South Korea Cuts Naver From Sovereign AI Project Over Originality Standards

South Korea is tightening its grip on AI autonomy by eliminating established tech giants from its national foundation model project. The Ministry of Science and ICT (MSIT) has advanced three consortia—led by LG AI Research, SK Telecom, and the startup Upstage—to the second stage of its Independent AI Foundation Model Project.

The decision, following evaluations by NIPA and TTA, saw the removal of Naver Cloud and NC AI. According to government reports, Naver was eliminated primarily for failing to meet strict originality standards regarding independent pre-trained components. The project mandates that models be developed from scratch using domestic data and architecture to ensure the nation is not reliant on foreign intellectual property.

The Push for AI Sovereignty

The initiative is a strategic effort to secure AI independence in critical sectors, including national defense, healthcare, finance, and public administration. The government's definition of "sovereign AI" requires full domestic control over both the training data and the model's architecture, explicitly forbidding dependence on foreign pre-trained weights.

Vice Minister Ryu Je-myung of the MSIT emphasized that true autonomy requires rigorous proof of independence. "Even if open-source components are used, sovereignty requires proof that the weights were trained independently," Ryu stated. "The ability to build and operate foundation models under full control is the essence of Korea’s AI autonomy."

A Shift in Innovation Power

The advancement of Upstage, a smaller startup-led consortium, over a behemoth like Naver signals a potential structural shift in South Korea's tech ecosystem. Industry observers suggest the project is prioritizing technical agility and talent over corporate scale. Jin Yoon-jung, Executive Director at SBVA, noted that "innovation speed and talent density now outweigh sheer capital" in the current AI race.

Upstage's contribution to the project, the 'Solar Open 100B,' utilizes a Mixture-of-Experts (MoE) structure, reflecting the move toward more efficient, specialized architectures rather than simply increasing model size.

Global Implications and Next Steps

By establishing a rigid national standard for "AI originality," South Korea is creating a blueprint for other nations attempting to avoid total dependence on a few global tech superpowers. The project serves as a test case for whether a mid-sized economy can maintain a competitive, independent AI stack without relying on the weights of dominant US or Chinese models.

The remaining three teams will now undergo further testing to prove their models can meet the government's performance and autonomy requirements. While the core narrative of the project's progression is confirmed, the industry remains focused on whether these domestic models can maintain parity with global benchmarks while adhering to the strict "from scratch" mandate.