Good morning. Nicholas Gordon here, reporting from Hong Kong. “Sovereign AI” is the buzzword of choice in the global tech sector—at least if you’re based anywhere other than the U.S. and China. Government officials and companies are increasingly worried about tying themselves too closely to one country’s AI systems.

But what “sovereignty” is meant to protect varies by geography. In Europe, it’s data, keeping personal information at home. In the Middle East and much of Asia, it’s local industry, part of a bet that a homegrown AI sector will pay economic dividends. And for smaller economies, it’s autonomy to ward off the fear of being cut off by a supplier in another country.

“AI has become such an essential need, and so you don’t want to be tethered to anybody else who can turn it off,” Pak-Sun Ting, the founder of Hong Kong-based Votee AI, recently told me. (Ting is joining our Fortune Leaders Forum in Macau on Sep. 8, where he’ll join a panel on how businesses can think about AI adoption and move beyond “tokenmaxxing.”)

Votee AI’s main product is an AI model that operates in Cantonese, the Chinese dialect spoken in Hong Kong and the surrounding Guangdong province. “The whole AI revolution is in English and Mandarin,” he told me. “Cantonese is used in education, health care, and police communications. If those don’t get covered, then AI is essentially useless.”

Votee AI is part of a growing group of developers, startup founders, and major companies trying to build AI models for the rest of the world. Indosat, one of Indonesia’s largest telecoms companies, is building Sahabat AI, a model that focuses on Indonesian languages like Bahasa. South Korean companies are currently taking part in the “AI Squid Game,” a government-sponsored competition to build the best homegrown AI model. And just yesterday, Humain, an AI company backed by Saudi Arabia’s Public Investment Fund, debuted an Arabic-language model, built by the Chinese AI developer MiniMax.

It’s important to remember that none of these languages are small. Korean and Cantonese each have around 80 million speakers. More than 200 million people speak Bahasa Indonesia. Yet these ‘low-resource languages’ don’t have the large text corpora that English and Mandarin Chinese do.

But sovereign AI takes more than political will. AI processors are expensive to buy, data centers are expensive to run, and tech talent is expensive to hire.

Ting pointed out that governments, the most likely customers for a sovereign AI model, don’t need the most powerful models. Those more restrained ambitions help to lower cost. Ting said that Votee trained its model for around $250,000—not a small amount but still significantly less than the tens of billions of dollars spent by AI developers like Anthropic and OpenAI.

Countries (and companies) can get their models, semiconductors, and processing power from different sources, then add their own local spin on the AI’s output to get a better result. They’ll be helped by a growing number of powerful Chinese open-source models that anyone can download and run for free.

It could be that “sovereignty” in AI is less about control and more about choice.