Nvidia has reduced the number of Asian customers to which it sells its chips by more than half to stop its chips entering China, according to the Financial Times.
Citing three people with knowledge of the matter, the publication reported Nvidia has created a new “white list” of approved companies that have cleared more stringent compliance checks.
Nvidia’s increased scrutiny of its Asia customers is intended to reduce the movement of its chips into China via third countries. It is understood that less than half of Nvidia’s previous customers were approved for the new list, which covers countries such as Japan, Malaysia, and Singapore.
The sources added that the changes primarily affect neocloud companies, specialised cloud providers focused on offering AI infrastructure as a service, with Nvidia staff now manually verifying contracts, visiting data centres, and interviewing customers to confirm they are legitimate.
Restricting China’s access to advanced AI chips is a core focus of the US government. The FT added that the Department of Commerce is assisting Nvidia with its new vetting efforts.
The process of training AI models is heavily dependent on Nvidia’s proprietary computing platform Compute Unified Device Architecture.
Without access to the latest chips, Chinese AI firms have resorted to making their model architectures as efficient as possible to train trillion-parameter models on older Nvidia chips and those made by domestic firms such as Huawei.
But US officials have repeatedly claimed that more advanced Nvidia chips are still making their way into the Chinese black market.
In February a senior Trump administration official told Reuters that the latest model by Chinese AI firm DeepSeek, V4-Pro, was trained on Nvidia’s most powerful chip, the Blackwell. Nvidia declined to comment on the allegation.
In March, the co-founder of the IT company Super Micro Computer were charged by the Department of Justice for allegedly conspiring to smuggle $2.5 billion worth of Nvidia chips into China via Taiwan.
Earlier this month, Singaporean prosecutors filed new charges against suspects in a case involving the export of controlled Nvidia chips to China.
Alongside the alleged use of Nvidia chips to train AI models, Chinese labs are also flocking to easily accessible domestic chips, driven in part by the Chinese government’s protection of home-grown companies.
Recent months have seen Beijing step in to spur domestic chip uptake, with the Chinese government banning the import of Nvidia’s China-approved H200 chip in January.
In May, the FT reported that the Chinese tech giant Huawei expects its AI chip sales to grow by at least 60 per cent this year as domestic demand accelerates. One of its flagship offerings is the Ascend 950PR, a chip that outperforms the Nvidia H20 chip which is the firm’s US government-approved offering in China.
To date, DeepSeek has used Huawei’s Ascend 910C chip for inference, the task of actually running AI models and processing their inputs and outputs. But on 8 July, Reuters reported that the firm is working on its own in-house AI chip to reduce reliance on both Huawei and Nvidia.
