The European Commission has officially selected Italian company Domyn to develop a sovereign European AI model with 400 billion parameters. In this article, NEWSROOM IN explains the corporate tasks the new system will solve, why the French market leader Mistral was excluded from the grant, and how much the project will cost. The article also examines whether the EU’s strict AI Act can protect citizen data, or if future security breaches—similar to a recent large-scale hack in Lithuania—could expose these localized databases on the darknet across all 24 official European Union languages.
European AI Model: An Answer to US and Chinese Tech
Developing an open-source model is a strategic step for Europe’s technological plans. Until now, models with over 400 billion parameters were developed almost exclusively by tech giants from the United States and China. The new European system will support all 24 official EU languages, reducing the region’s dependence on foreign advanced technologies.
The future model will be released as open-source software, providing equal technological opportunities for EU public institutions, researchers, and businesses. Domyn will receive infrastructure support from the EuroHPC Joint Undertaking. For one year, the company will have access to 2.5% of the computing capacity of European supercomputers optimized for AI algorithms.
Focus on Security and Strict Regulation
Unlike many commercial AI projects built for general academic or creative tasks, Domyn’s model is designed for highly regulated sectors. The model will specialize in financial services, defense, and advanced manufacturing. The system is being optimized to handle complex queries within structured databases.
Technical Foundation and Nvidia Support
Domyn has prior experience in scaling AI models. Its portfolio includes the 263-billion-parameter Domyn-Large, as well as Italia-10B and Colosseum-355B, which run as Nvidia NIM microservices. Developed with technical support from Nvidia, the new project will use a modular Mixture-of-Experts architecture to efficiently distribute the massive computing load.
Why the European Commission Bypassed France’s Mistral
The French company Mistral AI is a prominent European AI startup, but the Commission did not select it for three main reasons:
- A shift toward proprietary code: A key EU requirement was a fully open-source model. Mistral made its most powerful model, Mistral Large, closed-source and partnered with Microsoft, which runs contrary to the concept of European technological sovereignty.
- Product specialization: Mistral develops general-purpose generative models. The European Commission required a specialized tool for highly regulated and compliant industries, which is what Domyn offered.
- Political balance: France is already Europe’s main AI hub. Granting supercomputing capacity to an Italian consortium helps diversify technological influence and support the industry in other EU member states.
It remains unclear whether Mistral participated in the tender. Under EU grant rules, the European Commission only publishes the winner’s name, keeping the full list of applicants confidential. Mistral has not commented publicly on whether it applied. Given the startup’s current business strategy, it is highly likely the company chose not to submit a proposal.
Project Costs and Required Funding
Maintaining and developing a model with over 400 billion parameters requires significant funding, estimated at $80 million to $150 million annually.
The public grant providing supercomputing capacity offers a strong start, but it is currently limited to one year. The total amount needed to maintain and develop a 400-billion-parameter model could range from €70 million to €130 million per year.
These estimates cover several primary expense categories:
- Regular retraining and updates (from €35 million per year): Retraining a massive neural network requires substantial machine resources. Even with free access to EuroHPC supercomputers in the first year, subsequent compute rentals will cost tens of millions of euros.
- Inference infrastructure and APIs (from €22 million per year): Supporting the cloud infrastructure (inference) so European companies can seamlessly access the open model requires clusters of hundreds of modern GPUs. This also involves high electricity and server maintenance costs.
- Salaries (from €13 million per year): Maintaining and developing a system of this scale requires highly skilled engineers and researchers. To retain top European talent and compete with US companies like OpenAI or Google, salaries must be competitive with Silicon Valley rates.
To prevent the project from becoming outdated, the European Commission will need to budget approximately €100 million annually. Without continuous funding, the model risks becoming obsolete within a few years.
The Future of Data: Law vs. Reality
Although regular users will not interact directly with this specialized AI model, its implementation could affect daily life. Banks, medical institutions, and government agencies using the model are expected to process requests faster and improve services. The main benefit lies in personal data protection: European citizens’ data will be processed locally and in strict compliance with the EU AI Act, which is intended to minimize the risk of large-scale leaks. Additionally, the open-source nature of the model will allow local, independent developers to build consumer applications on top of this infrastructure free of charge.
However, as a recent large-scale database breach in Lithuania demonstrated, hackers do not consult European Parliament or European Commission directives before launching attacks. Concerns remain over whether localized, legally compliant personal data can be kept secure, or if it will end up on the darknet, translated across all 24 official EU languages.
Background: Large-Scale Database Hack in Lithuania
In May 2026, details emerged of an unprecedented cyberattack on Lithuania’s state Center of Registers. Attackers exfiltrated over 600,000 records, affecting nearly 20% of the country’s population. The stolen data included names, personal identification codes, birth dates, and property ownership details of individuals and legal entities, including politicians, military personnel, and political emigrants. The breach occurred via compromised accounts at the Migration Department. The incident led to the resignation of the head of the Center of Registers and highlighted vulnerabilities in state-managed databases.