Back to Blog
Industry

Nvidia Just Left China. Beijing's Chip Makers Should Send Flowers

Nvidia ending H200 sales did what Beijing's directives couldn't: it left Chinese AI labs with no CUDA hedge. Domestic chip makers now have captive customers, and the next open models get built for Ascend first.

OPC Community

OPC Community

Nov 19, 2026 8 min read
Nvidia Just Left China. Beijing's Chip Makers Should Send Flowers

Beijing spent two years trying to pry its AI labs off American silicon and got nowhere. Nvidia's board managed it in an afternoon.

The most valuable thing Nvidia ever sold China was a reason to wait.

On November 17, interim CEO Colette Kress pulled it from the shelf. She ended H200 sales to China and set up a board-level national security committee, all within days of Jensen Huang's exit. A petition to bring him back has passed 200,000 signatures. Palmer Luckey signed on, and his line got around: China is thrilled to have Jensen out.

He's right. China is thrilled. But most people heard that line and concluded that a Nvidia without its founder is a weaker Nvidia, and a weaker Nvidia is a gift to Beijing.

Wrong gift. The real one is smaller and a lot more useful. Beijing's chip makers finally have customers with nowhere else to go.

The Hedge Was the Moat

Picture the head of infrastructure at a Chinese AI lab in early 2026. She needs a new training cluster and she has two ways to get one.

The first is Huawei's Ascend. The hardware is decent, it's improving fast, and the state is behind it. Her problem is that her whole codebase assumes CUDA. Moving to Huawei's CANN stack means rewriting kernels, re-validating every training run, and chasing bugs nobody on Stack Overflow has ever seen. Call it six months of her best engineers doing work that ships no product.

The second is to wait for an H200 allocation. Washington cleared those sales in December 2025, with a 25% cut going to the US government. The chips sit a generation behind Nvidia's best. They also run everything she already has, unchanged, on day one.

She waits. Anyone would.

Now multiply her decision by every lab in the country. Each H200 that cleared customs was a pile of porting work that never got done, a bug in Huawei's toolchain that never got reported, a purchase order Cambricon never received. China's domestic stack had money. It had talent. What it never had was desperate users.

That's how a hedge works. Keep the better option even partly open and nobody commits to the worse one. And a chip platform nobody commits to stays worse, because platforms get better by being suffered through.

People talk about CUDA's moat as if it were the code. The moat was that everyone capable of building the alternative was busy using CUDA.

Beijing Spent Two Years Trying to Do This

American coverage keeps missing this part: China's government wanted exactly this outcome, wanted it badly, and couldn't get it.

In the summer of 2025, regulators summoned Tencent and ByteDance to explain why they were still buying Nvidia's H20. That September, the Cyberspace Administration told big tech firms to stop testing and ordering Nvidia's China-specific RTX Pro 6000D, according to the Financial Times. In January 2026, Reuters reported that customs agents had been told H200s were not permitted in.

The labs found a way anyway. They bought through intermediaries. They rented GPU capacity in Southeast Asia. They lobbied for exemptions. A directive from your own government is something you negotiate with, especially when the competitor down the road is quietly ignoring it and training faster than you.

A seller who refuses is a different animal. You can't lobby your way around a product that isn't for sale.

So Kress and her new committee pulled off what the Cyberspace Administration couldn't. They made the ban credible. All the Chinese state could do was tell its labs to stop buying. Nvidia's board stopped selling, and did it in the name of American national security, which hands every Chinese CTO a clean story for the board deck: we have no option, fund the migration.

I used to roll my eyes at Huang on this. For years he said that selling into China kept Chinese developers on the American stack, and that export controls would only build a rival. It sounded like exactly what a man with a $17 billion China business would say. In 2025 he told an audience that Nvidia's share of China's advanced AI chip market had gone from 95% to zero, and all I heard was a salesman mourning a territory.

It was self-serving. It was also correct. Those two go together more often than we like to admit.

What a Captive Customer Is Worth

Bad products rarely kill a chip company. Uncertain demand does.

A fab line. A packaging contract. A software team of two thousand people. Each one is a bet placed years ahead of revenue, and every domestic Chinese chip maker has stared at the same question: if we build it, and Washington loosens the rules again next quarter, do our customers walk?

Until this week the honest answer was yes. The rules flipped constantly. The H20 was banned in April 2025, which cost Nvidia a $4.5 billion charge, then un-banned that summer. The H200 was forbidden, then allowed. If you ran Cambricon, your customer base would leave the moment an American president changed his mind.

Even so, Cambricon's revenue grew more than fortyfold in the first half of 2025, on nothing more than the possibility of a cutoff. Huawei published an Ascend roadmap running through 2028. That was the appetite when nobody knew what came next.

Now everybody knows. A board-level national security committee is hard to dissolve quietly. It tells every buyer in Shenzhen and Hangzhou that Nvidia is gone for structural reasons, and you can write a five-year procurement plan against a structural absence.

