Rethinking Sovereignty: AI Dependencies Demand Flexible Governance

Sovereignty, as the Tony Blair Institute argues, won't be won with emergency decrees or flashy announcements. It's a long, structural fight about who builds the foundations of power — and who pays to maintain them. Follow the money.

Where the piece earns its keep is in its insistence that states stop chasing headlines and start building capability. Tactical measures buy front pages; enduring capacity buys options. The argument for a “long game” is right — but the real contest sits in the shadows of that phrase.

Because sovereignty in AI is not just a question of what governments choose to do.

It's a question of what they no longer control.

Private power, public promise

The Institute is correct to frame sovereignty as a strategic choice. But here's what they won't tell you: those choices are already narrowed by private actors who own the infrastructure, talent and data pipelines that matter most.

National capability is not just a procurement problem; it's a dependency problem.

Yes, governments can fund labs, set standards and recruit technical staff. They can pour money into compute, education and regulatory architecture. Yet when the core models, critical datasets and most experienced engineers sit in boardrooms and shareholder meetings, not ministries, sovereignty turns into a negotiated status, not an absolute condition.

The article nods to “structural dependencies” and then treats them like engineering glitches. They’re not glitches. They’re levers.

Ask who benefits when a state leans heavily on external suppliers for AI capacity. Private firms win contracts, stack proprietary tools on proprietary infrastructure, and gain bargaining chips that shape what “national interest” even looks like in practice. That’s not conspiracy; that’s contract law and market power doing what they do best.

Convenient, isn't it?

The Institute’s long-game prescription needs a sharper accounting of those power imbalances — and a clearer sense of how governments resist capture without suffocating the innovation they need.

Cooperation without capitulation

One of the article’s strengths is its pushback against short-termism. But embedded in the text is a false choice: build national capability or deepen international collaboration.

Why pretend those are rivals?

Autarky would slow technical progress and fracture standards. A naïve faith in global supply chains would hand initiative and resilience to whoever sits at the chokepoints. The harder, more honest path is selective integration: invest domestically in core competencies while shaping interoperable, enforceable rules with other states.

That demands diplomats who can read code as well as communiqués. Regulators willing to demand verifiable commitments, not glossy ethics pledges. Procurement rules that prize transparency and portability over the easy comfort of vendor lock-in.

It also demands a public debate about what sovereignty actually means when code and data move across borders faster than any customs officer can blink. Not in abstract slogans, but in concrete questions: which systems must be auditable on demand, which data flows must be reversible, which dependencies are acceptable — and which are simply reckless.

The missing muscle: enforcement and data

Where the Institute’s piece feels thin is on enforcement. Anyone can write a standard; the test is whether it bites.

Data governance should sit at the center of its argument and doesn’t. Sovereignty today rests on who collects, stores and governs data flows — the fuel that trains and tunes models. The article’s emphasis on long-term capability is welcome, but it skims over the tools that would actually change incentives: legal hooks, cross-border enforcement mechanisms, and penalties that matter more than a day of bad press.

Without enforcement teeth, what you have is not sovereignty. It’s wishful thinking stapled to a strategy document.

Inside democracies, there’s another uncomfortable layer the piece mostly walks past. Buyouts, subsidies and preferential contracts aimed at “building capacity” all too easily entrench the very concentration of power that erodes sovereignty. States trying to catch up risk hardwiring private monopolies into the core of their national infrastructure.

There are alternatives: backing open models where safe and appropriate, shared compute pools for public-interest research, public datasets with clear guardrails. None of these are silver bullets. But they at least push against the gravitational pull of permanent dependency.

A counter-argument — and a harder one

Critics of national AI investment have a case: centralized state programs can be slow, insular and prone to misallocating resources. Compared with nimble private labs, governments look lumbering and risk-averse. Cooperation with those labs, they argue, isn't just pragmatic — it's essential.

That’s not entirely wrong. But it carries an assumption that never quite gets said out loud: that the private sector’s trajectory will eventually align with the public interest if you just keep it close enough.

History suggests otherwise.

The answer is not for governments to try to out-build every frontier model or to shut out corporate actors. It’s to become smarter, more demanding partners. Use public funding to create credible alternatives that set norms; require operational transparency in exchange for access to public contracts; tie subsidies to interoperability, data-access commitments and exit options.

Follow the money, and then rewrite the rulebook it runs through.

A short historical memory

There’s a historical amnesia in the Institute’s vision. Communications networks, from telegraphs to undersea cables, have always been battlegrounds of sovereignty. States that left core infrastructure entirely to private consortia discovered, usually in crisis, that they had outsourced bargaining power along with maintenance.

Today’s cloud and model providers are not building Victorian railroads. But the pattern rhymes: whoever owns the rails decides what runs on time.

Ignoring that history doesn’t make the risk go away. It just means we repeat it in code instead of steel.

The long game they gesture at

The Institute is right about one thing: long games need continuity. And continuity is political. Ministers rotate, coalitions fracture, fiscal policy turns on a headline or a scandal. AI strategy that sits in a single department, or depends on a single political champion, is strategy already half-abandoned.

The only way their argument holds is if AI sovereignty gets wired into the ordinary machinery of state — defense, health, education, competition authorities, all making aligned, sometimes boring, choices about contracts, standards and data.

That’s where the long game actually happens: not in speeches about sovereignty, but in the slow, quiet normalization of the idea that some dependencies are negotiable — and some simply aren’t for sale.