Leonardo Foundation: Italy in the Era of AI - Part 1
On March 19, 2026, in the Queen's Hall of the Chamber of Deputies, a document was presented that stands out clearly from the category of institutional reports destined to collect dust on some ministerial shelf. "Italy in the Era of AI. Growth, Challenges, and Perspectives of an Ongoing Revolution" was born with a declared ambition: not a speculative exercise, not an optimistic manifesto, but an operational map of the state of artificial intelligence in Italy, with recommendations accompanied by measurable indicators, precise institutional responsibilities, and analysis of barriers to implementation. It is, in a literal sense, what the title promises: a compass.
Signing it are Luciano Floridi, President of the Leonardo Foundation ETS and Director of the Yale Digital Ethics Center, and Micaela Lovecchio, who handles Education and Training for the Foundation. Floridi is a figure who has crossed the digital ethics debate as a protagonist for twenty years, not as an external commentator, and it is precisely this combination—academic rigor with a pragmatic orientation—that characterizes the framework of the work. In the preface, the author is explicit about the approach: "Too many reports remain a dead letter because they formulate vague objectives without indicating who must do what, by when, and how success will be verified. We have tried to avoid this trap."
The Leonardo Foundation ETS guaranteed scientific independence and involved a top-tier institutional ecosystem: CINECA, the FAIR (Future Artificial Intelligence Research) project, the Artificial Intelligence Observatory of the Politecnico di Milano, AgID, the National Cybersecurity Agency, INPS, ISTAT, the Chamber of Deputies, the Senate, Sapienza, the Politecnico di Torino, the Italian Institute of Technology, the Scuola Normale Superiore, along with private sector champions like Bending Spoons, Domyn, Eni, Fastweb, and Intesa Sanpaolo, and startups like Datapizza, ASC27, Crossnection, and Metis. The result is a text built on real data, not marketing projections, organized into six parts that proceed methodically from context to action: market analysis, assets and strengths, challenges and risks, sectoral opportunities, strategic recommendations, and monitoring mechanisms.
A methodological note is necessary from the start: the authors themselves systematically point out the limits of the sources, distinguishing between official statistical surveys and market estimates, between verifiable data and qualitative inferences. The warnings that dot the report are not a weakness, but a choice of rigor that must be kept in mind when reading any of the numbers that follow.
The market accelerates, but at two speeds
The starting point is a fact that, read without context, would sound like a triumphant press release: in 2024, according to the Artificial Intelligence Observatory of the Politecnico di Milano, the Italian artificial intelligence market reached 1.2 billion euros, with a 58% growth compared to the previous year. A number that immediately deserves clarification: it refers to spending on AI software and services, not the entire technological ecosystem, and should not be confused with commercial projections from providers like Statista, which adopt broader boundaries and produce significantly higher figures. The policy objective set in the report is to reach 5 billion euros by 2030: it is a political direction target, not an economic forecast, and the distinction is not secondary.
On the adoption front, ISTAT data indicate that 16.4% of companies with at least ten employees formally adopted artificial intelligence solutions in 2025, double the 8.2% of the previous year. The report explains this acceleration with a plurality of causes: the collective familiarization effect produced by tools like ChatGPT, the maturation of ready-to-use solutions that no longer require custom development, the PNRR incentives related to Transition 4.0, and the regulatory clarity introduced by Law 132/2025. There is also a statistical factor not to be overlooked, and the authors point it out honestly: starting from a low base amplifies the percentage variation.
Italy's 16.4% falls below the European average, which according to Eurostat reached 20% in 2025. Germany at 26%, France at 19.5%, Spain at 19.1%: Italy ranks 18th in the EU-27 for adoption but shows one of the fastest growth dynamics on the continent. The data on individual use is even sharper: only 19.9% of Italian citizens used generative AI tools in 2025, among the lowest values in the European Union, although the report calls for caution because surveys in this area are not yet fully harmonized across countries.
