NextFin news, Chinese technology conglomerates such as Alibaba and ByteDance have increasingly resorted to training their cutting-edge artificial intelligence (AI) models in overseas data centers located in Southeast Asian countries like Singapore and Malaysia. This shift, reported as recently as November 2025, is a strategic response to stringent U.S. export controls instituted in April 2025 under the Trump administration, which prohibit sales of high-performance Nvidia GPUs—including the H100 and A100 models—to China directly.
Financial Times disclosed via informed industry sources that these companies lease compute resources operated by non-Chinese entities abroad, enabling them to access Nvidia's latest AI accelerators indirectly. This operational model allows them to evade U.S. constraints that restrict onshore GPU sales without violating current export laws, which, following revisions in mid-2025, permit overseas leasing of such hardware by third parties. Consequently, flagship Chinese AI models like Alibaba's Qwen and ByteDance's Doubao leverage these offshore resources for training, achieving performance metrics comparable to top U.S. and international AI research efforts.
Notably, the relocation of training clusters has primarily concentrated in Southeast Asia due to its favorable business environments, data center infrastructure maturity, and geographic proximity to China. Where private and sensitive Chinese user data cannot legally leave national borders, base models trained offshore are then fine-tuned or deployed domestically, often on AI chips supplied by indigenous manufacturers like Huawei. DeepSeek, another Chinese AI player, differs by stockpiling Nvidia chips prior to the embargo and conducting training operations exclusively within China, supplemented by partnerships aimed at advancing domestic AI chip capabilities.
This circumvention technique arises from the complexity of the U.S. export control regime: while prohibiting direct sales and usage within China, it currently lacks enforcement provisions against indirect foreign data center leasing. A previously contemplated Biden-era 'AI diffusion rule' designed to close this loophole was rescinded in May 2025, underscoring evolving U.S. policy ambiguity and the challenges of technological export management in a globalized context.
From an analytical viewpoint, this development reflects a multifaceted geopolitical technology standoff where China adapts rapidly through supply chain resilience and innovative operational strategies. The use of third-party data centers in neutral or allied countries serves as a tactical workaround that sustains China's AI R&D momentum despite strategic export controls. This approach also amplifies Southeast Asia's emerging strategic importance in global tech supply chains, potentially catalyzing increased foreign investment and regional tech ecosystem growth.
Looking forward, the persistence of such overseas training models could incentivize U.S. policymakers to revisit export frameworks to address loopholes that undermine intended technological containment. Conversely, China's parallel advancement of domestic AI chip design and partnerships with firms like Huawei may reduce reliance on foreign GPUs over time, increasing self-sufficiency amid geopolitical pressures. Additionally, restrictions on cross-border data flow will shape hybrid AI development workflows, where offshore training complements domestic fine-tuning, balancing technological innovation with data sovereignty concerns.
In conclusion, the relocation of AI model training by Chinese tech firms to Singapore and Malaysia exemplifies a sophisticated adaptation to U.S. technology embargoes under President Donald Trump's administration. This maneuver secures access to critical Nvidia GPU resources and sustains global competitiveness in AI while exposing the limitations and unintended consequences of current export controls within an interconnected digital economy.
Explore more exclusive insights at nextfin.ai.
