Four levels of control, not a single requirement
Most generative AI tools used daily, such as ChatGPT and Gemini, are developed and managed by large, predominantly U.S. technology companies. According to the source AI Watch Japan, when these systems enter the critical processes of companies and public administrations, questions arise about dependence on foreign suppliers: what would happen if prices or terms of use changed suddenly, and whether it is legitimate to entrust foreign companies with custody of confidential data or of national languages and cultures.
This concern gives rise to the notion of digital sovereignty applied to AI, which the source sets out at four levels. The first is data: establishing where information used for training and inference resides, who can access it and how it can be used. The second concerns models: choosing which ones to adopt and maintaining the capacity to train, adapt and update them. The third is computing infrastructure, namely GPUs and data centers. The fourth is organizational: training technicians capable of managing security, incidents and governance. Sovereignty therefore does not coincide with ownership of a system, but spans the chain of decisions and responsibilities.
Why local servers and national models are not enough
The source warns against a simplistic equation: having data centers on national territory does not automatically guarantee control. The example offered is that of a household safe entrusted to the management of a foreign operator. The legal reference cited is the U.S. CLOUD Act, which, subject to certain legal procedures, can allow data stored abroad to be requested from entities under American jurisdiction. It follows that physical localization and legal control are distinct issues.
Similarly, a language model developed in-house or local execution of inference is not enough to constitute sovereignty: a national model may in fact rely on foreign clouds, and a system run locally may still depend on outside parties for updates and maintenance. The source also notes the reverse case: open-weight models of foreign origin can increase autonomy if an organization is able to manage and modify them internally.
National initiatives and the principle of selective sovereignty
The international picture offered by the source includes Denmark, which has built the Gefion supercomputer for applications such as weather forecasting, and Switzerland, where Apertus was presented in 2025, a model that makes public its weights, training data and development procedures. For Japan, the source cites the AI law that entered into force in 2025, the strategic plan adopted in parallel, the ABCI 3.0 computing infrastructure launched by the national industrial research institute in January 2025, support programs of the Ministry of Economy and Industry, and domestic models such as NTT's tsuzumi, Preferred Networks' PLaMo and NEC's cotomi. For 2026, it also mentions the use of national models in the Gen'nai government environment and the launch of the FRONTia project for multimodal foundation models oriented toward so-called physical AI.
The source's conclusion is that sovereignty does not require total self-sufficiency, not least because many GPUs remain of foreign origin. The central question thus becomes selective: maintain control in sensitive areas such as security and administration, flexibly resort to the most advanced foreign systems for ordinary activities, and establish clearly what to delegate and what instead to decide directly.
