The loophole Chinese tech giants are using to train AI with restricted NVIDIA GPUs

Chinese technology companies have devised an ingenious way to circumvent Trump’s AI Chip export controls, opting to train their most powerful Artificial Intelligence models in data centers located outside of mainland China.

According to a new FT Report, major players like Alibaba and ByteDance have been leasing significant computing power in facilities across Southeast Asia, primarily in Singapore and Malaysia. This offshore training allows them to access high-end NVIDIA GPUs, such as the A100 and H100 platforms, which are restricted from direct sale to China under current U.S. regulations.

The offshore loophole

The shift highlights the adaptability of Chinese firms and a major loophole in the current export control regime.

The U.S. restrictions are primarily focused on controlling the physical sale and deployment of advanced chips within China's borders. By leasing compute capacity from non-Chinese entities that own and operate the data centers abroad, Chinese companies can utilize the necessary hardware for the most resource-intensive phase of AI development, model training, without violating the law.

The sources revealed that models such as Alibaba's Qwen and ByteDance's Doubao have been trained, at least in part, using these overseas GPU clusters.

The Chinese companies are reportedly employing a split strategy: performing the heavy training workloads offshore to leverage Nvidia's superior power, and then deploying the finished model weights back in China for inference (user interaction) using domestically sourced chips, including those from Huawei.

A Singapore-based data center operator noted, "Coming here is an obvious choice. You need the best chips to train the most cutting-edge models, and it's all legally compliant."

Geopolitics and the AI race

This circumvention strategy complicates Washington's efforts to slow China's technological progress. While the export controls have forced Chinese companies to expedite the development of domestic alternatives, the move abroad ensures they can maintain a global competitive footing in the rapidly advancing field of generative AI.

The practice has fueled a rapid expansion of data center capacity in Southeast Asia and other neutral regions, making these areas critical battlegrounds in the global tech rivalry.

Looking forward

Despite the move overseas, Chinese firms still face limitations. Domestic laws prohibit moving certain sensitive or private data out of the country, meaning that any AI model customization based on specific client data must still be performed inside China, relying on less-advanced domestic hardware.

The U.S. government has continually refined its export controls to close loopholes, but the latest reports suggest that the global and interconnected nature of the cloud computing market presents an ongoing challenge to enforcement.