudflare fixed a cross-tenant flaw in Containers and Sandboxes. Here is how storage reuse exposed residual data and what cloud teams should learn from it.
Nvidia: its real strength isn't just its chips | Trigenys Insights
