Nvidia Returns to China: Jensen Huang Got His Way (But There's a Catch)

Jensen Huang knows how to wait and, more importantly, how to convince. After months of endless shuttling between Santa Clara and Washington, Nvidia's CEO finally got what he fought for: official White House approval for selling AI chips to China. Seemingly a victory, but this story has too many "buts" to pop champagne right now. Nvidia is returning to a market it once considered its backyard, but now discovers not only barbed wire sanctions there, but also fierce competitors.

Let's recall how we got here. Just a couple of years ago, China provided Nvidia with nearly a quarter of all data center revenue. Then big politics intervened. The US government decided that supplying Chinese AI with powerful accelerators was a bad idea and imposed strict restrictions. First, top-tier A100 and H100 chips were cut off, then their China-specific versions A800 and H800. Jensen Huang directly stated back then that cutting American companies off from the Chinese market was like shooting yourself in the foot. And apparently, he was heard, albeit with serious caveats.

So what are they allowed to sell now? We're talking about H20 series chips. This is a kind of 'homeopathic' version of the flagship H100. Nvidia engineers achieved the opposite architectural feat: they artificially limited computing power so that the chip would pass all US export control standards while remaining useful for neural network training. The problem is that Chinese giants like Baidu, Tencent, and Alibaba are no longer the naive buyers they once were. They have learned to live in scarcity mode and began actively looking around.

And looking around, there's Huawei. While Nvidia was blocked, Huawei didn't waste time and actively promoted its Ascend 910B accelerators. By many tests, they're already approaching the capabilities of those very 'trimmed' chips from Nvidia. For Chinese companies, buying domestic hardware is not just patriotism but also insurance against Washington changing its mind tomorrow and turning their servers into expensive bricks. Huang will have to work very hard to convince old partners to return to American technologies that can be taken away at any moment.

The situation is complicated by the fact that Nvidia has to play on two boards at once. On one hand, pressure from American regulators who monitor every transistor. On the other, a capricious Chinese market that demands maximum performance. If the H20 turns out to be too weak, the Chinese will simply wave their hand and switch to Huawei or Biren Technology. If it's too powerful, people in suits from the Department of Commerce will visit Huang's office again. It's a tightrope walk over an abyss, where every step costs billions of dollars in market capitalization.

And we shouldn't forget about the software ecosystem. The CUDA platform has long been the main 'moat' around Nvidia's castle. Chinese developers have grown accustomed to writing code under Huang's architecture, and switching to new rails is painful for them. However, Huawei is actively building its software analogues. If sanctions pressure continues for another couple of years, Chinese engineers will simply get used to different tools. At that point, Nvidia won't be needed by China even with its most powerful chips, because retraining thousands of specialists is more expensive than buying servers from a local supplier.

The Nvidia case shows that technological sovereignty is no longer a slogan but a reality. The world of AI hardware is splitting into two camps. We are witnessing the birth of an alternative ecosystem in China that is forced to develop despite sanctions. Nvidia is trying to sit on two chairs while maintaining its status as a global monopolist, but with each new package of restrictions, this becomes increasingly difficult. Jensen Huang got permission to trade, but he did not get guarantees that his product is still as needed by China as before.

The key point: Nvidia has retained access to China's wallet but lost its technological halo of invincibility. Can a 'slowed down' chip defeat a 'fast' local competitor?