South Korea will launch a state-backed programme to develop a frontier artificial intelligence model beginning next year, backed by 4.7 trillion won (roughly $3.5 billion) in government equity investment, the Ministry of Science and ICT has announced. Second Vice Minister Ryu Je-myung told reporters in Seoul that the project will be written into next year's budget, with Reuters reporting the programme is slated to start from March 2027 once funding clears the National Assembly.
The announcement marks one of the largest single national commitments to frontier-scale AI development outside the United States and China, and signals that Seoul wants its model developers competing at the very top tier rather than a step below it. For more context on this story, see our ongoing AI trends.
How the $3.5 billion breaks down
The headline figure is split into two major line items. Around 3.9 trillion won is earmarked to secure approximately 10,000 Nvidia Vera Rubin GPUs — the computing backbone for training a frontier-class model. The remaining 800 billion won is allocated to purchasing training data, an increasingly decisive input as labs exhaust commonly scraped web corpora.
Unlike the grant-style support that characterised Korea's earlier AI programmes, the government plans to take an equity stake. Companies selected through open application and evaluation processes will be required to match the government's investment with private funding of their own, and officials say returns from the project could ultimately be shared with the public. One option under consideration is a special purpose company that would house the investment and involve the country's top AI model developers, according to the Korea Times.
Details such as the project's structure, anchor companies and startup participation have not been finalised. The ministry plans to publish the application criteria later this month.
Why Seoul is changing strategy
Ryu framed the decision as a response to a step-change in the global race. In remarks reported by the Korea Times, he said the government and domestic companies agreed that competition for advanced AI models "entered a completely different phase" after Anthropic previewed its Claude Mythos model in April, with rivals now competing at a scale that demands investment on an entirely different level.
"The question was whether we should target a mid-tier model or one capable of competing with the world's leading models," Ryu said, explaining that companies told the ministry that frontier-level competition would require pooling resources rather than distributing them across competing teams.
"The frontier AI project will begin next year," he added. "It will be reflected in next year's budget, and we will pursue a higher goal by breaking through current technological limitations."
The Korea Times reports the explicit ambition is to compete with leading Chinese open-source models, which have set the pace on cost-efficient capability throughout 2026.
The existing national model project continues
The new programme does not replace Korea's ongoing national AI foundation model project, a survival-style competition launched in 2025 to develop the country's most capable homegrown models. Three teams remain in the third stage of that contest, with two finalists due to be selected by the end of this year. Under the original plan, each finalist was to receive 1,000 GPUs through 2027 — support that will continue, but with a shifted emphasis.
Ryu said the foundation model project's focus will move toward models tailored to specific industries and real-world applications, while the frontier project takes on the race for raw capability. That division of labour effectively creates a two-track national AI strategy: one track commercialising applied models for Korean industry, the other chasing the global frontier.
From outline to commitment
The announcement is the culmination of a fast-moving policy arc. In July, Korean media reported that the government was outlining a plan of more than 3.5 trillion won to develop a homegrown frontier model as part of its sovereign AI push, alongside a security-focused model for the public sector. The figure unveiled this week — 4.7 trillion won in equity investment — goes beyond that outline, and the equity-based structure goes beyond the subsidy-plus-competition format Korea used to seed its foundation model race.
The matching-fund requirement is the mechanism that makes the sum larger than it looks. Because participating companies must contribute private capital on top of the government stake, the total pool available for compute, data and engineering could materially exceed the public commitment. It also disciplines the programme in a way pure grants do not: firms with their own money at stake have stronger incentives to ship a competitive model rather than a research artefact.
Why compute scale is the sticking point
The GPU line item is the clearest signal of intent. Nvidia's Vera Rubin generation is the company's latest data-centre platform, and 10,000 units places Korea's project in the same order of magnitude as the training clusters that Western and Chinese frontier labs are assembling. For comparison, the two finalists in the ongoing national foundation model competition were each originally slated to receive 1,000 GPUs through 2027 — support Ryu confirmed will continue, but an order of magnitude below what the new project contemplates.
Securing allocation will be its own challenge. Advanced Nvidia accelerators are sold out well into the future, and hyperscalers with deeper pockets are competing for the same supply. The government's willingness to commit 3.9 trillion won specifically for GPUs suggests it has been negotiating with that constraint in view.
A crowded sovereign AI field
Korea is far from alone in treating frontier AI as a matter of national capability. Governments from Saudi Arabia to Japan have committed multi-billion-dollar sums to domestic compute and model development, and Korean conglomerates have already struck some of the largest AI infrastructure deals on record. The new programme adds a purpose-built frontier model to that pipeline, with government money and private capital stacked together for the first time.
Whether the project survives next year's budget process intact remains the near-term question. But with 10,000 next-generation GPUs on the table and matching obligations that keep private participants financially committed, Seoul has made clear it intends to be a frontier contender rather than a fast follower. As the latest AI developments show, the countries that secure compute and capital this year will be the ones setting the terms of the race in 2027.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
