Chinese tech firms such as Alibaba and ByteDance are increasingly moving parts of their top AI model training to data centers in Southeast Asia. Reports suggest that these sites in Singapore and Malaysia use high-end Nvidia GPUs and are rented from non-Chinese operators.
This approach helps Chinese labs to work around US export limits by leasing compute capacity from foreign providers rather than buying the chips directly. Over the past year, Alibaba’s Qwen and ByteDance’s Doubao have climbed global LLM rankings. Reporters say both models were trained at least partly on Nvidia hardware in overseas clusters.
Demand rose after tighter US measures were put in place in April. Those measures targeted Nvidia’s H20 and similar chips. A proposed rule meant to block overseas leasing, called the “diffusion rule”, was later withdrawn under revised policy.
Current US export controls bar direct sales of the most advanced GPUs to China. China also bans foreign AI chips in state funded data centers. Leasing compute from foreign owned centers outside China is still allowed under the present rules.
A notice in May 2025 removed the proposed Biden era limits known as the “AI diffusion rule”. That means companies may use H100 and A100 class accelerators abroad if the hardware is owned and run by a compliant third party.
ByteDance and Alibaba are the most visible users of this route but not the only ones. This setup lets them train models to compete with Western labs. After training, the model weights are often run inside China for inference on local chips. Chinese firms now increasingly use chips from Huawei and other domestic suppliers to handle user facing workloads.
Leave a comment