The Short Answer
South Korea's Ministry of Science and ICT published second-phase scores for its sovereign AI foundation model project on 27 August 2026, with SK Telecom leading on 70.6 points ahead of Upstage on 69.9. Two teams chosen in December will supply a national assistant for all 51 million residents. The transferable part is the procurement method, not the model. On ibl.ai you own all the code and the data, so the evaluation stays yours to re-run.
Most governments announce a national AI champion and a budget. South Korea is publishing a ranked scorecard, and that difference is what other buyers should be studying.
What did South Korea actually announce?
South Korea released the second-phase evaluation results of its Sovereign AI Foundation Model Project, ranking the competing teams by score.
The published results put SK Telecom first on 70.6 points, followed by Upstage on 69.9, LG AI Research Institute on 69, and Motif Technologies on 65.8.
Two teams that win the final December evaluation become the primary technical suppliers for "AI for All" — a programme intended to give all 51 million Korean residents free, unlimited access to a national AI assistant.
The wider programme received Cabinet approval in May 2026 with roughly $5.7 billion allocated through the National Growth Fund, and targets models reaching at least 95% of the performance of leading global systems.
Why is the evaluation method more interesting than the model?
The evaluation method is more interesting because a national model is a capability while a published, contestable evaluation is an institution — and institutions outlast models.
Part of South Korea's process used a demographically weighted citizen lottery: a panel of ordinary residents scoring candidate systems across a four-day evaluation in August, with absolute-scoring results feeding into the decision on which team was eliminated.
It appears to be the first documented use of citizen scoring to govern national AI procurement.
Set aside whether a citizen panel is the right instrument for judging model quality — reasonable people will disagree, and the panel is only one input among several.
The structural point holds regardless: the criteria are declared in advance, the results are published, and the outcome is attributable.
Compare that to the default. A ministry announces a partnership with a national champion, the evaluation is internal, the scores are never released, and nobody outside the committee can assess whether the decision was sound. Both approaches produce a model.
Only one produces a record.
What does this mean for "sovereign AI" as a term?
It means sovereignty has at least two components, and most of the discussion only covers one.
The component everyone discusses is where the system runs — whose infrastructure, whose jurisdiction, whose keys.
That question is real, and it is the one we examined in the UK-Ukraine AI declaration, where a non-binding agreement committed the parties to respect each other's data sovereignty while specifying no architecture at all.
The component this announcement surfaces is whether the choice is auditable — whether a government can show which system it selected, on what evidence, against which alternatives, and whether it retains the ability to change its mind.
A state that hosts a model it cannot evaluate, cannot compare, and cannot replace has custody without control. That is a weaker position than it appears, and it is not what most people mean when they say sovereign.
For the broader definition and where the term is used loosely, see what is sovereign AI.
Does building a national model actually deliver sovereignty?
Building a national model delivers one part of it, and creates a dependency of its own if the surrounding layer is neglected.
The case for building is genuine: a model trained on Korean language and Korean context, under domestic governance, with no foreign vendor able to change terms, restrict access, or be compelled by another jurisdiction.
At $5.7 billion and a 95%-of-frontier target, the ambition is coherent.
The risk is that a national model becomes a national single point of failure.
If every public institution routes to one state-selected system, the state has replaced vendor dependency with sovereign monoculture — and the ability to evaluate, compare, and swap becomes more important, not less.
Which is exactly why publishing the scores matters more than picking the winner. A country that ranks four systems on declared criteria has built the muscle to do it again in three years, when the leaders have changed. A country that anointed one has not.
What transfers to organizations that will never build a foundation model?
What transfers is the procurement discipline, and it applies to any institution buying AI at scale.
Declare the criteria before the demos. Vendors optimise for whatever they think you will measure. Deciding afterwards guarantees you measure what they showed you.
Score against alternatives, not against a threshold. "Meets requirements" is not comparable across vendors. A ranked score forces the trade-offs into the open.
Keep the evaluation set. The single most valuable artifact of a selection process is the test corpus, because it is what lets you re-run the decision later. Vendors change, prices move, and capability leadership rotates within a year.
Make the decision reversible. Reversibility is a property of the architecture. If switching requires rewriting applications, the evaluation was decorative — you were always going to stay.
On ibl.ai you own all the code and the data, run it model-agnostic across any LLM, and pay with no per-seat pricing — so you can deploy anywhere, from your own cloud to a fully air-gapped network, and re-run the model decision without re-running the project. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
Frequently asked questions
Who is currently leading South Korea's sovereign AI evaluation?
SK Telecom led the second-phase scores published on 27 August 2026 with 70.6 points, ahead of Upstage (69.9), LG AI Research Institute (69), and Motif Technologies (65.8). Final selection is scheduled for December.
What is "AI for All"?
A South Korean government programme intended to provide all 51 million residents with free, unlimited access to a national AI assistant, to be supplied by the two teams that win the final evaluation.
Is a citizen panel a sound way to evaluate AI models?
It is one input among several, and it is better suited to judging usefulness and acceptability than technical capability. Its more durable contribution is procedural: it forces the criteria and results into public view.
The bottom line
The $5.7 billion and the 95%-of-frontier target will get the coverage. The scorecard is the part worth copying.
Sovereignty is usually framed as a question about where the weights sit. This announcement is a reminder that it is also a question about whether you can show your work — and whether, having chosen, you could still choose again.