IMAGE: GETTY IMAGES
IN FEBRUARY more than 100 governments signed something called the Bangkok Declaration, formally committing themselves to “AI sovereignty”. It is the sort of document that sounds grandly aspirational and is, in reality, an admission of anxiety. The anxiety has a cause. Two countries, America and China, currently control some 90% of the computing power needed to build and run frontier artificial intelligence, and between them own every single one of the 50 top-ranked AI models in the world, according to the Centre for a New American Security’s Sovereign AI Index (it tracks how self-sufficient countries are in AI). Everyone else is, to varying degrees, a tenant in someone else’s data centre.
What is striking about the response is how universal it has become. France has Mistral. The UAE has Falcon, built under the G42 banner. Saudi Arabia has HUMAIN, a subsidiary of its sovereign wealth fund. India has Krutrim, the country’s first AI unicorn, alongside the government-favoured Sarvam AI. Japan runs a consortium called LLM-jp through its National Institute of Informatics. Singapore built SEA-LION for eleven Southeast Asian languages. South Korea has HyperCLOVA X. By early this year more than 60 countries had published formal national AI strategies, and over 30 had put real money behind them.
The reasons given are consistent wherever you look. The first is linguistic and cultural: a model trained mostly on English-language internet text handles Hindi, Arabic or Bahasa Indonesia noticeably worse, which matters enormously to a country like India, with 22 official languages and a legal and administrative culture that foreign models do not understand. Krutrim’s second-generation model, trained specifically on 22 Indian languages, scores 0.95 on sentiment analysis against some 0.70 for competing general-purpose models, and Indian officials treat that gap as a national capability question.
The second reason is plainer: nobody wants their hospitals, courts or military relying on a chatbot that a foreign government could, in theory, switch off. The third is economic. AI is increasingly being talked about the way electricity or telecoms once were, as infrastructure too important to rent indefinitely from somebody else.
The trouble, and it is a serious one, is that almost none of this sovereignty is actually sovereign. The CNAS index, which tracks more than 130 national AI projects worldwide, finds that some 70% involve at least one foreign partner, and four out of five of those partners are American. The UAE’s Falcon, presented as a flagship of Gulf technological independence, was trained on Amazon Web Services. Singapore’s SEA-LION depends on GitHub, Hugging Face and soon IBM for distribution, all three American.
Even China, the one country building a separate AI stack, is locked into its own version of the same dependency on domestic chips because of American export controls on the most advanced ones. The true description of most “sovereign AI” programmes is that they are an attempt to manage a dependency, rather than eliminate one.
There is also a quieter problem hiding inside the headline investment numbers. The UAE and Japan alone account for more than two-thirds of all disclosed sovereign AI spending; almost everyone else is working with a few hundred million dollars rather than the tens of billions Gulf states and Washington can deploy.
And every country chasing this race needs people who can actually build the thing. China is short an estimated 5m AI professionals. India’s AI-specific roles go unfilled at a rate of 82%. A newly minted American AI PhD commands something like $185,000 a year against roughly $67,000 in China, which is the kind of gap that keeps pulling talent towards Silicon Valley regardless of how many national champions Beijing or New Delhi fund.
None of this means sovereign AI is theatre. Competition among more than 20 frontier-quality models from over ten countries has pushed API prices down by 90% since 2023, which is a real and useful outcome for anyone building on top of these systems. What it does mean is that the language of national independence is mostly aspirational. The really useful question for most governments is not how to own a frontier model outright, something all but a handful of them cannot do. It is which layers of dependency they can live with, and which partnerships make that dependency survivable rather than dangerous. ■
