China’s major tech firms, including Alibaba and ByteDance, are shifting AI model training to Southeast Asia to access Nvidia’s advanced hardware and bypass US export restrictions on H20 chips imposed in April. This relocation ensures compliance while leveraging top-tier semiconductors for cutting-edge large language models.
-
Key drivers include US limits on H20 sales, prompting offshore training in Singapore and Malaysia data centers equipped with Nvidia products.
-
Alibaba’s Qwen and ByteDance’s Doubao models rank among the world’s top LLMs, with Qwen’s open version gaining global developer adoption.
-
Data center expansions in Southeast Asia have surged 25% year-over-year, fueled by Chinese demand and high-end Nvidia chip availability, per industry reports.
Discover how China evades US AI chip bans via Southeast Asia: Alibaba and ByteDance train models on Nvidia hardware abroad. Explore impacts on global tech race—read now for key insights!
What is Driving China’s AI Firms to Relocate Development Overseas Amid Nvidia Restrictions?
China’s AI development is increasingly moving overseas as tech giants seek Nvidia’s high-performance hardware while navigating US export controls. Major firms like Alibaba and ByteDance are training large language models in Southeast Asian data centers to comply with restrictions on advanced chips. This shift, confirmed by sources with direct knowledge, follows the US ban on H20 sales in April, enabling continued innovation without legal violations.
The relocation highlights the global scramble for AI supremacy, where access to superior semiconductors is paramount. Alibaba’s Qwen and ByteDance’s Doubao have solidified their positions as leading LLMs over the past year, with Qwen’s open-source variant attracting widespread international use. By basing operations in regions like Singapore and Malaysia, these companies tap into robust infrastructure while avoiding domestic constraints.
Two sources familiar with the operations noted that offshore training has accelerated since the April restrictions. A Singapore-based data center operator emphasized the appeal: “It’s an obvious choice to come here. You need the best chips to train the most cutting-edge models, and it’s all legally compliant.” This strategy allows Chinese firms to maintain momentum in AI advancements despite geopolitical tensions.
How Are Chinese Tech Firms Complying with US Export Controls in AI Training?
Chinese tech companies are leasing foreign data centers to adhere to US export rules, particularly after the Biden-era “diffusion rule” was eliminated by the Trump administration earlier this year. This compliance enables access to Nvidia’s China-specific H20 chips abroad, where they power intensive AI workloads. Reports from industry insiders indicate a steady rise in such offshore activities post-April, with Southeast Asia emerging as a prime destination.
Data center clusters in Singapore and Malaysia have expanded rapidly, boasting facilities comparable to those used by US tech leaders. Equipped with high-end Nvidia products, these centers support the training of sophisticated models like Qwen and Doubao, which have performed exceptionally in global benchmarks. The growth, estimated at a 25% increase in capacity over the last year, is largely driven by Chinese demand, underscoring the region’s role in the AI ecosystem.
However, not all efforts are offshore. DeepSeek, a prominent Chinese AI developer, continues domestic training by utilizing Nvidia chips stockpiled before export curbs tightened. Partnering with local manufacturers like Huawei, DeepSeek is advancing indigenous chip technologies to reduce reliance on foreign hardware. This dual approach—offshore for immediate needs and domestic for long-term self-sufficiency—demonstrates strategic adaptability in the face of restrictions.
Beijing’s recent policies further emphasize localization. New regulations require government-funded data centers to use only domestically developed chips, excluding Nvidia’s offerings from significant market segments. Early-stage projects must remove foreign chips or abandon plans, while advanced ones face individual assessments. Despite these measures, advanced US-controlled models still enter China through unofficial channels, complicating enforcement.
The economic fallout from these policies has been notable. Nvidia reported billions in losses after President Trump blocked H20 sales in April. Later, in August, Trump conditionally approved limited sales of AI chips by Nvidia and AMD to China, stipulating a 15% US revenue share. This pivot treats export controls as economic leverage but has sparked debate over national security. Deborah Elms, head of trade policy at the Hinrich Foundation, critiqued the arrangement: “You either have a national security problem, or you don’t. If you have a 15% payment, it doesn’t somehow eliminate the national security issue.”
Trump has also hinted at allowing a restricted version of Nvidia’s flagship Blackwell processor for China, potentially easing H200 export limits. Analysts view this as a boon for Nvidia’s CEO, who advocates for flexible policies to compete with rising international rivals. Current rules effectively cede China’s data center market to competitors, but approval of H200 sales could soften US stances on curbing China’s AI progress. Washington insiders anticipate resistance from China hawks, balancing innovation against strategic risks.
Overall, the US maintains a slim one- to two-year lead in AI and semiconductor tech over China, per expert estimates. This narrowing gap pressures policymakers to refine export strategies without stifling global competition. Chinese firms’ offshore maneuvers illustrate the challenges of enforcing tech controls in an interconnected world.
Frequently Asked Questions
Why Are Alibaba and ByteDance Training AI Models in Southeast Asia?
Alibaba and ByteDance are relocating AI training to Southeast Asia to access Nvidia’s advanced H20 chips, restricted for direct sale to China since April. This complies with US export controls while enabling high-performance model development like Qwen and Doubao, which rank among top global LLMs, avoiding domestic limitations on foreign hardware.
What Impact Have US Chip Restrictions Had on Nvidia’s China Market?
US restrictions on Nvidia’s H20 chips have led to billions in losses and shifted market share to international competitors in China. While August’s conditional sales approval offers some relief, ongoing curbs on advanced AI hardware continue to challenge Nvidia’s dominance, prompting calls for policy flexibility amid a tightening US-China tech race.
Key Takeaways
- Offshore AI Training Surge: Chinese firms like Alibaba and ByteDance are increasingly using Southeast Asian data centers for Nvidia-powered model training to sidestep US H20 chip bans, ensuring legal access to cutting-edge hardware.
- Domestic Chip Push: Beijing mandates local semiconductors for government-funded projects, boosting partnerships like DeepSeek with Huawei, though foreign chips persist via stockpiles and informal routes.
- Policy Shifts and Risks: Trump’s August deal for 15% US revenue from AI chip sales to China raises security concerns; easing H200 curbs could aid Nvidia but face pushback, highlighting evolving export leverage.
Conclusion
China’s AI development amid Nvidia restrictions reveals a dynamic interplay of innovation, compliance, and geopolitics, with firms leveraging Southeast Asian hubs for advanced training. As US policies evolve— from April’s H20 blocks to potential H200 relaxations—the global AI landscape intensifies, narrowing technology gaps. Stakeholders should monitor these shifts closely, as they could redefine semiconductor access and foster new collaborative opportunities in the coming years.
