- Sovereign AI is AI that an organization owns and controls: its agents, models, data and results, run in any cloud, on premises or air-gapped.
- The phrase entered policy in a Tony Blair Institute report on 22 February 2023, and NVIDIA CEO Jensen Huang made it widely known from November 2023.
- Gartner predicts that by 2028, 65% of governments worldwide will introduce some technological sovereignty requirements.
- Data residency, sovereign cloud and air-gapped AI each cover one part of sovereignty; sovereign AI covers who owns and controls the AI itself.
- Sovereignty is a matter of degree: most organizations aim to own their models, data and agents rather than build their own chips or data centers.
What is sovereign AI?
Sovereign AI means the organization that uses AI also owns it. Its agents, its models, the data they learn from and the results they produce belong to that organization. It also decides where they run: a public cloud, its own data center, or a network with no internet connection. Nothing depends on a provider keeping an API available, priced the same, or unchanged.
The term started with nations and now covers organizations too. Gartner defines AI sovereignty as "the ability of a nation or organization to independently control how AI is developed, deployed, and used." The idea is the same at both sizes: whoever owns the AI controls it. For what this looks like in a platform, see how H2O.ai delivers sovereign AI.
Where the term comes from
Sovereign AI grew out of digital sovereignty, the idea that a state should control the digital technology it depends on. CNAS calls it "an offshoot of digital sovereignty."
The phrase entered policy on 22 February 2023, when a Tony Blair Institute report warned of "the erosion of the UK's sovereign AI capabilities." Heesoo Jang's study of 485 news articles dates the first press use to the same day and traces the spread to India, South Korea, Germany, Australia and France.
NVIDIA brought the term into wide use. A 2025 review in Natural Language Processing credits CEO Jensen Huang, who "elevated the visibility of the concept, starting with an Nvidia earnings call in November 2023."
There is still no agreed definition. CNAS says a consensus "remains elusive," and no standards body has published one. One older meaning is unrelated: in AI-safety writing, a "sovereign" is an AI that acts without human steering, from Nick Bostrom's typology of advanced agents.
Why sovereign AI matters
Three pressures pushed the term from policy papers into IT budgets.
- Dependence on a provider. A rented model can be retired, repriced or cut off. McKinsey names the "kill switch" risk: service denied because of a contract dispute or a geopolitical conflict.
- Law that follows the provider, not the data. Lawfare notes that US providers' promises to keep data in Europe "cannot stand in the way of their legal obligations" under the US CLOUD Act.
- Regulation. Gartner predicts that by 2028, 65% of governments worldwide will introduce some technological sovereignty requirements.
Organizations feel it as well as states. In a Linux Foundation survey, 79% of respondents called sovereign AI a strategic priority, and 47% said it matters at the organizational level.
How sovereign AI works
Sovereign AI comes down to four things an organization owns, plus the freedom to run them anywhere.
| What you own | What it means in practice |
|---|---|
| Agents | The workflows, tools and permissions your AI agents use are yours to inspect, change and move. |
| Models | You run open or fine-tuned models with weights you hold. A provider can't retire or reprice them under you. |
| Data | Training data, prompts and documents stay inside your boundary. Nothing goes to a third party to be processed. |
| Results | Outputs, logs and audit trails stay with you, so you can show a regulator how a decision was made. |
Where it runs is a separate choice. The same stack can run in a cloud account you control, in your own data center, or fully air-gapped. Air-gapped is one end of that range, not the definition.
Sovereignty is a matter of degree, not all or nothing. Brookings and CEPS describe "a spectrum of strategies" rather than self-sufficiency, and CNAS frames it as "managing rather than eliminating dependencies." Few organizations build their own chips or data centers. Most aim to own the layers that decide what their AI does: the models, the data and the agents.
Ownership also changes how you pay. A rented large language model is billed per token. An owned model, often a small language model fine-tuned for one job, costs the compute it runs on.
Sovereign AI vs data residency, sovereign cloud and air-gapped AI
These terms get treated as one idea. They overlap, but each answers a different question.
| Term | The question it answers | What it leaves out |
|---|---|---|
| Data residency | Where is the data stored? | A model hosted somewhere else can still process the data. |
| Data sovereignty | Whose law governs the data? | The models and agents that act on it. |
| Sovereign cloud | Who operates the infrastructure, and where? | Whether you own or can change the models that run on it. |
| Air-gapped AI | Is the system cut off from the internet? | An air-gapped system can still run a model you don't own. |
| Open-weight models | Can you download and run the model? | Where it runs and who governs the data and agents around it. |
| Sovereign AI | Who owns and controls the AI? | Nothing: agents, models, data and results, wherever they run. |
Each of the first five can be part of sovereign AI. None is enough on its own. As McKinsey's Ali Ustun puts it, "you can actually have data sovereignty, but you may not have sovereign AI."
Sovereign AI examples
Nations
- India approved its IndiaAI Mission in March 2024, with public compute of 10,000 or more GPUs and investment in its own foundation models.
- Canada launched a Sovereign AI Compute Strategy in December 2024, with up to C$2 billion.
- The European Union announced InvestAI in February 2025, mobilizing €200 billion for AI, including €20 billion for AI gigafactories.
- Switzerland released Apertus in September 2025, an open model from EPFL, ETH Zurich and CSCS. Swisscom offers it to business customers on its sovereign Swiss AI platform.
Organizations
- A federal research agency runs an AI assistant with no public-internet connection. At NIH, it serves more than 8,000 employees across 28 institutes and deflects up to 10,000 IT service requests a year.
- A telecom operator replaces rented models with its own. AT&T cut call-analytics costs by 90% with fine-tuned small language models.
How to evaluate a sovereign AI platform
Ask these before you sign. Each one maps to a way control can leak.
- Can you switch models? Gartner advises "model agnostic workflows" that can move between LLMs, regions and vendors.
- Who holds the weights? If a provider can retire or reprice the model, you rent it.
- Where does it run? Check that the same platform runs in your cloud, on premises and air-gapped, not three different products.
- Who operates it, and under whose law? A local data center run by a foreign provider can still fall under foreign law.
- Is there an off switch you don't control? Look for license checks or cloud dependencies that stop the system if the vendor relationship ends.
- Can you audit every result? Logs, lineage and model decisions should stay with you.
- How is it priced? Per-token billing grows with use. Capacity pricing tied to your own compute does not.
Sovereign AI in H2O AI
H2O AI Cloud, the trusted Sovereign AI Platform, is built on this idea. H2O AI Super Agent orchestrates predictive, generative, and agentic AI in one system, with observability built in. Open source at the foundation, any model, no lock-in.
H2O AI Cloud is available as H2O AI Managed Cloud (H2O.ai operates it), H2O AI Hybrid Cloud (a private cloud you operate, in your own cloud, on-premises or air-gapped), or H2O AI Cloud for Government (FedRAMP High certified).