Nvidia Drops H200 Production for China, Hands the Factory Slots to Vera Rubin

Nvidia built about 250,000 H200 chips for Chinese buyers. It has now stopped making more and handed that manufacturing capacity to its next-generation Vera Rubin platform, according to the Financial Times.

The standoff isn't coming from Washington anymore. The US granted Nvidia export licenses for the H200 to China in January, and the Trump administration loosened restrictions further after that. Yet Commerce Secretary Howard Lutnick says Nvidia still hasn't sold a single one. Chinese customs authorities are blocking imports, and Beijing's regulators are steering buyers toward domestic silicon from SMIC and Huawei.

So Nvidia is taking the off-ramp on its own terms. The existing 250,000 units aren't being scrapped — they're sitting in inventory, ready if the deadlock ever breaks. But no new wafers are going toward them, and TSMC's lines are now aimed at a platform with real buyers lined up.

Meanwhile, US officials are still haggling over a trade that isn't happening. The administration is weighing a 75,000-unit cap on advanced chip sales to each Chinese customer for Nvidia and AMD, according to Bloomberg. About ten Chinese firms already hold licenses under that rough ceiling — Alibaba, Tencent, ByteDance and JD.com among them — with Lenovo and Foxconn cleared as distributors.

None of it matters much if the chips can't land. Beijing has told Chinese companies to keep any approved H200 use overseas, which kills the point for buyers who wanted the hardware in their own data centers. Licenses exist. Customers exist. Zero chips have shipped.

That's a deliberate choice in Beijing, not a loophole snapping shut. China would rather run its AI on slower domestic chips than lean on an American part it can't fully control, export terms and all — a bet on self-sufficiency over raw performance. Nvidia, for its part, didn't need much convincing: Vera Rubin is expected to hit volume production in the second half of 2026, and demand from US, Gulf and European hyperscalers is already outrunning what TSMC can supply. Every wafer freed from an unsellable H200 goes to a customer who'll actually take delivery.