The sovereign AI gamble: Why nations risk wasting billions chasing homegrown AI

There is also a fear that embedded values in the training data of foreign AI models can now spread at scale. This risks erasing cultural and linguistic nuances at a time when these tools are increasingly relied on by everyday citizens for search, drafting emails or completing homework assignments.

These sensitivities are especially prominent across Asia, where even the names of major bodies of water are heatedly contested. Many smaller nations are also wary of having to pick a side that may further entrench the supremacy of US or Chinese tech giants, which could lock in their dominance for decades to come.

But the dream is also a trap. Building foundation models, massive AI systems that are trained on enormous amounts of data, requires billions of dollars, scarce chips and vast engineering talent. Only a handful of global firms have succeeded. For most countries, this moonshot risks becoming an expensive exercise in futility.

South Korea recently launched an ambitious initiative to develop a foundation model. A pubic-private partnership is sponsoring a Squid Games-like competition among five local tech companies to create a domestic AI system that can compete with leading-edge rivals from the US and China.

It’s very much a gamble—most tech companies recognize the market reality that building a new foundation model from scratch that can take on the likes of OpenAI, Alibaba or even DeepSeek is a near-impossible task. It will only divert immense capital, computing resources and talent for a largely symbolic payoff.

Mongolia’s Egune AI illustrates the challenge more starkly. Founder and CEO Badral Sanlig told me the project initially stemmed from frustrations over foreign chatbots stumbling with nuances of the local language. “Nobody helps us, because our market is too small," he told me. “We had to build AI for Mongolia ourselves."

Their first model released in 2023 could write Mongolian poetry very well, Sanlig said, “But its general knowledge was almost zero." The latest model, launched earlier this year, was much larger (70 billion parameters), had fewer issues with hallucinations and understands Mongolian text and culture like a local, Sanlig said, but its growth is restrained by limited resources.

Accessing chips from Nvidia, he said, has been very hard, given “our difficult neighbours." Due to limited resources, it has only amassed some 20,000 daily active users. ChatGPT, by comparison, has around 700 million weekly active users.

Sanlig is confident that his company will have large appeal within Mongolia, thanks to the model’s pitch that it is “located within our borders." But it will be hard to move beyond a symbolic project that boosts national pride to make it a competitive player in a winner-takes-it-all industry.

The ultimate irony is that the biggest benefactor of these scattered sovereign AI efforts is likely to be massive foreign firms, especially chipmakers like Nvidia, that stand to gain the most. “The term ‘sovereign AI’ originated from Nvidia, possibly as a motive to sell more chips," James Landay, co-founder of the influential Stanford Institute for Human-centered Artificial Intelligence, said in a recent interview with Korean media.

He’s correct: Massive orders for data-centre buildouts from the UAE and Saudi Arabia, which are also pursuing sovereign AI goals, can help offset lost sales from China for Nvidia, Bloomberg Intelligence analysts said in a note last week. Smaller nations’ quest for independence may end up strengthening the very firms seek to escape.

A smarter strategy may be to channel resources not on reinventing foundational AI tech, but by focusing on the application layer. Building domain-specific tools in healthcare, finance or public services is also far more likely to deliver economic returns. Governments concerned about sovereignty would be better off concentrating on infrastructure at the deployment level: It matters less that an AI model was trained elsewhere if it can be run locally.

Countries worried about being left behind in AI should focus more on diffusion. Historical tech revolutions tell us that it isn’t necessarily the countries that develop a new general purpose technology first that benefit the most, but those which deploy them most effectively. Nations that adapt their economies, industries and institutions to make the most of AI tools will capture the greatest gains.

When I asked Sanlig for advice for smaller nations trying to develop sovereign AI ecosystems, he sighed: “It is not so easy. But it’s not impossible." ©Bloomberg

The author is a Bloomberg Opinion columnist covering Asia tech.