China eases limits on Nvidia H200 chips as AI race escalates
Stay informed with free updates
Simply sign up to the Chinese business & finance myFT Digest — delivered directly to your inbox.
Small batches of Nvidia’s H200 chips have been allowed to enter mainland China, as Beijing tries to aid its leading AI companies in the race to catch up with US rivals.
ByteDance and Tencent have each received about 10,000 H200 processors in recent weeks, while a few other Chinese tech groups could soon obtain approval for shipments of similar size, according to two people with knowledge of the matter.
The H200 is at least two generations behind Nvidia’s most powerful chips, which Chinese customers are not allowed to buy due to US export controls.
While the Chinese groups have been cleared by the US to buy up to 100,000 H200s apiece, Beijing wants them to keep most of these chips outside the mainland to support domestic chipmakers, one of the people said.
In addition to the small number of H200 chips permitted to enter the mainland, Chinese regulators have told companies that they can ship such processors to Hong Kong, which operates as a territory outside its customs border, and use them there.
The US licences permit shipments of Nvidia’s H200 chips to mainland China and Hong Kong.
Allowing the chips into Hong Kong still presents a challenge to Chinese tech groups as the former British colony lacks enough data centres to house them. That situation is unlikely to change soon as Hong Kong doesn’t have sufficient capacity due to power supply issues.
“It’s a dilemma. Everyone needs the chips but struggles to find a way to use them in Hong Kong,” said a person familiar with the situation. “The hope is for the control to loosen up gradually.”
Nvidia has about half a million H200 chips in stock mainly for Chinese clients. Sales were held up due to Beijing’s restrictions, according to multiple people with knowledge of the situation.
Some Nvidia partners such as Lenovo told Chinese customers last week that they could resume placing orders for their products containing H200 chips, one person said. These partners integrate Nvidia chips with CPUs and other networking kit to sell as AI servers.
Purchases would still require separate approval through an application process overseen by the state planner National Development and Reform Commission (NDRC), according to a person with direct knowledge.
China’s continued import controls reflect its desire to support domestic chipmakers led by Huawei, which aim to dominate the world’s second largest semiconductor market where Nvidia is mostly shut out due to export controls.
However, China’s lack of the most advanced chipmaking equipment will restrict its ability to meet surging domestic demand in the near future, meaning some imports of Nvidia chips will continue to be needed — especially for AI training.
Chinese regulators have recently started loosening restrictions on US chips to ensure that leading domestic AI developers can continue training cutting-edge models. Chinese groups are vying to compete with advanced models such as Anthropic’s Mythos 5, said one of the people familiar with regulatory thinking.
Last month Chinese AI lab Moonshot released its K3 model, whose performance almost rivalled the most advanced US models.
Alibaba, DeepSeek and Z.ai all followed with models of similar intelligence level in a sign that the gap between leading Chinese and US labs has narrowed.
Chinese labs are increasingly using domestic chips for AI inference work that involves generating live responses; most still use Nvidia chips for the more complex training phase to build knowledge and look for patterns, according to people with knowledge of the matter.
Nvidia, ByteDance, Tencent and the NDRC did not respond to requests for comment.
