[Yoo Choon-sik] Sovereign AI and the limits of autonomy

South Korea’s government-led project to develop sovereign foundation artificial intelligence models capable of competing with — or even surpassing — global leaders such as ChatGPT, Gemini and DeepSeek has cleared its first major hurdle with the completion of its initial evaluation round.

Out of five competing consortia, three — led by LG AI Research, SK Telecom and Upstage AI — advanced to the second round, which ultimately aims to designate two teams as national champions with full-scale government backing. The first-round results, announced late last week, clarify the project’s direction while exposing tensions that will shape its future.

Although four consortia other than the NC AI team met the basic quantitative performance criteria, the exclusion of the Naver Cloud team came as a surprise. Given Naver’s long-standing dominance in the domestic internet and AI industries, the team had widely been viewed as a front-runner.

The Ministry of Science and ICT cited Naver Cloud’s failure to meet the “independence requirements,” pointing to the use of pre-trained vision and audio encoders linked to Chinese entities. Officials argued that such components conflicted with the goal of building sovereign AI systems “from scratch.”

The decision immediately sparked heated debate, not only because of Naver Cloud’s stature but because it crystallized a deeper unresolved question: What does “sovereign AI” mean in a world where no meaningful AI system is built in isolation?

At the news conference announcing the results, Vice Minister Ryu Je-myung struck a conciliatory tone, emphasizing that all five first-phase models had been recognized as “notable AI models” by Epoch AI, a US-based research organization. He added that Epoch AI’s latest compilation lists eight South Korean models, ranking the country third globally behind the United States, with 43, and China, with 30.

If a firm with Naver’s data, infrastructure and experience can be disqualified on independence grounds, doubts naturally arise about the feasibility of building truly sovereign AI in a deeply interconnected global environment.

There are concerns that reliance on pre-trained components developed abroad — particularly in China — could embed long-term dependencies that undermine autonomy in sensitive areas such as national security, cultural representation and strategic decision-making.

This fear has been a central driver of the project from the outset, with the government repeatedly warning that abandoning domestic model development would leave South Korea dependent not only commercially but also strategically on US and Chinese AI developers.

At the same time, critics have raised a different but equally uncomfortable argument: that devoting national resources to foundation model development may be a misallocation of time and capital precisely because the global AI landscape is already moving beyond them and foundation models are becoming commoditized.

Harder questions lie ahead

By the time South Korea produces fully sovereign models using its own resources, critics warn, the world may already have moved on.

This critique cannot be dismissed lightly because foundation models are no longer the final destination of AI innovation but the starting point. As global leaders build layered ecosystems by integrating tools, memory, autonomy and multimodal perception, a national race to replicate foundation models risks looking like yesterday’s battle.

Yet the argument cuts both ways. Foundation models may no longer be sufficient, but they remain necessary. Without control over underlying models, downstream innovation is constrained by external licensing terms, strategic priorities and opaque governance structures. Application-layer creativity alone cannot compensate for a lack of foundational autonomy.

In this sense, developing sovereign models is an effort to preserve strategic options in an increasingly concentrated AI landscape. The real danger lies in mistaking foundation model development for an end in itself.

If South Korea builds models that score well on benchmarks but fails to translate that into widely adopted services, industrial transformation or effective public-sector tools, the strategic payoff will be limited. Quantitative benchmarks offer a useful common language but cannot capture everything that matters.

Robustness, safety, domain-specific performance and cultural alignment — especially in the Korean language — are far harder to measure.

The disqualification of Naver Cloud also highlights the need for clearer rules. Modern AI development relies on a global pool of ideas, techniques and components. Without clear and consistently applied criteria, similar disputes are likely to recur, undermining confidence in the project’s governance.

Despite these challenges, the project’s broader strategic context remains compelling. South Korea has set its sights on becoming a leader among the next tier of AI innovation powerhouses after the United States and China. Recognition by Epoch AI supports this ambition symbolically, suggesting that South Korea remains relevant in the global AI conversation even if it is not at the very frontier.

What remains uncertain is whether that relevance can be sustained. Long-term competitiveness will require sustained investment not only in larger models but also in efficiency, multimodality and human-centered design.

It will also demand difficult policy choices about openness. Broad availability could accelerate innovation but raises concerns about security and misuse, while tight control may protect strategic interests at the cost of ecosystem growth.

Regulatory and ethical frameworks must evolve. Issues of data privacy, bias, accountability and safety will become unavoidable. Public trust will ultimately determine whether these systems are embraced or resisted.

The first-round evaluation should therefore be seen neither as a triumph nor as a distraction. It is a necessary step in a longer and far more complex process. The controversy over Naver Cloud’s exclusion, the debate over sovereignty and claims that foundation models may already be outdated all point to the same conclusion: the hardest questions lie ahead.

The challenge for policymakers is to ensure that this project does not become an exercise in symbolic self-reliance but a foundation for sustainable innovation. In AI, as in economic policy, early validation is easy to celebrate. Strategic clarity, execution and adaptation are far harder to achieve. That is where this project will ultimately be judged.

Yoo Choon-sik

Yoo Choon-sik worked for nearly 30 years at Reuters, including as chief Korea economics correspondent, and briefly as a business strategy consultant. The views expressed here are the writer’s own. — Ed.