Korea's AI Survival Game: Two Giants Cut, One Surprise Casualty, and Only Two Models Will Make It Past 2026
Korea is running a national tournament to build homegrown AI that can stand next to US and Chinese models. Here is who is still in, who got cut, and why it matters globally.
What if a government ran an AI contest the way a TV network runs an elimination show, with scores, public rankings and a cut after every round? That is not a pitch for a reality series. It is how South Korea is trying to build its own large language models.
The program is called the Sovereign AI Foundation Model project, run by the Ministry of Science and ICT. After the second evaluation, one of the best-scoring models in the field was sent home. Keep reading, because the reason is the most interesting part.
1The Rules of the Game
The idea is simple. Korea wants AI models built at home, trained with domestic control over data and infrastructure, that can compete with the best from the United States and China. Teams get government support, including access to scarce GPUs, and in return they are judged repeatedly.
According to the Korea Herald and Korea Times reports from August 18, 2026, the second-stage evaluation was scored out of 100 points. The scoring split is worth remembering because it shaped who survived.
The gaps between teams were small. Benchmarks varied by 4 points from top to bottom, expert review by 2.4, and user assessment by 5. A different team ranked first in each category, and the ministry did not publish individual rankings.
2Scoreboard: Who Is Still Standing
Yes, Naver, Korea's biggest search company, was already out before this round. For a country that likes to say it has its own Google, that was a headline in itself.
3The Plot Twist: The Model Cut Was Not a Weak One
Here is the twist. Motif 3, from the eliminated team, scored 47 on the Artificial Analysis Intelligence Index, which the Korea Times reports ranked 10th worldwide and highest among models from outside the US and China. The ministry itself said Motif "showed outstanding technological capabilities."
So why the exit? The ministry said Motif rated lower on usability and applicability, which carried substantial weight. In plain terms, a great lab result was not enough. The judges wanted proof that real people and real businesses could use the model.
If you follow AI outside Korea, this is the lesson. Leaderboards reward raw capability, but a national program funded with public money asks a different question: can this actually be deployed in a hospital, a bank or a factory?
4Why Compute Is the Real Currency
Training a frontier model needs enormous computing power, and the project measures support in B200 GPU equivalents. Allocation rises from roughly 768 chips in the first half of 2026 to about 1,000 in the second half, and the government says it will step up support for the three remaining teams.
That may sound small next to the data-center budgets of the biggest American companies, and that is the honest challenge here. Korea is betting that focused national backing, strong hardware partners and a deep local user base can close part of the gap. Whether that is enough is still an open question.
What the teams are building
LG's K-EXAONE, SK Telecom's A.X and Upstage's Solar families all target Korean-language strength plus industrial use. For readers outside Korea, that matters because Korean is a relatively low-resource language for global models, and local systems often handle honorifics, mixed Korean-English text and regional context better.
How to read the cuts
It is tempting to treat each elimination as a verdict on a company, but the format is closer to a relay race than a final exam. Teams that leave still keep their research, their engineers and, in many cases, their commercial products. The people who lose are rarely the ones whose models disappear. What changes is who receives public GPU time and who carries the national label, which matters for government contracts, for recruiting, and for how customers perceive the winners.
For anyone tracking Korean tech from abroad, the practical move is to watch the finalists' product launches rather than only the ministry's score sheets. A model that ships inside a messaging app, a call center or a car dashboard will tell you more about real adoption than any single benchmark number.
5What Happens Next
The Herald reports the field will shrink from three teams to two finalists by the end of 2026. The ministry says the third-stage evaluation will take place in the second half of 2026, with scoring details to be settled after an objection period and consultation with the remaining teams. It is also reviewing a frontier-level AI model initiative that could be linked to, or reshape, the current project and a separate AGI effort.
These are the latest official details I could verify, from mid-August. If you are reading this later in the year, check the Ministry of Science and ICT for the latest cut, because the schedule is still moving.
6Why Non-Koreans Should Care
Three reasons. First, it is a live experiment in whether a mid-sized country can build its own AI stack rather than rent one. Second, it shows how governments may judge AI in the future, with usability weighted alongside benchmarks. Third, the winners could become the Korean-language engines behind phones, cars and appliances you may already use, since Korean tech firms sit at the center of this effort.
My personal take: the tournament format is almost Korean by nature. It is public, competitive and a little dramatic, and it turns national ambition into a scoreboard anyone can follow.
Key Takeaways
- LG AI Research, SK Telecom and Upstage advanced after the August 18 evaluation; Motif was cut, and Naver Cloud and NC AI left in January.
- Scoring was 40 percent benchmarks, 35 percent expert review and 25 percent user assessment, so usability counted.
- The field is due to shrink to two finalists by the end of 2026, with GPU support rising to about 1,000 B200 equivalents.
Your Turn
Should every country build its own national AI model, or is it smarter to share global ones? Drop your view in the comments, and tell me which Korean team you would back.
Sources: Korea Herald and Korea Times reports of August 18, 2026, and the Ministry of Science and ICT announcement of phase-one results.