China's tech giants, led by Alibaba and ByteDance, are shifting the training of their most advanced large language models (LLMs) to data centers abroad. The strategic move lets them keep using Nvidia's powerful GPUs while staying within the legal framework of US export control rules. It has turned Southeast Asian hubs such as Singapore and Malaysia into key centers for China's AI industry. The workaround became necessary after Washington tightened controls on high-end AI accelerators, including the H20 chip designed for the Chinese market.
Southeast Asia: A Destination for Nvidia Computing Power
Chinese companies have found a significant loophole in US regulations. The law bans the direct export of advanced Nvidia GPUs (such as the related H100 and A100 lines) to China, but it does not stop non-Chinese data center operators in other countries from buying those chips and renting out access as a cloud service. Chinese AI labs are making full use of this mechanism by signing long-term rental contracts with providers overseas.
The shift was reinforced when the Trump administration rescinded the AI "diffusion rule" before it could take effect. The rule had been expected to close this rental loophole. As a result, Chinese companies can still use dense computing clusters with high-bandwidth interconnects (comparable to those of leading US AI labs) to train models such as Alibaba's Qwen and ByteDance's Doubao.
There is one exception, however. DeepSeek accumulated enough Nvidia accelerators before the latest ban took effect, which allows it to keep running full-scale training jobs at home. DeepSeek is also working closely with domestic chipmakers such as Huawei to adapt hardware and software, preparing for a self-reliant future.
Alibaba and ByteDance's Dual Strategy: Train Abroad, Run Inference at Home
The approach of China's tech heavyweights is now a dual model: train overseas, run inference domestically. The companies still favor Nvidia's top products for model training because of their stability, software ecosystem, and superior performance, which reduce the engineering effort needed to scale up to hundreds of billions of parameters.
Once the models are trained, by contrast, Chinese companies increasingly rely on domestic chips for inference, the stage that handles production traffic and real-world queries. Chinese vendors are rolling out accelerators optimized for specific memory bandwidth and power levels to handle this production traffic at lower cost and with less geopolitical risk. The facilities in Singapore and Malaysia serve as temporary "training factories," with equipment rented from foreign operators to stay in compliance with US law.
