Sovereign AI gives countries and institutions independent capacity to develop, host and manage AI systems, relying on local infrastructure, data, models and human expertise. The approach has gained prominence as technology use has expanded and concerns have grown over national security, privacy, intellectual property and dependence on foreign providers.
Sovereignty goes beyond keeping data within borders
In this field, sovereignty is not limited to keeping data within national borders. It also covers computing resources, model operation, update management, access permissions and regulatory compliance. It also aims to ensure the continuity of vital services during geopolitical disruptions, the imposition of trade restrictions or the withdrawal of support by a foreign provider.
Sovereign systems differ from global services in the scope of legal authority, the movement of data and the degree of autonomy. International platforms may transfer information across borders or place it under foreign legal frameworks, while sovereign systems operate within a defined jurisdiction and restrict the transfer and reuse of sensitive data.
This approach also enables the design of models that account for local languages, dialects, laws and social norms. Models can be trained or fine-tuned using national datasets, making them better suited to public-sector and educational applications, among other fields that require a precise understanding of the cultural and legal context.
National infrastructure for model operation and data protection
Sovereign AI relies on computing, storage and network resources within the national or regional sphere, whether in local data centres, high-performance computing clusters or cloud services subject to the relevant jurisdiction. This includes protecting data during input, model training, inference operations and backups.
This does not necessarily mean creating a separate data centre for each institution. Sovereign cloud services provided by local operators or entities subject to national law can be used, provided that data and computing resources remain within the specified legal jurisdiction. Confidential computing and trusted execution environments may also be used to protect information while it is processed and reduce unauthorised access.
Model governance and options for applying sovereignty
Implementation options include full sovereignty, which keeps data, models, infrastructure and personnel within the country, and hybrid sovereignty, which assigns national resources to sensitive tasks while using global platforms for other tasks. Sovereignty can also be applied within a regional bloc or limited to vital sectors such as defence, health and financial services.
- 1
Germany’s Sofi project
Germany is working on the Sofi project to develop a sovereign, open-source foundation model with 100 billion parameters, with technical support from Deutsche Telekom and T-Systems. The project uses about 130 systems from Nvidia and more than 1,000 graphics processing units, targeting complex applications including robot operation.
- 2
Switzerland’s Apertus model
In September 2025, the Swiss AI Initiative launched the multilingual Apertus model, making its training architecture, data, source code and model weights available. It was trained on 15 trillion tokens in more than 1,000 languages, including Swiss German and Romansh. Poland also launched its local language model, PLLuM, in February 2025, designed to handle the characteristics of the Polish language.
- 3
Europe’s Alia project
In January of the same year, the Barcelona Supercomputing Center launched the Alia project, an open, multilingual European infrastructure providing data, models and tools in Spanish, Basque, Catalan and Galician.
