Recently, competition surrounding artificial intelligence has shifted beyond performance to issues of sovereignty. A symbolic event illustrating this is the U.S. government's export controls on Anthropic’s top-tier AI models, Top5 and Fable5. In June 2026, following a directive from the U.S. government, Anthropic immediately suspended these services. Although the control period was relatively short, overseas users were unable to access the services due solely to the unilateral decision of the U.S. government.
This incident raises fundamental questions about sovereign AI. Even if services are freely accessible under normal circumstances, if a foreign government can unilaterally activate a kill switch during a crisis and block access, can this truly be considered a nation’s own AI capability? As artificial intelligence becomes embedded as a core foundation for defense, finance, administration, and industry, the interruption of access rights can lead not merely to inconvenience but to disruptions or even paralysis of national functions.
Similarly, it is no coincidence that the European Union strengthened its digital sovereignty policy around the same time. Conflicts between the U.S. and Europe have expanded beyond tariffs and trade to questions of who will hold the right to develop and control technology. The competition between the U.S. and China further clarifies this trend. Both governments are moving beyond merely regulating or subsidizing AI companies, toward direct participation in AI enterprises and infrastructure. While protecting strategic industries is not new, it represents a different level of change for governments to become shareholders in advanced AI companies and intervene in ownership and governance structures. This reflects a shift in viewing AI companies not as ordinary software firms but as national core infrastructure.
So what are the minimum requirements for sovereign AI? It is often said that AI sovereignty involves controlling domestic data and infrastructure and possessing models tailored to the nation’s language and culture. However, it remains unclear whether every country must develop its own super-large-scale models or whether controlling only data and infrastructure while utilizing foreign models would suffice. It is also not clear how much of semiconductors, cloud services, data centers, data, and models must be secured domestically.
It is understandable that domestic discussions focus on developing independent foundation models. Even if a model is developed domestically, excessive reliance on advanced computing chips, data center infrastructure, and cloud platforms abroad would make stable operations difficult during crises. Conversely, even with foreign models, if they are operated domestically, subject to domestic laws regarding critical data, and capable of switching to other models when needed, a certain level of sovereignty can be considered achieved. Independent models are an important component of sovereign AI but not the entirety of it.
From this perspective, the minimum requirements for sovereign AI can be summarized in three points. First, the nation must be able to maintain its core AI functions for a certain period even under foreign government export controls or corporate service suspensions. Second, overall data management and ethical and legal responsibilities regarding AI models must fall under domestic laws and policies. Third, systems must be operable without foreign technical support, and capable of switching to other models or infrastructure when necessary.
It is neither realistic nor necessarily desirable for Korea to independently build its entire AI ecosystem. Forcing the use of domestically developed AI with inferior performance in the local market could constrain corporate productivity improvements and weaken global competitiveness. Therefore, the starting point of policy should not be localization but rather distinguishing between core and non-core areas. In core areas such as defense, finance, and administration, data and system operation rights must be controlled domestically, and minimum independent computing capabilities and alternative means must be secured to prepare for crises. Meanwhile, in commercial and general-purpose application areas, global technologies should be actively utilized, but without dependency on specific suppliers, ensuring the ability to switch to other models or infrastructure when needed. Securing sovereignty and control in core areas while guaranteeing openness and choice in non-core areas represents the most realistic approach Korea can take.
in OPINION
