SAIL Interests
What do people actually want from sovereign AI?
When a minister, a CIO, or an industrial-policy lead says "sovereign AI," they rarely mean the same thing. This reference guide maps the most common interests behind the slogan. For the narrative introduction, read the featured essay.
We are less interested in settling what sovereign AI is than in reading why decision-makers care. The same demand for “sovereignty” can be a bid for vendor independence, a play for chips and jobs, a data-residency requirement, or a cultural statement. Treating these as one thing is how policy debates go in circles.
Security and weaponization
Security and weaponization
"We cannot run critical national systems on AI that a foreign government or vendor could read, alter, or switch off."
What they want
Assurance that no single foreign actor can disable, surveil, or tamper with the systems they depend on. The classic answer is source-code escrow and infrastructure control — Microsoft's offer to keep a copy of source in a hardened Swiss facility is this interest made literal.
Typical instruments
Source escrow, on-soil hosting, kill-switch audits, classified-environment deployments, supply-chain review of hardware and weights.
Where SAIL speaks to it
Compute control and the absence of unilateral kill-switches, plus legal authority to override or retire systems.
What it cannot solve
Escrow proves you could read the source; it does not give you the talent, data, or compute to rebuild and run the system independently. Security control is necessary but not the same as capability.
Industrial policy
Industrial policy
"AI is the next growth engine. We need the chips, the fabs, the data centers, and the talent here, building national champions."
What they want
Domestic economic capacity: compute hardware, energy, skilled workers, and a flagship lab or two. "Sovereign AI" is often the headline; the underlying pet interest may really be minerals, energy contracts, or jobs in a region.
Typical instruments
Subsidies and tax breaks, sovereign compute clusters, immigration lanes for researchers, procurement preferences for domestic vendors, energy and minerals deals.
Where SAIL speaks to it
Compute and infrastructure capacity is the layer that maps most directly; model and training capacity matter when the goal is a genuine domestic lab rather than a press release.
Coordination risk
High. If every country races to onshore the same chips and poach the same talent, middle powers compete instead of pooling. The backlash to Cohere and Aleph Alpha shows how quickly "national champion" becomes "now we don't have to depend on our neighbors."
Enterprise and procurement
Enterprise and procurement
"Our data cannot leave the jurisdiction, and we need a contract we can exit without rebuilding everything."
What they want
Control over the data supply chain and the commercial relationship. This is the most concrete and tractable interest: it is mostly about residency, provenance, auditability, and contractual exit rights, not national identity.
Typical instruments
Data-residency clauses, provenance and deletion rights, model-agnostic architectures, vendor-substitution clauses, observability and audit logging.
Where SAIL speaks to it
This is the part of sovereign AI that SAIL was built to assess. Application portability, orchestration independence, and data control are directly scorable.
Why it travels well
Because it is about capability and contracts rather than symbolism, this interest is the one where productization and certification genuinely help.
Cultural identity and values
Cultural identity and values
"A French model should be and feel French. It should reflect our language, our law, and our values."
What they want
A system that represents a community's language, culture, and norms — not just one hosted on home soil. This is a question of legitimacy and representation, not infrastructure.
Typical instruments
Domestic data and evaluation sets, language and cultural fine-tuning, public oversight of training objectives, participatory governance.
Where SAIL speaks to it
Partially. Owning data, evaluations, and training objectives is relevant, but these are inputs to legitimacy, not legitimacy itself.
What it cannot solve
A model feeling French cannot be certified into existence. No checklist confers cultural legitimacy; that comes from who governs the system and whether communities recognize themselves in it. Treating values as a productization problem is the most common category error in the sovereign AI debate.
Middle-power alliance
Middle-power alliance
"We cannot match the hyperscalers alone, so we pool compute, data, and models with trusted partners."
What they want
Strategic autonomy without autarky. Rather than build everything domestically, they share capacity through federated compute, joint models, and allied governance — reducing dependence on any single hyperscaler or superpower.
Typical instruments
Federated compute agreements, jointly governed open models, shared evaluation infrastructure, multilateral funding (the "Airbus for AI" idea).
Where SAIL speaks to it
SAIL explicitly credits federated and allied arrangements as legitimate sovereignty, provided dependencies are acknowledged. This is the one interest the spec actively rewards over going it alone.
Coordination risk
This is the strategy that nationalist industrial policy quietly undermines. Every country chasing its own champion erodes the trust a joint middle-power approach needs. The interest is fragile precisely because it depends on others resisting the same temptation.
Decoder: is this about refusing the OpenAI for Countries deal?
A government declining a hyperscaler's national-AI offer is read as a single act of sovereignty. It rarely is. The same refusal can mean very different things:
- Vendor diversification — they want optionality and exit rights, and would happily sign with two vendors instead of one. This is the enterprise interest in disguise (Layers 1–3).
- Domestic champion building — they are protecting a national lab or a planned compute build-out. This is industrial policy, and the deal is a competitor (Layers 4–6).
- Security posture — they cannot accept a foreign kill-switch on critical systems regardless of price (Layers 6–7).
- Electoral or values signaling — the refusal is the point; the policy substance is secondary. No layer of the spec addresses this, and pretending otherwise wastes everyone's time.
Before designing a response, decode which of these is actually driving the decision. The strategy that follows is completely different in each case.
When sovereignty and LLMs pull apart
It is worth saying plainly: some sovereignty claims sit awkwardly with how large language models actually work. LLMs have steep fixed costs, benefit enormously from scale and shared data, and are cheap to copy once trained. That profile suits a globally provisioned public good with some decentralization far better than it suits dozens of self-sufficient national stacks.
For several of the interests above — especially middle-power alliance and the capability dimension of industrial policy — the most sovereign outcome may be a shared, openly governed system rather than a national one. That is the public AI frame: treat the model as public infrastructure, govern it accountably, and let many parties draw on it. Sovereignty becomes a question of governance and access, not of ownership and borders.