• Alibaba and ByteDance are training their next-generation LLMs at data centers in Southeast Asia to access high-end Nvidia chips — something they struggle to do domestically due to US export restrictions.
  • This trend intensified after April 2025, when the Trump administration tightened the sale of the H20 line (a China-only version). Due to the requirements for training state-of-the-art models, “going offshore” is seen as a necessary and legal option.
  • Data center clusters in Singapore and Malaysia are booming due to demand from Chinese businesses; many centers are equipped with Nvidia chips of quality comparable to those used by US Big Tech.
  • Chinese enterprises often sign contracts to lease infrastructure operated by non-Chinese third parties — complying with US export regulations after the “diffusion rule” was abolished by President Trump.
  • Alibaba’s Qwen model and ByteDance’s Doubao model have risen to the ranks of the world’s leading LLMs; Qwen is also widely used by the international community due to its “open-source” nature.
  • An exception: DeepSeek still trains domestically because it had accumulated a large number of Nvidia chips before the ban and is collaborating with Huawei to optimize a new generation of AI chips. Huawei is sending engineers to work directly at DeepSeek’s headquarters in Hangzhou, viewing this as a strategic project to boost the domestic semiconductor-software ecosystem.
  • Due to the massive computational requirements of LLM training, Nvidia chips remain the most preferred; meanwhile, Chinese chips are increasingly used for the inference stage — the part that accounts for a large proportion of the AI workload.
  • Chinese enterprises also use Southeast Asian data centers to serve international customers and expand their cloud market share, simultaneously making inroads in the Middle East.
  • However, they are prohibited from taking private data of Chinese customers abroad, so projects requiring training on internal data must be conducted domestically.

📌 Alibaba and ByteDance are moving AI training to Singapore–Malaysia to access Nvidia chips amid US export restrictions. Qwen and Doubao are becoming leading models; DeepSeek is the exception, training domestically thanks to accumulated Nvidia chips and collaboration with Huawei. Chinese enterprises also use Southeast Asian data centers to serve international customers.

Share.
© 2025 Vietmetric
Exit mobile version