What it would actually take for a European frontier model

No country has all four ingredients

The data exists in Europe, but it is spread across languages, states and institutions that do not pool it. Public compute exists, but it is fragmented across national programmes with distinct priorities. The talent exists, and it is excellent, but it is scattered across laboratories that do not work together. The market exists, and it is considerable, but today it mostly buys from non-European suppliers.

Each of these four resources, taken on its own at the scale of one country, sits below critical mass. Taken together at the scale of the continent, they are comfortably above it. That is the whole question, and it is one of organisation before it is one of funding.

The demonstration by proof

The argument is not theoretical. A complete assistant was assembled in a few weeks by Swiss, Finnish and French players, once each brought what it knew how to do: an open base model on one side, speech processing capabilities on another, integration and operations from a third.

The result was not a laboratory prototype but a usable system. What made that assembly possible was not exceptional funding: it was the fact that each component was open, documented and therefore combinable without prior negotiation.

That is the most underestimated property of digital commons. They do not only cut cost, they cut delay, because they remove the slowest step of any European project: agreeing before starting.

What happens when common infrastructure is missing

In the absence of open common infrastructure, the trajectory of national champions is predictable. They raise funds, reach a respectable size, hit the wall of training costs, then get acquired by or aligned with a non-European player that does hold all four ingredients.

This is not a charge against those companies: it is the mechanical consequence of an environment that forces them to fund alone what ought to be shared. A private champion is acquirable by construction, and its sovereignty lasts as long as its capital.

A common, by contrast, cannot be bought. It can wither if nobody maintains it, which is a real and different risk, but it does not change owner. For infrastructure that public administrations and essential operators will rely on for twenty years, that difference is decisive.

What Europe already has, and underestimates

The dominant narrative describes a gap. It describes the assets poorly, and they are real, and they are not the ones usually cited.

Multilingualism first, presented as a handicap when it is a rare training resource: few datasets in the world cover as many languages with as many aligned documents of controlled editorial quality.

The public corpus next, considerable and largely underused: official journals, case law, parliamentary debates, archives, publicly funded scientific output. This corpus has a property few others share, that of being legally clear.

The engineering culture of free software, finally, which is not a cultural footnote. It gives the habit of working in common between competing organisations, of documenting for third parties and of maintaining over time. That is exactly the skill producing a digital common demands, and it cannot be bought.

Four concrete levers

A European project dedicated to open frontier models, bringing existing initiatives together under collective governance rather than creating yet another structure. The building blocks already exist, what is missing is the frame that makes them converge.

Reserving public compute capacity for fully transparent projects: published weights, documented training data, reproducible method. Public money is entitled to require the openness of what it funds, and it is the most immediately available lever.

A common evaluation infrastructure, against European criteria and on European tasks. You do not steer what you do not measure, and evaluating European models with test sets designed elsewhere means competing in someone else's category.

Public procurement as an outlet, finally, in essential services. A model without usage does not improve. Public demand is the only instrument capable of creating, within a few years, the volume of usage that is missing.