No European companies are asking for sovereign AI

Yesterday I sat down with one of Europe's most important system integrators, people who wire AI into some of the continent's most critical white-collar industries.

Their Chief AI Officer said something that would horrify half my LinkedIn feed: "Up until now, we haven't seen any important companies ask for sovereign AI to replace American labs."

He's right.

And I co-lead a company betting on European AI infrastructure, so let me explain why I agree.

Nobody buys flags. Value in AI accrues overwhelmingly at the frontier, and until very recently, asking for "sovereign AI" in Europe usually meant asking for worse AI (sorry, Mistral, I'm rooting for you). Serious companies don't order worse on purpose; they order it when they can get away with it to get the job done, but the most important tasks still need frontier intelligence.

So European buyers did the rational thing: they bought American frontier intelligence and kept quiet about the geopolitics. The absence of demand for the sovereign label was never the absence of the problem; it was evidence that capability beats ideology in almost every procurement meeting when you are competing globally in certain industries.

To see where this goes next, you need the best mental model, and one of the best things I have read recently on the AI market comes from the excellent Giovanni Cattani.

To his point, everyone models AI supply, and the supply is spectacular: roughly $1 trillion of AI investment in 2026 (equal to France's GDP), with estimates heading far higher over the coming years. But as Cattani says, almost nobody models demand.

To do so, he does something smart, which is to build a taxonomy of AI work on two axes:

1. How big is the task? Every model can complete work up to a certain size, and according to some, the horizon has/will roughly doubled every seven months for six years.
2. And does the work ever end? Bounded tasks are finite: taxes, a discharge note, a compliance check. There's a "done." Unbounded tasks have no obvious ceiling, and the more intelligence you throw at them, usually, the more value you can accrue: research, trading, building software, improving strategy. More intelligence can keep being worth more.

When you face a bounded task, especially a repeatable one, like writing a transcript over and over again, you start optimizing heavily for cost and speed of execution. That means work converges on the cheapest capable model, and here, small open-weight models a step behind the frontier can win it on price.

But where the marginal value of additional intelligence remains high, the frontier keeps its pricing power. Imagine wanting to compete in quant financial trading against the likes of Jane Street, allegedly a major Anthropic customer, with a materially weaker model. You are toast.

According to Cattani, this creates a reflexive spiral which, simply put, goes something like this: tokens generate more revenue at the frontier, and the revenue then buys the frontier more tokens/compute to expand the advantage (especially through more training of bigger and better models), driving more tokens being sold. He estimates such loops drive roughly half of frontier lab revenue and, therefore, much of the capex buildout. He adds a fair warning that reflexive loops can also run in reverse, though I agree with him that this doesn't look particularly likely anytime soon, as CAPEX spend on frontier models keeps delivering capabilities the market wants to pay for.

Now, let's place Europe on this map: Our adoption is allegedly "world-class", with 20% of EU enterprises using AI in 2025, and Denmark leading at 42%. But according to Eurostat, our number one workload is text mining. Documents, summaries, extraction, and other relatively bounded tasks, most done within the $20/month subscriptions. According to this model, almost all of that will increasingly run on commodity models in places where it can be consumed easily.

Meanwhile, Europe holds roughly 5% of global high-end AI compute against America's roughly 75%, and much of the compounding, strategically decisive work still runs on the other side of the Atlantic.

That's why (almost) nobody asked for sovereign AI in Europe. There was nothing at the frontier to ask for in Europe. We were selling sovereignty bundled with a capability downgrade.

Here's what's changing, and why I'd take the other side of that Chief AI Officer's observation as a forecast: The frontier is no longer just a model. It's a system.

For years, delivered intelligence and raw model quality were almost the same thing: whoever pre-trained the biggest model shipped the smartest product, and pre-training is increasingly a capital monopoly.

But capability gains now come from somewhere else too.

Test time compute: the same weights think longer on hard problems and produce materially better answers, meaning capability is now partly bought per task rather than entirely baked in at training. Harness engineering: Claude Code is a harness, not a model, and it has become one of Anthropic's fastest-growing businesses. Agentic engineering: decomposing work, routing steps, tool use, state management, governance. The same model can complete a task inside a well-built loop or fumble around naked without it. Inference optimization: costs are collapsing quickly, which converts directly into more thinking per euro. Post-processing: validators, retrieval, structured outputs, PHI scrubbing, domain-specific checks. The unglamorous layer that turns an impressive demo into something a regulated enterprise can actually depend on. We live this daily at Corti: in specialized domains, a well-harnessed open model can beat a generic frontier.

If we put these factors together, the strategic picture starts to change, and the gap between open weights + scaffolding vs. closed models has narrowed dramatically across a growing share of commercially relevant unbounded workloads, and more of what closes the remaining gap comes from harnesses, orchestration, verification, inference-time compute, and domain engineering, versus just from more compute.

For the first time, running near-frontier intelligence on European soil, under European law, does not necessarily require a major capability sacrifice, and the share of "frontier" capability that lives above the weights grows every quarter.

We believe that the moment sovereignty stops coming at the cost of intelligence, the demand lying dormant inside "compliance requirements," "price predictability," "data residency," and "vendor risk" gets translated into Euros.

We rebuilt our own stack on that thesis and opened it as Corti Models: near-frontier open weights, harnessed, orchestrated, and served on European infrastructure. The results across our workloads: 2.5x the speed of the frontier APIs we replaced, quality on par with them in our domains, and materially lower build costs because a system tuned to the task doesn't waste thinking, nor do we force you never to KV cache anything ;).

That's the systems dividend, collected in production, under European law, so no, I don't think many companies will walk into a procurement meeting asking for "sovereign AI" at low quality. They'll ask for the best AI they're allowed to depend on, and when the offering becomes frontier-near, the European AI workload will really kick off, and that is the actual opportunity.

Sovereign AI wins when sovereignty stops being the product and becomes a property of the best infrastructure.

‍