US soft-pedals AI exports in Asia as China's open models gain regional footing

The US AI Exports Program, launched last summer to counter Chinese dominance in Southeast Asia, has received only 78 applications so far—well below expectations, according to Politico. Meanwhile, at the Asia-Pacific Economic Cooperation Digital Weeks in Chengdu this month, the US presence was notably subdued: only Google and Meta maintained public booths, and neither emphasized their flagship LLMs. In contrast, Chinese companies highlighted their open-source models, particularly Tencent's Hunyuan, while President Xi Jinping announced 5,000 government-funded AI training opportunities and AI application cooperation centers across Southeast Asia.

China's open-weight model strategy is gaining traction with regional governments and enterprises where cost and technological sovereignty matter more than frontier capability. Beijing's APEC endorsement of 'open-source AI with strong security' legitimizes China's approach in emerging Asian economies facing billions in AI infrastructure spending. The US currently lacks a centralized open-source offering to counter Alibaba's Qwen, Baidu's Ernie, or the newly announced Kimi K3 from Moonshot AI. US companies are fragmenting their approach: Google emphasized AlphaFold and business applications; Meta focused on SMB adoption rather than competing for large enterprise seats.

For practitioners deploying in Southeast Asia, the geopolitical split is reshaping procurement: the US maintains advantages in full-stack integration (chips + models + cloud), but Chinese alternatives offer 80% capability at 20% cost. Localization is becoming the key wedge—teams building for Thai, Vietnamese, or Filipino workloads are increasingly forced to integrate multiple model families or fine-tune on cheaper open weights. The Exports Program's low uptake suggests the US strategy is not resonating with buyers prioritizing vendor diversity and total-cost-of-ownership over integration risk.