Europe's AI sovereignty must propel growth, not gatekeep

Look: McKinsey & Company’s “Accelerating Europe’s AI adoption: The role of sovereign AI capabilities” hits all the right political receptors — control, resilience, European agency. I’ll be honest: if you’re a policymaker who’s watched cloud, chips, and platforms drift offshore, “sovereign AI” sounds like a reset button you’ve been waiting years to press.

But the argument quietly smuggles in a big assumption: that sovereignty is an accelerator by default, rather than a drag if you get the architecture wrong.

McKinsey’s core claim is straightforward enough — build national or regional AI capabilities, and adoption across Europe speeds up. There’s a piece of truth there. When governments fund infrastructure, clarify liability, and commit to buying AI services, they derisk adoption for everyone else. Public procurement has always been a decent on-ramp for new tech; ask any cybersecurity vendor that landed its first big contract with a defense ministry.

Where the piece wobbles is in treating “sovereign capabilities” as a neutral good, as if the only variable is how much money you pour in.

Sovereignty sounds singular, but in practice it usually fragments. If every country spins up its own models, its own “trusted” infrastructure, its own certification scripts, Europe stops looking like a single AI market and starts looking like a loose federation of semi-compatible stacks. That’s not an abstract worry — we’ve seen the same pattern in payments and identity systems, where national schemes proliferated and only later had to be stitched together at great cost.

AI does not love duplication. Training and maintaining strong models is capital-intensive and talent-hungry; you want those resources compounding, not copied. Scatter them across a dozen lightly coordinated sovereign projects and you get smaller systems, thinner expertise, and fewer network effects. The bill rises, the quality drops, and the very companies you want to help — the ones without in-house ML teams — end up paying more for worse tools because economies of scale never kicked in.

McKinsey nods at coordination but doesn’t linger on the trade-off: every extra increment of control has a price in lost scale and added friction. Sovereignty isn’t just a shield; it’s a tax line.

Now, McKinsey is right that trust is the real accelerant. The piece argues that if citizens and firms can see that “Europe runs this,” adoption will follow. That makes political sense. After years of debates around data protection and platform power, a European-branded AI stack feels psychologically safer than a black box from somewhere else.

Sure, but trust doesn’t only come stamped with a national logo.

Standards bodies, cross-border regulatory frameworks, common technical baselines, open-source reference implementations, and accredited audits can all build confidence without putting every country in the model-building business. Think of it as shifting from “trust our flag” to “trust this shared protocol.” That approach keeps interoperability intact while still giving regulators hard levers to pull when systems misbehave.

The article also underplays who usually spreads new tech: not ministries, but companies that move fast because they have to. Banks, retailers, manufacturers, hospitals — they’re already experimenting with AI to shave costs and create new services long before there’s a state-backed model in sight. Public “sovereign” projects can be a catalyst here, especially where markets underinvest. The risk is when they become a gatekeeper instead.

If procurement rules implicitly favor only state-blessed models or infrastructure, you don’t just create a comfort blanket — you create a moat. The players best positioned to work inside that moat are typically large incumbents who can navigate long certification cycles, not scrappy startups pushing the technology in new directions. That’s how you end up with safe, compliant, mediocre tools locked into public contracts while more capable systems orbit outside the fence, unavailable to the very public agencies that need them.

There’s also a quieter assumption embedded in the sovereignty pitch: that only a nationally controlled stack can really protect sensitive data and guarantee compliance. That’s a strong claim, and politically it’s catnip.

But protection and ownership are not the same thing. You can keep control of data flows without nationalizing every layer of the stack. Techniques like federated learning, strict data residency rules, well-audited secure enclaves, and binding contractual regimes for model providers all let Europe tap global innovation while keeping its crown-jewel datasets on home soil. You get trust and scale — not perfect, but a better trade than going it alone model by model.

History offers a gentle warning here. Europe has been here before with telecom. Every country once insisted on its own standards and tightly controlled networks; only when GSM gained real cross-border traction did the continent suddenly look like a unified market with global pull. Fragment first, harmonize later is an option, but it’s also why some industries took decades to reach their potential.

The funny thing is, “sovereign AI” doesn’t have to mean “national everything.” It could mean shared European infrastructure where it matters — compute, foundational models, benchmarking labs — combined with clear interoperability rules that any provider, domestic or foreign, can plug into if they meet the bar. Imagine a common certification layer recognized across member states, rather than parallel national seals of approval that all say “trust me, I’m different.”

McKinsey’s framing taps into a real fear of dependence and a real hunger for control. That’s why it will resonate in ministries and boardrooms. But if sovereignty is pursued as a badge rather than a design constraint — if the optics of owning the stack matter more than how that stack connects across borders — Europe may find that its “accelerator” feels more like a handbrake.

My bet: we’ll see at least one flagship sovereign AI program quietly pivot in a few years, from building its own full vertical stack to anchoring shared services and standards that others can build on — not because Brussels suddenly gets less anxious about control, but because the economics of scale won’t give them much choice.