The U.S.-China battle over advanced semiconductors has entered a new phase. Despite Washington’s tightening export controls on Nvidia’s most powerful AI chips, Chinese tech giants like Alibaba and ByteDance are quietly training their cutting-edge models overseas, accessing restricted hardware through data centers in Singapore, Malaysia, and other Southeast Asian hubs.
This offshore strategy has become a critical lifeline. Since the U.S. expanded bans in early 2025 to include even downgraded chips designed for the Chinese market, domestic training on top-tier Nvidia GPUs became nearly impossible. Rather than slow down, companies shifted their heaviest computational workloads abroad, where Nvidia hardware remains freely available.
Alibaba has been particularly aggressive. Its Qwen series of large language models, now among the strongest open-source alternatives globally, relies heavily on foreign clusters for pre-training. ByteDance, the parent company of TikTok and one of Nvidia’s largest former customers in China, follows the same playbook with its Doubao models. Both companies built massive stockpiles before restrictions tightened, but those reserves are finite. Moving training abroad extends their runway significantly.
The workaround is clever but complex. Chinese data laws prevent exporting sensitive user information, so companies split their workflows: foundational model training happens overseas on unrestricted Nvidia systems, while fine-tuning and real-time inference shift back to domestic servers using Chinese-made chips from Huawei, Cambricon, or Biren. This hybrid approach keeps development speed high while gradually building independence from Western hardware.
Southeast Asia has become the unexpected winner. Data center construction is booming in Singapore and Malaysia as Chinese firms sign long-term leasing deals. Some facilities now rival the scale of major U.S. cloud providers. At the same time, Beijing is pouring resources into its semiconductor industry, with state-backed projects aiming to close the gap within years rather than decades.
For Nvidia, the situation is bittersweet. The company continues to dominate global AI infrastructure, but its China business—once a major growth driver—has been severely curtailed. Even compliant chips face delays and restrictions, pushing Chinese developers to explore alternatives faster than anticipated.
The bigger picture reveals a deepening technological divide. Instead of one global AI ecosystem, two parallel universes are emerging: one powered by Nvidia and Western cloud giants, the other built on Chinese hardware and increasingly sophisticated domestic software stacks. Developers in China still prefer Nvidia’s mature tools when they can get them, but the gap is narrowing quickly.
This isn’t just about chips—it’s about who controls the future of artificial intelligence. U.S. restrictions have undeniably slowed China’s access to the very best hardware, but they’ve also accelerated its push toward self-sufficiency. Offshore training buys time. Massive state investment buys capability. And determination buys momentum.
In the end, export controls may delay China’s AI ambitions, but they haven’t stopped them. If anything, they’ve lit a fire under an industry already known for moving fast. The race isn’t over—it’s just moving to new territory, one data center at a time.
