American regulators have closely begun examining a practice that allows Chinese artificial intelligence developers to circumvent export sanctions on advanced Nvidia GPUs. The essence of the scheme is radically simple: instead of physically importing processors into the PRC, companies rent computing power in data centers outside the country, maintaining remote access to the hardware.
The equipment may be located in Malaysia, Thailand, or Indonesia. Chinese engineers connect to the servers remotely, using their performance to train highly complex models. This creates a legal vacuum that Washington is now trying to close.
The Mechanics of Circumvention and the First Signs
Since 2022, the United States has consistently tightened export restrictions on the most powerful AI chips. However, current rules regulate the supply of equipment specifically, not remote access to it. Chinese technology corporations are actively exploiting this loophole. Analysts at the Carnegie Endowment note that even giants such as Alibaba and ByteDance have obtained computing resources in this manner.
The example of Alibaba's collaboration with the Singaporean company Megaspeed is telling. Formally, the Chinese firm gains access to Nvidia chips installed in Malaysia. However, the ownership chain is deliberately convoluted: the scheme involves a structure from the Cayman Islands that ultimately controls Alibaba. This allows the Chinese side to deny direct supplies, while the American side can only record indirect use of the technology.
Kimi K3 as a Catalyst for Conflict
This topic gained particular resonance after the release of the open model Kimi K3 from Moonshot AI. American officials, including Michael Kratsios, director of the White House Office of Science and Technology Policy, stated that the startup allegedly used servers with Nvidia GB300 in Thailand for training. No public technical evidence was presented, and the Chinese side called the accusations slander. Notably, to optimize GPU cores for Kimi K3, developers did use Nvidia H200, which fuels interest in their operational activities.
A similar story previously arose around DeepSeek. Representatives of the U.S. State Department claimed that the startup attempted to use shell companies in Southeast Asia to gain access to American chips. Nvidia then countered that DeepSeek used legally purchased H800, which only underscores the complexity of controlling end-use.
Legislative Response and Nvidia's Position
To close the loophole, the U.S. House of Representatives approved the "Remote Access Security Act." The document proposes extending export controls to cloud computing and the use of technology over the internet. If the law is passed by the Senate, restrictions will apply not only to physical supplies but also to the ability to utilize chip computing power from China.
Nvidia strongly opposes such an expansion. The company reminds that current rules allow the creation of cloud infrastructure outside controlled jurisdictions, and its partners undergo strict vetting. The manufacturer also warns that sanctions have already deprived it of its second-largest sales market, and costs associated with H20 licensing are estimated at $5.5 billion. Notably, in January 2026, Washington nevertheless allowed supplies of the more powerful H200 to select clients in the PRC under certain conditions, which points to the inconsistency of the containment policy.
My comment: This situation demonstrates the fundamental vulnerability of export controls in the digital age. As long as chips remain a physical commodity, they can be restricted, but once computing power becomes a service, geography loses its significance. The United States is trying to patch holes in legislation, but the AI arms race has long since shifted into the realm of a "cloud war," where the winner will be the one who adapts faster to new realities, not the one who tightens the screws more rigidly.