Chinese Tech Firms Offshore AI Training to Access Nvidia Chips Amid US Restrictions

Chinese tech giants Alibaba and ByteDance are increasingly training their large language models in Southeast Asian data centers to circumvent US export controls on high-performance Nvidia chips. The practice gained momentum after the Trump administration restricted sales of Nvidia's China-specific H20 semiconductors in April.

Singapore and Malaysia have seen rapid expansion of data center clusters equipped with Nvidia's advanced chips, mirroring the infrastructure used by US tech firms for AI development. "You need the best chips to train cutting-edge models, and it's all legally compliant," noted a Singapore-based data center operator.

Alibaba's Qwen and ByteDance's Doubao have emerged as top-performing LLMs globally over the past year, with Qwen's open-source model gaining particular traction among international developers. While most Chinese firms lease foreign facilities, DeepSeek continues domestic training using Nvidia chips stockpiled before export restrictions took effect.