Beyond Emma: Europe and Italy are actually building their own AI sovereignty

In a handful of days a small chatbot drew the irony of half the internet. Emma, presented as "the Italian artificial intelligence", answered that dogs can fly and that drinking acid is good for you, and it was pulled offline amid the memes. From there the public conversation slid toward a comfortable, false conclusion: the idea that Italy is unable to do AI. That story closes on itself, and the facts contradict it.

It is worth telling what actually happened — once, with precision — and then turning the page. Because the story that matters is a different one: everything that, in the same month, Italy and Europe put on the field for real.

European technological sovereignty is already under construction: a decisive chapter is led by the Italian company Domyn, with the open model of the EUROPA consortium.

What happened with Emma, briefly

Two distinct things. The first is technical. Emma is a model of roughly 550 million parameters, with a context window of a few thousand tokens and with its safety alignment still switched off. It belongs to the class of experimental models from a few years ago, an order of magnitude smaller than the open models that run on a laptop today. For that size, it does exactly what it can do. Its own model card described it as experimental and ruled out critical uses.

The second is about communication, and it is the one that weighs. An experimental prototype reached the public wrapped in the language of "critical infrastructure for the economic, cultural and democratic future" of the country. The gap between the product and the story did the rest: the public judged "the Italian AI", rather than a lab experiment, and every stumble became the failure of a national ambition.

The lesson, said with respect for everyone, is a single one: the framing weighs more than the model. The very same artifact, called "an experiment", would have passed unnoticed; called "national sovereignty", it became a case. It is a problem of positioning, ahead of engineering — and a mistake is something to learn from, ahead of mocking it.

Why a sovereign model is a hard build

That said, a technical truth explains why "building sovereign AI" stays an exceptional undertaking: at the frontier it is mainly a question of capital and compute. Training a competitive model means thousands of last-generation GPUs and investments on the order of a billion. That is a scale very few players in the world can afford, and one the European private market alone struggles to assemble.

To this add a real gap, worth acknowledging honestly: today European and open models shine on linguistic and document tasks, while the more complex autonomous workflows — the "agents" that use tools and chain several steps — remain a terrain where the big international models keep an edge. Acknowledging it is realism, and it is the premise for playing the right game well.

What Italy and Europe are actually doing

And here is the right game. While the Emma memes circulated, on 19 June 2026 the European Commission chose the EUROPA consortium, led by the Italian company Domyn (formerly iGenius, born in Milan), as the sole winner of the Frontier AI Grand Challenge. The mandate: to build from scratch an open-source model of more than 400 billion parameters, reproducible and deployable for free by businesses and governments, covering all 24 official languages of the Union, compliant with European rules by design. Alongside Domyn, heavyweight partners such as Germany's Fraunhofer and technology collaborations with the main compute providers.

In plain terms: the sovereign foundation model that many dream of building in-house, someone else is already building — with European public legitimacy, and with a commitment to keep it open and free. That is the news. And it carries an Italian signature.

Domyn is the tip of a broader and serious map:

Minerva, the first large language model trained from scratch in Italian, born at Sapienza University on the Cineca-Leonardo supercomputer: open, and able to beat open competitors of the same size on Italian-language benchmarks.
Almawave, a listed Italian group, with its Velvet model family, also trained on Leonardo and aimed at the enterprise.
Fastweb, which trained its own model (MIIA) on infrastructure located on Italian soil.
Dedagroup with Istella, the Italian-data search and generative-AI engine it entered with a 39% stake, in a vertical-integration play: technologies already serving institutions such as Treccani, the Italian Senate, Infocamere and Cerved.

The honest picture: on foundation models we are looking at serious industrial bets, still seeking commercial confirmation. The Italian successes already proven, instead, live in the application and integration layer — where Italian skills are recognized and sold across half of Europe. It is concrete excellence, ahead of a wish.

The real money on the table, and how to read it

On the EUROPA figure precision matters, because a misunderstanding hides here. The European prize is made of compute, ahead of cash: a strategic access of up to 2.5% of the overall European supercomputing capacity (EuroHPC) for one year, on public systems such as Leonardo and Jupiter, plus a roughly 6,000-chip NVIDIA Blackwell cluster in development. Europe lends the supercomputers to take the risk off whoever builds the model. At this scale, the compute cost covered alone is worth tens of millions.

The "real" capital is private and ambitious: Domyn has already raised hundreds of millions and aims to mobilize on the order of a billion, with the declared goal of building one of Europe's largest AI supercomputers right in Italy. And the public context is imposing: the EuroHPC programme is worth around 10 billion over the seven-year period, while the InvestAI initiative aims to mobilize around 200 billion by 2030, with 20 billion earmarked for four-to-five European AI "gigafactories".

It is a precise industrial-policy model — the public puts in infrastructure and legitimacy, private capital puts in the risk money — and it says one thing, clearly: Europe is on the field.

"Sovereignty" is one word for four different things

Here lies the heart of the matter, and it is also where the Emma case muddied the waters. "Sovereign" describes four different things, and keeping them distinct is what separates seriousness from a slogan:

Model sovereignty — training the foundation model from scratch. It is the terrain of billions and thousands of GPUs: Domyn/EUROPA and the big European labs.
Infrastructure sovereignty — data centers and supercomputing, from the EuroHPC plan to the gigafactories.
Deployment sovereignty — running open models on European cloud or your own servers, with data residency, governance and auditability.
Vertical-application sovereignty — building agents and systems on a sector's or a client's own data, with the rules and controls of the domain.

Emma's mistake was presenting itself at the first level while staying, in fact, an experiment. Maturity, by contrast, lies in knowing which level you play on — and playing it well.

Synthos Logic's position: sovereign pragmatism, beyond the flag

As observers of the sector, the reading we propose is simple and grounded in the facts. For the vast majority of European companies and institutions the right direction lies in using open European models — and extra-European ones too — specialized for verticals, run on local machines or on European cloud when the data is sensitive, ahead of chasing a foundation model of one's own. On many tasks dominated by the Italian language and domain, a smaller specialized model holds its own against the giants; and beyond a certain volume threshold, running in-house becomes cheaper too.

With three firm points, which keep us at a distance both from uncritical enthusiasm and from denialism.

The first: the model itself is turning into a common good. Open weights spread at great speed, and with EUROPA a 400-billion-parameter model will soon be available for free. The value lies in the vertical data, the workflows, the controls and the trust built around it, ahead of owning "your own model" — that remains Emma's trap on a smaller scale.

The second: "European, open and local" works as a compass, ahead of a dogma. The serious answer is hybrid, chosen task by task based on the sensitivity of the data and the capabilities required. Imposing "all-European" even where an alternative is simply better hands the client a worse product — and at that point the flag, on its own, counts for little.

The third, the most important and the most human: the rules of language matter. The Emma case, ahead of being a question of parameters, is a question of words. A system that speaks with people carries a responsibility — over safety, over the truth of what it states, over the limits it declares. The sovereignty worth building takes charge of these rules, ahead of merely making noise. And it is exactly the ground Europe, with its rules and its investments, chose to move on first.

Italy, in all this, is much more than the country of a botched communication. It is the country leading the largest European public project for open AI, training models in its own universities, selling integration and expertise across half of Europe. Emma is a footnote. The book is being written by others, and it speaks Italian.