China bypasses US chip curbs using overseas data centers and smuggling

The GPU bottleneck

Nvidia's graphics processing units, or GPUs, have become the essential engine of the modern AI era. While these chips were originally designed to render 3D graphics for video games, their ability to perform massive amounts of simultaneous calculations makes them perfect for training large AI models. The 2022 launch of ChatGPT sparked a global investment boom, making these processors the most sought-after components in the industry.

Because of this demand, Nvidia has reached a market valuation of approximately $4.5 trillion, placing it alongside Apple as one of the world's most valuable companies. This high value reflects the fact that access to advanced computing power is now a primary driver of national and economic competitiveness. Consequently, the semiconductor supply chain has become a central pillar of the ongoing technology rivalry between the United States and China.

To protect its technological lead, Washington has imposed strict export restrictions on Nvidia's most powerful products, including the Blackwell and Rubin platforms. The goal is to prevent Beijing from using high-performance chips for potential military or national security applications. However, these restrictions are creating a complicated landscape for Nvidia, which faces the challenge of balancing regulatory compliance with the desire to maintain a major market.

Remote access and cloud loops

One primary way Chinese companies are obtaining restricted power is by renting computing resources from data centers located in third countries. Instead of physically importing the chips into China, companies can access the servers remotely from within Chinese borders. This allows them to use the high-end processors while the hardware remains physically situated in a country where its presence is permitted.

Industry analysts note that US export controls primarily restrict where a chip is shipped rather than who is allowed to use its computing capacity. For instance, a data center in Malaysia's Johor region could house Nvidia GPUs that are accessed by a customer in Shenzhen. This geographical separation creates a significant hurdle for regulators who are trying to monitor the actual end-use of the technology.

Reports have already surfaced regarding specific companies using these methods, with some allegations suggesting that firms like Moonshot AI accessed high-end chips in Thailand. While the legality of these arrangements depends on the specific conditions governing the cloud services, they provide a viable workaround for Beijing. This trend highlights the difficulty of policing a decentralized digital infrastructure where hardware and users are often separated by thousands of miles.

Smuggling and dismantling tactics

Beyond remote cloud access, there are reports of more direct and clandestine methods to move hardware across borders. One notable allegation suggests that the AI firm DeepSeek developed a new model using thousands of chips that were dismantled overseas before being transported into China. By breaking down the hardware, it may become easier to move restricted components past customs and security checkpoints.

Nvidia has responded to these specific reports by stating they have not seen evidence of such smuggling and described the method as appearing far-fetched. Nevertheless, the possibility of physical smuggling remains a major concern for US officials who want to ensure that high-end chips do not reach Chinese laboratories. These allegations add a layer of complexity to the security debate, as establishing the origin and final use of smuggled parts is difficult without detailed evidence.

The situation is further complicated by the fact that Chinese AI laboratories are also being accused of training systems specifically to imitate the capabilities of American models. This creates a dual challenge for the US: preventing the flow of hardware while simultaneously trying to protect the intellectual property of its software. As smuggling networks potentially evolve, the effectiveness of physical export controls may continue to be called into question.

The rise of domestic hardware

In response to these restrictions, China is investing heavily in its own domestic semiconductor industry to reduce its reliance on American suppliers. Huawei has emerged as a primary competitor in the AI chip market, though its products currently face challenges regarding performance and technology compared to Nvidia's top-tier offerings. Some Chinese firms are already beginning to plan their next major AI models using these locally manufactured chips.

To overcome the performance gap of domestic chips, Chinese companies may use a strategy of deploying larger numbers of processors to achieve the same results as a single high-end Nvidia chip. While this approach requires more electricity and more complex engineering for cooling and networking, China's relatively inexpensive electricity and state subsidies can help offset the costs. This strategy allows developers to maintain progress even if they cannot access the most efficient individual components.

Huawei has recently accelerated its timeline, announcing that its next-generation AI chip will be ready by the first quarter of 2027, which is three quarters earlier than originally planned. This move signals a long-term shift toward a self-sufficient Chinese AI ecosystem. As domestic chips improve rapidly and collaboration between Chinese tech firms grows, the reliance on Western hardware may decrease, regardless of how effective the current export bans are.