Korean Public Sector AI Policy Direction: Stargate, DeepSeek, and the EU AI Act

In January 2025, President Trump announced the "Stargate" project, a joint initiative with OpenAI, SoftBank, and Oracle. The plan calls for up to USD 500 billion in investment over four years to build AI data centers and supercomputing infrastructure across the United States. The first phase — construction of ten NVIDIA GB200-based data centers in Abilene, Texas, targeting completion by 2026 — is already underway.

Stargate goes beyond a private investment — it is directly tied to the US government's AI hegemony strategy. If US AI companies such as OpenAI, Anthropic, and Microsoft come to dominate the foundational infrastructure of global AI services, public agencies and corporations worldwide will become increasingly dependent on US cloud and AI services. Korean public institutions are no exception. MOEF explicitly identifies finding the balance between reducing this dependency and accessing state-of-the-art AI capabilities as one of the central challenges of public sector AI policy.

In January 2025, China's DeepSeek shocked the global AI industry with its R1 model. The claim of achieving reasoning performance matching OpenAI's o1 model at a training cost of just USD 6 million — combined with openly releasing the model weights for free use by anyone — fundamentally disrupted established assumptions. The V3 model delivered benchmark results surpassing GPT-4o in coding, mathematics, and natural language understanding, marking an inflection point where "democratization of AI costs" became real.

DeepSeek carries dual implications for Korean public sector AI policy. On the positive side, installing the open-source DeepSeek model on-premise within a public agency eliminates dependence on US cloud APIs while enabling high-performance AI at comparatively low cost. On the security and data sovereignty side, however, there are concerns about feeding public data into a Chinese AI model or using services connected to Chinese servers. MOEF's policy direction leaves open the possibility of using open-source models like DeepSeek, but presents rigorous security validation as a prerequisite.

The EU AI Act (Artificial Intelligence Act) entered into force in August 2024, with prohibited AI use rules beginning to apply from February 2025. Full obligations for high-risk AI systems take effect in August 2026, and with certain exception provisions expiring in August 2027, an effectively comprehensive enforcement regime will be in place. While the EU AI Act applies only to AI systems operating within the EU, Korean export companies and public agencies face direct compliance obligations when they provide AI-based services to EU counterparts or target EU markets.

The EU AI Act compliance need for Korean public agencies arises through two pathways. First, when KOTRA, the Korea Trade Insurance Corporation, and similar agencies provide AI-based services — buyer matching, risk analysis, and so on — to customers or partners in EU member states, the extraterritorial application principle of the EU AI Act makes those AI systems subject to regulation. Second, when Korean companies operating in EU markets use KOTRA's AI support tools to transact with EU buyers, the AI compliance of the KOTRA platform may indirectly come under verification requirements. MOEF's policy direction actively recommends using EU AI Act standards as the reference framework for Korean public sector AI governance.

Following the inauguration of the new government, MOEF announced the "AI Utilization Policy Direction for Public Institutions," presenting three core pillars: active utilization (accelerating AI adoption), safe management (minimizing risk), and productivity innovation (improving operational efficiency). This direction was formulated with direct reference to the three global variables — Stargate, DeepSeek, and the EU AI Act — targeting a balance that keeps pace with global AI competition while maintaining public trustworthiness.

The most notable shift in the policy direction is the "AI-First" principle. If the previous default for introducing new work systems was traditional IT-based approaches with AI as an option, the new procedure requires that every new system or service design begin with the question: "Can AI handle this?" KOTRA's Navi AI buyer matching and TriBIG market analysis platform are emerging as flagship examples of this AI-First principle and are becoming a model for AI adoption by other public agencies.

One of the core elements of MOEF's policy direction is the mandatory AI safety management framework. From the first half of 2026, public agencies using high-risk AI must conduct a pre-deployment AI Impact Assessment and publish the results. The standard for "high-risk" is broadly aligned with the EU AI Act, with additional Korea-specific criteria covering administrative service AI and AI processing large volumes of personal data.

The three global forces of Stargate, DeepSeek, and the EU AI Act are exerting pressure on Korean public sector AI policy from different directions simultaneously. Stargate deepens Korean dependency as US AI infrastructure dominance consolidates; DeepSeek raises the possibility of public agencies running their own AI at low cost; and the EU AI Act introduces new compliance burdens. MOEF's AI-First principle and safety management framework represent a structural response designed to manage these three forces in balance. Trade and investment public agencies including KOTRA must accelerate AI adoption on this policy foundation while upholding three principles: data sovereignty, algorithmic transparency, and human oversight. Overseas outposts such as the KOTRA Dhaka Trade Center serve as the field-level embodiments of this AI infrastructure — taking on the role of proactively adapting to a shifting global regulatory environment.