Europe built the compute. The question now is who gets to use it.

For most of the past five years, the European conversation about artificial intelligence infrastructure was a conversation about absence. The largest training runs happened on American clusters, leased by American companies, under contracts European institutions could not match.

That framing is now out of date. The EuroHPC machines in Kajaani, Bologna and Barcelona represent genuine, publicly owned training capacity, and the newer AI factories attached to them were specified for model work rather than for classical simulation. The hardware argument has largely been won.

The allocation argument has not. Public supercomputing time in Europe is still awarded through review processes designed for computational chemistry and climate modelling: proposal windows measured in months, scientific merit panels, and reporting obligations that assume a research paper at the end.

A model team building a Nordic language model does not work on that clock. It needs a block of compute within weeks, the freedom to discard a run that is going badly, and no obligation to describe the failure in a journal.

What the queue looks like in practice

Three research groups interviewed for this piece, in Helsinki, Trondheim and Ghent, described the same sequence. An application accepted, an allocation granted for a window six to nine months out, and by the time the window opened, a change in the open weight landscape that made the planned run redundant.

None of them described the hardware as the problem. All three described the calendar as the problem.

The comparison that matters

The relevant benchmark is not total petaflops against the United States. It is time from decision to first token of training. On that measure, a European team using public infrastructure is operating at a structural disadvantage against a team with a commercial cloud contract, regardless of how much public silicon exists.

Some of this is being addressed. Fast-track allocation lanes for small industrial and startup runs now exist at several sites, and the AI factory model explicitly contemplates shorter cycles. The lanes are small relative to total capacity.

What to watch

Three signals will tell you whether the policy is working. The share of total allocated hours that goes to runs booked with less than sixty days of notice. The number of commercial and startup users as a fraction of all users. And whether European model teams stop describing their public allocation as a supplement to a commercial contract and start describing it as their primary environment.

Until those three numbers move, Europe has built the factory and kept the old booking system.