Think Nvidia’s losing China because of export controls? Think again.
While everyone’s been obsessing over Washington’s chip restrictions, something more interesting is happening: Chinese hyperscalers are simply building around Nvidia. Not because they have to—because they want to.
The numbers tell a story that Jensen Huang probably doesn’t love. Despite shipping H200 chips to China in January 2026 and ramping up production specifically for that market, Nvidia’s grip on China’s AI accelerator server market is slipping. Not collapsing, mind you—slipping. There’s a difference, and it matters.
The Real Competition Isn’t Who You’d Expect
Here’s what’s actually happening: Chinese companies are building custom inference solutions. Not training chips—inference chips. That’s the unglamorous work of actually running AI models after they’re trained, and it’s where the real money lives at scale.
Nvidia still dominates training. The H200 is a beast for that. But inference? That’s where hyperscalers can optimize for their specific workloads, cut costs, and—here’s the kicker—not depend on a foreign supplier whose products might get export-restricted next Tuesday.
This isn’t about nationalism or politics. It’s about business sense. When you’re running inference at the scale of Alibaba or Tencent, even small efficiency gains translate to massive cost savings. Custom silicon starts looking pretty attractive when you’re burning through millions of inference requests per second.
The H200 Gambit
Nvidia’s response? Ship more H200s. Huang announced they’re “firing up manufacturing” specifically for Chinese customers. It’s a smart move—the H200 is still the gold standard for training workloads, and China’s AI ambitions aren’t going anywhere.
But here’s the uncomfortable truth: this is a defensive play dressed up as offense. Nvidia is essentially saying “we’ll give you our good stuff” while Chinese companies are saying “thanks, but we’re also building our own stuff for the parts that matter most to our bottom line.”
The company claims to see $1 trillion in demand for its AI systems in 2026. That’s an absurd number, and it’s probably real. But how much of that comes from China going forward? That’s the question investors should be asking.
What This Actually Means
Nvidia isn’t dying in China. Let’s be clear about that. They’re still shipping chips, still making money, still the default choice for serious AI training work. But “default choice” is different from “only choice,” and that gap is widening.
The shift to custom inference solutions is happening globally, not just in China. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft is working with AMD. Everyone with enough scale is asking: “Do we really need to pay Nvidia prices for inference?”
China’s just asking that question louder and moving faster on the answer.
For Nvidia, this creates an interesting problem. They can’t exactly complain about losing market share in a segment (custom inference) they never really dominated. But they also can’t pretend it doesn’t matter when that segment is growing faster than traditional GPU sales.
The Bigger Picture
What we’re watching isn’t Nvidia losing China—it’s the AI chip market fragmenting exactly the way mature tech markets always do. Training stays centralized around a few high-performance options. Inference splinters into custom solutions optimized for specific use cases.
Nvidia will be fine. They’ve got the training market locked down, and that’s not changing anytime soon. The H200 is a technical marvel, and Chinese companies will keep buying them because they need them.
But the days of Nvidia owning the entire AI compute stack in China? Those are ending. Not because of export controls or geopolitics, but because the economics of custom inference silicon finally make sense at hyperscale.
That’s not a crisis for Nvidia. It’s just reality catching up to hype. The company will still print money. Just maybe not quite as much money from China as investors were hoping.
And honestly? That’s a more interesting story than “export controls bad” or “China building knockoffs.” This is about market evolution, not market collapse. Nvidia’s China problem isn’t that they’re losing—it’s that they’re learning to share.
đź•’ Published: