Brussels Chooses Italian Consortium EUROPA to Build Open Frontier LLM in 24 EU Languages

The European Commission selected the EUROPA consortium, led by the Italian company Domyn, as the winner of the Frontier AI Grand Challenge on June 19. The group will receive up to 2.5% of the global computing capacity of EuroHPC for twelve months to train an open frontier model with over 400 billion parameters covering the 24 official languages of the bloc. The public launch of the program, which occurred in February of this year, promised a selection within six months, and the Commission met this deadline.

Domyn, based in Milan, already operates the Colosseum, which the company describes as the largest enterprise supercomputer in Europe. The facility, located in northern Italy, has 6,000 NVIDIA Blackwell GPUs installed and contracted capacity for an additional 4,000 by the second half of the year. The company is in a funding round of approximately 1 billion euros to scale what it calls "AI gigafactories," according to a report from Tech Funding News published on June 17.

What's at Stake

The selection defines more than just a winner. It positions the Frontier AI Grand Challenge as the first program in the EU with scale comparable to the U.S. Project Stargate, albeit with fractions of the private budget. The only publicly known contender on the shortlist was the consortium led by Mistral, based in France and Germany. The choice of EUROPA, in addition to the multilingual agenda, reflects the Commission's position of favoring open-source code and portability among AI-optimized supercomputers already in operation (LUMI in Finland, Leonardo in Italy, JUPITER in Germany).

The agreement outlines three contractual requirements that matter for any European company considering replacing Claude or GPT in regulated stacks: open weights under a commercial-use compatible license, complete technical documentation, and demonstrated capability for local fine-tuning for public sector use cases. "We’re talking about making advanced AI accessible to businesses, researchers, and public institutions across the linguistic diversity of the Union," stated Henna Virkkunen, Executive Vice President for Technological Sovereignty, in the announcement.

The Counterargument that Needs Consideration

Ana Paunescu, an analyst at the think tank Bruegel, published a column in the European Express on June 20 arguing that the bet on open-source underestimates the real cost of large-scale inference. According to Paunescu, the EU confuses sovereignty with self-reliance: having the model doesn't guarantee economic competitiveness if the data centers running the model continue to rely on American GPUs and cloud services from the same three hyperscalers that the Commission aims to regulate under the EU AI Act. This argument is legitimate and weakens the triumphalist narrative from Brussels, celebrating independence at the model layer while the infrastructure layer remains under foreign control.

A counterpoint that appears in the Commission's document is that support for multilingualism in 24 official languages is not a demand from the North American market, thus creating differentiation. Models like GPT-5 and Claude Mythos continue to perform significantly worse in Finnish, Hungarian, Lithuanian, and Maltese than in English. EUROPA could occupy this niche with structural advantages, although the market size may be smaller.

Cross-Reading: Where the Game is Decided Beyond Italy

In Germany, SAP has already signaled that it will evaluate EUROPA for implementation in Business AI stacks under BaFin's data sovereignty requirements. The company has not committed, but a report from Handelsblatt in February quoted SAP executives indicating that replacing Claude with a European alternative in regulated financial services workloads was "a matter of when, not if." For Capgemini, which operates delivery centers in Poland serving much of the DACH banking sector, the arrival of EUROPA reduces the risk of exposure to American export controls on frontier models.

In Japan, the reading is indirect but relevant. NTT, KDDI, and Sakura Internet, which are building national alternatives to language models with METI funding, see the institutional design of the Frontier AI Grand Challenge as a precedent for their own AI Bridging Cloud Infrastructure 3.0. Brazil, which currently has no equivalent public-private consortium, stands outside a wave that is beginning to shape what will be the model-infrastructure axis of the non-American world. SUDENE and BNDES have been discussing financing lines for a sovereign data center in Pernambuco since March, but the institutional design remains undefined.

The operational conclusion for the global CIO is that the LLM supplier option has shifted from a duopoly (Anthropic-OpenAI with Google in the background) to a three-geography board with open models competing for technical parity. EUROPA has 18 months to demonstrate that this is more than just a political signal.