On 19 June the European Commission named the EUROPA consortium — led by the Italian company Domyn — winner of the Frontier AI Grand Challenge. The brief is specific: an open-weight frontier model, north of 400 billion parameters, fluent across all 24 official languages of the Union. Henna Virkkunen, the Commission's Executive Vice-President for tech sovereignty, framed it in one line: "Europe can lead in advanced AI on its own terms." The compute and the public money follow the mandate.
I want Europe to win this decade. I have spent it building AI inside a European company, under European rules. So read what follows as a builder's objection, not a cynic's. Europe just committed public money and the scarcest thing it owns — political attention — to the one layer of the AI stack that gets cheaper with every model generation. The model. The layer that commoditizes itself on a schedule set in San Francisco and Hangzhou, not in Brussels.
This is not a sovereignty play. It is catch-up, in the only part of the stack where catch-up does not compound. By the time EUROPA ships a tuned 400-billion-parameter model and clears it across 24 languages, the capability it represents will be rentable by anyone, for a price that only falls. The value will have moved two layers up — to a place no consortium was assembled to reach.
What Europe actually bought
The fear underneath the decision is real, and it deserves its due. In mid-June a US export-control order forced Anthropic to cut access for foreign nationals, and Claude Code and Fable 5 went dark outside the United States. Every European team that had wired those tools into production learned, in an afternoon, what rented means. A brain you rent is a brain someone else can switch off. Europe noticing this is correct.
EUROPA is the wrong answer to the right fear. What the Commission bought is an open-weight model in the 400-billion-parameter class, trained for European languages, designed to sit at the global frontier. As infrastructure, that is a fine thing to have. But infrastructure is not a moat, and a model is the most rapidly depreciating infrastructure in the entire stack. The thing you build at enormous cost this year is the thing the market gives away for free in eighteen months.
A model is not a cathedral. It does not stand for a century. It is closer to a phone — frontier on release, mid-tier within a year, a museum piece soon after. You do not achieve sovereignty by building one. You achieve a maintenance bill.
The layer that gets cheaper without you
I made this argument three weeks ago about companies, in The company is the product. It applies to a continent without changing a single variable.
A company that sells the model is structurally short the frontier labs. Every generation makes the same capability cheaper to produce, until the lab ships it natively and the margin collapses. Replace "company" with "Europe" and the sentence still holds. EUROPA is structurally short OpenAI, Anthropic, Google, and DeepSeek — and short its own depreciation schedule on top.
The timeline makes it worse. The Grand Challenge opened in February. EUROPA was selected in June. Training a 400-billion-parameter model, aligning it, evaluating it across 24 languages, and getting it safe enough to ship is not a 2026 deliverable. By the time it lands, the open-weight frontier it is chasing will have moved again. DeepSeek, Qwen, and the Llama lineage already hand Europe open weights at or near the frontier for the cost of the download. Building one more adds a unit to a commodity pile that grows whether Europe contributes to it or not.
This is the trap of competing on a depreciating asset. The frontier is not a finish line you cross. It is a treadmill someone else controls the speed of. You can run hard, hit the target you set in February, and arrive to find the target two layers up the stack and a year down the road.
Where the value already moved
Read the companies that are actually capturing the AI budget right now. None of them built a model.
Harvey sells legal work to BigLaw firms — the memo, the diligence summary, the contract review. Sierra sells resolved customer-service tickets, priced per resolution, and already runs in production across 40% of the Fortune 50. Cognition sells pull requests. Each one rents the brain from a frontier lab and captures value in the layer above it: the integrations into the systems where the work lives, the workflow knowledge encoded as prompts and evaluation harnesses, the trust to be let near the work at all.
The public money and the political oxygen behind EUROPA went to the brain. Not one comparable instrument exists for the body — the layer where Harvey, Sierra, and Cognition actually live and bill. There is no Grand Challenge for the work layer. No consortium, no headline, no Executive Vice-President assigned to it.
Ask the simple question and the gap is obvious. Where is the European Harvey? The European Sierra? The European Cognition? They do not exist at scale yet, and the instrument that might have funded them was spent one layer down, on a brain that the market is busy commoditizing for free.
Open weights are table stakes, not a moat
Here is the objection I expect, because I half-believe it myself: it is open-source, that is the whole point, that is how Europe stays sovereign.
I argued in February, in Open is the moat, that open beats closed in this cycle. I still believe it. But read what that argument actually says. Open weights are a moat for a frontier lab running an ecosystem play — Meta open-sourcing Llama to commoditize its rival's complement, to deny anyone a monopoly, to make its own layer the default surface everyone builds on. Open is a weapon when you already own a layer above the model where you capture the value the open weights unlock.
Europe has the weapon and not the layer. Open weights you build with public money and then give away are not a moat. They are a public good. A fine thing to fund — like a road. But nobody captures the value of a road by having paid to build it. They capture it by what they drive on it. Europa está pagando la carretera y no tiene flota. / Europe is paying for the road and owns no fleet.
The road gets used. By Harvey. By Sierra. By the next American or Chinese work-firm that points a cheap, capable, open European model at a European market and bills for the output. Open weights distribute capability. They do not capture value. Confusing the two is how you spend a sovereignty budget subsidizing everyone else's margin.
What sovereignty actually requires
If Europe wanted to capture value and not just compute, it would fund the layer above the model.
That layer is concrete. It is the integrations into the systems where European work actually happens — the ERPs, the hospital systems, the court e-filing stacks, the public-sector platforms. It is the data signal those systems throw off, which no American lab can legally touch the way a European operator can. And it is the one asset Europe already owns that nobody else does: regulatory standing. The right to be in the loop on high-risk work.
Here is the irony that should sting. Europe's best moat is already built, and it is not a model. It is the AI Act. The same regulation everyone files under "compliance tax" is, read correctly, a context-layer advantage — it defines who is allowed to deliver high-risk work, under what audit, with which human in the loop. I wrote in May about the line the AI Act drew through HR. That line is a moat for whoever stands on the right side of it with the integrations and the trust to operate. The model is rented from a lab. The right to operate is not.
This is the layer I spend my days on. At Shakers I am not trying to out-train Anthropic — the model is rented, and that is the correct factoring. What we build is the context: the matching engine, the signal the marketplace throws off about how talent and demand actually meet, the workflow knowledge of hiring, the standing to operate inside the AI Act's high-risk perimeter for recruitment rather than around it. None of that commoditizes when the next model ships. All of it compounds. Spain's own regulatory sandbox, Sandbox IA España, is closer to the right instrument than a Grand Challenge: it builds the rules of the road and the standing to drive, not another brain to be obsolete by 2028.
If I had the Commission's budget and its mandate, I would not have run a contest to build a model. I would have run one to build ten work-firms — European Harveys, European Sierras — each renting the best brain available and capturing the work above it. The brain can come from anywhere. The work has to be owned.
The company is the product. So is the country.
A sovereign model is a commodity with a flag on it. The capability it carries will be rentable by anyone within two years, for a price that only falls, and the open weights guarantee that everyone — including the firms that will out-compete European operators — gets it for free.
What does not commoditize is the layer above: the integrations, the data signal, the trust, the regulatory right to be in the loop on the work that matters. That is the layer with no consortium, no Grand Challenge, no Executive Vice-President. Europe funded the brain and left the body to whoever shows up. The market that wins it will not be European by default.
Europe did not need its own brain. It needed to own the work that the brain does. On 19 June it funded the first and forgot the second.