A flagship, and a compute problem
In late June 2026, the European Commission named EUROPA — a consortium led by Milan-based Domyn (formerly iGenius) with Germany's Fraunhofer-Gesellschaft — winner of its Frontier AI Grand Challenge. The prize: up to 2.5% of EuroHPC's total AI-optimized supercomputing capacity for a year, to build a fully open-source model exceeding 400 billion parameters, trained in all 24 EU official languages (European Commission; Heise). Domyn CEO Uljan Sharka has said the model will ship within a year, fully open-weight and reproducible, so European firms and governments can run it on their own infrastructure (Il Sole 24 Ore).
That framing — open, sovereign, no vendor lock-in — is the whole political sell. It is also on a collision course with the EU AI Act's compute math.
Why 400B parameters probably means systemic risk
Under Regulation (EU) 2024/1689, a general-purpose AI model is presumed to carry "systemic risk" — triggering the Act's toughest tier of obligations — once its cumulative training compute exceeds 10^25 floating-point operations. That is a rebuttable presumption, not an automatic designation: a provider can submit evidence to the Commission that the model doesn't actually pose systemic risk, and the Commission decides (EU Artificial Intelligence Act, Art. 51).
The threshold isn't hypothetical at this scale. Meta's Llama 3.1 405B — a similarly sized, 405-billion-parameter model trained on 15.6 trillion tokens — used an estimated 3.8×10^25 FLOPs, roughly four times the threshold, and Meta had to notify the Commission under the Act's systemic-risk process (DigiAlps). EUROPA is being built to be genuinely competitive at frontier scale, in 24 languages simultaneously — which mechanically demands more training tokens, not fewer, than a comparable English-only model. There is no serious scenario in which Domyn trains a 400B+ multilingual frontier model on meaningfully less compute than Meta used for 405B in one language.
The exemption that disappears exactly when you need it
This matters because the Act's open-source carve-out is narrower than the EUROPA announcement's messaging implies. Article 53(2) exempts providers releasing models under a free and open license — with published weights, architecture, and usage information — from the Act's technical-documentation and downstream-information duties. But the text is explicit: "this exception shall not apply to general-purpose AI models with systemic risks" (EU Artificial Intelligence Act, Art. 53). Cross the FLOP line, and open-weight status stops buying any relief — the provider owes model evaluations, adversarial testing, incident reporting to the Commission, and cybersecurity safeguards under Article 55, on top of the copyright and training-data-summary duties that apply regardless of license.
Italy's own AI statute doesn't create a separate escape hatch. Legge 23 settembre 2025, n. 132, in force since October 10, 2025, is explicit that it complements rather than displaces Regulation 2024/1689 (Gazzetta Ufficiale; Norton Rose Fulbright). Rome's national law leaves GPAI classification entirely to the EU regulation. So the "open-source exemption" the EUROPA project is implicitly banking on isn't an Italian carve-out at all — it's the same Article 53(2) exemption available to any provider in the bloc, and it is the one provision in the whole Act least likely to survive contact with a 400-billion-parameter model built to be state of the art.
The steelman, and where it runs out
The compute threshold isn't arbitrary regulatory reflex. Brussels needed some measurable, hard-to-game proxy for "capability jump large enough to warrant scrutiny," and qualitative capability tests invite endless litigation over what counts as dangerous. A rebuttable presumption tied to a public, auditable number is more predictable than a discretionary case-by-case call — and Meta's own notification shows the process can run without halting a release.
But a training-FLOP proxy is a poor fit for exactly the kind of model the Commission itself funded. Training 24 languages to competence inflates compute without proportionally inflating dangerous capability in any single language — the threshold penalizes linguistic breadth, not risk. And a publicly funded, fully open-weight model that European regulators, universities, and competitors can inspect line by line is arguably more auditable than a closed frontier lab's black box, yet the Act treats both identically once compute crosses the line. Commission Executive Vice-President Henna Virkkunen framed EUROPA as proof Europe can build "openness, trust and strategic autonomy" together; a compute rule that strips openness of its regulatory benefit the moment a model is good enough to matter cuts against that goal.
None of this kills EUROPA — Article 55 obligations are compliance burdens, not a ban, and Meta's precedent shows a systemic-risk-classified model can still ship. But Brussels and Rome should stop selling this project as free of the Act's heaviest tier and start building the evaluation and reporting pipeline now, rather than let a genuinely European achievement get overshadowed by a rushed retrofit six months from now.