Sovereign AI for defence is not simply a question of where a model is hosted. It is a system-control question spanning compute, data, identity, software dependencies, assurance evidence, continuity, interoperability and accountable decision authority.
Artificial intelligence is moving from isolated experimentation into the operating fabric of institutions. In defence and other high-consequence environments, that shift changes the meaning of technology sovereignty. Control over an AI-enabled capability depends not only on ownership of a model, but on the infrastructure and dependencies that determine whether the system can be operated, secured, updated, audited, recovered and governed on acceptable terms.
01 Executive summary
Data residency is useful, but it is not sufficient to establish operational sovereignty. An organization may keep information within national borders while still depending on foreign-controlled identity services, opaque software supply chains, external model APIs, inaccessible administrative keys, proprietary update channels or infrastructure that cannot be restored independently. Sovereignty therefore has to be evaluated across the complete technical and organizational system.
For defence-oriented environments, the relevant questions are practical: who can administer the platform; where workloads can execute; which suppliers are critical; how keys and credentials are controlled; whether software and model changes are visible; how the system behaves when connectivity is constrained; whether evidence can be reconstructed after an incident; and which decisions remain under accountable human authority.
02 Operational sovereignty beyond data residency
A residency-only model can create a false sense of control. Physical location does not by itself establish administrative control, continuity, portability or the ability to challenge a supplier. A stronger operating model distinguishes several forms of control: jurisdictional control over where data and workloads reside; technical control over identities, keys and administrative functions; operational control over configuration and recovery; commercial control over supplier obligations and transition rights; and governance control over acceptance, change and risk decisions.
These dimensions do not require every component to be domestic. Modern defence and critical systems are inherently connected to allied, commercial and global technology ecosystems. The objective is instead to understand which dependencies are acceptable, which must remain replaceable, which require contractual or technical safeguards, and which capabilities must continue under degraded external support.
03 Six control domains for a sovereign AI stack
Compute and workload control
Sovereign compute begins with knowing where workloads execute and who holds privileged authority over the environment. This can include national data centres, accredited cloud environments, private cloud, edge compute and deployable infrastructure. The design question is not whether every workload must run in one place, but whether placement, failover, encryption keys, privileged administration and continuity are governed intentionally.
Data, provenance and policy
AI systems depend on data pipelines, sensor feeds, reference datasets and operational context. Sovereign design requires visibility into where data originates, how it is transformed, who can access it, how long it is retained and which downstream systems receive it. Provenance becomes especially important when outputs influence consequential decisions because reviewers need a basis for reconstructing what information shaped the result.
Identity and authority
Users, devices, services and autonomous agents require explicit identities. Authority should be bounded, revocable and observable. As software agents become capable of invoking tools and initiating actions, identity moves from a conventional enterprise-control function to part of the AI control plane itself.
Model and software supply chain
Models are only one dependency among many. Libraries, containers, orchestration tools, firmware, APIs, data services and third-party platforms can all affect security and continuity. Sovereign architecture therefore needs inventories, version awareness, change control and a practical understanding of which components can be replaced, isolated or operated without continuous external dependency.
Assurance evidence
Trust cannot rest on a one-time acceptance test. Evidence should accumulate across development, integration and operation: evaluation results, security testing, model and configuration versions, incident records, change approvals, runtime telemetry and exception handling. This creates a reviewable technical record rather than relying on broad statements of trustworthiness.
Continuity and human authority
Sovereign capability includes the ability to continue or degrade safely when dependencies are disrupted. It also requires clarity about which decisions remain human decisions, who can suspend automation, and what intervention mechanisms exist when the system behaves outside expected boundaries.
04 Dependency mapping and concentration risk
One of the most useful sovereignty exercises is to build a dependency map before a major platform decision. The map should identify cloud regions, identity providers, model providers, software repositories, update services, data sources, network dependencies, encryption and signing services, monitoring systems, and specialist suppliers. Each dependency can then be assessed for criticality, substitutability, geographic concentration, administrative access, support assumptions and recovery requirements.
This approach makes concentration visible. A program may appear diversified because it uses multiple applications while all of them depend on the same identity tenant, cloud control plane or external model endpoint. Conversely, an external supplier may present limited sovereignty risk if the dependency is well-bounded, portable, contractually transparent and technically replaceable.
