The message from Draghi, beyond the headline

Mario Draghi, inaugurating the academic year at the Politecnico di Milano, stated something very simple and very uncomfortable: with the revolution of artificial intelligence, over the past year the United States has produced about 40 large foundational models, China 15, and the European Union only 3. If we do not bridge this gap and do not adopt these technologies on a large scale, Europe risks a future of stagnation. (ANSA.it)

He added a point that should make any economy minister tremble: if AI were adopted with the same trajectory as digital technology in the United States, productivity growth could increase by just under 0.8% per year; if the impact were comparable to the electrification of the 1920s, we could exceed an additional 1% annual growth. This would be the most significant acceleration seen in Europe in decades. (Sky TG24)

Translated into very concrete terms: either AI becomes a real infrastructure of the European economy, or in the next 20–25 years we will find that our “normality” is flat growth, a structurally unmanageable public debt, and a gradual, silent shift of innovation elsewhere.

In this piece, I do not want to repeat the headline from ANSA; I want to try to answer a different question: what does it mean, in practice, to “develop AI” for Europe? And where, today, are we telling ourselves a reassuring but misleading story.

The European delay: the numbers are the symptom, not the cause

The numbers that Draghi cites – 40 US models, 15 Chinese, 3 European – are a good headline because they summarize the perception of delay. But in itself, the “number of foundation models” is not the real problem; it is just an indicator of something else:

• the critical mass of capital and talent concentrated in certain hubs;

• the ability to transform frontier research into industrial platforms;

• the availability of data, hardware infrastructure, and product pipelines that transform models into scalable services.

Research tells us that, on rapidly evolving technologies, Europe is less efficient than other areas of the world: scientific productivity measured by impact (not by the number of articles) is below the global average and significantly behind the United States. (arXiv)

If we combine this data with Draghi's framework on the “28th regime” – a truly unique internal market for innovative companies, without 27 micro-regulatory regimes that fragment everything – we see the outline of the problem: Europe does not lack brains, it lacks continuity between research, product, and market. (ANSA.it)

When Draghi speaks of stagnation, I do not read a purely technological alarm, but a structural alarm: an institutional machine that works well for regulation, poorly for enabling.

It is not a race to see who has more LLMs: the blind spot of language models

There is also a misunderstanding that, as a technician, I feel the duty to dismantle: the AI revolution does not coincide with the chatbot revolution.

General-purpose large language models (the ones we all use every day, myself included) are the most visible part of the iceberg. But the real economic transformation occurs in another category of systems, which I have long called “non-linguistic models”:

• neural networks that replace manual quality on a production line;

• artificial vision systems that analyze X-rays, CT scans, industrial images;

• models that predict energy loads, failures, maintenance;

• agents that orchestrate data and software flows in factories, hospitals, banks.

These models rarely make it to the newspapers, do not issue press releases, and are not branded as “GPT-something.” Yet they are the ones that, in a real company, change margins, eliminate waste, reduce errors and waiting times.

If Europe focuses the AI discourse on regulating generative “consumer” models and neglects the construction of a non-linguistic AI industrial base, it risks making a double mistake:

1. rigidifying the regulatory context precisely where technology is most visible but less structurally determinant for GDP;

2. substantially leaving uncovered the field where AI generates deep competitive advantages, but is less “Instagrammable.”

When Draghi says that every day the technological frontier moves further away, I think precisely of this: not the distance between a European LLM and yet another American proprietary model, but the distance between a European factory with four AI systems embedded in its processes, and a factory that continues to operate as it did ten years ago. (ANSA.it)

From “gadget” AI to “productive infrastructure” AI

In recent years, I have seen two different Europes:

• one that puts chatbots on the company website, “AI-powered” marketing campaigns, conference demos;

• one that brings AI inside the operational backbone: supply chains, predictive maintenance, risk management, quality control, dynamic pricing, energy management.

The second Europe is much smaller than the first, but it is the one that matters.

When Draghi speaks of a possible economic boom linked to AI – a +0.8% or +1% annual productivity growth – he is not talking about writing emails faster. He is talking about:

• redesigning logistics chains with models that simulate scenarios and optimize stocks, routes, times;

• automating the generation of production plans, estimates, complex offers;

• ensuring that an industrial plant “talks” in real time with an AI system that understands its drifts and anticipates problems.

This is AI as productive infrastructure. For years, the work I do with Algoretico has gone exactly in this direction: bringing AI where it generates measurable value, with models that usually do not speak, but see, classify, predict, decide.

If AI remains a cosmetic layer over old procedures, Draghi's warning is already a reality: stagnation, covered by a veil of technological rhetoric.

Private AI, European data, and the regulatory paradox

Then there is the great unspoken of the speech: how to reconcile the expansion of AI and the protection of fundamental rights, in a continent that has built part of its identity on GDPR and now on the AI Act.

Here the European position is paradoxical:

• on one hand, we have the most advanced legislation in the world on privacy and data protection;

• on the other hand, we continue to use models trained elsewhere, with data we do not control, on infrastructures we do not own.

It is like building the best road code on the planet and then only allowing buses from other countries with drivers we do not hire.

My answer to this paradox, in recent years, has been a very clear trajectory: Private AI.