The same week Kress made her announcement, 4Paradigm said it had trained two base models end-to-end on domestic compute. Read that as a status report. A project like that starts many months before the press release, so the migration was already under way while H200s were still legally on sale. The serious players had started hedging in the other direction.

Last month those early movers looked like eccentrics. Nvidia's exit makes them the standard.

The Honest Cost, and Why It Doesn't Save the Argument

Kress's decision deserves a fair hearing, because the case for it isn't stupid.

Chinese labs will have less compute in 2027 than they would have had with Nvidia's chips. Domestic accelerators trail on performance per watt. SMIC's advanced capacity is tight, and high-bandwidth memory is a real bottleneck. If you believe the next eighteen months decide everything, then denying compute for eighteen months is worth almost any price.

I don't believe that, and the record since 2022 backs me up.

Every round of controls was supposed to open a gap that would compound. What compounded was Chinese efficiency. DeepSeek trained competitive models on a fraction of the hardware its American peers used. In August 2025 it shipped a release with a number format tuned, in its own words, for the next generation of domestic chips. One line in a release note, and it marked the first time anyone had shaped a frontier-adjacent model around non-Nvidia silicon on purpose.

Scarcity made those teams better at the exact skill you need when your chips are weaker. Put a million engineers to work optimizing for a weaker chip and it stops falling behind at the old rate.

So here is the trade Nvidia's board just made. America gives up a dependency it controlled and gets a competitor it doesn't. The dependency came with a kill switch, a revenue stream and a view into what Chinese labs were building. The competitor comes with none of those.

“A customer you can cut off is leverage. A customer you have cut off is a rival with a head start on hating you.”

Why This Lands on Your Desk

Maybe you build on AI and have never spent a minute thinking about Chinese semiconductors. This week is still yours.

Most of the open-weight models indie builders run in production come from Chinese labs. Qwen, DeepSeek, Kimi. For three years those models were trained on Nvidia hardware and ran best on Nvidia hardware, which happened to be the hardware you rent. Nobody planned that alignment. It was an accident of the hedge.

The accident is over. The next generation of those models gets designed for Ascend clusters first: their precision formats, their memory limits, their interconnects. The CUDA port becomes somebody's second priority. You'll notice it as a quantization that behaves oddly, an inference speed that disappoints, a release that lands on your stack three months late.

The world is getting a second AI stack, complete from the transistor up, and the cheapest capable models on earth will run on it.

There's a plainer lesson in here for anyone who sells anything. Your customers put up with you, and you shouldn't mistake that for loyalty. They've done the math and switching costs more than staying. Make staying impossible and you've paid their switching cost for them. They'd never have found the nerve on their own.

Every founder who has deprecated an API, killed a free tier, or locked out a region "for strategic reasons" has run a small version of this experiment. Most learned what Nvidia is about to learn. The customers you push out don't vanish. They go and build the thing that replaces you, and they bring a grudge.

Send Them to Santa Clara

Luckey wants Huang back, and 200,000 people agree with him. Maybe they get their way.

It won't change any of this. You can reinstate a CEO. You can't reinstate a hedge. Once that infrastructure lead has spent her six months on the port, once her team knows the Ascend toolchain cold and her models run on chips no American committee can revoke, she isn't coming back for a discount. The reason to wait was only worth something while people were still waiting.

So yes, Beijing's chip makers should send flowers.

Address them to the board.

Frequently Asked Questions

Why did Nvidia stop selling H200 chips to China?

On November 17, interim CEO Colette Kress ended H200 sales to China and set up a board-level national security committee, days after Jensen Huang's exit. The article argues that the committee makes the exit structural rather than temporary, because a board-level body is hard to dissolve quietly. Chinese buyers can now write five-year procurement plans on the assumption that Nvidia is gone.

How does Nvidia leaving China help Huawei Ascend and Cambricon?

While H200s were available, Chinese labs kept waiting for Nvidia allocations instead of spending roughly six months of their best engineers' time porting from CUDA to Huawei's CANN stack. That left domestic chip makers with money and talent but no committed users, and with demand that could vanish whenever Washington loosened the rules. With Nvidia no longer selling, those labs have nowhere else to go, which gives Huawei and Cambricon the captive demand they need to justify long-term bets on fabs, packaging and software.

What does Nvidia's China exit mean for developers using Qwen, DeepSeek or Kimi?

Most of the open-weight models indie builders run in production come from Chinese labs, and until now they were trained on Nvidia hardware and ran best on it. The article expects the next generation to be designed for Ascend clusters first, with the CUDA port a second priority. Builders may notice quantizations that behave oddly, slower inference, or releases that reach their stack months late.

Share this article

Join the OPC Community

Connect with solo founders worldwide. Get daily insights, curated opportunities, and peer support.

Join free