.png)
Last week we published a piece that looked at the current state-of-play in sovereign AI research with the goal of adding an operational or executional layer to the analysis. You can find that post here. In case you missed the research we drew from, it is below:
McKinsey - sovereign AI is now treated as a strategic or existential priority among decision-makers
Tony Blair Institute - AI as governing infrastructure shaping state capability
Brookings Institution - sovereign AI as a dimension of geopolitical competition
World Economic Forum - AI as a driver of national productivity and influence
Oxford Internet Institute - structural concentration of compute and limited global distribution
The feedback on the post was outstanding and many of the conversations encouraged us to go a little further and create a playbook for sovereigns to incorporate into their own discussions/strategies.
This following draws from the literature while providing a concrete set of recommendations. Needless to say, every sovereign will be dealing with their own considerations but this four part framework offers a practical pathway to genuine operational independence.
Stage 1: Foundations (Months 0–12)
Define the sovereignty baseline. Classify workloads by the two relevant dimensions: sensitivity (what are the consequences of foreign access?) and criticality (what are the consequences of unavailability?). This classification drives the tenancy requirements and the required sovereignty tier for each workload category. Crucially, define the 'Sovereignty Break-Point': the specific conditions under which the nation must be able to operate the system without any external vendor connectivity.
Simultaneously, address the energy constraint. Identify sites with access to dedicated generation capacity or grid connections that can be hardened and expanded. Behind-the-meter solutions require longer lead times than hardware procurement - starting here is starting at the right place. Energy is the physical anchor of sovereignty; without domestic power, the control plane is irrelevant.
Procure initial sovereign AI infrastructure - not as a commercial cloud subscription, but as capital infrastructure, financed at infrastructure cost of capital, with domestic asset ownership, and a contractual capability transfer plan. The initial deployment should be large enough to be operationally meaningful but structured for expansion.
Architect for Administrative Isolation. The control plane must not only be locally deployed but must be capable of Full Air-Gap Operation. If the system requires a 'phone-home' to the vendor for a license check or a patch to function, it is not operational independence. These are foundational concepts. They are required to build big.
Stage 2: Operational Transition (Months 6–24)
Begin the structured capability transfer process. National operators move from observers to co-pilots. Every configuration change or security patch must be executed by the sovereign team using the vendor’s tooling, ensuring the 'knowledge transfer' is an 'action transfer' from day one. Monthly competency assessments with defined pass/fail criteria, not training completion certificates.
In parallel, begin building the data ecosystem. The Oxford research notes that meaningful AI sovereignty requires not just compute but data - the ability to train models on nationally relevant datasets. This is the governance work: public sector data release with clear licensing, trusted data-sharing frameworks, curated domain datasets for priority sectors (healthcare, national security, finance, critical infrastructure).
Adopt a layered model strategy. The WEF/Bain framework correctly identifies that few nations need to build frontier foundation models from scratch. The productive path is leveraging open-weight or internationally derived models for general workloads while developing domestic fine-tuning capability for domain-specific and culturally specific applications. The infrastructure platform must explicitly support both workflows - fine-tuning on domestic data and, for nations with the resources and strategic requirement, pre-training from scratch.Lastly, secure the rights and physical access to the control plane’s source code and local instances. In a 'relationship break' scenario, the sovereign must have the legal and technical keys to keep the lights on.
Stage 3: Full Operational Independence (Months 18–36)
Complete capability transfer. National operators run primary on-call. Vendor provides escalation support only. Define and test the disconnected operation mode - schedule a planned connectivity blackout and verify full operational capacity. This is the 'Sovereignty Audit.' If the AI stops dreaming because it can't talk to a server in another jurisdiction, the project has failed its primary objective.This is not a crisis simulation; it is a certification requirement.
Expand the sovereign infrastructure footprint based on demonstrated utilization and demand signals from government and regulated industry. The McKinsey research notes that effective sovereign ecosystems treat the government as an anchor customer - bundling public sector demand into multi-year frameworks that justify and sustain the investment into operational independence.
Stage 4: Ecosystem Development (Years 3+)
At this stage, domestic operational capability is established. The nation can operate its AI infrastructure independently. The focus shifts to what the WEF/Bain and Tony Blair frameworks correctly emphasize: application development, talent cultivation, and ecosystem building.
This is where most of the economic value is captured. Brookings notes that "the upper layers of the stack, particularly applications and domain-specific systems, offer far greater scope for differentiation, competition, and local value creation." The infrastructure sovereignty achieved in stages 1–3 is not the end goal - it is the foundation that makes application-layer sovereignty possible.
A nation that has achieved genuine operational independence over its AI infrastructure can make credible, durable commitments to domestic and international AI ecosystem partners. It can host sensitive workloads that foreign-dependent infrastructure cannot. It can fine-tune models on genuinely sovereign data. It can attract applications and model development that require sovereign compute. The infrastructure independence is the platform from which every other element of AI competitiveness is built.
The Vendor Evaluation Framework
The research literature frames the problem well but stops short of telling a procurement officer what to actually ask. The following questions should be posed to any vendor making sovereign AI claims. The answers should be specific, demonstrated, and contractually binding - not described in marketing materials.
On Operational Independence
- Can your system operate production workloads at full capacity with all vendor-initiated outbound connectivity severed? This means no mandatory callbacks, no license telemetry, no undocumented dependencies on your infrastructure. Demonstrate it in a controlled test with network-level logging to verify.
- Show us your control plane architecture diagram. Label every component. For each component, identify: what jurisdiction it runs in, who operates it, and what happens operationally if connectivity to that component is severed.
- Provide a bill of materials for every external service your platform calls. For each service: is it optional or mandatory for core operation? What is the degradation path if it becomes unavailable?
On Capability Transfer
- What is your capability transfer timeline? Provide a phased plan with specific milestone definitions - not training completion, but demonstrated competency: hardware replacement without assistance, network troubleshooting, system upgrade execution, security incident response, tenant provisioning with strict isolation requirements.
- Are these milestones contractual or aspirational? Show us the contract language.
On Supply Chain Resilience
- What is your platform's dependency on vendor-specific accelerator software (e.g., CUDA)? If we needed to migrate to an alternative accelerator architecture, what does that transition path look like, and what is the contractual mechanism for vendor support during that migration?
- Can we replace the networking vendor without rearchitecting the platform?
- Can we replace the storage vendor without rearchitecting the platform?
- What is your software bill of materials? For each major component, identify the license type, the upstream maintainer, and the exposure to export controls or sanctions.
On Economics
- What is your cost of capital for this deployment? If you are financing it on your balance sheet as a tech product, we are paying for your venture returns embedded in our infrastructure pricing. Seek infrastructure-grade economics.
- Provide a total cost of ownership model for five years including power, cooling, and operations - not just hardware and software.
If a vendor cannot provide clear, demonstrated, contractually binding answers to these questions, what you are purchasing is data residency, not sovereignty - regardless of how it is presented.
The Questions Governments Are Not Asking
The policy literature has done excellent work framing the strategic landscape. What it has not answered is the operational question: when the strategy is set, the investment is made, and the data center is built - what happens next? Who controls the scheduling system? Whose key management service holds the encryption keys? What happens when the vendor's telemetry pipeline goes down? Can the national team actually run the system on a Tuesday morning when the vendor's SRE is in a different time zone?
These questions are not glamorous. They do not appear in strategy documents or ministerial speeches. But they are the questions that determine whether a sovereign AI program delivers strategic independence or strategic dependency wearing a sovereign label.
The answer to "who controls the control plane" is the answer to "who has sovereignty." Everything else is preparation.
