Sovereign AI

Sovereign AI is a nation's or organization's ability to build, deploy, and govern artificial intelligence on compute, data, and models that it controls and that operate under its own laws. NVIDIA, which popularized the term in 2024, defines it as a nation's capability to produce AI using its own infrastructure, data, workforce, and business networks. The motivations are dependence on a few foreign providers for frontier models and GPUs, legal exposure when data sits with a company subject to another country's laws, models that underserve local languages and institutions, and the economic value of AI capacity itself.

The Sovereign AI Stack#

Sovereignty is decided layer by layer, and a gap in one layer limits the others. Compute and energy come first: data centers and accelerators such as GPUs that a government or company owns or can contract for under its own jurisdiction. The European Union's InvestAI initiative, launched in February 2025, aims to mobilize €200 billion for AI, including a €20 billion fund for four AI gigafactories, large AI factories with around 100,000 advanced AI chips each.

Data is the second layer: public records, local-language text, and sector data kept within the jurisdiction and curated into training data that reflects local languages, laws, and institutions. Models are the third, either trained domestically or adapted from open-weight foundation models by fine-tuning. Switzerland's Apertus, released in 2025 by EPFL, ETH Zurich, and the Swiss National Supercomputing Centre, is a large language model in 8B and 70B parameter sizes, trained on 15 trillion tokens covering more than 1,000 languages, with its weights, training data, and recipes published. Small language models and compact computer vision models make self-hosting practical without a national supercomputer. The last layer is people and rules: engineers who can operate the stack, and regulation such as the EU AI Act that sets the terms for AI used within a jurisdiction.

Sovereignty, Residency, and Privacy#

Data residency is where data is physically stored. Data sovereignty is which country's laws govern that data, and the two can diverge: under the US CLOUD Act of 2018, US authorities can compel US-based providers to produce data they control even when it is stored abroad. Sovereign AI extends the question beyond storage to the full pipeline, including who operates the hardware, who can update or withdraw the model, and where the model runs. Data privacy and data security protect personal information and systems within any of these arrangements, and confidential AI adds hardware protection for data in use.

Organizations usually decide workload by workload how much sovereignty they need:

Residency is the first step on that path. Ultralytics Platform asks each account to choose a data region at onboarding, US, EU, or Asia-Pacific, and keeps its datasets, trained models, and managed training data in that region, which the Ultralytics Academy lesson Privacy, Regions, and Compliance covers in more detail.

Applications#

• Satellite image analysis for defense: Governments run imagery and document processing on air-gapped or government-operated systems, because classified data cannot leave national control.
• Healthcare: Hospitals and national health systems keep patient scans within the jurisdiction and run diagnostic models on infrastructure they or approved domestic providers operate.
• Language and culture: National language models serve languages and legal systems that global models cover poorly, for example in public administration and education.
• Edge AI in industry: Manufacturers run visual inspection models on factory hardware, which keeps production images on-site and removes the dependency on an external service.

Sovereignty has costs. Domestic compute is expensive to build and staff, open-weight models usually trail the largest proprietary ones, and full independence from foreign hardware is rare, since nearly all advanced AI accelerators are designed by a handful of companies and manufactured by TSMC in Taiwan.