Italy stands to benefit from open-source AI models

The letter published on 24 July 2026 by Nvidia’s CEO, Jensen Huang, and co-signed by Microsoft, Meta, Hugging Face, IBM, Mistral and twenty other signatories, sent shockwaves through the entire digital technology sector. Entitled ‘Open Weights and American AI Leadership’, the document calls on Washington not to impose premature restrictions on downloadable artificial intelligence models. Twenty-four hours later, the number of signatories had risen to fifty, with the addition of OpenAI and Google, which had not appeared on the initial list. Even before taking a firm stance in the ideological battle between ‘open’ and ‘closed’ AI models, Huang makes a very simple point: in artificial intelligence, strategic advantage does not depend solely on who possesses the most powerful system, but on who builds the ecosystem upon which others choose to innovate. It is a difference in terminology that signals a difference in strategy. Europe and Italy would do well to take this seriously.

This letter is not a libertarian manifesto, nor is it an act of corporate philanthropy. It is an economic policy document written by individuals with very concrete interests (and associated conflicts). But it is precisely for this reason that it deserves attention. When the leading players in the American AI sector, amidst fierce competition with China, argue that ‘open weight’ models – that is, models that can be downloaded, inspected, adapted and run on one’s own infrastructure – are a decisive component of technological leadership, they are saying that the AI race will not be won solely by cutting-edge laboratories. Above all, it will be won by those who make it easier to roll out AI across ordinary sectors of the economy.

It is important to be precise here. ‘Open weight’ does not automatically mean ‘open source’. In public debate, the two terms are often used interchangeably, but under European law they are not the same. The AI Act distinguishes between general-purpose systems and models and recognises specific exceptions only under certain conditions; for general-purpose models, even if they are open source, obligations regarding copyright and the synthesis of training data remain, and these are stricter where systemic risks are present. We must therefore stop treating ‘open weights’ and ‘regulatory free zone’ as synonyms. That said, Huang’s point remains valid. If the weights are available, innovation is not confined to a handful of proprietary APIs. Universities, start-ups, businesses, hospitals, factories and public administrations can adapt a model to their own domain, run it in their own environments and use their own data. Lock-in is reduced, and competition is increased. And for a Europe that is unlikely to match the budgets of American frontier labs or China’s aggressiveness in the short term, this is good news.

Europe’s real interest does not lie in training the world’s largest model from scratch every time. It lies in entering the most economically attractive part of the value chain: fine-tuning, industrial integration, on-premises deployment, vertical applications, AI for manufacturing, healthcare, professional services, supply chain software and public administration. In other words: utilising the division of labour rather than dreaming of costly digital autarky. Open weights make this participation far more plausible, because they allow even non-giant players to create value without having to constantly seek permission from the API owner.

But here we must avoid the opposite ideological shortcut. Defending open weights does not mean denying the risks. The letter itself acknowledges this: once released, the weights escape their original control, and modified versions may be difficult to trace or revoke. The European Commission, in its guidelines on general-purpose AI models, emphasises that open-sourcing the most advanced models can make it easier to circumvent or remove risk mitigations. The question to ask is whether the response should be a blanket crackdown on open models or a more targeted regime focusing on risks and genuinely sensitive uses.

This is where much of the European criticism misses the mark. Reducing AI to an object to be classified, documented and brought into compliance as soon as possible is an understandable, yet insufficient, bureaucratic temptation. The AI Act is not as crude as its critics sometimes portray it: it distinguishes between models and systems, provides for exceptions, sandboxes, codes of conduct and phased implementation. The problem lies in the culture in which it will be implemented: in Europe, and often in Italy, any margin of discretion risks becoming a new administrative burden, any grey area a precautionary brake, and any risk a reason to extend ex ante control.

The most sensible approach is not to ‘do away with the rules’, but rather to avoid blanket restrictions on open-weight models; stricter requirements and more thorough checks are, however, legitimate when dealing with high-risk cases, essential services or contexts in which automated output could cause systematic and potentially irreversible harm.

For Italy, the implication should be quite clear. There is no reason to add layers of domestic caution to an already complex European framework. The state’s role is not to prevent open weights from circulating. It is to distinguish where strong caution is required and where, instead, there should be scope for competition, testing, sandboxes and sector-specific adaptation.

What is interesting for Italy, however, lies in a historical pattern regarding general-purpose technologies: they do not generate value where they are invented, but where they are adapted. Italy will not be developing cutting-edge models, and that is no reason to be disheartened. General-purpose technologies are monetised through co-invention, and co-invention requires precisely those qualities in which Italian specialisation is competitive: small markets, many vertical niches, and margins protected by tacit knowledge and long-standing customer relationships. This is the opposite profile to that needed to build a generalist model, and it is the right profile on which to build.

The most obvious example is the instrument-making industry. Italy is one of the world’s leading exporters of machine tools and equipment for packaging, bottling, paper and food processing, and every machine sold generates a stream of data that is currently largely lost. A small model, trained on faults specific to that particular family of machines, which runs inside the electrical cabinet and flags up a bearing before it fails, is worth more than a generalist system that knows everything and understands nothing. And no serious customer accepts that their process data should pass through a third party: local execution is an industrial requirement. It is also the path by which a machine manufacturer can become a provider of recurring services, making that transition from sales to subscription fees that the Italian engineering sector has been pursuing for years.

The second scenario concerns industrial districts: many small businesses, distributed knowledge, and a technical lexicon not found in manuals. None of the companies in Belluno, Arzignano, the Cusio area or the Emilia-Romagna packaging sector can afford an AI department; a consortium, however, can afford to fund training. This is the case with the Ucima project for AI in the packaging supply chain: a shared resource, paid for once and used by all.

The third point concerns language and documents. General-purpose models speak Italian well but understand it less well in the contexts that really matter: tender specifications, clinical reports, court judgements, safety data sheets and technical standards. Adapting an open-source model to these corpora is within the capabilities of a university, a professional practice or an association; retraining one from scratch is not. The same applies to healthcare and public administration, where data restrictions often make local processing essential. And it applies – with a comparative advantage that remains largely untapped – to archives and cultural heritage.

For once, the maths is spot on. IT4LIA, the Italian AI Factory hosted at the Bologna Technopole and managed by Cineca, has been providing services since April 2025 using the Leonardo supercomputer, and has a dedicated system worth 290 million euros in place, co-funded in equal parts by EuroHPC and the Italian government, with stated focus areas including manufacturing, agritech, cybersecurity and meteorology. Access is free for small and medium-sized enterprises, start-ups, research organisations and public administrations. The bottleneck, therefore, is not the infrastructure: it is the skills required to use it and an industrial policy that continues to prioritise the purchase of hardware rather than the adaptation of software that makes it more effective. A tax credit that funds the lathe rather than the two hundred hours of specialist training that make it reliable is investing in the least valuable part of the transition.

There is a tendency in Italy and across Europe to believe that sovereignty is built by controlling everything. But in general technology, true sovereignty is built by choosing wisely where to position oneself in the value chain, under what rules and at what pace. For years we have discussed ‘human in the loop’. Huang’s letter proposes something different: ‘market in the loop’ – control derived from a distributed ecosystem. This does not mean eliminating accountability and safeguards in high-risk contexts. It means recognising that the best process for collective discovery remains the competitive market. Read without naivety but also without Pavlovian reflexes, Huang is saying this: AI can spread productive capacity or become a new fiefdom administered by a few operators and numerous compliance modules. If Europe and Italy wish to avoid the latter outcome, they must not worship open platforms. They must do something more difficult: stop being afraid of them.