This means at least three things:

1. Local models, executable on controlled infrastructures (on-premise, edge, European cloud), not just APIs towards transatlantic stacks.

2. Data that does not leave the perimeter of the organization, with techniques for anonymization, reduction of the information surface, and – where necessary – selective deletion of sensitive concepts in the models (the work I have called “Brain Surgery” for GDPR compliance). (arXiv)

3. Orchestration architectures that separate roles and responsibilities: who controls the data, who controls the models, who controls the decisions.

If Europe wants to avoid stagnation without becoming a regulated appendage of foreign big tech, the only way is this: stop seeing GDPR and the AI Act as an abstract constraint and start systematically designing native technical solutions for this context.

This is why I insist so much on concepts like “Brain Surgery” or “Private AI kit”: they are attempts to translate legal principles into replicable engineering patterns, allowing European companies to use AI without having to rely on black boxes outside the EU or on creative interpretations of the regulations.

A real European agenda for AI (seen from the field)

If I take Draghi's words seriously and cross them with what I see every day, the European agenda on AI – the real one, not the one in programmatic documents – should shift to some very concrete axes.

Standardize the “craft” of doing industrial AI

Today, every industrial AI project in Europe is almost a unique piece: different stacks, ad hoc integrations, artisanal data management. This kills scalability.

There is a need for standardization of the craft, not another monolithic platform. Shared frameworks for:

• data pipelines (ingestion, cleaning, labeling, quality control);

• MLOps cycles adapted to European SMEs, not just to big clouds;

• model governance (versioning, audit, explainability, logging).

My work on Software-Production-as-a-Service (SPaaS) and on AI agent orchestrators goes precisely in this direction: creating factories of AI-by-design software that transform business problems into iterative development cycles, with control, measures, rollback.

We do not need “another European LLM”; we need European production lines of AI solutions, robust and replicable.

Shift the center of gravity to the edge: Edge AI as a competitive factor

A good part of the truly transformative use cases do not happen in the cloud, but at the edge of the network: in an industrial machine, on an assembly line, inside a clinic, in an energy control system.

Here Europe has a chance: strong manufacturing tradition, strong hardware base, evolved SME ecosystems.

Working on Edge AI – small optimized, compressed, distilled models to run close to the data – means:

• reducing dependence on extra-EU infrastructures;

• improving privacy and security;

• reducing latency on critical applications (healthcare, energy, transport).

Projects like those I am developing with “AI-ready” boards and local kits go exactly in this direction: bringing cognitive capability where things happen, not just in a distant data center.

Invest less in “generic AI,” more in domain-specific AI

Another widespread illusion is that a “large general model” applicable everywhere is sufficient. Experience says otherwise: the best results come from domain models, trained on specific data, designed with the close collaboration of industry experts.

This is the spirit behind works like Adaptive Meta-Domain Transfer Learning (AMDTL): transferring knowledge between different domains without destroying specializations, reducing negative transfer and catastrophic forgetting. (ANSA.it)

For Europe, this means: less mantra on “AI for everything,” more targeted investments in:

• AI for advanced manufacturing;

• AI for health (diagnostics, therapeutic pathways, clinical research);

• AI for energy and transition;

• AI for logistics and the continental supply chain.

Truly innovate: research, not just implementation

Draghi's speech fits into a context where many countries are equipping themselves with high-level scientific reports on the safety and development of AI, such as the International Scientific Report on the Safety of Advanced AI, which aggregates global expertise to address systemic risks and governance. (arXiv)

If Europe wants to avoid stagnation, it cannot limit itself to importing models and rules, it must contribute to new architectures and new methodologies:

• techniques for concept erasure to make models compatible with the right to be forgotten and data minimization principles;

• new paradigms of adaptive transfer learning (like AMDTL) to address contexts where labeled data are few, fragmented, sensitive;

• frameworks of agentic AI that do not limit themselves to composing APIs, but define structures of responsibility, accountability, and human-in-the-loop control.

Here there is, for Europe, an enormous window: to bring together its legal tradition, its engineering culture, and its industrial heritage to invent new ways of doing AI, not just to “catch up” with others.

Stagnation or leap of phase: what really depends on us

Draghi's warning is deliberately dramatic: stagnation or boom. Personally, I do not believe in binary scenarios, but I recognize the truth that lies beneath that polarization.

If nothing structural happens, the scenario is quite clear:

• a little more office automation, a few more chatbots, a few rebranded analytics platforms;

• very modest real growth;

• increasing dependence on extra-EU technological stacks;

• a gradual, silent slipping towards the margins of value chains.

If, on the other hand, we take seriously the idea that developing AI in Europe means:

• building robust, sovereign, and compliant Private AI;

• shifting attention to non-linguistic models that truly change production;

• standardizing the “craft” of doing industrial AI, not just celebrating its outcomes;

• embedding European research in new architectures, not just in implementations,

then the future is not written.

In this sense, I share the heart of Draghi's message, but I read it this way:

we do not need “a European AI” as a slogan: we need a Europe that makes AI its way of working, producing, caring, designing, while respecting rights as it truly innovates.

If we do not do this, stagnation will not be a surprise but a logical consequence. If we do, AI will not be the cause of our fear, but the powerful and delicate tool with which to redesign the pact between technology, economy, and people on this continent.

And this, beyond the headlines, is a choice that depends on us.