Few governments have compressed as much AI policy into as short a window as South Korea has since mid-2025. In roughly fourteen months, the country has stood up a presidential AI secretariat, passed and enforced what it calls the world’s first comprehensive AI law, committed tens of billions of dollars to compute and chips, launched a national sovereign foundation model contest, and formally merged its civilian and defence AI strategies under a single banner. The banner itself is a number: G3. South Korea’s stated ambition is to rank among the world’s top three AI powers, alongside the United States and China, and nearly every policy move since President Lee Jae-myung took office has been justified, explicitly, in those terms.
A PRESIDENTIAL PUSH BUILT AROUND ONE NUMBER
The AI G3 framing did not emerge from a single strategy document so much as from a rapid sequence of institutional moves. In June 2025, the new administration created a senior presidential secretary role dedicated exclusively to AI, appointing Naver AI Innovation Center head Ha Jung-woo to the post, and tasking the office with coordinating policy across the Ministry of Science and ICT, known as MSIT, and a newly empowered National AI Strategy Committee. That committee was formalised through a Presidential Decree approved in September 2025, with officials framing the move as necessary to marshal the country’s full national capability at a moment when the global AI contest had shifted from rivalry among individual companies to an all-out contest among nations. The comparison points used internally were blunt: the United States’ Stargate infrastructure programme, valued at roughly KRW 700 trillion, and the European Union’s own AI investment push, estimated at around KRW 300 trillion, were both cited as the scale South Korea now considers itself competing against.
The financial commitments that followed matched that framing. A KRW 100 trillion public-private investment fund, worth close to USD 74 billion, was announced to expand AI infrastructure, talent, research and development, and data capacity, complemented by a KRW 1.8 trillion supplementary budget specifically earmarked for AI capability building. A National AI Capability Enhancement Plan set a target of securing over 50,000 GPUs in 2025 alone, including 10,000 high-end units, with a longer-range MSIT goal of reaching 260,000 advanced GPUs by 2030. By May 2026, the ministry’s own year-one progress report under the Lee government credited itself with laying the groundwork for the G3 push, restoring the country’s research and development ecosystem, and guaranteeing basic communications rights, while pointing to a second-year agenda built around expanding services on homegrown AI models, launching a so-called K-Moonshot project, and shifting toward a dedicated AI security posture.
THE WORLD’S FIRST COMPREHENSIVE AI LAW
South Korea’s most closely watched policy achievement is legal rather than financial. The AI Framework Act, more formally titled the Framework Act on the Development of Artificial Intelligence and the Establishment of Trust, or AI Basic Act, was passed by the National Assembly in December 2024 after more than four years of legislative debate that had first begun in 2020. It took effect on 22 January 2026, following a one-year grace period during which enforcement decrees and detailed guidelines were drafted and revised through public hearings. Korean officials have been explicit that the timing is strategic. While the European Union’s AI Act technically arrived first, its key provisions do not take full effect until 2027, leaving Seoul positioned to become the first government anywhere to actually enforce a comprehensive, unified AI law in real terms rather than on paper.
The law consolidates nineteen separate bills into a single governance framework, establishing a National AI Committee, an AI Master Plan, and an AI Safety Institute, while adopting what officials describe as an innovation-first principle: companies may develop and deploy AI systems without seeking prior government approval, a deliberate break from precautionary regulatory models. That permissive default comes paired with obligations. The Act defines a category of high-impact AI, covering systems deployed in sensitive domains such as medical devices, energy infrastructure, hiring algorithms, and nuclear facility controls, and requires operators of such systems to conduct internal risk assessments, maintain ongoing monitoring, and report safety protocols to government bodies. A separate transparency provision requires all AI-generated content, whether image, video, or audio, to carry visible labelling as well as machine-readable, invisible watermarking, aimed squarely at curbing deepfakes and disinformation. Penalties for noncompliance, which can reach KRW 30 million, were suspended during the one-year grace period specifically so companies would have room to adjust rather than face immediate punitive action.
Implementation has not been entirely smooth. Development of the Act’s subordinate enforcement decrees was reportedly paused for a period following cabinet and leadership reshuffles inside the new administration, and a Startup Alliance survey found that roughly 98 percent of local AI startups had yet to initiate formal compliance steps as the law took effect, many citing unclear guidance on what precisely counts as high-impact AI. Industry sources and legal scholars analysing the law have flagged genuine jurisdictional overlap between MSIT and Korea’s other digital regulators, along with concern that ambiguous definitions could leave companies caught between rushed timelines and vague standards. Officials have acknowledged the rollout will not be seamless, framing the one-year grace period, and the law’s built-in capacity to evolve through executive decree, as the mechanism meant to absorb that friction over time.
COMPUTE, CHIPS, AND THE SOVEREIGN MODEL CONTEST
Behind the legal architecture sits a large, fast-moving infrastructure build. MSIT’s National AI GPU Infrastructure Initiative, announced in July 2025, committed roughly KRW 1.46 trillion, close to USD 1.1 billion, to procure 13,000 high-performance GPUs distributed across three domestic cloud providers, Naver Cloud, NHN Cloud, and Kakao, with a unified online support platform allowing universities, research institutions, and companies to apply for allocations based on project priority. Minister Bae Kyung-hoon described the effort as a first step toward building an AI superhighway for the country. Smaller, targeted allocations have followed, including 264 high-end GPUs set aside specifically for small and medium enterprises and startups through the Ministry of SMEs and Startups, split between manufacturing-focused quality control projects and an early-stage founder incubator.
