South Korea’s frontier AI model program is planned as a 4.7 trillion won, or roughly $3.5 billion, effort beginning in March 2027. The government wants to concentrate computing chips, data and technical talent behind a model capable of competing near the industry’s performance frontier.

The Ministry of Science and ICT expects to choose a lead developer through a competitive tender after parliament considers the 2027 budget in December. A winner could be selected as early as February, leaving a short transition from procurement to the proposed March launch.

The planned program has four defining elements:

• A 4.7 trillion won public-private funding structure.
• A competitive tender for one lead developer.
• Large-scale support in chips, data and talent.
• A March 2027 target for work to begin.

South Korea Frontier AI Model Program Starts With a Tender

Reuters reported on October 6 that the ministry plans to combine state equity investment with private financing rather than fund the entire program as a conventional government grant. The 4.7 trillion won figure converts to about $3.5 billion at the exchange rate used in the report.

Participation is intended to be broad. A single company, consortium or special-purpose vehicle could bid, and the ministry said startups, established firms, universities and individual researchers should be able to take leading roles. Detailed eligibility, ownership and intellectual-property rules have not yet been published.

The lead-developer structure is designed to reduce fragmentation. Instead of dividing scarce accelerators and researchers across many equal projects, Seoul would put a concentrated package behind the successful bidder. That could accelerate training, but it also makes tender design, independent evaluation and access for smaller participants unusually important.

The schedule remains conditional. Parliament must first approve the proposed 2027 budget, after which the ministry can run the tender. A February selection and March start are planning targets, not completed milestones, and any delay in appropriations or contracting could move the launch.

10,000 GPUs Anchor the State-Backed Compute Pool

The ministry’s official 2027 budget proposal allocates 5.5937 trillion won to top-tier AI models and related technologies, up from 2.6146 trillion won in 2026. That wider category includes the frontier-model effort, next-generation AI research and security work.

Within the sovereign-model initiative, MSIT says public and private teams will receive support from a pool of 10,000 GPUs, along with data and talent. The ministry has not identified the accelerator models, cloud operators, allocation rules or how much capacity the eventual lead developer will control.

Those details will determine what the headline budget can buy. Frontier training depends not only on the number of chips, but also on networking, memory, storage, power, model architecture and the efficiency of the software stack. A nominal GPU count does not by itself establish training throughput or model quality.

State equity adds another governance question. If public money buys an ownership interest, the tender will need to define who holds the stake, how private investors participate, and whether taxpayers receive financial or licensing rights if the model becomes commercially valuable.

The Existing Sovereign Model Contest Will Continue

The new frontier effort is separate from South Korea’s existing home-grown foundation-model program. That competition is expected to continue and select two finalist teams during a third-stage review in February 2027, according to the ministry.

MSIT issued an October 3 clarification stating that the independent foundation-model project would continue while the country also pursues a top-tier model. An earlier clarification said the creation of a special-purpose company had not been decided.

That distinction matters because the two programs address different policy goals. A multi-team sovereign initiative can preserve domestic capability and experimentation, while a concentrated frontier bid can direct more compute toward a single performance target. Running both may broaden the technology base, but it can also create overlapping spending and competition for the same engineers.

Clear evaluation criteria will therefore be essential. The government has not publicly specified which benchmarks, safety tests, Korean-language capabilities or efficiency measures will define “frontier” performance. Without fixed criteria, a large budget could produce a capable domestic system without demonstrating parity with leading global models.

Budget Approval and Governance Are the Next Tests

The December parliamentary vote is the first hard checkpoint. If funding passes, the tender documents should reveal the public-private capital split, decision rights, compute allocation, data rules, model-access policy and the obligations attached to state equity.

February would bring two separate decisions: a possible lead developer for the new program and two finalists in the existing foundation-model contest. Keeping those selections transparent will help show whether the initiatives are complementary or simply bidding for the same national resources.

The March 2027 start would still mark the beginning of development, not the delivery of a finished frontier model. Training schedules, evaluation results, safety testing and release terms will determine whether the project creates a broadly useful national platform, a commercial product, or a more restricted research asset.

South Korea already has major semiconductor, cloud and consumer-electronics companies, but assembling those strengths into a competitive model is a separate engineering and governance challenge. The proposed program gives Seoul scale; the tender and evaluation framework will decide whether that scale converts into measurable performance.

Related Research

• Hyundai Unveils Atria AI Roadmap With Nvidia Systems Due in 2028
• Google Invests €13 Billion in Finland AI Infrastructure and Nuclear Power