Two very different groups spent this week trying to reduce their dependence on the same company, and both ran into the same wall. In China, AI labs blocked from Nvidia’s top chips keep quietly using them anyway. In the United States, the giant cloud providers that can buy all the Nvidia hardware they want are busy designing their own alternatives to avoid it. The escape routes point in opposite directions, but they lead to the same discovery: the chip was never the hard part.
Start with China. As the South China Morning Post reports, the biggest obstacle to switching off Nvidia is no longer the silicon but everything built around it. Nvidia’s CUDA software has been woven so deeply into how labs train models that moving to a domestic alternative like Huawei’s Ascend chips can feel less like swapping a machine and more like rebuilding the factory. One researcher estimated the shift could add at least 50 percent to a project’s time and cost. It depends heavily on what you are moving: an open-source model such as DeepSeek can be migrated by a handful of engineers in a few weeks, because the code is open to modify, while a closed system without accessible source can tie up ten engineers for more than half a year.
There is real progress underneath the frustration. Once a model is trained, running it day to day, the inference stage, is far easier to move onto Chinese hardware. And domestic chips are starting to handle training too: Meituan says its large LongCat-2.0 model was built on a 50,000-chip Chinese cluster. But the pattern holds. The hardware gap is closing faster than the software habit, and habit is the part Nvidia cannot be forced to hand over.
Now flip to the hyperscalers. Google, Amazon and Microsoft are not banned from anything; they simply resent the “Nvidia tax,” the premium paid for proprietary hardware. Their answer is custom silicon, application-specific chips hard-wired for the exact maths that transformer models need, stripped of the general-purpose overhead that makes a GPU flexible but power-hungry. Google’s TPUs were the early blueprint; the approach is now spreading across the cloud industry as inference workloads balloon and electricity bills become the binding constraint. Even the CPU is contested: analysts expect agentic AI, which leans harder on general-purpose processors, to push the server-CPU market past $210 billion by 2030, with AMD and ARM eating into territory Intel once owned outright.
Nvidia’s response is to refuse to stand still. It still commands something close to 90 percent of the data-centre accelerator market, and its newest Vera Rubin systems bundle GPUs, CPUs and networking into one architecture that customers plan whole data halls around. It is even pushing into CPUs with its own Vera chip, closing the one flank rivals hoped to exploit. When SpaceX committed to Nvidia exclusively this month, Elon Musk called Vera Rubin simply “the best AI computer.” The lesson running through all of it is that Nvidia’s real moat was never a single fast chip. It is the years of software, tooling and developer muscle memory built on top, and that is the thing no export ban or in-house chip project can replicate on a schedule.