[2026-07-23] Korea Unveils ‘Agentic AI Initiative’ — 3,352 GPUs Committed as Government Targets Global AI Top-3 Status

On July 23, 2026, Korea’s Ministry of Science and ICT (MSIT) unveiled the “Agentic AI Initiative” at the 11th Science and Technology Ministerial Meeting, alongside a decision to allocate 520 additional GPUs to national AI projects. This builds on 2,832 GPUs already committed in March 2026 to 52 projects across 25 agencies, bringing the cumulative total to 3,352 GPUs. The initiative rests on three strategic pillars: leading agentic AI safety and trust, building an open execution ecosystem, and driving demand-side adoption.

What the initiative covers

• An “Agentic AI Safety and Trust Guideline” to be finalized within 2026, covering personal-data protection, permission management, and prevention of goal deviation
• Development of an AI agent marketplace and core “AI OS” technology to connect models, inference, agents, and tools
• 512 of the newly allocated Nvidia B200 GPUs go to the “AI for Everyone” project, aimed at giving all citizens low-friction access to AI agents
• An “AI Agent Flagship Project” spanning foundation-model firms, infrastructure providers, service developers, and demand-side companies
• Expanded real-world testing through the ICT regulatory sandbox, with evidence-based policy adjustments to follow

Why this policy matters for the United States

MSIT’s own announcement explicitly cites Anthropic, OpenAI, and Google as the reference points for how agent safety features — permission scoping, action logging — are being designed into agentic AI systems from the outset. Korea’s forthcoming safety guideline is taking shape around practices already standard at these US labs, which signals continued reliance on US-origin agent frameworks and models even as Seoul funds domestic alternatives. At the same time, Korea’s rising GPU demand (3,352 units committed across two ministerial rounds in 2026) adds to regional demand that touches US chip and cloud supply chains. For US AI safety-tooling vendors and foundation-model companies, Korea’s guideline process is also a live case study in how a US-aligned but distinct legal framework treats agentic AI risk. That framework is Korea’s AI Basic Act, in effect since January 22, 2026.

International comparison

Singapore’s Infocomm Media Development Authority (IMDA) launched the world’s first Model AI Governance Framework for Agentic AI at the World Economic Forum in January 2026, updating it in May 2026; it remains voluntary with no enforcement mechanism. Japan’s AI Promotion Act, in effect since June 2025, takes a non-binding, innovation-friendly soft-law approach with no penalty framework. Korea, by contrast, already has a binding, risk-based AI Basic Act in force since January 22, 2026, and is now layering a sector-specific agentic AI guideline on top of it — a combination that industry watchers say could offer more legal predictability than Singapore’s voluntary model, if the detailed rules materialize on schedule.

Industry and consumer impact

For industry, the initiative responds to a real adoption wave already under way in Korean enterprises, where AI agents are increasingly handling repetitive tasks and reshaping hiring patterns. The government’s focus on “token-efficient execution” technology reflects industry concern about the high compute cost of running agentic workloads at scale. For consumers, the clearest near-term change is the “AI for Everyone” service, expected to launch within 2026 through two to three private-sector partners, aiming to make AI as accessible as a calculator. Because agentic AI can take real-world actions on a user’s behalf — payments, bookings, data handling — the initiative also flags unresolved questions around liability when an agent malfunctions or is misused, which MSIT says will be addressed through the Agentic AI Alliance’s work on agent identity verification and accountability.

Source: https://www.korea.kr/news/policyNewsView.do?newsId=148968687