China is quietly telling Alibaba, ByteDance and DeepSeek they can buy Nvidia's H200 chips again, just not enough to settle the bigger fight over whose silicon runs the country's AI industry.
Beijing plans to approve H200 purchases for a handful of its biggest AI developers, according to a report from The Information published Wednesday. Chinese officials have told the same three companies in recent weeks that permission to buy the chips could come soon. The catch: total approvals are expected to cap out at fewer than 200,000 units, according to the report, less than half of what those companies originally asked for.
That's a real reversal. For most of the past two years, Beijing has leaned on its biggest tech companies to buy Chinese-made accelerators instead of Nvidia's, betting that pressure plus time would let Huawei and the rest of China's chipmakers - Cambricon, SMIC among them - close the gap on their own. Nvidia shares rose about 1% on the news, a modest bump that tells you Wall Street has learned not to get too excited about China headlines until chips actually cross the border.
It has good reason to be cautious. Bloomberg reported back in January that China planned to approve H200 imports as soon as the first quarter. The U.S. Commerce Department cleared a path for the sales soon after, allowing exports subject to a 25% duty and inspection in the United States. And yet by late February, according to Bloomberg, Nvidia had sold zero H200s into China, two months after Washington first opened the door. Approvals on paper and chips on a loading dock have been two very different things all year.
May brought another checkpoint: the Commerce Department cleared roughly ten Chinese companies, including Alibaba, Tencent, ByteDance and JD.com, to buy the H200. Now, in July, it's Beijing's side of the transaction that's finally catching up, with a specific number attached and a specific set of buyers named.
The approvals coming, per The Information, apply narrowly. The H200s are being reserved for training the largest and most complex models, the stage where raw compute matters most and where Huawei's Ascend lineup still trails Nvidia by a wide margin. For everyday inference work, the tasks that actually serve answers to users, Chinese firms are still being pushed toward domestic processors. Beijing isn't opening the market. It's making a calculated trade: buy the training horsepower you can't get at home, but keep the volume business with our own chipmakers.
That distinction matters more than the headline number. It undercuts the idea, pushed hard in recent coverage of Huawei and DeepSeek, that China's AI industry is marching steadily toward local-silicon-only. Reuters reported this week that DeepSeek is separately developing its own inference chip, aiming to cut its dependence on both Nvidia and Huawei at once. DeepSeek trained its earlier models on Nvidia's H800 and shifted much of its recent work, including its V4 release in April, onto Huawei's Ascend chips. Building a third, in-house option isn't the move of a company that trusts either supplier to be there long term.
Huawei, meanwhile, holds roughly half of China's $50 billion domestic AI chip market, built almost entirely on the back of the U.S. export ban that kept Nvidia's best chips out. That position is already softening. Alibaba and Baidu are both building their own accelerators and taking share from Huawei inside China, before Nvidia even re-enters the picture. Industry surveys now point to Chinese firms shifting 46% of their AI-accelerator budgets to domestic suppliers within a year, up from around 30% today. Nvidia getting a green light doesn't reverse that trend. It just means the trend has room for both.
Nvidia's math is still constrained. Fewer than 200,000 H200s split across a handful of buyers is a rounding error against the volumes the company ships to U.S. hyperscalers each quarter, and every one of them carries a 25% duty across the border. Washington gets a fresh test of how far case-by-case export licensing goes before critics on either side call it too generous or too restrictive - and Beijing gets a quiet admission that even with Huawei ramping up, the biggest models still run faster on someone else's silicon.
None of that settles the argument the headline seems to promise. It just moves it forward by one round, with a number attached this time instead of a rumor.
Also read: Metal Turns Fundraising From a Networking Problem Into a Data Problem • Iluvatar CoreX seeks up to $850 million just as its IPO lockup expires • Zhipu AI Quietly Raises the Default Concurrency Limit on Its GLM API