The real structural knot is the internal distribution. Companies with over 250 employees reached 53.1% adoption, more than double the 32.5% of the previous year. Small and medium-sized enterprises stop at 15.7%. This is what the report calls the "dimensional digital divide," and the causes are not reduced to the generic formula of lack of skills and high costs. They are structural and multiple: fixed implementation costs, from consulting to integration to training, are similar regardless of company size but weigh proportionally more on the turnover of an SME; the absence of dedicated figures like CTOs or data scientists makes artificial intelligence one priority among many equally pressing ones; information asymmetry is high, with SMEs struggling to distinguish valid solutions from commercial hype; in many realities, organized and adequate quality data archives are lacking. Since SMEs are the backbone of the Italian production system, the report is clear: generic incentives are not enough. Trusted intermediaries like trade associations and European Digital Innovation Hubs, pre-configured solutions for specific sectors, and low-risk experimentation environments are needed.
The supercomputing triad
If there is an asset where Italy has no inferiority complex in Europe, it is the high-performance computing infrastructure. The report documents what it calls the "supercomputing triad": in the November 2025 TOP500 ranking, two Italian systems occupy positions of absolute excellence at the continental level. HPC6, Eni's industrial supercomputer, is sixth in the world and second in Europe with 477.9 PetaFLOPS of real power (Rmax), with a peak power of about 606 PetaFLOPS. Leonardo, the system managed by CINECA in Bologna, is tenth in the world and fifth in Europe with 241.2 PetaFLOPS. Added to these is davinci-2 from Leonardo S.p.A., 123rd in the world ranking, with HPE Cray XD670 architecture equipped with Intel Xeon Platinum 8568Y+ processors and NVIDIA H200 accelerators, for a real power of 14.2 PetaFLOPS. Italy is the only European country with two systems in the continent's top 5, combining public and private capacity in a way that has no equal.
The dimension of energy sustainability of this infrastructure is not an incidental detail: CINECA's Leonardo supercomputer operates at 100% with renewable energy and reaches a PUE (Power Usage Effectiveness) indicator close to 1.1, where 1.0 represents the perfect theoretical limit and the industry average hovers around 1.5. In an era where the debate over artificial intelligence's environmental footprint is becoming increasingly urgent, every training cycle of a large language model consumes significant amounts of energy and emits CO₂, and energy efficiency stops being an environmental virtue to become a direct competitive advantage. The report dedicates a specific section to the environmental sustainability of artificial intelligence as a systemic criticality to be managed, and the fact that Italian infrastructures are already ahead on this front means anticipating regulatory constraints that will inevitably arrive.
The LISA Project, started in May 2025, is upgrading the GPU partitions of the Leonardo supercomputer specifically for generative artificial intelligence. On the future development front, Domyn's Colosseum project proposes an ambitious goal of 100 AI ExaFLOPS in low-precision computing, while the IT4LIA AI Factory initiative involves a 420 million euro investment, co-financed equally by the Italian Government and the European Commission, to develop skills and build a bridge between computational infrastructure and human capital. The National Strategic Hub responds to the need for sovereign cloud for the public sector, but dependence on large international cloud service providers remains an open strategic issue.
The report does not stop at celebration. There is what Floridi calls the sovereignty paradox: we own the engines, but the hardware supply chain remains dependent on foreign sources. NVIDIA holds an estimated share between 80% and 95% of the market for GPU accelerators for training artificial intelligence models. Export restrictions to China have shown that the United States is willing to use hardware as geopolitical leverage. Germany, meanwhile, has activated JUPITER, the first European exascale system with one ExaFLOPS of power. The structural response involves active participation in the European Chips Act and monitoring the development of European alternatives to accelerators currently dominated by a single supplier.
Speaking Italian: sovereign models
If supercomputers are the brute force, Italian language models are something subtler: a matter of cultural identity, legal precision, and autonomy in an era where words are data. A model trained predominantly on English texts brings with it cultural biases, gaps in understanding the Italian regulatory context, and inaccuracies in the language that for a doctor or a judge can have concrete consequences.
