Ever feel like you're playing a high-stakes game of hide-and-seek, but instead of finding a friend, you're trying to get your hands on the world's most coveted AI chips? That's precisely the game Chinese tech companies are playing, and the stakes couldn't be higher. But how are they managing to do it?
The Essentials: AI Training's New Geography
Chinese tech giants like Alibaba and ByteDance are increasingly turning to data centers in Southeast Asia to train their advanced AI models, according to recent reports. This strategic move allows them to bypass U.S. export restrictions on high-performance Nvidia chips, which are essential for developing powerful AI. By leasing data center capacity in countries like Singapore and Malaysia, these companies can access the computing power they need without directly violating U.S. regulations. It's like needing a rare spice for your dish, but instead of buying it directly (which is forbidden), you rent a kitchen in another country where it's readily available.
This workaround has emerged as a direct response to U.S. export controls, particularly those implemented in April 2025, which restrict the sale of Nvidia's H20 chips to China. These controls aim to slow China's technological and military advancement by limiting access to critical semiconductor technology. The strategy is considered legally compliant because the hardware remains outside Chinese territory. Alibaba's Qwen model and ByteDance's Doubao model are among those reportedly being trained in these offshore locations. According to The Guardian, the initial export ban on Nvidia's H20 chips was estimated to potentially impact Nvidia's revenue by $5.5 billion. Is this a calculated risk or a necessary adaptation for survival in the AI race?
Beyond the Headlines: A Calculated Circumvention
The move to train AI models offshore isn't just a simple workaround; it's a strategic adaptation to a rapidly changing geopolitical landscape. By leasing data center capacity in Southeast Asia, Chinese companies maintain access to Nvidia's high-end GPUs like the H100 and potentially even the B200 Blackwell, which are crucial for training large language models (LLMs). This ensures they can keep pace with global AI development despite the restrictions.
How Is This Different (Or Not): Echoes of the Past
This isn't the first time companies have sought creative solutions to navigate trade restrictions. Throughout history, businesses have adapted to tariffs, embargoes, and other barriers by finding alternative routes, establishing foreign subsidiaries, or developing substitute products. What makes this situation unique is the speed and scale at which it's happening, driven by the relentless pace of AI development and the strategic importance of semiconductors.
While some companies, like DeepSeek, managed to stockpile Nvidia chips before the export bans, others are focusing on developing domestic alternatives. China is encouraging the use of chips from local players like Huawei, Alibaba, MetaX and Cambricon. One Chinese startup even claims its GPTPU chip, dubbed "Chana," can outperform Nvidia's A100 GPU while using less energy and costing less, according to Tom's Hardware. Is this a sign of true innovation, or just wishful thinking?
Lesson Learnt / What It Means for Us
The offshore AI training trend highlights the complex interplay between technology, geopolitics, and economic strategy. It demonstrates the lengths to which companies will go to maintain their competitive edge in the AI era. The U.S. restrictions, intended to slow China's progress, may inadvertently accelerate its push for technological self-sufficiency. Will this cat-and-mouse game between nations spur innovation or simply create more convoluted supply chains?