Korea's Sovereign AI Race: Why Naver Lost Its National Team Spot, and What It Did Next
Updated: Oct 24
Few countries take "sovereign AI" as literally as Korea. Since 2025, the government has been running a national contest to select homegrown AI foundation models, with GPUs, data and funding as the prize. Teams are scored and cut round by round, like a sports league.
The biggest surprise came in January 2026. Naver, the search giant that had championed sovereign AI for years, was eliminated. Not for weak performance, but for not being independent enough. What Naver did next shows that Korea now has two very different ideas of what "sovereign" means.
1. How Korea picks its national AI team
The Ministry of Science and ICT launched the Sovereign AI Foundation Model project with five teams: Naver Cloud, SK Telecom, LG AI Research, Upstage and NC AI. Models are scored on benchmarks (40 points), expert review (35) and user review (25).
There is also a strict test of independence. The government defines a sovereign model as one designed and pre-trained in Korea, not a fine-tuned foreign model. Open-source code can be used, but outside model weights must be reset and retrained from scratch.
2. Why Naver was cut
On January 15, 2026, the ministry announced the first-round results. Naver Cloud and NC AI were both eliminated. NC AI fell short on performance. Naver failed on independence: according to Korean media, its multimodal model used the vision encoder of Alibaba's open-source Qwen model without retraining it from scratch.
The contest continued. Motif Technologies was added as a replacement team in February. In the second round, announced on August 18, SK Telecom (70.6 points), Upstage (69.9) and LG AI Research (69.0) advanced, and Motif was eliminated even though its model scored highest on an independent benchmark. Motif has appealed, and the ministry has said it is reviewing how the next stage will be run.
3. Naver's other definition of sovereign AI
Naver did not try to win back its place. It went the other way. In June 2026, it joined NVIDIA's Nemotron Coalition as the first Korean member and said it would build future versions of its HyperCLOVA X model by fine-tuning NVIDIA's open Nemotron model with its own data.
At the same time, it is expanding its GAK Sejong data center on NVIDIA's AI factory platform, starting at 55MW with plans to reach gigawatt scale, and offering "data-sovereign" AI services to governments and companies in Korea, Europe and the Middle East.
In other words, Naver now sells sovereignty as control over data, infrastructure and deployment, not over every model weight. The export record so far: digital twins of Mecca, Medina and Jeddah for Saudi Arabia (a second-phase contract was signed in May 2026), a Thai-language model built with Thailand's Siam AI, and new alliances with France's Mistral AI and Vietnam's FPT. At home, it closed its standalone CLOVA X chatbot and Cue: AI search in April 2026 to focus AI inside Naver search.
4. The finalists: SK and LG
The remaining teams show how differently Korea's large groups approach AI.
SK Telecom built A.X K1, a 519-billion-parameter model, and is turning it into a broader business: an AI data center in Ulsan with AWS, a robot training platform that works with any robot maker, and above all SK hynix, the world's leading supplier of the high-bandwidth memory that AI chips need. SK hynix has started building a roughly US$4 billion advanced packaging plant in West Lafayette, Indiana.
LG AI Research released K-EXAONE 2.0, a 750-billion-parameter model, under the Apache 2.0 license so anyone can use it commercially, and is using its EXAONE models across LG's factories and robots. Upstage, the only startup in the final group, focuses on enterprise document AI.
What this means for startups, investors and EDOs
Expect two kinds of Korean AI customers. The national-team companies need tools for training from scratch: data, evaluation, safety and efficient training infrastructure. Naver and others building on open models need fine-tuning, inference and deployment tools. The sales pitch is different for each.
Korean AI is going abroad. Naver is selling sovereign AI to the Middle East, Southeast Asia and Europe, often with partners. Foreign companies with local relationships, from data center operators to system integrators, can be useful co-sellers.
Follow the infrastructure money. SK's Indiana plant has more than 100 suppliers under evaluation, and SK's chairman has signaled tens of billions of dollars of new US AI investment, with no sites announced yet. For economic development organizations, Korea's AI build-out is as much about chips, memory and data centers as about models.
Watch the rules. The independence test, and the debate it has sparked, will shape which partnerships Korean companies can sign without risking public funding.
The bottom line
Korea's contest made "sovereign" mean "built from scratch at home." Naver, cut from the contest, is betting that customers care more about where data lives and who runs the system. Both models will coexist, and both will need foreign partners. Knowing which definition a Korean company is working under is the first step to selling to it.
If you are an AI or infrastructure company, an investor or an economic development organization looking to work with Korean AI players, we would be glad to help. Contact us at info@globalrika.com.
About the data
Figures are from government announcements, company releases and Korean media reports through October 2026. The contest's next stage and schedule may change. Model sizes and scores are as reported by the companies and the ministry.
Sources
Ministry of Science and ICT, first-round results (January 15, 2026) and second-round results (August 18, 2026)
Newsis and Money Today Broadcasting, first- and second-round coverage (January and August 2026)
NVIDIA newsroom, NAVER and NVIDIA expand AI infrastructure (June 7, 2026)
Dealsite, Naver's overseas sovereign AI business (2026)
ZDNet Korea, CLOVA X and Cue: closure (February 2026)
SK Telecom newsroom, A.X K1 technical report (January 2026)
The Register and SK hynix SEC filing, Indiana packaging plant (2026)
LG AI Research, K-EXAONE 2.0 release (July 31, 2026)
Comments