What if the thing locking you into Nvidia was never the silicon?
I spend most of my time here poking at AI tools to figure out what actually holds up under load and what falls apart the moment you ask it to do real work. And the lesson that keeps repeating is that hardware specs almost never decide who wins. Switching costs do. Which is why the numbers coming out of China right now deserve a closer look than the usual geopolitical hot takes give them.
Bernstein Research estimates Nvidia held roughly 40% of China’s AI chip market in 2025, running neck and neck with Huawei at about the same share. Their projection for 2026 has Nvidia dropping to around 8% while Huawei climbs to roughly 50%. That is not erosion. That is a market reorganizing itself in a single year.
Comparable is the magic word
Huawei’s Ascend chips are now considered roughly comparable to Nvidia’s H200. Read that carefully, because “comparable” is doing a lot of work. It does not mean better. It does not mean cheaper per FLOP. It means good enough that the comparison is no longer embarrassing.
That is the threshold that matters in tooling. I have reviewed enough alternatives-to-the-default to know the pattern: nobody switches for a 5% gain, and everybody switches once the gap closes to the point where other factors start deciding. Availability. Supply certainty. Political risk. Price. Once the hardware is in the same class, those factors stop being tiebreakers and start being the whole decision.
U.S. export restrictions turned “other factors” into the only factors that mattered for Chinese buyers. If you cannot reliably get the best chip, the second-best chip you can actually buy becomes the best chip you can actually use. China’s push toward a self-sufficient semiconductor industry did the rest.
The DeepSeek migration is the part that should worry Nvidia
Here is the detail I would put at the top of any honest analysis: DeepSeek has moved from Nvidia’s CUDA ecosystem to Huawei’s CANN framework. And DeepSeek’s next-generation model, V4, released April 24, 2026, debuted on the Huawei Ascend platform.
Chip market share is a lagging indicator. Framework adoption is a leading one.
CUDA was never just an API. It was close to two decades of accumulated tutorials, Stack Overflow answers, kernel optimizations, library support, and muscle memory in the fingers of every ML engineer who ever trained anything. That is the real moat, and it is the kind of moat that makes hardware comparisons irrelevant. You can hand a team a faster chip and they will still pick the slower one if the slower one means their existing code runs tonight instead of next quarter.
Which is exactly why a serious lab migrating off CUDA matters more than any share percentage. Somebody paid that migration cost. Somebody ported the kernels, hit the weird edge cases, filed the bug reports, and figured out which operations behave differently. That work does not stay private for long. The second team to make the same move will find a shorter path. The tenth will find it close to routine.
Analysts have described Huawei’s chip progress as another DeepSeek moment, and I think that framing is right for reasons that have nothing to do with hype. The first DeepSeek moment was about proving an assumption wrong in public. This one works the same way. The assumption being tested is that the CUDA ecosystem is permanent.
What this means if you pick tools for a living
I am not going to tell you to rewrite your stack. For most teams outside China, Nvidia plus CUDA is still the pragmatic default, and the supply story there has not changed. But a few things are worth adjusting in how you think:
- Treat framework lock-in as a real line item. If your inference and training code assumes one vendor’s runtime at every layer, you own that risk whether you have priced it or not.
- Watch where new models debut, not just where they run. A model launching first on non-Nvidia hardware tells you more about the direction of the ecosystem than a quarterly earnings call does.
- Abstraction layers earn their keep in exactly this scenario. The teams who kept a portability layer around their hardware-specific code are the ones with options now.
- Discount any claim that a software ecosystem is unassailable. Jensen Huang has publicly acknowledged how far Nvidia’s position in China has slipped. When the vendor stops arguing, the argument is over.
The uncomfortable read for anyone building on a single vendor’s tooling is that export controls did not protect Nvidia’s moat in China. They funded a competitor’s effort to build a bridge across it. Restrictions bought time, and that time got spent on exactly the problem Nvidia least wanted solved.
Good tools win on merit. Default tools win on inertia. Those are different things, and the difference only shows up when someone finally pays the switching cost and tells everyone else how much it hurt.
🕒 Published: