The U.S. authorities have increased scrutiny on a scheme allowing Chinese AI companies to use advanced Nvidia chips, bypassing direct export restrictions. Instead of importing processors into China, these companies rent computing power from overseas data centers, reports CNBC.
Physically, the equipment remains in locations like Malaysia or Thailand. Chinese developers remotely access it to utilize computing power for training models.
How the Loophole Works
Since 2022, the U.S. has steadily restricted the supply of high-performance AI chips to China. The rules primarily regulate the export of the equipment itself, leaving remote access to processors outside the country as a separate issue for some time.
As a result, Chinese firms have started renting servers with Nvidia chips in Southeast Asia. According to Carnegie Endowment, companies like Alibaba and ByteDance have obtained computing power this way.
The U.S. Bureau of Industry and Security is now examining these schemes more systematically. The agency is compiling separate lists of countries through which banned chips are believed to physically enter China and those where Chinese companies gain remote access. The latter option does not always violate current export rules.
One example is Alibaba’s collaboration with Singapore-registered Megaspeed. The Chinese firm accesses Nvidia chips located in Malaysia. However, the agreement is not direct: a Singaporean entity linked to a Cayman Islands company, ultimately owned by Alibaba, is involved in the chain.
Kimi K3 Intensifies the Debate
The issue of China circumventing restrictions gained attention from U.S. authorities following the release of Kimi K3 by China’s Moonshot AI.
In July, Michael Kratsios, Director of the White House Office of Science and Technology Policy, stated that the startup acquired servers with Nvidia GB300 and accessed such processors in Thailand, presumably for training its models. He did not provide public technical evidence.
Kratsios also accused Moonshot AI of using Anthropic Fable’s capabilities in creating Kimi K3. The Chinese side called these claims defamatory.
A few days earlier, the project introduced an open model with 2.8 trillion parameters. In certain tests, it performed on par with leading American systems. Developers used Nvidia H200 among other resources to optimize GPU cores.
A similar situation arose with DeepSeek. In June 2025, a U.S. State Department representative told Reuters that the startup attempted to use shell companies in Southeast Asia to bypass export controls and sought access to data centers in the region to remotely work with American chips.
Nvidia responded at the time that DeepSeek used legally acquired H800 chips.
Congress Proposes Extending Control to Cloud Services
The “Remote Access Security Act” aims to close the loophole. The House of Representatives approved the bill in January this year.
The document extends export control to the use of technologies via the internet and cloud services. This means restrictions will apply not only to the physical supply of chips but also to the ability to use their computing power from China. The bill still needs to pass the Senate.
Nvidia opposes further expansion of restrictions. The company points out that current rules specifically allow for the creation of cloud infrastructure outside controlled jurisdictions and that its partners undergo checks.
The manufacturer also warned that export bans have already effectively deprived it of its second-largest commercial market and are helping competitors from other countries strengthen their positions.
In April 2025, the U.S. introduced licensing for the H20 processor, which the company specifically developed for China following previous restrictions. Nvidia estimated the related costs at approximately $5.5 billion.
Later, Washington eased its approach several times. In January 2026, authorities allowed the supply of more powerful H200 chips to certain Chinese clients under specific conditions.
