What sovereign AI means

Sovereign AI is control across the whole workflow.

It is not a hosting badge. It is the ability to decide where AI runs, follow where information travels and keep operational authority with the teams accountable for the service.

Three primary control layers

Sovereign AI begins with decisions your organisation can verify.

Infrastructure, information flows and operational authority must be designed together. Controlling only one leaves dependencies elsewhere in the workflow.

01

Infrastructure

Choose an architecture for on-premises, customer-controlled private-cloud or verified air-gapped operation—using Docker-compatible containers, Kubernetes, virtual machines or another process approved by your organisation.

02

Data path

Know where prompts, uploads, retrieved passages, model calls, outputs, logs, exports and backups are stored or transmitted. A local deployment keeps these paths inside your boundary only when every configured provider, connector and support path does the same.

03

Operations

Keep identity, access, retention, audit, administration, updates, recovery and operator access with named teams. These controls can support GDPR and EU AI Act governance work; they do not create automatic legal compliance.

Self-hosting is a starting point

Control must survive every connection.

A locally installed application may still send information to external models, identity providers, logging services, connectors or support systems. Each path needs an owner, policy and evidence.

Model boundary

Define which local, private or hosted model endpoints may receive prompts, retrieved context and generated content.

Identity and permissions

Connect sign-in, user lifecycle, groups and resource permissions to accountable organisational owners.

Knowledge and files

Decide which information may be uploaded, indexed, retained, shared, exported or deleted.

Tools and actions

Restrict which people and agents can call connectors, execute tools or change external systems.

Logs and recovery

Control what enters logs, where backups live, how long records remain and who can restore them.

Support and exit

Document operator access, update paths, support boundaries, exports and the handover needed to leave.

End-to-end boundary

Follow information from the user to the evidence record.

The deployment review should show what each stage can access, where it runs and which team is authorised to change it.

  1. IdentityAuthenticate the user and resolve current roles and group membership.
  2. ApplicationApply the permissions and retention rules of the organisational workflow.
  3. KnowledgeRetrieve only authorised files, records and contextual passages.
  4. ModelRoute prompts and context only to approved inference endpoints.
  5. ToolsConstrain connectors and consequential actions with policy and human approval.
  6. EvidenceRetain the permitted events, decisions and sources inside the agreed boundary.

Deployment patterns

The operating model follows the required boundary.

These patterns are not interchangeable. The selected version, operators, dependencies, update route and support process must be verified for the chosen environment.

01

Customer premises

Applications, storage and approved model inference operate inside the customer’s physical or virtual environment.

02

Customer private cloud

The organisation controls tenancy, region, identity, networking, keys and administrative access.

03

Verified air gap

Use this designation only after installation, models, licences, updates, support, exports and recovery have been tested without a required network path.

Evidence, not labels

Make the boundary reviewable.

A sovereignty claim should lead to deployment-specific architecture, policy, technical tests and named operational responsibilities.

Minimum deployment record

  • Architecture and end-to-end data-flow diagrams
  • Software, model, dataset and licence inventory
  • Operators, administrators and support-access paths
  • External endpoints, telemetry and subprocessors
  • Identity, permission, retention and audit policies
  • Update, backup, recovery and incident procedures
  • Export, migration and knowledge-transfer artefacts

Public-source direction

Software designed to be inspected and operated.

SovereignEU applications are being developed for source-available release and customer-controlled operation under branding-protected Community terms. No application makes an environment sovereign by itself; the configured workflow and operating model determine the real boundary.

Explore the platform

Intended Enterprise Licence

Identity rights and expertise for the boundary you choose.

Executed Enterprise agreements are intended to grant scoped rebranding or white-label rights and add deployment engineering, integration, configuration, hardening, documentation, training and support.

Review planned Enterprise licensing

A practical first step

Define the boundary before selecting the stack.

Map the infrastructure, information flows, operators and exit requirements for one concrete organisational workflow.

Assess your control boundary