Chinese Tech Giants Shift AI Training Abroad to Navigate US Chip Curbs

Major Chinese technology corporations are reportedly establishing artificial intelligence training operations outside mainland China to secure access to advanced computing hardware restricted by US export controls. By utilizing overseas data centers, these companies aim to maintain competitiveness in the global AI race despite tightening geopolitical constraints.

According to a report by the Financial Times published on November 27, 2025, leading Chinese firms have ramped up efforts to train large language models (LLMs) in offshore facilities, primarily located in Southeast Asia.

Industry heavyweights, including Alibaba Group Holding Ltd. and ByteDance Ltd. , are leveraging data centers in Singapore and Malaysia to access high-end Nvidia Corp. semiconductors. The report indicates that this strategic shift accelerated after the US administration moved to restrict sales of Nvidia’s China-specific H20 chips in April 2025. By leasing computing power from non-Chinese entities abroad, these companies can legally utilize top-tier hardware to train advanced models like Alibaba’s Qwen and ByteDance’s Doubao, effectively bypassing current US export bans.

This offshore strategy became viable after President Trump scrapped the Biden-era "diffusion rule" earlier this year, which had previously closed loopholes regarding overseas data center usage.

However, not all major players are moving offshore. The report highlights DeepSeek as a notable exception. The company continues to train its cost-effective models domestically, relying on a stockpile of Nvidia chips acquired before the sanctions took effect. Furthermore, DeepSeek is collaborating closely with Huawei Technologies Co., Ltd. to optimize the next generation of Chinese AI chips, with Huawei engineers reportedly stationed at DeepSeek’s headquarters in Hangzhou.

While high-intensity model training is moving abroad, Chinese firms are increasingly adopting domestic chips for "inference"—the process of generating responses from AI models. Nevertheless, data sovereignty remains a significant constraint. Due to strict Chinese regulations prohibiting the export of private domestic data, any model customization requiring local sensitive data must still be conducted within China, limiting the scope of offshore operations.