Foreign media reports that the term "sovereignty AI" has been frequently mentioned in the global tech industry recently, but this concept varies from region to region. Europe places a greater emphasis on keeping data locally, while some markets in the Middle East and Asia focus more on the development of their own industries. Smaller economies are particularly concerned about maintaining autonomy in key technologies to avoid being constrained by external suppliers.

Local language becomes the entry point

The founders of Hong Kong startup Votee AI – Pak to Sun Ting – stated that AI has gradually become a fundamental capability. Users do not want this system to fall into the hands of others and even face the risk of being shut down at any time. The company focuses on Cantonese language models, targeting regions in Hong Kong and Guangdong where Cantonese is spoken.

He pointed out that the current AI ecosystem is mainly centered around English and Mandarin, but Cantonese is still widely used in education, medical treatment, and police communication. If these scenarios are not supported, the actual usability of AI will be significantly limited.

Promoting local models in multiple places

Similar attempts are emerging in more markets. Indonesia's major telecommunications operator Indosat is developing Sahabat AI that focuses on local languages such as Indonesian. Korean companies are participating in government-supported local AI model competitions. The AI company, supported by Saudi Arabia's Public Investment Fund, also recently launched an Arabic model, which was built by a Chinese AI developer named MiniMax.

The article mentions that these languages are not niche markets. There are approximately 80 million speakers of Korean and Cantonese, and over 200 million speakers of Indonesian. However, compared to English and Mandarin, the amount of text corpus available for training models in these languages is still limited, which is why they are often classified as “low-resource languages.”

Cost remains a practical barrier.

However, promoting "sovereignty AI" does not rely solely on policy will. Chip procurement, data center operations, and recruitment of technical talent all require continuous investment, which are also practical barriers faced by many countries and enterprises.

Ting believes that governments are often the most likely customers for such models, but they may not necessarily need the most powerful general-purpose models. Once the objectives are achieved, the costs will also decrease. According to him, the cost of training Votee AI's model is about $250,000, which is far lower than the billions of dollars invested by companies like OpenAI and Anthropic.

The article argues that countries and enterprises do not necessarily need to build an entire system from scratch; they can also obtain models, chips, and computing power from various sources, and then adapt them to local data and linguistic capabilities. As more open-source models from China become available for download and deployment, "sovereignty" may not necessarily mean complete control, but rather the retention of the right to choose in key technologies.