What is sovereign AI?
A plain definition, the five layers it covers, how governments use the term, and how to check a system against it.
Sovereign AI is artificial intelligence whose data, models, compute and operations stay under the control, and the laws, of the country or organization that depends on it. It comes in degrees: the European Commission's 2025 Cloud Sovereignty Framework grades services from SEAL-0 to SEAL-4. Picking a local cloud region does not make AI sovereign on its own.
Sovereign AI describes who is in charge of an AI system: who holds its data and keys, who can run and change its models, where its compute sits, who operates it, and whose laws can reach it. The more of those answers point to you, the more sovereign the system is.
What are the layers of sovereign AI?
Sovereign AI has five layers, and a system can be sovereign in some and dependent in others.
| Layer | The question it answers | What control looks like |
|---|---|---|
| Data | Where are prompts, documents, outputs, logs and backups stored and processed, and who holds the encryption keys? | Data stays in places you choose, and only you can decrypt it |
| Models | Can you keep running the model if a vendor changes terms, prices or versions? | You hold the weights, under a licence you accept, and choose when to upgrade |
| Compute | Whose hardware runs training and inference, and can it be withdrawn? | GPUs in your data center, a provider you choose, or capacity reserved for you |
| Operations | Who installs, patches, monitors and supports the system, and from where? | Your staff or vetted operators, with logged and approved access |
| Jurisdiction | Which laws can compel access to the data or force the service to stop? | The provider and operators answer only to the laws you accept |
The European Commission's Cloud Sovereignty Framework (version 1.2.1, October 2025) spells out the data and AI layer in procurement terms. Its "Data & AI sovereignty" objective asks that only the customer, not the provider, has effective control over cryptographic access to its data; that storage and processing stay in European jurisdictions with no fallback to third countries; and that AI models and data pipelines are developed, trained, hosted and governed under EU control.
How do governments and standards bodies use the term?
Governments mostly use "sovereign AI" for national capability, while procurement rules turn it into criteria that can be scored.
- National capability. NVIDIA's explainer, updated in July 2026, describes countries building and running AI with national infrastructure, data, talent and business networks.
- Public investment. The EU's AI Continent Action Plan names "sovereignty, security and democracy" as one reason for the plan, which targets at least 13 operational AI factories by 2026 and AI gigafactories backed by €20 billion from InvestAI. The UK's Sovereign AI programme, funded by the UK Government, runs a £500 million fund to scale British AI startups.
- Procurement criteria. The Commission's framework scores cloud services on eight objectives, from strategic and legal sovereignty to supply chain and technology, and grades each on five Sovereignty Effectiveness Assurance Levels, from SEAL-0 ("No Sovereignty") to SEAL-4 ("Full Digital Sovereignty"). In its April 2026 tender, providers had to reach SEAL-2 to qualify, and most of those awarded reached SEAL-3.
For a company, the same idea shrinks to the organization: the business, or its client, rather than the state, is the party that must stay in control.
Who is sovereign AI for?
It matters most to organizations that would be harmed if someone else could read, stop or change their AI.
| Who | What they need to control |
|---|---|
| Governments and public bodies | Citizen data, continuity of public services, dependence on foreign suppliers |
| Regulated organizations, such as banks, insurers and hospitals | Where personal and confidential data is processed, and who can be compelled to hand it over |
| Industrial and robotics companies | Sensor data, designs and trained models that are their intellectual property |
| AI product companies and service providers | Installs inside each client's environment, which the client owns and governs |
What is sovereign AI not?
It is not a cloud region, an air gap or a promise to build everything yourself.
- Not a region setting. Location is only one layer. The US CLOUD Act, for example, requires covered providers to preserve or disclose data in their control "regardless of whether" it is stored inside or outside the United States (18 U.S.C. § 2713).
- Not necessarily air-gapped. An air-gapped system has no external network path at all. A sovereign system can keep governed connections.
- Not self-sufficiency. The framework's own levels accept "meaningful but not full" influence at SEAL-3. Most organizations will depend on some foreign chips or software, and the goal is to know and limit those dependencies.
- Not the same as data residency. Residency is about where data sits; sovereignty adds who can reach it. The blog post on data residency vs data sovereignty covers the difference.
How can you check whether an AI system is sovereign?
Ask one question per layer and write down the answer and its evidence.
- Where are prompts, outputs, logs, embeddings and backups stored and processed, including during support incidents?
- Who generates and holds the encryption keys, and can the provider decrypt anything?
- Do you hold the model weights, and does the licence let you keep using them if the relationship ends?
- If the vendor's service stopped tomorrow, could you keep serving the model on other hardware?
- Who has administrative access, from which countries, and is each access logged and approved?
- Which laws apply to the provider, its parent company and its subcontractors?
- What does the system call out to: licence servers, telemetry, update services, hosted model APIs?
- How would you move the workload, its data and its configuration somewhere else?
Gaps are normal. The useful result is a list of dependencies you have chosen on purpose.
Where does Hyperlake fit?
Hyperlake describes itself as a sovereign AI operating platform. It assembles, deploys and governs the data, models, applications and tools behind AI workloads in a customer's own environment or their clients' (AWS, Azure, Google Cloud, OVH, private cloud or on premises), so data, models, applications and keys stay in infrastructure the customer controls. The Hyperlake FAQ explains what it covers.
Frequently asked questions
Does sovereign AI mean building your own foundation model?
No. A country may fund national models, but most organizations reach sovereignty by deploying existing models, often open-weight ones, in infrastructure they control. What matters is that you can keep running the model under your own rules, not that you trained it.
Does sovereign AI require open-weight models?
Not strictly, but it is much easier with them. A model you can only reach as someone else's service depends on that provider's operations, location and retirement schedule. Weights you hold under a licence you accept can be run, moved and audited without asking anyone.
Can sovereign AI run in a public cloud?
Partly, depending on the layer. Compute in a public cloud account can still leave you holding the keys and the model weights, but the provider's jurisdiction and its operators' access remain. The EU framework's levels exist because this is a spectrum, not a yes-or-no test.
What is sovereign AI infrastructure?
The physical and platform layers underneath: GPUs and other compute, storage, networking, and the software that deploys and governs models and data on them. It is necessary for sovereign AI but not sufficient; data handling, keys, operations and legal exposure also count.
Sources
Sources for the facts on this page, last checked October 9, 2026.
- What Is Sovereign AI? (NVIDIA blog, updated July 2026) checked October 9, 2026
- Cloud Sovereignty Framework, version 1.2.1 (European Commission, October 2025) checked October 9, 2026
- Commission advances cloud sovereignty through strategic procurement (European Commission, 17 April 2026) checked October 9, 2026
- AI Continent Action Plan (European Commission) checked October 9, 2026
- UK Sovereign AI (funded by the UK Government) checked October 9, 2026
- 18 U.S. Code § 2713, required preservation and disclosure (US CLOUD Act) checked October 9, 2026
Related guides
- How do you deploy an LLM on premises?Deploy an LLM on premises in seven steps: model and licence, GPU sizing, serving engine, gateway, security, monitoring and offline updates.
- Data residency vs data sovereignty for enterprise AI
- Air-gapped AI: the strictest enterprise standard
- A sovereign Physical AI stack on KubernetesA vendor-neutral reference architecture for running a Physical AI stack on Kubernetes that you control: layers, data flow, sovereignty controls and decisions.
Start with a workload. Build the environment around it.
Tell us what you need to deploy, whose environment it must run in, and what it needs to connect to.