According to the Financial Times, major Chinese technology companies are relocating parts of their AI model training to data centers in Singapore and Malaysia to retain access to Nvidia H20 chips restricted within China.

Models and deployment

FT reports that foundation models including Qwen and Doubao are trained outside China using external computing clusters, enabling continued use of advanced Nvidia hardware.

Companies accomplish this by operating on clusters that are formally not owned by the Chinese firms, which allows training workflows to run where export controls do not prevent chip access.

Exception: DeepSeek

One notable exception is DeepSeek, which, according to FT, had previously accumulated its own Nvidia clusters and therefore did not need to relocate core training tasks overseas.

Context

The moves respond to U.S. export restrictions that limit access to certain high-end accelerators inside China, and firms are adapting infrastructure and ownership arrangements to preserve performance for large-scale model training.

The Financial Times provides the primary reporting for these operational shifts and the specific examples of models and firms involved.