Chinese tech giants Alibaba and ByteDance are pivoting to Southeast Asia for training their advanced AI models, Qwen and Doubao, respectively. They're using data centers in Singapore and Malaysia to access high-performance Nvidia GPUs, sidestepping U.S. export controls. This strategy emerged after tighter restrictions on chip exports were imposed by the Trump administration. Following a rollback of the "AI diffusion rule" under the Biden administration, these firms can legally lease computing power from foreign-owned data centers, even though the end users are Chinese. The demand for these overseas resources has surged since April. The existing U.S. export controls prohibit Nvidia from selling its latest GPUs directly to China, while China has banned foreign chips from state-funded data centers. However, leasing from compliant third-party data centers abroad remains permissible. Companies like Alibaba and ByteDance are leveraging this loophole to develop models that can compete with Western counterparts. Once trained, these models can be utilized within China using local chips from Huawei and others. DeepSeek, another Chinese firm, is taking a different approach. It stockpiled Nvidia components before the U.S. ban and still conducts training inside China. The company reportedly uses shell companies to navigate restrictions and has partnered with Huawei to enhance its operations with local silicon. Despite the shift of training efforts abroad, Chinese firms face a significant limitation: they cannot move private data out of China. Fine-tuning or retraining models with Chinese user data must occur domestically, which complicates their AI development strategy.