Chinese AI dominance is not Asia’s destiny

East Asia Forum

Source published: Site published:

Original article · فارسی

In January 2025, India's Ministry of Electronics and Information Technology announced that DeepSeek's models would be hosted on servers in India. This was a rare approval for a Chinese artificial intelligence (AI) model from a government that has banned many Chinese apps. Also in 2025, AI Singapore released the latest version of SEA-LION, its open-source language model for Southeast Asian languages. It was built using Alibaba's Qwen model instead of Meta's Llama. Thailand's OpenThaiGPT is also based on Qwen2.5, which was further trained for its needs.

These examples give us some idea of what is happening, because there is little data about AI use across Asia. Around the world, by August 2025, versions of Qwen made up more than 40 per cent of the new language models based on other models on the open-source platform Hugging Face. Llama-based models made up 15 per cent. Chinese models now also account for more than 80 per cent of the use of open models on OpenRouter.

There are good reasons why people are using open-weight models. These are AI models whose trained settings are made available for anyone to download, change and run on their own computers or servers. They are often cheaper, which makes them attractive in markets where customers are careful about costs, such as India. Chinese AI companies have released many powerful models that can run on relatively basic computers. An Indian or Singaporean startup trying to keep its AI service bills low, or a government that wants its citizens' data to stay inside the country, may find a Chinese model a better choice.

The important question for most Asian countries is whether using these models could eventually make them dependent on China. Once downloaded, open-weight models cannot easily be taken back. They can also be replaced with other models. So the models themselves are not where the biggest long-term dependence is likely to develop. The greater risk lies in other parts of the AI system, especially the computing power and cloud services needed to run the models.

A Chinese model could encourage businesses to use Chinese cloud services and computing equipment, but nobody is forced to do so. Chinese models can also run on Amazon Web Services or in local data centres instead of Alibaba Cloud. However, running a model independently is not affordable for every business. Companies need skilled engineers and powerful graphics processing units (GPUs). Many businesses therefore pay for models as managed services. The company that created a model can also offer its newest versions first and provide tools designed to work with them.

Alibaba is building infrastructure across the region to attract this business. In 2025, it opened a third data centre in Malaysia and announced plans to build a second one in the Philippines. It also opened an AI centre in Singapore that offers businesses and developers credits for using AI and access to computing power. However, there is still little evidence that demand for Chinese models is turning into cloud-service contracts across Asia. Many companies run the models on their own systems to avoid becoming dependent on outside providers.

Chinese AI models could also help increase demand for computer chips made in China. DeepSeek's V4, released in April 2026, was adapted to work with Huawei's Ascend chips. Huawei says its chips were also used for part of the model's training.

Dependence on China is therefore not certain, but the choices countries make early on are important. Developers tend to keep using models they already know. The more deeply a model is built into a company's products, the harder and more expensive it becomes to replace. If one model becomes the usual choice across a region, the tools, different versions and tests used to judge new models may also be designed around it.

AI models can also reflect the political views and language habits found in the information used to train them. In January 2025, DeepSeek's chatbot refused to answer when asked whether Arunachal Pradesh is part of India. Perplexity decided to retrain DeepSeek's R1 model to remove censorship imposed by the Chinese state. This suggests that such political limits are built into the model itself, rather than being added only by the app through which people use it.

Using open-weight models does not give China a way to force other countries to follow its wishes. Once a model has been downloaded, Beijing cannot take it back. The open-source system also makes it harder for China to threaten countries by refusing to release future models. However, this advantage is becoming weaker for the most advanced models. Moonshot's Kimi K3 and Alibaba's Qwen 3.8-Max now require commercial agreements with large companies that want to provide them as services. This turns something that was available for free into a paid business relationship. China still has some influence through the models people use by default, the values built into them and the attraction of Chinese cloud services. But that influence is real and is still being challenged.

Open-weight models from the United States, including OpenAI's gpt-oss and Google's Gemma, compete for the same users. Many companies also use both Chinese and American models so they do not become dependent on a single provider. A more immediate problem may come from Washington. The US government has banned DeepSeek from federal government computer systems and is considering wider restrictions. These could create problems for Asian companies that have customers in the United States.

India shows how a country can try to reduce these risks, as well as the limits of this approach. While Indian companies are using Chinese models, the government in New Delhi is also investing in other options. The US$1.2 billion IndiaAI Mission is providing support for the use of around 38,000 subsidised GPUs. Indian company Sarvam AI has also introduced its own models with 30 billion and 105 billion parameters. Having computing power inside India allows local developers to adapt and run different AI models within the country. This reduces the attraction of Chinese cloud services.

But India's approach moves its dependence to another place rather than removing it. The GPUs are imported, mostly from US chipmakers. This means India is reducing its reliance on Chinese models but becoming more exposed to American hardware and US export restrictions. Sarvam's earlier model, Sarvam-M, was built using France's Mistral Small. The company's newer models were trained from the beginning rather than being built on existing models, but they are still not as powerful as the world's most advanced systems. India's plan reduces the risk of becoming locked into one provider, but it does not make the country fully independent. Even this partial approach is too expensive for many Asian countries.

Chinese open-weight models are becoming a common building block for AI development across Asia. But using a model is not the same as depending on the infrastructure behind it. Beijing cannot decide on its own whether Asian countries will become dependent. That depends on what countries in the region build around these models: different sources of computing power, the ability to test and check models, and reliable local alternatives.

Chinese models are becoming part of the region's AI systems faster than Asian countries are building the ability to examine them carefully and replace them. Whether China's early lead turns into long-term dependence is a decision for Asia to make.

Karthik Nachiappan is a Non-Resident Senior Fellow at the Institute of South Asian Studies, National University of Singapore.