If Nvidia has taught the world anything, it is not that the best chip always wins. That would be too simple, and history dislikes simple explanations almost as much as purchasing departments like them. Nvidia’s real lesson is more annoying: if you control enough of the stack, the stack begins to control the market.
At first glance, this remains faintly absurd. The great engine of modern artificial intelligence was not born as a solemn temple of cognition. It was born as a way to make explosions, shadows, dragons, leather jackets, wet cobblestones, and anatomically overconfident elves look more convincing. The GPU was a graphics device. It was meant to move pixels around very quickly, not to simulate fragments of human reasoning at planetary scale. And yet here we are: civilisation now asks its most important questions to machines trained, in no small part, on the descendants of hardware once optimised for Quake, Lara Croft, and reflective puddles.
The joke, of course, is that Nvidia made the joke stop being funny. CUDA, introduced in the late 2000s and now described by Nvidia as its platform for accelerated computing, became the software layer that lets developers use GPUs through C++, Python, Fortran, PyTorch, libraries, compilers, profilers, and a whole apparatus of industrial convenience. The hardware mattered enormously, but the software made it habitable. A GPU without CUDA is not useless, but it is like a magnificent city with no roads, no plumbers, no restaurants, and a municipal website last updated in 2009.
This is why Nvidia’s advantage has always been slightly misunderstood. It is not merely a chip company. It is a chip company that convinced programmers, researchers, datacenter operators, framework maintainers, and nervous executives that its ecosystem is the least painful place to live. The GPU became not just an accelerator, but a jurisdiction. Once your code, tooling, kernels, training recipes, debugging rituals, hiring pipeline, and institutional folklore all assume CUDA, switching vendors becomes less like buying a different graphics card and more like moving a medieval monastery stone by stone.
The results are visible everywhere. CUDA is not just an API; it is an accumulated civilization. It includes fast libraries, documentation, tools, examples, habits, forum posts, Stack Overflow debris, and the precious confidence that when something breaks, somebody else has already suffered first. That last point is worth more than any marketing benchmark. Engineers do not merely buy performance. They buy the probability that they can go home before midnight.
Naturally, everyone else noticed.
China noticed most urgently, because export controls transformed dependence from an inconvenience into a strategic humiliation. Since October 2022, the United States has progressively tightened restrictions on advanced computing components for China, and recent analyses argue that such controls have raised costs while also strengthening the incentive for Chinese self-reliance. The intended effect was to constrain capability. One side effect was to make the phrase “domestic software stack” sound less like bureaucratic wallpaper and more like national infrastructure.
Huawei’s Ascend ecosystem is the most obvious example. Huawei has long framed Ascend as a full-stack AI platform, including CANN, MindStudio, MindX, and related tools. CANN — Compute Architecture for Neural Networks — sits in roughly the same strategic position for Ascend that CUDA occupies for Nvidia GPUs: it is the layer that compiles, optimises, and runs AI workloads on the hardware. MindSpore, meanwhile, is presented as part of Huawei’s AI full stack for deep-learning tasks. In 2025, Huawei announced that CANN operators and core Ascend software components would be open-sourced, explicitly trying to build a broader developer ecosystem around Ascend.
This is the correct move, but not a magic spell. Software ecosystems cannot be conjured by press release, even a very confident one. Developers have complained that CANN remains less mature and harder to use than CUDA, and reports have described instability, difficult documentation, crashes, and slower adaptation work for large Chinese customers. That does not make Huawei irrelevant. It makes Huawei interesting. The first stage of stack sovereignty is rarely elegance. It is usually inconvenience plus compulsion.
Other Chinese companies are pursuing similar logic. Moore Threads has its MUSA software stack, promoted as a CUDA alternative for its domestic GPUs, including tooling to help port CUDA code. Recent research on MusaCoder even presents a full-stack training framework targeting CUDA and MUSA backends, with native kernel generation and execution-feedback training on Moore Threads GPUs. Cambricon, Hygon, Biren, MetaX, Iluvatar CoreX, Alibaba’s T-Head, and others are all part of the same broader movement: do not merely fabricate silicon; build the habitability layer around it. China has also reportedly added several domestic AI chips to secure-and-reliable procurement lists, including Huawei Ascend, Alibaba T-Head, Biren, Hygon, Iluvatar, MetaX, and Moore Threads products.
The strategic direction is clear: the chip is no longer the product. The stack is the product. The chip is the ceremonial object at the centre of the ritual, but the ritual includes compilers, kernels, interconnects, schedulers, model formats, quantisation paths, inference servers, cloud APIs, documentation, and the reassuring presence of a bored engineer who has seen this error before.
Even model builders are adapting. Reuters reported in July 2026 that DeepSeek is developing its own inference chip to reduce reliance on Nvidia and Huawei, while still facing the brutal realities of chip design, foundry access, memory supply, and export restrictions. Another report said China may allow top AI firms to buy limited numbers of Nvidia H200 chips, which shows the awkward transitional truth: sovereignty is desirable, but H200s are still H200s. Nobody wants dependence. Everybody wants delivery dates.
This is where the amusement becomes slightly dark. The U.S. tried to restrict access to the best chips, and in doing so helped turn domestic alternatives into patriotic infrastructure. Recent academic work argues that American policy shocks may have unintentionally accelerated China’s open AI ecosystems, increasing the strategic value of locally adaptable models and infrastructure. Containment may still matter, especially at the frontier, but it also teaches the contained party what to build.
Nvidia’s achievement was to make graphics cards useful for everything. China’s response is to make “everything” less dependent on Nvidia. The first effort produced CUDA. The second is producing CANN, MUSA, domestic procurement lists, chip startups, awkward porting tools, and an entire generation of engineers discovering that sovereignty means writing drivers.
In the end, the moral is not that Nvidia had the best hardware, although it often did. Nor is it that CUDA is perfect, because nothing used at scale by programmers can remain morally pure. The moral is stranger and more durable: whoever owns the boring layers owns the exciting future. The glamour is in the model. The money is in the accelerator. But the power is in the stack.
And somewhere, deep inside the machine room, a graphics card is still wondering why it was promoted from drawing goblins to training civilization’s next oracle.
No comments yet