Capital cities that previously relied on the private sector for technology are exhibiting an odd new trend. All of these cities—Paris, Tokyo, Delhi, Bern, Kuala Lumpur, and even Ottawa—quietly write checks for their own AI systems. In the middle of a sentence, some of these models can transition between Bahasa Indonesia and Lao. Others have been taught the distinction between the German “ß” and the Swiss German “ss,” a detail that no Silicon Valley engineer would ever worry about. It’s the kind of thing that almost makes you laugh, but then you realize how much money is involved.
The most frequently mentioned figure is France’s €100 billion pledge, but the global total is actually unknown. After accounting for chips, data centers, electricity contracts, and the lengthy tail of subsidies, estimates reach the hundreds of billions. It appears that governments can’t afford to wait this out. It’s another matter entirely if they can afford to sit in it.
| Topic Snapshot | Details |
|---|---|
| Subject | The global rise of sovereign AI initiatives |
| Notable Models | Singapore’s Sea-Lion (11 languages), Malaysia’s ILMUchat, Switzerland’s Apertus |
| France’s Pledge | Over €100 billion in national AI spending |
| Japan’s Framework | The AI Promotion Act, in full effect since September 2025 |
| India’s Footprint | Second-largest market for OpenAI users, 100M+ ChatGPT downloads |
| Largest Project | OpenAI’s planned 5-gigawatt campus in Abu Dhabi |
| Core Tension | National security and cultural identity vs cost and feasibility |
| Year of Shift | 2025, after the so-called “DeepSeek shock” out of China |
Much of this anxiety dates back to early 2025, when China’s “DeepSeek shock” upended preconceived notions about who could create a competitive model and how cheaply. In a matter of months, Japan—never the strongest voice in tech policy—passed the AI Promotion Act, the EU began softening parts of its own AI Act in the name of competitiveness, and the United States unveiled its AI Action Plan. Tokyo’s proposal is remarkably straightforward: become the most AI-friendly nation in the world, maintain lax regulations, and serve as a sort of link between Brussels’ more circumspect tendencies and Washington’s deregulatory rush.
Then there is India, which strangely sits in the middle of all of this. With over 100 million downloads of ChatGPT, India is now the world’s second-largest market for OpenAI users. However, local developers like Soket AI persistently and somewhat impatiently point out that local languages, contexts, names, and presumptions continue to cause imported models to falter. It is difficult to ignore the fact that the same grievance is raised in Jakarta, Riyadh, and Nairobi. The technology is portable. Its internal cultural memory does not.
It is more difficult to ignore the physical reality that lies beneath the speeches and press conferences. You can already see the outlines of an industry that resembles steel more than software outside the new hyperscale data centers in locations like Abu Dhabi, where OpenAI is planning a 5-gigawatt campus spread over about ten square miles. Electricity is consumed by servers. Water is consumed by cooling towers. Supply chains for chips travel via Taiwan, the Netherlands, and a few mines located from Inner Mongolia to the Congo. According to a recent IMF essay, the competition for algorithms is also a competition for minerals, energy, and land. Very few of these are present in the majority of middle-power governments.
As a result, the awkward question keeps coming up, usually in whispers rather than directly. Are these autonomous AI initiatives a sincere wager on independence, or are they merely a nostalgic impulse that gave rise to state oil companies and national airlines in the past? The Atlantic Council’s analysts have cautioned that smaller powers run the risk of squandering massive sums of money creating something they could license, collaborate on, or just regulate. Some contend that even a flawed domestic model is preferable to giving your court systems, school curricula, and civic data to a foreign vendor whose prices could triple the following year.
Observing all of this from a distance gives the impression that no one is certain which camp is correct just yet. Neither GPT nor Gemini will be overtaken by the models being introduced in Singapore and Switzerland. They really aren’t supposed to. They are supposed to maintain a position at the table, protect a language, and retain a small amount of power in a market that is becoming more and more like an oligopoly. Around the end of the decade, the world will likely find out if that is worth the expense. The checks continue to be signed for the time being.
