US mulls capping Nvidia H200 sales at 75,000 per Chinese customer

[HONG KONG] US officials are considering caps on the number of artificial intelligence (AI) accelerators Nvidia can export to any one Chinese company, which would further constrain the chipmaker’s re-entry into a crucial market.

The Trump administration has talked about limiting Chinese firms to buying 75,000 of Nvidia’s H200 chips each, according to sources familiar with the matter. Shipments of Advanced Micro Devices’ (AMD) MI325 chips, which have similar capabilities, would also count towards a customer’s cap, the sources said. These accelerators, a prized commodity in the tech world, are used to develop and run AI models.

Total shipments to China could still reach as many as a million units, the sources said, citing an upper bound that Trump’s team set earlier in the regulatory process. But the lion’s share of current applications come from a small number of Chinese tech giants, which under per-customer caps could collectively receive hundreds of thousands at most. The 75,000 limit is less than half of what firms such as Alibaba Group Holding and ByteDance privately told Nvidia they’d like to purchase.

AMD and the Commerce Department’s Bureau of Industry and Security, which oversees chip export licenses, declined to comment. Nvidia did not respond to a request, while Alibaba and ByteDance did not respond outside of normal business hours.

The per-customer restrictions would complicate an already murky picture of when – and to what extent – the world’s most valuable company can return to the largest market for its products. Nvidia said last week that it still is not getting any data centre revenue from China and does not know whether Beijing will allow imports, even if the US gives permission. Washington has so far authorised only a small number of H200 exports.

China’s government, which must sign off on any Nvidia sales into the country, is balancing AI developers’ desire for best-in-class American semiconductors against years-long efforts to boost use of Chinese-made chips from the likes of Huawei Technologies. US President Donald Trump said in December that Chinese leader Xi Jinping had responded positively to the H200 offer, and Chinese regulators have told companies to start preparing orders.

Much hinges on Trump’s planned meeting with Xi in a few weeks’ time, when the US president is hoping to strike an agreement for H200 exports to nonmilitary Chinese companies, according to sources familiar with preparations for his trip. The White House did not respond to a request for comment.

The H200 is the most powerful chip from Nvidia’s previous generation of products. It was the industrial standard for training and operating AI software such as ChatGPT until Nvidia debuted its current-generation Blackwell line last year. It has six times the computational capabilities of what Trump’s team had previously authorised for sale to China – and is far better than anything Huawei can make.

But Beijing rejected Trump’s earlier attempts to export Nvidia’s less-advanced H20 chip, even though AMD was ultimately able to sell some units of an equivalent processor. Trump then considered Blackwell shipments, but ultimately decided against them, at least for now, at the urging of several senior advisers. That made the H200 something of a compromise.

China hawks in and outside of the Trump administration worry that H200 exports will only help China develop and deploy better AI models, with no benefit to the US. But Nvidia’s Huang convinced the president that such exports would contribute to “a positive economic relationship with China” and be “positive for us overall”, as Commerce Secretary Howard Lutnick recounted on the All-In Podcast in January.

Nvidia’s logic is that keeping Chinese AI companies hooked on American technology will prevent Huawei from building the revenue and developer ecosystem necessary to compete on the global stage.

“There are plenty of people who disagree” with Nvidia’s reasoning, Lutnick said, without expounding on his own position. “But that was Jensen’s argument,” he said, and “this is the president’s deal”.

While Trump’s H200 announcement in December was a major win for Nvidia, implementation details matter. Over the past few months, US officials have added a series of constraints designed to limit what many see as the negative impact of allowing H200 sales, while still following the president’s overarching directive.

Nvidia, meanwhile, has grown increasingly frustrated, accusing bureaucrats of undermining Trump’s vision and at times complaining that US export conditions are so onerous as to dissuade Chinese firms from purchasing H200 chips at all.

Those conditions include a requirement for exporters to assert that chip sales to China won’t cut into availability for American companies, echoing a goal of failed legislation that Nvidia staunchly opposed last year. There’s also a mandate for exporters to certify that Chinese customers will perform “rigorous” due diligence checks to ensure AI chips sold to China do not benefit the country’s military, a concern Nvidia has dismissed.

It’s unclear how Trump’s team will navigate this provision when making license decisions, given that Washington has designated or plans to designate three of China’s four biggest tech companies as Chinese military companies – a claim the firms reject. BIS declined to comment.

Lutnick, asked in a February congressional hearing how he’d enforce restrictions on military use, said that Nvidia “must live with and must hew to” detailed license terms. He was also asked whether he trusts China will do the same. “I will leave that opinion to the president,” Lutnick said.

Still, some officials view volume as the most important constraint. National security hawks notched an early victory in setting an overall limit for China of one million H200s – more than what many officials would like, though far lower than the figure Nvidia originally proposed. But that many chips, if clustered together, would still allow China to build one of the world’s largest supercomputers.

That’s one reason why conversations have also included limits for individual firms. The 75,000-chip total allowed under the potential cap would be enough for a company to stand up a midsize data centre of roughly 100 megawatts.

By comparison, OpenAI aims to build multiple facilities of at least one gigawatt, 10 times larger, as part of its US$500 billion Stargate initiative. A gigawatt is roughly the power capacity of a nuclear reactor.

Another crucial question was whether to allow the use of the chips outside the Asian country. US export controls, first imposed under President Joe Biden, restrict sales of advanced AI chips not just to China itself, but also to Chinese companies operating overseas.

Firms there can rent advanced Nvidia hardware located in other countries – Alibaba, for example, teams up with a South-east Asian company called Megaspeed International for this purpose – but they cannot buy the chips themselves without Washington’s permission.

Those constraints, along with US curbs on equipment exports that have limited China’s domestic AI chip production, have made it hard for Chinese companies to expand their own cloud computing businesses. “When AI chips are in short supply, we prioritise internal use instead of renting it out externally,” Tencent Holdings president Martin Lau said in November. “If there weren’t an AI chip supply constraint, our cloud revenue would be growing more quickly.”

The ability to purchase H200 chips provided a potential opportunity to change that paradigm – and Chinese firms’ export license applications reflected their global ambitions. The majority of H200 license requests were for overseas use, according to sources familiar with the matter, such as an application by Alibaba to deploy H200 chips in Malaysian data centres. Alibaba did not respond to questions on the matter.

But Trump’s team took that off the table. The US rules do not “allow Chinese companies to use this chip to build data centres overseas to compete with American hyperscalers,” White House Office of Science and Technology Policy Director Michael Kratsios told lawmakers in mid-January. AI infrastructure spending plans by US hyperscalers, the largest data centre operators, dwarf those of their Chinese competitors.

It’s unclear how that restriction may affect the total number of H200 chips sold. Alibaba, Tencent and ByteDance did not answer questions about whether they’d apply to buy more H200 chips for domestic use, given that overseas deployment is not an option. BLOOMBERG