Chinese tech giants, led by Alibaba and ByteDance, are shifting the training of their most advanced large language models (LLMs) to overseas data centers. This strategic move aims to continue utilizing powerful Nvidia GPUs while complying with the regulatory frameworks of US export control rules. This action has turned Southeast Asian regions like Singapore and Malaysia into key hubs for the Chinese AI industry. This is a necessary solution after Washington tightened controls on high-end AI accelerators, including the H20 chips designed for the Chinese market.
Southeast Asia: A Destination for Nvidia Computing Power
Chinese companies have found a significant loophole in US regulations: while laws prohibit the direct export of advanced Nvidia GPUs (such as the relevant H100 and A100 series) into China, they do not prohibit non-Chinese data center operators in other countries from purchasing those chips and leasing access as a cloud service. Chinese AI labs are fully leveraging this mechanism by signing long-term lease contracts with overseas providers.
This shift is reinforced by the fact that the “AI diffusion rule,” which was expected to close this leasing loophole, was rescinded by the Trump administration before it could take effect. In this way, Chinese companies ensure they can still utilize high-density, high-bandwidth interconnected computing clusters (equivalent to top US AI labs) to train models like Alibaba’s Qwen and ByteDance’s Doubao.
However, there are exceptions: DeepSeek accumulated enough Nvidia accelerators before the latest ban took effect, allowing them to continue running full-scale training workloads domestically. DeepSeek is also working closely with domestic chip manufacturers like Huawei to optimize hardware and software, preparing for a future of self-reliance.
Alibaba and ByteDance’s Dual Strategy: Overseas Training, Domestic Inference
The strategy of Chinese tech giants is currently a dual model: training overseas, inference domestically. Companies still prioritize Nvidia’s advanced products for the model training phase due to their stability, software ecosystem, and superior performance, which minimizes technical effort when scaling up to hundreds of billions of parameters.
In contrast, once models are fully trained, Chinese companies are increasingly relying on domestic chips for inference tasks—the stage of processing production traffic and real-world queries. Chinese providers are rolling out accelerators optimized for specific memory bandwidth and power levels to handle this production traffic at a lower cost and with less geopolitical risk. Facilities in Singapore and Malaysia serve as temporary “training factories,” utilizing a structure of leasing equipment from foreign operators to maintain compliance with US laws.
