South Korea Bets $3.5B on a Homegrown Frontier AI Model
3 min readSouth Korea plans to put 4.7 trillion won, about $3.49 billion, into building a homegrown frontier AI model, and it will reopen the competition for that money rather than hand it to the teams already in its national model program. Second Vice Minister of Science and ICT Ryu Je-myung set out the plan at a briefing in Seoul on Friday, as reported by The Korea Herald on October 5.
Background: the Dokpamo program
Korea’s independent foundation model project, known as Dokpamo, has been running a multi-stage contest among consortia led by LG AI Research, SK Telecom and Upstage. Each surviving team receives government computing support worth roughly 1,000 high-end GPUs, with two finalists due to be picked after a third-stage evaluation that runs through January.
The program is the centerpiece of Seoul’s push for sovereign AI, the idea that Korea should own a frontier model rather than depend on US or Chinese labs. Korea’s chip industry already supplies much of the world’s AI memory, but its model builders have lagged far behind.
What changes for South Korea frontier AI funding
The government will split the program into two tracks from 2027. The first is a frontier model project backed by the 4.7 trillion won equity investment in the 2027 budget proposal, with private capital expected to top up the public money. The second track will fund the deployment of Korean models across industries, giving smaller and specialized models a route to market without frontier-scale compute.
The frontier track will be a new open competition. The two Dokpamo finalists will still be chosen in mid-February, and Ryu said the ministry is weighing incentives for the resources they have already committed, but the new funding will be open to other companies, including startups. The investment vehicle is undecided; a consortium and a special-purpose company are both under review.
Ryu was blunt about the reason. Investment on the scale of 1,000 GPUs is not enough to catch up with frontier-class models, he said, and the scale has to be completely different. The ministry is also preparing to distribute about 29,000 GPUs the government has bought or is bringing online, with an allocation plan due by year-end.
Why it matters
Ryu conceded that going head to head with the largest US labs is unrealistic. The stated target is a model competitive with the leading Chinese open models, which have become the benchmark for every national AI program outside the US. That is a more candid goal than most sovereign AI plans admit to, and $3.5 billion is still a fraction of what a single frontier lab spends in a year.
The budget needs National Assembly approval, expected in December, and the new contest could open as early as March 2027. Watch whether LG, SK Telecom and Upstage accept a reopened race, and which startups step forward to claim a share of the GPUs.
Korea is doubling its bet on a national model while admitting the first round was underfunded. The next six months will show whether a bigger check changes the outcome.