South Korea is concentrating 4.7 trillion won, about $3.5 billion, in proposed government money on building a single frontier AI model, a sharp shift from spreading compute across competing domestic teams, with the program potentially starting as early as March 2027. The figure sits in the government’s 2027 budget proposal now before the National Assembly, and the science ministry confirmed the design at an October 2 briefing. Private capital is expected to supplement the public outlay.
The money breaks down into three parts, according to reporting by Maeil Business: roughly 3.9 trillion won for about 10,000 Nvidia next-generation Vera Rubin GPUs, 800 billion won for AI training data, and 25 billion won to recruit 20 top overseas researchers. Those component figures come from reported budget documents rather than a finalized procurement contract, so they can still move during the assembly’s review.
Why the change of course
Until now Korea ran the Sovereign AI Foundation Model Project, a government contest launched in June 2025 in which companies race to build homegrown large models in elimination rounds, with winners receiving government-owned GPUs and support. Five teams started; Naver Cloud and NC AI were cut after the first evaluation in January, Motif Technologies fell out in August, and three teams remain: SK Telecom, LG AI Research and Upstage. Their models are SK’s A.X K2, LG’s K-EXAONE 2.0 and Upstage’s Solar Open 2, with published second-stage scores of 70.6, 69 and 69.9 points respectively.
The default plan had been simple: the two surviving teams would keep receiving compute support through 2027. The government dropped that. Instead of automatically extending support, it will run a new competition for the frontier funding, open to other companies including startups with strong technology and specialist talent. Second Vice Science Minister Ryu Je-myung said officials will “select two. That is certain” from the current field, per The Korea Herald, but those two get no guaranteed road into the new program.
The problem with the old design, ministry officials now argue, is scale. Ryu said current support, roughly 1,000 high-end GPUs per remaining team, is not enough to compete with frontier-class models. “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,” he said. So the government split its AI strategy into two tracks: keep the existing contest through its final round, and build a separate frontier program with an order of magnitude more compute concentrated on one model.
| Program | Status | Funding |
|---|---|---|
| Sovereign AI contest | Running, two finalists due February 2027 | Existing GPU support continues |
| Frontier AI program | Budget pending assembly approval | 4.7 trillion won for 2027 |
| GPU line item | Allocation undecided | 3.85 trillion won, about 10,000 Vera Rubin |
| GPUs already secured | Government-held plus private leases | 7,000 owned and 8,000 leased |
The open questions
Several things remain unsettled, and they are bigger details than they sound. The ministry has not decided how the roughly 3.85 trillion won of GPU funding will be distributed or which developer or developers get the hardware. The investment structure is also undecided, with a consortium or a special-purpose company under review, and the ministry is weighing roles for major companies, startups and other investors. Requiring participating companies to provide matching private funding is under consideration, which would stretch the public money but also raise the entry cost for younger firms.
The frontier competition will be open to new entrants including startups, which means the three current contestants could face each other again in a fresh contest, this time for a far bigger prize. The ministry is considering incentives for the two finalists, recognizing the resources they have already committed, but has not confirmed them. Earlier in the month there was public confusion over whether the sovereign contest would simply be scrapped; the October 2 briefing was partly a clarification exercise. Both tracks proceed, but separately. Deputy Prime Minister Bae Kyung-hoon first floated the frontier plan in July with a more than 3.5 trillion won estimate, and the budget proposal concretized it.
The timetable, as reported by Technology Org and others: the sovereign contest’s third stage runs through January, with final two teams selected around mid-February 2027. The assembly is expected to pass the 2027 budget in December. Applications for the frontier program open as soon as the budget is finalized, and development could begin as early as March. Businesses chosen will get support before the new GPUs arrive, using the government’s existing hardware and leases.
What Korea is actually buying
The stated aim is a domestic model that reaches parity with leading Chinese open-source systems, not the top US labs, an honest calibration that says a lot about how Seoul sees its position. Korea hosts world-class chip manufacturing but has no frontier lab of its own, and its consumer internet companies have watched foreign models take the software layer. The equity-investment structure, rather than straight grants, signals the government wants commercial discipline attached to the money and expects an asset it controls.
The scale comparison explains the urgency. A frontier training run at the top labs now involves compute in the hundreds of thousands of accelerated GPUs. Ten thousand Vera Rubin units is serious compute, but it is a one-model bet against campaigns, not a standing racing program. That is why the ministry’s language focuses on concentrating chips, data and talent on a single project instead of spreading resources across many teams.
The calendar that decides everything
Passage of the budget in December is the gating event. If it slips, or the GPU allocation changes, the whole timetable moves with it. The budget sits inside a larger political fight over next year’s spending, and AI has supporters and skeptics across parties. The debate is expected to continue through the assembly’s review, per Korea JoongAng Daily.
For Korea’s AI industry the two-track design answers one question and opens another. Whether the two finalists complain about competing against the government that just funded them, and whether any startup can credibly assemble a matching-funding share, the answers arrive with the December vote. Watch December, then March.