South Korea Bets Big on Home-Grown AI Models

On 4 August 2025, South Korea’s Ministry of Science and ICT (MSIT) named five “elite teams” that will each build a proprietary large-scale AI foundation model. The government will pour GPUs, datasets and talent grants into the project, then narrow the field to two national models by mid-2027.

Why South Korea Wants Its Own AI

The global AI race is dominated by U.S. and Chinese tech giants. Without a native model, Korean services—from search to smart devices—risk permanent dependence on foreign platforms. MSIT therefore set a pragmatic target: reach at least 95 percent of the performance of the newest global models while leaving room to leapfrog later. That goal frames the effort as a sovereignty play rather than a vanity project.

Meet the Five Winning Consortia

Naver Cloud and LG AI Research come with proven Korean-language giants HyperCLOVA X and EXAONE, giving them massive local data troves. NC AI plans to fold interactive-character know-how into a multimodal model, and start-up Upstage brings deep open-source fine-tuning chops. Yet the most ambitious roadmap belongs to SK Telecom, whose “full-stack” pipeline stretches from custom chips to mass-market services.

Inside SK Telecom’s Full-Stack Strategy

SK Telecom has been training its A.X large-language-model series on TITAN, the company’s in-house supercomputer; this year alone it unveiled A.X 4.0 (standard / light) and A.X 3.1 (standard / light) variants that benchmark near GPT-4o on overall quality and lead on Korean-context reasoning.  Future training will add thousands of GPUs plus Rebellions’ AtomMax neural-processing units, cutting both latency and energy draw for inference at scale.

The consortium’s chip-to-service stack links AtomMax silicon, the A.X models, and telecom-scale data streams from more than ten million A.Dot virtual-assistant users, giving the team a uniquely broad corpus for continual pre-training and real-time reinforcement.  Academic muscle comes from Seoul National University, KAIST and the University of Wisconsin–Madison, whose faculty join core research on model reliability and rapid inference.  In total the consortium counts 800+ research papers, 736 patents and 270 open-source projects in its war chest, signalling depth far beyond a single corporate lab.

Industrial partners extend that reach: game studio Krafton validates the models inside next-gen titles, mobility firm 42dot pushes on-device compression, Liner fine-tunes search LLMs, and SelectStar curates high-fidelity datasets—turning the project into a miniature AI supply chain.  The end goal is an “omni-modal” AI that reads text, sees images, hears speech and analyzes video in one go, ready to automate office workflows, factory QA, in-car assistants and even game characters.

Government Support Package

Each team starts with roughly 500 high-end GPUs and can scale past 1 000 if milestone tests are met. Data support combines a ₩10 billion (≈ US $7.6 million) joint-purchase fund, ₩2.8 billion per team for bespoke curation, and an extra ₩20 billion media-dataset program to lift multimodal accuracy. Matching grants help consortia recruit overseas Ph.D.s through 2027, anchoring hard-to-replace talent.

What to Watch Next

• December 2025 prototypes will reveal how close Korean models are to GPT-4-class systems.
• AtomMax rollout could slash inference costs, testing whether domestic NPUs rival NVIDIA in real workloads.
• Open-source releases—A.X Encoder, A.X 4.0 VL Light and more—will show whether global developer communities adopt Korean models or stick with Llama/Mistral.
• Semi-annual “moving-target” reviews will trim the roster, forcing consortia to out-innovate not just American giants but their local peers as well.

In short, Seoul’s bold wager could shift the country from AI follower to independent innovator—but only if teams like SK Telecom’s turn their chip-to-cloud ecosystem into real-world services that users everywhere choose.