South Korea wants a frontier AI model that can compete with the best Chinese open-source systems, and it plans to put 4.7 trillion won, about $3.48 billion, into building one. Second Vice Minister of Science and ICT Ryu Je-myung laid out the plan at a Friday press conference, according to The Korea Times, which reported it on October 5. The project starts next year. The striking detail is how little of the money is for anything except chips.

Most of the money is for hardware

The Korea Times reports that 3.9 trillion won is expected to go toward 10,000 Nvidia Vera Rubin GPUs, with 800 billion won set aside for training data. That makes hardware about 83 percent of the 4.7 trillion won, and data roughly 17 percent. The budget bets that scarce compute, not scarce data, is what holds Korean labs back.

Some arithmetic shows how big the line is. At the exchange rate implied by the Korea Times figures, about 1,350 won to the dollar, 3.9 trillion won is roughly $2.9 billion. Spread over 10,000 GPUs, that is about $290,000 per chip. The line item probably covers servers, networking and installation as well as the processors, but even so it is a large sum for hardware that, as we reported after its MLPerf debut, was still not shipping when it posted its benchmark results. The Korea JoongAng Daily puts the GPU share at 3.85 trillion won and notes that no specific plan exists yet for how the processors will be distributed.

The state wants a stake, not just a subsidy

The government will invest equity, not hand out grants, and a special purpose company is one option for carrying it out, Ryu said, as the Korea Times reported. The reasoning is that returns can eventually be shared with the public. Companies that join must match the government's money with private funds, and the ministry plans to publish application criteria later this month.

Much remains undecided. The Korea Times says the structure, the anchor companies and the role of startups are not finalized. Technology.org reports that a lead developer would be chosen by competitive tender after parliament passes the 2027 budget in December, possibly as early as February, with the program starting in March 2027. Nothing here is binding until lawmakers vote.

Why the earlier contest was not enough

Korea already runs a national foundation model competition, launched in 2025. Three teams are in its third stage, and two finalists will be picked by the end of this year. The original plan gave each finalist 1,000 GPUs for 2027, or 2,000 in total. The new project's 10,000 GPUs are five times that total.

Ryu admitted the survival-style format created a stigma for eliminated companies and public controversy, according to the Korea Times. He said participating companies told the ministry that frontier-level competition needs pooled resources, and that officials saw the global race change after the Claude Mythos preview in April. The two finalists may now choose between joining the frontier project and focusing on industry-specific models. The sequence says something about Seoul's thinking: it spread GPUs thinly across rivals, saw that the approach could not reach the frontier, and is now concentrating resources instead.

The benchmark is Chinese open-source models

The stated target is notable. The Korea Times says the model is meant to compete with leading Chinese open-source models, not with closed American systems. That is a measurable goal, because open-weight rankings are public and updated often. It also reflects where the pressure comes from. Chinese labs have made open weights a diplomatic product, as Xi's proposed BRICS open-source AI zone showed last month.

There is an irony in the budget. A project meant to build Korean independence is expected to spend more than four of every five won on a single US supplier's hardware. That is a defensible trade if compute is the binding constraint, but it ties Korean ambitions to Nvidia's delivery schedule.

What could still go wrong

Three risks stand out. The first is the December budget vote. The second is the unresolved structure, including who anchors the special purpose company and how GPUs are shared, which the JoongAng Daily describes as a source of confusion. The third is timing. The March 2027 start assumes deliveries of a chip generation that had not yet shipped when this plan was announced, which is my inference from the benchmark coverage, not something the ministry has said.

If those pieces hold, Korea gets a large, state-aligned compute pool. If one slips, the money will sit in a budget line while rival labs keep training.