Pragmatic AI Governance Over Sovereignty, Not Isolation

Sovereignty as a Silencer

Lawfare is right to ring the alarm: AI sovereignty — the push for states to control and localize models, data, and compute — carries serious downside risk. The logic is straightforward: once you turn code into a jurisdictional artifact, you get multiple, overlapping rulebooks. That splinters markets, raises compliance costs, and hands governments blunt instruments to throttle access. The piece sketches that risk; where it undershoots is in treating sovereignty as a mainly technical policy choice. It's political power dressed up as risk management.

When rules fragment, friction compounds. Firms face divergent deployment standards, conflicting certification regimes, and localized retraining demands. That raises marginal costs and shrinks the set of models that are commercially viable in any given market. Consumers pay through higher prices and fewer options, startups face steeper barriers to entry, and researchers lose easy cross-border collaboration. So yes — sovereignty can harden an uneven playing field where incumbents with deep pockets consolidate advantage and smaller players quietly disappear.

That’s the economic story. The political story is worse.

Sovereignty as a Weapon

The Lawfare piece gestures at a second danger but doesn’t fully follow it through: sovereignty as a foreign-policy instrument. When states can wall off models, demand backdoors, or set bespoke “national” architectures, they aren’t just regulating privacy and safety — they’re exporting their values via technical constraints and using access to AI as a bargaining chip. That shifts AI from a technology policy problem to a geostrategic asset.

This cuts both ways for national security. Autocracies can mimic democratic-sounding safeguards as cover for censorship or surveillance. Democracies can respond with reciprocity — restricting models, toolchains, or APIs from certain countries — and call it “responsible sourcing.” That’s not just regulation; that’s controlled decoupling.

Once AI infrastructure becomes a bargaining chip, interoperability is collateral damage. And once interoperability erodes, the cheap diffusion of innovation that has historically underpinned global tech growth goes with it.

I spent a decade at Goldman watching markets misprice this kind of thing. You can discount earnings for higher regulatory risk, but you can’t cleanly model the long-run cost of closed technical ecosystems. Risk management without credible multilateral governance isn’t prudence; it’s permanent crisis mode.

The Real Gap: Institutions, Not Intentions

Where Lawfare feels thin is on what comes next. Warning that “AI sovereignty is dangerous” is not a policy plan. The core problem isn’t that states want a say; it’s that we lack durable, interoperable institutions to set and enforce cross-border norms on safety, provenance, and auditability.

A single global regulator is fantasy. What’s plausible are coalitions of states that are politically aligned enough to trust each other’s processes. Think smaller blocs that agree on mutual recognition of certifications, shared audit frameworks, and basic dispute-resolution mechanisms. That’s slow, messy work. Let’s be real: unilateral digital borders, export controls dressed up as ethics, and hard localization mandates are politically easier — and they lock in the very fragmentation everyone claims to fear.

The “Necessary Sovereignty” Objection

Critics have a fair point: some sovereignty is non-negotiable if you care about citizens’ data and democratic processes. States do have legitimate reasons to assert control over models trained on their populations or deployed in critical infrastructure. Absolute openness isn’t a serious proposal.

But absent cross-border mechanisms, those protections morph into a shield for protectionism and repression. The line between safeguarding citizens and consolidating domestic political power is thin. Incentives do the heavy lifting. If local firms gain from strict localization because it keeps out foreign competitors, the logical endpoint is closed markets marketed as “ethical AI.”

What Lawfare Underplays: Three Frictions

There are at least three concrete frictions that get less attention than they deserve:

1. Compliance arbitrage: companies will gravitate toward jurisdictions with laxer AI oversight, creating regulatory havens for the riskiest models. Once that happens, enforcement becomes a game of chasing the weakest link.
2. Audit opacity: fragmented, locally forked models break global audit trails. Oversight turns into bilateral political bargaining instead of verifiable technical review.
3. Innovation concentration: only incumbents can afford teams of lawyers, policy staff, and engineers to tune models for every jurisdiction. Startups get priced out before they ship.

Those are engineering headaches with political causes — and they require political fixes, not clever model architectures.

A Useful Historical Echo

We’ve been here before, at least in outline. Think about early telecom and banking: bilateral deals, incompatible standards, and constant frictions. Things improved when states realized that persistent fragmentation was bad for everyone with actual capital at risk. They didn’t create a single global regulator; they built overlapping arrangements, standards bodies, and mutual recognition schemes that let money and messages move without rewriting the rules in every port.

AI sovereignty is replaying that script, but with far higher stakes because it touches information, labor markets, and security all at once. The math doesn’t lie: every national fork of the stack adds cost and subtracts transparency.

A Practical Constraint — and a Prediction

Any credible coalition model will live or die on one detail Lawfare barely touches: technical reproducibility and third-party audits that are portable across borders. Without that portability, certifications are just national flags on PDF files.

If AI sovereignty keeps advancing without matching progress on these shared mechanisms, the next “AI safety” announcement you see from a government is more likely to be about market control than model risk.