China’s AI champions shift training to Southeast Asia, turning the region into a geopolitical workaround for U.S. chip controls

BEIJING: Beijing’s tech giants are quietly heading offshore to keep training their most advanced artificial intelligence (AI) systems — and they’re doing it without breaking US export rules.

Southeast Asia becomes an unlikely AI hotspot

A growing number of China’s leading tech companies are packing up some of their biggest AI training jobs and sending them abroad. The destinations? Singapore and Malaysia — two places that, until recently, weren’t exactly known as global hubs for cutting-edge AI development.

However, that’s changing fast. People familiar with the shift say firms like Alibaba and ByteDance have begun routing key stages of training their newest large language models to data centres there, where they can access Nvidia’s most powerful chips. Those same chips have become harder to come by back home due to tightening US export restrictions.

The workaround is simple but effective: Instead of buying the graphics processing units (GPUs) directly — which is restricted — Chinese AI labs rent time from foreign-owned data centres that bought the hardware legally. One executive at a Singapore facility put it bluntly: “For Chinese customers who want the best performance and want to stay compliant, this is an obvious choice.”

A US policy that might have blocked this practice, known as the “AI diffusion rule,” was dropped earlier this year before it ever took effect, and with that roadblock gone, these offshore GPU clusters have become essential for Chinese companies trying to stay competitive with American AI leaders.

Overseas compute gives China’s new models a boost

This strategy seems to be paying off. Over the past year, models like Alibaba’s Qwen and ByteDance’s Doubao have climbed into the top tier of global AI benchmarks. Insiders say big chunks of the training work behind these models now happen in Southeast Asia, where the infrastructure — dense racks of Nvidia accelerators hooked together by ultra-fast networking — looks a lot like what leading US labs rely on.

Not every Chinese AI company is heading offshore, though. DeepSeek, known for building high-performing models on lean budgets, managed to stockpile a large number of Nvidia chips before the most recent export restrictions. That early planning means it can still run full training cycles within China. The company is also working closely with Huawei’s chip teams, who have dispatched engineers to DeepSeek’s Hangzhou offices to fine-tune local hardware for future training runs.

Even so, Nvidia remains the gold standard for training massive AI models. Its software ecosystem and performance advantages make it hard to replace — especially when training systems with hundreds of billions of parameters.

However, once the training is done, it’s a different story. More Chinese companies are relying on homegrown chips for inference — the process of running the trained models in everyday applications. Local accelerators come with fewer geopolitical headaches and much lower costs, making them well-suited for large-scale deployment.

Meanwhile, the data centres in Singapore and Malaysia are racing to keep up with demand. They’re expanding GPU capacity, upgrading networking, and tailoring their facilities to support the huge distributed training jobs Chinese firms need, and because these companies usually lease the compute rather than own the hardware, the arrangements stay within the boundaries of US export rules.

In other words, Southeast Asia is becoming a crucial link in the global AI supply chain — and a lifeline for China’s fastest-moving AI contenders.