Photo by Nvidia has stopped making AI chips meant for China and is redirecting that capacity into its next platform, Vera Rubin, betting that regulation in Washington and Beijing will keep blocking China sales.
A report from Financial Times alleges that Nvidia has moved manufacturing slots at Taiwan Semiconductor Manufacturing Company (TSMC) away from the H200 and toward Vera Rubin. Foundry time is limited, so the swap signals Nvidia is not counting on meaningful H200 volume in China soon.
The company has faced months of uncertainty around US export approvals and the risk of Chinese restrictions.
The H200 is an older Nvidia AI processor and has been presented as compliant with US export controls. Vera Rubin, unveiled earlier this year, is the newer architecture, built for heavier AI systems that need faster compute, more memory bandwidth, and better scaling across clusters. Demand is strong from US tech groups like OpenAI and Google.
In Washington, the Trump administration has been talking about limiting Chinese companies to buying 75,000 H200 chips each.
The same per-customer cap would also include Advanced Micro Devices’ MI325 accelerators because they offer similar capability. These accelerators are used to build and run artificial intelligence models.
Vera Rubin gets priority as Trump and Xi near talks
Even with caps, total shipments into China could still reach as many as one million units. Most applications come from a small set of Chinese tech giants, so per-buyer limits would squeeze the totals.
Under that structure, those companies could collectively receive only hundreds of thousands at most. The 75,000 limit is less than half of what companies like Alibaba and ByteDance privately told Nvidia they wanted to buy.
The next few weeks matter because Trump is planning a meeting with Xi. The US president wants a deal that allows H200 exports to Chinese companies classified as nonmilitary. Enforcement stays tricky because advanced chips can be redirected after arrival.
Technically, the H200 is the most powerful chip from Nvidia’s previous generation. It was the industrial standard for training and operating AI software like ChatGPT until Nvidia debuted the current Blackwell line last year. It delivers about six times the computational capability of what Trump’s team had previously cleared for China, and it beats anything Huawei can make.
Beijing rejected earlier efforts to export Nvidia’s less-advanced H20, even though AMD was able to sell some units of an equivalent processor. Trump also weighed Blackwell shipments, then decided against them for now after senior advisers pushed back, leaving H200 as the compromise.
In a February congressional hearing, Commerce Secretary Howard Lutnick said enforcement would rely on detailed license terms Nvidia must follow and declined to state whether he trusts China to comply.
Last week, CFO Colette Kress said small China approvals have brought in no revenue yet, and Nvidia does not know if any imports will be allowed into China.