The most consequential infrastructure race, however, runs through model development rather than raw compute. South Korea’s Sovereign AI Foundation Model project, coordinated by MSIT, is designed to fund the creation of a domestically built large language model capable of competing with leading American and Chinese systems, and the competition among candidate consortia has already produced a significant shakeout. In the project’s first-round evaluation in January 2026, Naver Cloud and NC AI were eliminated, Naver reportedly over concerns that its model’s vision encoder had been built on a Chinese-origin foundation rather than developed independently, undercutting the project’s sovereignty rationale. LG AI Research, whose K-EXAONE model consortium had scored highest in the initial evaluation, advanced alongside SK Telecom and Upstage. Naver has indicated it does not intend to reapply, a striking outcome given that the company’s HyperCLOVA X model family had previously been the most prominent example of Korean-built generative AI. The government has said only two consortia are expected to survive as officially designated national AI champions by 2027, a competitive-elimination approach that closely mirrors how Korea historically built its dominant position in semiconductors and telecommunications, backing domestic rivals against each other before any of them competed globally.
DEFENCE AND CIVILIAN AI FORMALLY CONVERGE
In May 2026, South Korea took a step that had been discussed for some time but not formally executed: merging its public-sector and defence AI strategies under a single coordinated framework. The Korea Institute for Defense Analyses, known as KIDA, and the National Information Society Agency, or NIA, held a joint seminar in Seoul under the banner of strategic linkage between public and defence AI transformation in pursuit of the national G3 goal, and signed a memorandum of understanding formalising cooperation on AI policy and infrastructure between the two sectors. As part of that agreement, KIDA plans to expand its existing Defense Artificial Intelligence Policy Research Office into a larger body, tentatively named the Defense AI Policy Center, intended to oversee defence AI strategy, data planning, and verification while maintaining a standing coordination channel with the NIA’s civilian AI policy division. Officials involved described the goal as building a single nationwide AI ecosystem linking private industry, government, and the military, rather than continuing to pursue AI transformation separately across those domains, as had largely been the case until this point.
That convergence has produced concrete downstream activity. In July 2026, South Korea’s Agency for Defense Development launched an internally built, secure AI system named Add-i, developed entirely in-house over three months using open AI models specifically because classified defence networks had previously limited the use of outside commercial AI services; a more advanced, agent-based successor, Add-i 2.0, built to support the Model Context Protocol standard for connecting AI agents, is already in development. Separately, Naver and Korea Aerospace Industries announced a partnership to build a defence-specific sovereign AI foundation model, explicitly framed around reducing dependence on foreign technology and strengthening what both companies called Korea’s defence technology sovereignty. OpenAI has also moved to position itself inside this ecosystem, announcing a Korea Cyber Action Plan in May 2026 to expand AI-based cyber defence cooperation with government agencies including the science ministry, the Foreign Ministry, and the National AI Strategy Committee, drawing explicitly on its prior experience working alongside the AI safety institutes of the United States and Britain.
AN OUTLIER APPROACH BY DESIGN
Comparative analysis of AI governance across the Indo-Pacific consistently singles out South Korea as an outlier, not because its ambitions are unusual but because of how centralised and legally formalised its approach has become relative to neighbours that lean more heavily on industry-led voluntary guidance. That posture has been traced to South Korea’s particular combination of strong domestic digital platforms, in contrast to markets more dependent on foreign ones, and a long-standing pattern of close coordination between government and the country’s major conglomerates, or chaebols, which gives Seoul both the incentive and the institutional capacity to regulate assertively in ways that protect domestic industry. Researchers examining the region’s governance spectrum place South Korea closer to the centralised, top-down end than most of its peers, while noting it still falls short of a fully state-directed, system-wide model. The AI Basic Act’s high-impact AI category, covering healthcare, transportation, finance, water systems, and nuclear safety, was cited by that same research as a clear marker of this posture, alongside industry criticism that portions of the law remain vague and difficult to implement in practice.
Nine months into enforcement of the world’s first comprehensive AI statute, and roughly a year into the presidentially chaired push toward G3 status, South Korea’s strategy reads less like a single coherent document and more like a coordinated sprint across every lever the state has available at once: legislation, procurement, chip policy, model competition, and now defence integration, all pointed at the same numerical ranking. Whether that sprint produces a genuine top-three AI power or simply the world’s most closely watched regulatory experiment likely depends on the same unresolved question industry groups have been raising since January: whether Korea’s enforcement decrees, model competitions, and infrastructure buildout can mature fast enough to match the speed at which the government has already declared its ambition.
KEY REFERENCES AND FURTHER READING
- National AI Strategy Policy Directions — Ministry of Science and ICT, Republic of Korea
- South Korea Launches Public-Defense AI Alliance to Pursue ‘AI G3’ Goal — UPI
- Why South Korea Is Vying to Be First to Regulate AI — The Korea Herald
- South Korea’s Approach to AI Governance: Mitigating Risks and Ensuring AI Safety — The National Bureau of Asian Research
- South Korea National AI GPU Infrastructure Initiative — OECD.AI Policy Navigator
- Sovereign AI in Asia 2026: Every Country’s LLM — Digital in Asia
- MSIT Says It Has Laid Groundwork for South Korea to Become a Top 3 AI Power — Digital Today
- AI Security Strategy and South Korea’s Challenges — Center for Strategic and International Studies
