The Sovereignty Vocabulary Problem
AI sovereignty
is broad enough to justify almost any technology purchase. Analysts debate its meaning while ministers announce multi-year funding packages. Finance ministries are caught in the middle, unsure what they’re actually buying.
Sovereignty can mean control over infrastructure. It can mean jurisdiction over data. It can mean freedom to change suppliers. These objectives often conflict, and no single budget line addresses all three. Without defining which matters most, governments end up funding everything and securing nothing.
The Uncomfortable Truth: Full Self-Sufficiency Isn’t Possible
The Tony Blair Institute made this clear in January. Achieving full AI self-sufficiency across the entire stack is unrealistic. The stack includes infrastructure, platforms, tools, and services. The Tech Policy Design Institute’s June report, Expanding AI Sovereignty to AI Agency,
identifies 103 separate capabilities across six layers, from compute to governance.
No treasury can fund them all. Nations must choose. Which capabilities warrant homegrown investment? Which should be pooled with regional partners? Which can be purchased from the market with clear exit terms?
Nvidia CEO Jensen Huang has urged governments to own the production of its own intelligence.
Huang sells the chips needed for such ownership. Ministers should not rely solely on vendor advice.
The Real Question: What Leverage Does This Money Buy?
Henry Farrell and Abraham Newman offer a useful framework. Bargaining power comes from occupying positions in a network that others cannot leave at acceptable cost. That is leverage. A capability justifies its budget only if it changes what a government can demand from suppliers, partner states, or international bodies.
Such positions weaken with overuse. The squeezed party will build alternatives. The strongest leverage is rarely exercised. This means most AI investments should sit idle, deployed only when they matter.
Before approving an AI budget line, a finance ministry should ask three questions:
- Which national objective does this capability serve? Growth, security, and influence over global rules require different investments. Define the chain of steps from spending to outcome.
- How much bargaining power does it generate? A data center with no laboratory users produces only electricity bills. Can the country actually deploy this capability?
- What would the same money produce in its next-best use? Consider options inside or outside the AI sector. The comparison matters.
What Actually Works: Case Studies in Smart and Wasteful Spending
Switzerland’s Apertus Model (September 2025): Trained on public supercomputing time, designed for open release. Bern has no proprietary asset to trade. This investment belongs in research or standards-setting budgets, not sovereignty. Publishing a model sets standards, but that is a separate objective with its own costs.
Japan’s Photoresist Play (2024): Japan Investment Corporation took photoresist maker JSR private for $6 billion. Photoresists are critical for advanced chipmaking. But Tokyo had already tested its position with export restrictions in 2019. Those controls cost no public money and created leverage. The $6 billion purchase secured equity in a position Japan already governed. Leverage that is for sale is already half spent.
Smarter alternatives: Power supply, water, and permitted land are increasingly crucial for data center location. A country offering these has a scarce resource to trade. Generic datasets lose value as models improve. Data retains value when it is current, legally usable, and hard to assemble. Court archives, health records, or unique language recordings qualify. New Zealand’s Te Hiku Media shows how Māori data can be governed with community benefit.
Regional Pooling and Exit Options
Pooled procurement or shared compute facilities can achieve scale for negotiation without full national ownership. ASEAN, the African Union, and the Gulf Cooperation Council could pursue this. The EU already pools for supercomputing through EuroHPC.
For outsourced capabilities, purchase terms dictate retained bargaining power. Multi-vendor contracts preserve exit options. The proposed Cloud and AI Development Act encourages public bodies to avoid single-vendor reliance.
Treat Spending Like Options, Not Assets
Ministers correctly note that capabilities get used in unforeseen ways. Some spending acts as a financial option: a small payment now for the right to act later. Options have prices and expiry dates. Ministries can fund small, staged bets with capped downside. Each should have a renewal or closure date.
Fund a capability at home when its generated leverage toward a national objective exceeds the return from the next-best use of public money. Otherwise, pool it regionally or buy it from the market while maintaining an exit option.
The Budget Question That Matters
A sovereignty claim becomes one line in a cost-benefit table, sitting alongside objectives like job creation and influence over standards-setting. The next sovereign AI budget should answer one core question: How much leverage does this money buy, and over whom?
If ministers cannot answer that clearly, the spending is not strategic. It is just expense.
Follow Hashlytics on Bluesky, Facebook, LinkedIn, Telegram and X to Get Instant Updates
