‘Sovereign AI’ is really about having a choice: ‘You don’t want to be tethered to anybody else’

“Sovereign AI” has become the preferred phrase for governments and companies that don’t want their AI future tied too tightly to Washington or Beijing. The idea sounds lofty, but the fear underneath it is plain: dependence. In Europe, the concern is keeping personal data at home. In much of the Middle East and Asia, it’s building local industry. For smaller countries, it’s simply not wanting a remote vendor to flip the switch.

That last worry comes through clearly in Hong Kong. Pak-Sun Ting, founder of Votee AI, says AI is now too important to leave in someone else’s hands. His company builds a model in Cantonese, the language used in Hong Kong and nearby Guangdong province, and he argues that’s not a niche indulgence. Cantonese is used in education, health care, and police communications. If AI misses those use cases, he says, it misses a lot of real life.

Votee is one of several efforts trying to make AI work for languages and markets that don’t dominate the big model race. Indosat is building Sahabat AI for Indonesian languages such as Bahasa. South Korean companies are taking part in a government-backed “AI Squid Game” to produce the best homegrown model. And Humain, backed by Saudi Arabia’s Public Investment Fund, recently unveiled an Arabic-language model built by MiniMax, a Chinese AI developer.

The numbers behind those languages matter too. Korean and Cantonese each have around 80 million speakers, while Bahasa Indonesia has more than 200 million. But size alone does not solve the problem. These are still “low-resource languages,” which means there isn’t nearly as much text available to train models as there is for English or Mandarin Chinese.

Then there’s the bill. AI chips cost a lot, data centers cost a lot, and talent costs a lot. Ting says governments do not always need the most powerful models, which helps keep costs down. Votee says it trained its model for around $250,000, far below the tens of billions spent by companies such as Anthropic and OpenAI. The emerging playbook is not total independence. It’s mixing sources: models from one place, semiconductors from another, processing power from somewhere else, then adding local tuning on top. In that sense, sovereign AI looks less like a fortress and more like a seat at the table.