Loading page…
Most of the proposed frontier-program funding would expand GPU capacity. Who receives that computing power remains undecided as lawmakers review the budget.
Listen to this story
The proposed 2027 frontier-AI program would add a second track to South Korea’s domestic effort, not replace its existing model contest. Of the proposed 4.7 trillion won, 3.85 trillion is earmarked for GPU capacity, but the ministry has not decided how that computing will be allocated. Parliamentary approval is still pending, as are incentives for the two contest finalists; developers therefore lack clarity on access to resources and how the two tracks will work together.
SK Telecom, LG AI Research Institute and Upstage are competing for two final places in the existing contest.
The contest targets models reaching about 95% of leading systems’ performance; the ministry says faster AI progress prompted a separate frontier effort.
Applications for the frontier program are expected after the budget is finalized at year-end, with development potentially starting in February 2027.
South Korea will keep its competition to build homegrown AI models while pursuing a separate frontier-model program, the Ministry of Science and ICT said at an October 2 briefing. The clarification follows confusion over whether the contest would be scrapped. The new program carries a proposed 4.7 trillion won budget for 2027, subject to parliamentary review, Korea JoongAng Daily reported.
Second Vice Science Minister Ryu Je-myung described the redesign as two tracks: frontier-model development, aimed at cutting-edge AI systems, and industrial deployment. He also said the ministry was considering incentives for the two companies ultimately selected in the existing competition. Those incentives are under consideration, not a confirmed award.
Launched in June 2025, the Sovereign AI Foundation Model Project uses elimination rounds to choose Korean model developers for government-owned GPUs and other support. Its target is about 95 percent of the performance of the world's leading systems. Three teams remain—SK Telecom, LG AI Research Institute and Upstage—competing for two final places.
That ambition helps explain the criticism of the redesign: the existing contest already sought performance close to frontier models. Ryu argued that faster advances had widened the gap and required a new response. Asked whether the original project had been poorly designed, he said the course of AI development had been difficult to predict.
The separate frontier effort was first floated by Science Minister Bae Kyung-hoon in July. In August, the ministry promised a complete overhaul of the sovereign project. The subsequent budget proposal placed the new funding under the frontier program rather than the existing contest.
The proposed 2027 frontier-program budget remains subject to National Assembly review.
The proposal earmarks this amount for expanding GPU capacity.
Most of the proposed funding would go toward GPUs, the processors used to develop and run AI models. But no specific plan has been set for distributing them. That leaves a central part of the redesign unsettled: which developers will receive the computing resources.
Rep. Kim Sun-min of the Rebuilding Korea Party challenged the separation, saying the government described two distinct programs but submitted a combined plan to the National Assembly. The ministry said it had allocated no separate GPU budget for the sovereign project this year, while personnel costs were included in next year's proposal.
Applications for the frontier program are expected to open after the budget is finalized at year-end. Development could begin as early as February 2027, when the sovereign competition's final teams are due to be decided, the ministry said. Both the funding and the development schedule remain prospective.
Loading discussion...
Join the conversation
Explain whether speed or a clear allocation plan should come first.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.