South Korea Bets 4.7 Trillion Won on a Frontier AI Model, With an Open Tender Starting in March 2027

South Korea is preparing its single largest bet yet on staying inside the frontier AI race. The Ministry of Science and ICT said Tuesday that a 4.7 trillion won (about $3.5 billion) program to develop a frontier-class AI model will begin in March 2027, pending parliamentary approval of the 2027 budget expected in December. Reuters first reported the plan, and the ministry laid out the details in a briefing in Seoul.

The mechanics are deliberately open-ended. Rather than extending support to the teams already in its national foundation-model competition, the ministry will select a lead developer through a competitive tender, with the winner potentially named as early as February. The state plans to combine its own equity investment with private capital and to concentrate computing chips, data and talent on a single push. Participation is open to single companies, consortiums and special-purpose vehicles, and the ministry explicitly wants startups, companies, universities and researchers to take leading roles rather than playing bit parts.

The announcement also redraws the boundary of the government's existing homegrown model effort, known in Korean as Dokpamo. That competition, which has been whittling down teams led by LG AI Research, SK Telecom and Upstage, will run to completion as planned, with two finalists selected after a third-stage review in February 2027. Second Vice Minister of Science and ICT Ryu Je-myung rejected any suggestion that the state program was being abandoned, saying the goal of securing independent foundation-model technology "has not changed, and it will not be discontinued."

The numbers behind the shift explain why. Ryu said current support — roughly 1,000 high-end GPUs per remaining team — is simply not competitive: "Investment on the scale of 1,000 GPUs is not enough to catch up with frontier-class models. The scale has to be completely different." The ministry is also preparing to allocate around 29,000 GPUs it has purchased or is bringing online, with a distribution plan expected by the end of this year.

Ryu was candid about the ceiling of the ambition. He said competing head-on with the largest US AI companies is unrealistic, but argued that with the planned public and private investment, Korea could build a model competitive with the leading Chinese open-weight models — the tier set by DeepSeek, Z.ai and Moonshot AI. A second, separate track will fund the deployment of Korean-developed models across industries, giving smaller and specialized models a path to market without requiring frontier-scale compute.

The plan lands amid a broader sovereign-AI scramble in Asia, where governments are trying to buy model independence rather than rent it from US labs. It also arrives as private frontier rounds have inflated to sums no single ministry can match — OpenAI is reportedly courting investors for a $30 billion round at a $1.4 trillion valuation. Korea's wager is that state equity plus concentrated national compute can buy a seat at a table where the entry ticket keeps rising. Whether the December budget holds, the February tender lands, and the March 2027 start date survives contact with a change of government will determine if the bet pays off.