05 Architecture before procurement
Sovereignty requirements are most effective when defined before procurement. If they remain implicit, selection can optimize for near-term functionality while creating long-term dependencies that are expensive to reverse. A pre-procurement architecture should identify trust zones, administrative control, data movement, integration points, recovery expectations, evidence requirements and components that must remain portable.
Procurement questions can then become more concrete: Can logs and configurations be exported in usable form? What evidence is available about model and service changes? Who can access customer data and under what conditions? Can workloads move between environments? What happens when a service is unavailable? Which interfaces are standards-based? How are subcontractors disclosed? What support is required to restore operations?
06 Resilience and degraded operations
Operational sovereignty is tested most clearly when normal assumptions fail. A trusted architecture should define which functions can continue when external connectivity, central identity, cloud services, data feeds or update channels are unavailable. Degraded operation may mean reduced functionality, local-only execution, stricter authorization, cached data, manual fallback or a deliberate safe state.
The key is to make these transitions explicit and testable. Continuity planning should identify recovery priorities, minimum viable services, synchronization rules after reconnection and the evidence required to confirm that restored components return to an approved state.
07 Interoperability without surrendering control
Defence systems must often interoperate with allies, partners, suppliers and legacy platforms. Sovereignty should not be interpreted as isolation. Interoperability can strengthen resilience when interfaces, identities, data exchange and assurance requirements are explicit. The objective is to avoid hidden coupling: integration should not silently transfer administrative control or make one external service an unexamined single point of failure.
Open interfaces, documented data models, portable identities, clear trust boundaries and tested federation mechanisms can allow collaboration while preserving national or organizational control over sensitive decisions and critical infrastructure.
08 Evidence as part of the architecture
Sovereignty claims should be demonstrable. Evidence can include asset and dependency inventories, architecture diagrams, access reviews, cryptographic-key ownership, software bills of materials, model and configuration records, testing results, service-level commitments, incident exercises, recovery tests and documented acceptance decisions. The purpose is not paperwork for its own sake. Evidence provides the basis for challenging assumptions and verifying that the system still matches the approved design.
This aligns with the broader direction of AI risk management. NIST’s AI RMF is designed to help organizations manage AI risk across the lifecycle, and NIST continues to develop profiles for specific contexts, including work announced in 2026 around trustworthy AI in critical infrastructure. For defence environments, the same principle points toward evidence that connects governance claims to the actual system.
09 Canadian and allied policy context
The Department of National Defence and Canadian Armed Forces published their first Artificial Intelligence Strategy in 2024. The strategy identifies fielding AI capabilities, change management, ethics, safety and trust, talent and partnerships as lines of effort, and links AI adoption to data and information-management foundations. NATO’s revised 2024 AI strategy similarly emphasizes responsible use, interoperability, testing and evaluation, and protection against adversarial use.
These public strategies do not prescribe one sovereign architecture. They do reinforce a consistent institutional point: AI adoption is not only a model-selection problem. It depends on data, people, infrastructure, assurance, interoperability and governance working together.
10 A practical sovereignty review
A leadership team evaluating an AI-enabled capability can begin with ten questions: What function is mission-critical? Who holds privileged access? Where can the workload run? Which data must remain under defined jurisdictional control? Which dependencies are externally controlled? Can the system operate in degraded conditions? How are changes detected and approved? What evidence supports trust in the current configuration? Which components can be replaced? Who has authority to accept risk, suspend automation or invoke fallback procedures?
The answers create a more useful baseline than a generic claim that a system is “sovereign.” They identify which controls matter, which dependencies need negotiation and where resilience investment should be concentrated.
11 Research view
Canadian AI Defence defines trusted sovereign AI as an AI-enabled capability whose critical dependencies, authority boundaries, data flows, operating state and change mechanisms are sufficiently controlled and observable for the operator to govern the system under its own legal, security and risk requirements.
This definition is intentionally technology-neutral. Sovereignty can be achieved through different combinations of domestic infrastructure, allied technology, commercial platforms and open systems. The important question is whether the resulting architecture gives the operator meaningful control, reviewable evidence, continuity options and credible authority over consequential decisions.
12 Selected public references
- Department of National Defence and Canadian Armed Forces Artificial Intelligence Strategy
- NIST Artificial Intelligence Risk Management Framework
- Summary of NATO’s Revised Artificial Intelligence Strategy
- Canada’s National Cyber Security Strategy
This publication is independent analysis based on public information. It is not a statement of Government of Canada, DND/CAF, NATO or allied policy and does not imply affiliation, endorsement, security accreditation or procurement status.


