U.S. export controls push Chinese AI training offshore

According to the Financial Times, groups like Alibaba and TikTok owner ByteDance are renting computing power in overseas data centers, mainly in Southeast Asia. On paper, these facilities belong to foreign entities, which allows the companies to skirt U.S. export restrictions on advanced GPUs going to China. These graphics processors remain the key resource for training large language models and other cutting-edge AI systems.

Earlier this year Washington tightened export rules again and restricted shipments to China even of the “cut-down” Nvidia H20 chips that were designed specifically for the Chinese market. As a result, several Chinese AI projects ran into a shortage of computing power, and prices for available GPUs climbed.

Against this backdrop, some players are trying to reduce their dependence on foreign suppliers. FT reports that DeepSeek stocked up on Nvidia chips in advance, before the latest bans came in, and is working with local manufacturers, including Huawei-linked firms, on its own solutions for AI training. The companies themselves are not publicly commenting on the situation.

It is also telling that the scramble for chips coincides with Chinese AI models joining global financial experiments. In November, the Alpha Arena platform launched Season 1.5 of its AI-driven algorithmic trading competition, where eight language models, including DeepSeek and Qwen 3 Max, trade crypto with $10,000 in capital each.

The result is a paradoxical picture. On the one hand, Chinese developers are forced to move AI training abroad just to keep access to Nvidia hardware. On the other, these same models are already taking part in global financial experiments and competing with Western systems on international markets. Export controls are turning not only into a story about geopolitics and tech sovereignty, but also into a factor that influences how future AI-driven profits will be distributed in the global financial system.