Policy
Asian governments adopt Chinese AI models despite geopolitical tensions
Open-weight systems from DeepSeek and Alibaba gain traction across India and Southeast Asia, raising questions about future dependence.
Key takeaways
- Chinese open-weight AI models like DeepSeek and Qwen now account for over 80% of open-model usage on some platforms, with adoption spreading across Asian governments despite geopolitical tensions.
- Real lock-in risk sits in cloud infrastructure and compute resources rather than the models themselves, as Alibaba expands data centers across Malaysia, Philippines, and Singapore.
- India's $1.2 billion IndiaAI Mission shows hedging is possible but relocates dependence to US hardware rather than eliminating it, as domestic models still trail frontier capabilities.
- Models carry embedded political assumptions from training data—DeepSeek declined to answer whether disputed Arunachal Pradesh is part of India, prompting competitors to retrain the system.
- Whether adoption hardens into dependence depends on regional capacity to build diversified compute, audit models, and develop credible local alternatives around Chinese inputs.
Asian nations embrace Chinese AI despite security concerns
Governments across South and Southeast Asia are increasingly deploying Chinese artificial intelligence models, even as they maintain wariness toward Beijing in other domains. India's Electronics and Information Technology Ministry approved hosting DeepSeek models on domestic servers in January 2025—a notable exception in a country that has banned numerous Chinese applications. Singapore's AI Singapore built its SEA-LION language model for Southeast Asian languages on Alibaba's Qwen foundation, while Thailand's OpenThaiGPT uses the same Chinese base.
The trend reflects pragmatic economics rather than political alignment, according to analysis first reported by the East Asia Forum. By August 2025, derivatives of Qwen accounted for more than 40 percent of new language-model variants on Hugging Face, the open-source platform, compared to 15 percent for Meta's Llama. Chinese models now represent over 80 percent of open-model usage on OpenRouter.
Karthik Nachiappan, a non-resident senior fellow at the Institute of South Asian Studies at the National University of Singapore, notes that open-weight models—systems whose trained parameters can be freely downloaded and modified—offer compelling advantages in price-sensitive markets. Chinese labs have released capable models that run on modest hardware, making them attractive to startups and governments wanting to keep citizen data within national borders.
Why it matters
The distinction between adoption and dependence will shape Asia's technological sovereignty for years. While open-weight models can be swapped relatively easily, the infrastructure built around them—cloud services, compute resources, developer tools—creates stickier relationships that could give China influence over regional AI development paths.
Where lock-in actually occurs
The open nature of these models limits Beijing's coercive leverage. Once downloaded, they cannot be revoked, and developers can switch between alternatives. The real risk of dependence sits elsewhere in the technology stack: cloud computing and processing power.
Alibaba is building regional infrastructure to capture demand generated by its models. The company opened a third Malaysian data center in 2025 and announced plans for a second facility in the Philippines. It also launched an AI center in Singapore offering compute resources to businesses and developers. However, evidence that adoption is converting into cloud contracts remains limited, as many organizations choose to self-host to avoid vendor lock-in.
Dependence also accumulates through path dependence. Developers tend to stick with familiar models, and switching costs rise as systems become more deeply integrated into products. When a model becomes the regional standard, the tools, benchmarks, and variants used to evaluate alternatives are built around it.
Models also carry embedded assumptions from their training. In January 2025, DeepSeek's chatbot declined to answer whether Arunachal Pradesh is part of India—a response that Perplexity's decision to retrain the model to remove Chinese state censorship suggests is built into the system itself.
India's hedge strategy shows promise and limits
India demonstrates both the potential and constraints of hedging against dependence. While Indian firms adopt Chinese models, New Delhi is investing $1.2 billion through its IndiaAI Mission to deploy approximately 38,000 subsidized GPUs. Sarvam AI has released its own 30- and 105-billion-parameter models.
Domestic compute infrastructure allows Indian developers to fine-tune and host any model within national borders, reducing the pull toward Chinese cloud services. But India's approach relocates rather than eliminates dependence—the GPUs are imported overwhelmingly from US chipmakers, trading exposure to Chinese models for exposure to American hardware and export controls. Sarvam's earlier model was built on France's Mistral Small, and its newer systems still trail frontier capabilities.
The analysis notes that Chinese open-weight models are becoming common inputs for AI development across Asia, but inputs are not infrastructure. Whether China's early advantage hardens into dependence depends on what the region builds around these systems: diversified compute resources, capacity to evaluate and audit models, and credible local alternatives.
The details were first reported by the East Asia Forum in analysis by Karthik Nachiappan.