The models of Alibaba and Bytedance bypass the US sanctions by renting data centers

Chinese tech giants like Alibaba and ByteDance have found a way to get access to Nvidia’s most powerful chips. Despite tough US sanctions banning direct sales of high-end graphics processors to China, these companies rent data centers in Singapore and Malaysia to train their AI models powered by American chips.

Big Chinese language models such as Qwen (owned by Alibaba) and Doubao (owned by Bytedance) have recently shined in global benchmarks and become serious competitors to Western models. According to the report of the Financial Times, an important part of the training of these models has been done in clusters located in Southeast Asia; Where the data centers are equipped with the latest hardware. The main reason for this is that the United States has banned the export of powerful chips such as Nvidia’s H100 and A100 to China.

Renting data centers in Singapore and Malaysia by Chinese companies

Chinese companies are currently using this legal loophole. Although exporting chips to China is prohibited, leasing processing power from foreign data centers (even if the end user is Chinese) is free under current rules. In May 2025, a law called the “Artificial Intelligence Release Act” that was intended to stop this work was repealed. This means that as long as the physical hardware does not enter Chinese territory and its ownership is in the hands of non-Chinese companies, there is no obstacle to using it.

Although the basic model training takes place outside of China, there is one big issue: China’s own strict rules. The Chinese government does not allow the release of private user data and sensitive information. Therefore, companies have to train the initial model externally, but the “fine-tuning” process is done with internal data on servers inside China and using weaker chips (including from Huawei).

Among the Chinese giants, Dipsik has followed a different path. The Shanghai-based company stockpiled a large stockpile of Nvidia chips before the sanctions began and continues to train its models inside China. It is also said that Dipsic, in cooperation with Huawei, is optimizing its models to run on native chips so that it does not depend on foreign hardware in the future.