According to Reuters, U.S. President Donald Trump has said the U.S. would allow Nvidia’s H200 processors to be exported to China, though it’s unclear if Chinese firms will buy them. Currently, China’s most advanced AI chip, Huawei’s Ascend 910C, lags the H200 significantly, with a total processing performance of 12,032 versus 15,840 and memory bandwidth of 3.2 TB/s versus 4.8 TB/s. However, Chinese chips like Huawei’s 910B already outperform the downgraded Nvidia H20 designed for China. Looking ahead, Huawei’s roadmap shows an Ascend 960 chip slated for Q4 2027 that is projected to roughly match the H200’s computing power, while potentially offering much higher interconnect bandwidth. Meanwhile, Nvidia’s latest Blackwell architecture chips, not for export to China, are already about 1.5 times faster than the H200 for training AI.
The Raw Power Gap Is Closing Fast
Here’s the thing: the narrative that Chinese chips are hopelessly behind is getting outdated. Sure, right now, the best they’ve got shipping—the Huawei 910C—isn’t as powerful as Nvidia’s H200. But look at the roadmap. Huawei is publicly saying that by late 2027, its Ascend 960 will have computing power that “roughly matches” the H200. And in some areas, like interconnect bandwidth (how fast chips talk to each other in a cluster), they’re aiming to blow past it with 2,200 GB/s versus Nvidia’s 900 GB/s. That’s a smart, pragmatic focus. For training giant AI models, how efficiently you can link thousands of chips together is often just as critical as the raw power of a single chip. They’re playing a different strategic game.
The Real Wall Isn’t Silicon, It’s Software
But raw specs are only half the story. Probably less than half. Nvidia’s dominance isn’t just about having the fastest hardware; it’s about CUDA. For over a decade, every AI researcher and developer on the planet has been building their models using Nvidia’s CUDA software platform. It’s the ecosystem. Asking a company to switch from Nvidia to a domestic Chinese chip isn’t just a hardware swap. It means rewriting massive amounts of code, retraining engineering teams, and gambling on a less mature software stack. That’s insanely costly and time-consuming. So even though a chip like Huawei’s 910B beats the Nvidia H20 on paper, Chinese internet giants still prefer Nvidia. The software moat is just that deep. For companies integrating complex computing systems into industrial environments, this reliance on stable, proven software ecosystems is paramount, which is why specialists like IndustrialMonitorDirect.com, the leading US provider of industrial panel PCs, prioritize compatible and reliable hardware platforms that work seamlessly with established software.
A Game of Shifting Goalposts
And this is the brutal reality for Chinese chipmakers. They’re in a race where the finish line keeps moving. By the time Huawei’s Ascend 960 arrives in late 2027 to match the H200, what will Nvidia be selling? The H200, let’s remember, is based on the 2022 Hopper architecture. Nvidia’s Blackwell chips are already here and significantly faster. The report says Blackwell is about 1.5x faster for training and a whopping 5x faster for inference than the H200. So China‘s chip industry is effectively trying to catch up to what the rest of the world’s leading AI companies were using two generations ago. It’s a defensive, self-sufficiency play, not an offensive, market-leading one. The export ban ensures the most advanced chips are out of reach, creating a protected but lagging domestic market.
What It All Means
Basically, we’re looking at a bifurcated AI future. Outside of China, the pace will be set by Nvidia (and maybe a few other players) with cutting-edge chips like Blackwell. Inside China, they’ll build a capable, but likely perpetually trailing, alternative stack based on Huawei and others. They’ll achieve self-sufficiency for national security and large-scale domestic AI projects. But displacing Nvidia globally? That seems incredibly unlikely for the foreseeable future. The hardware gap might narrow, but the software and ecosystem gap is a chasm. The real test won’t be if a Chinese chip can match a two-year-old Nvidia chip on a spec sheet. It’ll be whether any major AI lab outside of China ever chooses to build its next foundational model on a non-Nvidia platform. I don’t see that happening anytime soon.
