China’s major tech firms are relocating their AI development overseas to leverage Nvidia’s hardware and circumvent US restrictions on cutting-edge technology. According to two sources with direct knowledge of the matter, more AI model training has been taking place abroad following the US move to limit H20 sales in April.
Alibaba and ByteDance are among the tech groups training their latest large language models in data centers across Southeast Asia. Reports indicate that training in offshore locations has been steadily rising after the Trump administration moved in April to restrict sales of the H20, Nvidia’s China-only semiconductors.
“It’s an obvious choice to come here,” said one Singapore-based data centre operator. “You need the best chips to train the most cutting-edge models, and it’s all legally compliant.”
For the past 12 months, Alibaba’s Qwen and ByteDance’s Doubao models have emerged among the top-performing LLMs worldwide. Qwen’s free and available “open” model has made it widely adopted outside China by developers.
Data centre clusters have expanded rapidly in Singapore and Malaysia, catalyzed by growing demand from China. Many of these data centers are equipped with high-end Nvidia products, similar to those used by US Big Tech groups to train LLMs.
China is also pushing for firms to use locally developed chips
Chinese tech firms generally lease foreign data centers in compliance with US export controls, as the Biden-era “diffusion rule” designed to close this loophole was scrapped by US President Donald Trump earlier this year. But DeepSeek, which had amassed Nvidia chips before the US restricted exports, runs its AI training domestically. DeepSeek is also partnering with domestic chip manufacturers, including Huawei, to develop and refine China’s next-generation AI chips.
Recently, Beijing has also introduced rules mandating that any new data centre receiving government funding must rely solely on locally developed chips. Early-stage data centers must pull out foreign chips or scrap purchase plans, whereas projects already past the 30% completion mark will undergo case-by-case reviews. As a result, Nvidia and its AI chips are largely excluded from a lucrative portion of the market, despite advanced models under US controls still appearing in China via informal channels.
President Trump had first blocked Nvidia from selling its H20 chips in April, resulting in the company incurring billions of dollars in losses. In August, Trump approved the sale of some AI chips by Nvidia and AMD to China, on the condition that the US receives 15% of the proceeds, effectively recasting export controls as a means of leverage. However, the deal raised alarms, with critics raising concerns about security risks and questioning the Trump administration’s approach to handling private businesses.
“You either have a national security problem, or you don’t. If you have a 15% payment, it doesn’t somehow eliminate the national security issue,” said Deborah Elms, head of trade policy at the Hinrich Foundation.
Trump also signaled he might let Nvidia release a limited version of its flagship Blackwell processor in China.
Trump explored easing H200 export curbs for Nvidia
Reports suggest President Donald Trump’s team may permit Nvidia to export its H200 artificial intelligence chips to China. Analysts say the move could be a positive step for Nvidia’s CEO, who has argued for more flexible export policies.
Nvidia noted that current regulations make it impossible to sell a data-centre chip in China, effectively handing the market to its rapidly expanding international competitors. If the H200 chips are allowed to be sold to China, it would signal a substantial softening of US rules aimed at limiting China’s AI advances, but China hawks in Washington are expected to push back.
So far, analysts estimate that the US now holds only a one- to two-year edge over China in AI and semiconductor technologies.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. It's free.
