Seoul splits one programme into two
South Korea is separating its national AI effort into two tracks, the second vice minister of science and ICT, Ryu Je-myung, told reporters. One track chases a frontier model. The other pushes Korean models into industry. Behind the first sits a proposed government equity investment of 4.7 trillion won, about $3.49 billion, with private capital expected on top.
“We will evolve the technology development goal toward frontier models, while separating industrial deployment into a new track,” Ryu said.
The split is an admission about what the existing competition was and was not doing. Korea has been running a staged contest to pick national champions, and the labs in its third stage — LG AI Research, SK Telecom and Upstage — are competing to be among two finalists. That stage runs through January, with results expected in mid-February.
The money is proposed, not appropriated
The important qualifier is the budget calendar. The new frontier competition opens only after the National Assembly approves the 2027 budget, expected in December, which puts a plausible start around March. A plan for allocating GPUs is expected by the end of this year.
That makes the 4.7 trillion won a request rather than a commitment, and it makes the schedule contingent on a vote. It also means the new contest — which will be open to companies and startups beyond the three now in the running — begins roughly a month after the current one finishes, giving the winners of the old process very little time before the new one starts.
Korea’s position in the field is the reason for the urgency. Chinese labs have spent 2026 holding the open-weight frontier with DeepSeek, Alibaba’s Qwen line and Moonshot AI’s Kimi models, and the Korean entries have been competitive at the second tier rather than the first. Buying a seat at the frontier with state equity is a different proposition from subsidising adoption, and Seoul has now decided it wants both, in separate budget lines.
What a sovereign model buys
The case for a national frontier model is not usually made on benchmarks. It is made on control: a model trained domestically, hosted domestically, and not subject to another government’s export rules or another company’s terms of service. France made a version of this argument with Mistral, the EU makes it with its sovereignty language, and Korea is now making it with a number attached.
What the announcement does not settle is the hardest part. A frontier training run needs accelerators, and the GPU allocation plan is still to come. Until that lands, 4.7 trillion won is a budget line looking for hardware.
The dates to watch are the December budget vote, the GPU allocation plan before year end, and the mid-February result of the contest now running. Those three, in that order, decide whether the March start is real.
