Chinese tech companies have started training their newest artificial intelligence models in Southeast Asia. The Financial Times reported that Alibaba and ByteDance now use data centres in Singapore and Malaysia to get access to top Nvidia chips. Reuters reported that these companies want to work around United States measures that block direct access to advanced hardware.
The Financial Times said that these companies use lease agreements from operators that are not Chinese. These operators own and run the data centres. The report said there has been an increase in training activities in overseas locations after the United States blocked the sale of the H20 chip in April.
What Do The Companies Get Training From Overseas?
The Financial Times said Alibaba’s Qwen and ByteDance’s Doubao models have climbed into the leading group of global benchmarks. Both models were trained, at least in part, on Nvidia accelerators outside China. Singapore based operators told the Financial Times that requests from Chinese companies rose after the Trump administration brought in tighter rules on the H20 chip.
Those rules would have stopped companies from leasing overseas compute. A later policy change removed that plan, which allowed the overseas training to carry on. The Financial Times report explained that United States export controls stop Nvidia from selling its most advanced chips directly to China, but overseas leasing still fits under the law.
China also blocks foreign AI chips from its state financed data centres. This leaves overseas training as the option that fits both sets of rules. The Financial Times said the trained model weights can then be used inside China on local chips once they return home.
How Do The Rules Influence This Behaviour?
The Financial Times report said a notice in May 2025 cancelled the earlier plan known as the AI diffusion rule. This plan would have treated the overseas leasing of compute as a form of indirect breach of the export ban. Once the notice removed that idea, companies could use H100 and A100 class accelerators outside China as long as the operator of the hardware met the conditions set under United States rules.
This created space for companies like Alibaba and ByteDance to build large models with performance levels that match Western labs. Chinese companies then run those models inside the country using Huawei chips and chips from other local suppliers. The Financial Times said these local chips now handle a growing share of the day to day running of models.
DeepSeek stands out as a different case. It gathered a large supply of Nvidia chips before the bans and still trains inside China. The Financial Times reported that it works with Huawei so it can adjust and improve future chips made in China.
What Does This Do To China’s AI Sector?
The Financial Times said these overseas arrangements allow companies to carry on training powerful models even when direct access to Nvidia’s best chips went away. This gives them a way to keep development going under the current rules. Reuters reported that Alibaba, ByteDance, DeepSeek and Huawei did not respond to requests for comment.
The Financial Times said these arrangements help companies reach training goals that match Western labs without breaking current United States rules. This also allows China based teams to use local chips for user facing tasks once the trained model weights return home.
One result is that Southeast Asian data centres have become important for companies that want to keep building large models. The Financial Times report said operators in that region have noticed growing interest from China. This interest has changed the flow of training work toward Singapore and Malaysia.