The report documents an ecosystem of Italian language models that has no equivalent in Europe for a language with about 65 million speakers. Minerva, developed by Sapienza NLP as part of the FAIR project, was released in 2024 with 350 million, 1 billion, and 3 billion parameter models trained from scratch on Italian language data, followed in November 2024 by Minerva 7B with 7 billion parameters. Almawave's Velvet, in the 25 billion parameter version, is optimized for 24 European languages and for regulated sectors, trained entirely on Italian infrastructure, with data not leaving national territory. Domyn's Italia is a 9 billion parameter model released with Open Weights in July 2024. FastwebMIIA is based on the NeXXt AI Factory infrastructure with 248 H100 GPUs. ASC27's Vitruvian is oriented towards critical professional and institutional contexts. Colosseum 355B completes the picture with its large-scale architecture.
A fundamental technical and political distinction runs throughout the report: most of these models are Open Weights, not Open Source in the full sense of the term. An Open Weights model makes the trained parameters public, allowing it to be inspected, downloaded, and used locally, but does not necessarily share the training data or the complete training code. For public administration and regulated sectors, this distinction is crucial: inspectability and the possibility of running the model in controlled environments are indispensable ethical and legal requirements that a closed proprietary model cannot satisfy. The open source and open weights strategy for Italian artificial intelligence, to which the report dedicates a specific section, is presented not as an ideological choice but as a strategic positioning consistent with the country's vocation towards responsible and verifiable artificial intelligence.
Law 132: Italy first in Europe
In the race to regulate artificial intelligence, Italy made an unexpected move for those accustomed to considering the country chronically behind on regulatory infrastructure. On October 10, 2025, Law 132/2025 came into force, making Italy the first member state of the European Union to adopt a comprehensive law on artificial intelligence, ahead of the full application of the European AI Act, which has been in force since August 2024 but with progressive application. The Italian law does not just anticipate it: it flanks and integrates it with provisions specific to the national context.
The report identifies three strategic pillars of particular relevance. Article 7 introduces a global unicum by imposing the obligation of inclusive artificial intelligence for people with disabilities, an accessibility requirement that no other jurisdiction had yet codified with this explicitness. Article 18 promotes public-private partnerships for cyber defense, recognizing that the security of artificial intelligence systems cannot be delegated to the public sector alone. Article 23 authorizes a 1 billion euro fund for artificial intelligence, cybersecurity, and quantum computing, with the latter emerging in the report as one of the next strategic frontiers: the convergence between artificial intelligence and quantum computing could redefine capabilities in optimization, cryptography, and scientific simulation in the country's leading sectors, and the report treats it as a technological horizon to invest in today to avoid chasing tomorrow.
The regulatory advantage can translate into a competitive advantage through the compliance-by-design mechanism: companies that develop compliant solutions from the design stage can export pre-certified products into the European single market, reducing time-to-market compared to competitors who must adapt solutions born in different regulatory contexts. However, the conditional is mandatory: the report specifies that the concrete translation of this leadership into actual advantage will depend on the effective implementation of the law. Without the implementing decrees, legal certainty remains incomplete and the competitive advantage remains on paper. At the time of publication, no other EU country appears to have approved a similar measure.
Public perception accompanies this regulatory framework with mixed signals. According to Special Eurobarometer 554 from 2024, 55% of European citizens expect artificial intelligence to have a positive impact on daily life in the next twenty years, against 35% who predict negative effects. In Italy, trust in public administration remains generally lukewarm, with only 9% of Italians declaring full trust in the PA. But a 2025 Salesforce and The European House – Ambrosetti survey of a sample of 750 Italians shows specific signs of openness for digital services: 58% are in favor of using intelligent agents to book appointments, and almost half would use them to request or renew licenses and permits. It is a willingness that is growing, albeit slowly.
In the next episode, we will conclude the analysis of the report by covering: the private ecosystem, public administration, human capital, global positioning, and the organizational challenge.