Skirting US Curbs: China's Tech Majors Shift AI Training Offshore For Nvidia Hardware

Top Chinese firms are training their AI models overseas in order to access Nvidia's graphical processing unit chips to circumvent the restrictions on their use due to legal mandates, a Financial Times report said.

The US had legally enforced measures that prevented Nvidia from shipping out its most advanced AI chips to its geopolitical rival, with China responding with it's own regulations making it mandatory for Chinese firms to use domestically-produced AI chips.

This has lead to companies like Alibaba and ByteDance training their AI models in Southeast Asia, to use high-end data centres with Nvidia's latest adavanced AI GPUs. According to the report, the firms were leasing the compute systems from operators in Singapore and Malaysia.

The LLMs have allegedly been trained, wholly or in part, through Nvidia accelerators situated in offshore areas.

In the past year, Alibaba’s Qwen and ByteDance’s Doubao large language models has topped the charts in the list of top global LLM benchmarks, according to reports.

Financial Times reported that Singapore-based data centre operators had reported a rise in demand from such firms for their services since April 2025, when the Trump administration had mandated a stricter embargo on Nvidia's H20 chips among others.

Nvidia's H20 chips were GPU chips with less powerful capabilities in accordance with the US' regulations around not supplying cutting edge AI processing hardware to China.

According to the report, the companies would train their domestic models overseas using this hardware with performance targets on par with those of Western countries' AI labs.

The weights would then be run in China for inference on domestically-sourced silicon hardware.