What Emma Got Right About Sovereignty, and Why Italy Should Be Embarrassed Anyway

Egomnia pulled its Italian chatbot Emma-5 offline on 25 June 2026, five days after launch, after tens of thousands of users spent a weekend asking it simple questions and watching it fail them. A model built to prove Italy could field its own artificial intelligence (AI) could not reliably say whether a kilogram of bread weighs more than a kilogram of feathers. It does not, and Emma said it does.

Being Italian, I am not happy to write what follows. I think the instinct behind Emma was correct. I think the embarrassment around it was deserved. I do not think those two things contradict each other, and the distance between them is the actual subject of this piece.

Egomnia is not a startup operating in obscurity. The company is listed on Euronext Growth Milan, founded by Matteo Achilli, an entrepreneur whose earlier venture, a recruitment platform connecting graduates with employers, earned him a nickname from the BBC and a feature film about his career. Achilli named the model after his daughter and launched it with a manifesto describing artificial intelligence as critical infrastructure for a nation's economic, cultural, and democratic future.

Emma-5, the version the public actually tested, runs on roughly 550 million parameters and a 2,048-token context window, a scale closer to an educational exercise than to a generalist assistant capable of carrying that description. The mismatch between the language and the product is where this story actually starts.

The instinct nobody should mock

Start with the part that is easy to get wrong if the appetite is to laugh and move on. Wanting a domestic AI capability, hosted on national infrastructure, governed by national and European law rather than a foreign company's terms of service, is not a vanity project. It is closer to a baseline requirement for any country that intends to keep a say in how AI gets built, deployed, and restricted.

The argument for that baseline has stopped being theoretical. A few days before Emma's launch, the United States Commerce Department ordered Anthropic to suspend its Fable 5 and Mythos 5 models, citing a national security concern it never fully explained in writing. Anthropic's own statement describes a directive that arrived as a letter, with no prior technical notice, no independent verification, and consequences severe enough that the company had no compliant option except a global shutdown.

Mythos 5 has since been granted a narrow, partial return, restricted to a defined set of critical infrastructure operators inside the US, while Fable 5 stays dark for everyone else, with no date attached and no public account of what changed.

Sit with what that means for a company, a hospital system, or a government office that had built a workflow around Fable 5. None of those organisations did anything wrong. None of them had any input into the decision that took their tool away. They learned, in the space of a single news cycle, that the capability they depend on can be switched off by a government they do not vote for, acting on a justification they will likely never see in full.

That is not an abstract governance debate. It is the precise, concrete risk that "sovereignty" is supposed to address: the gap between using a technology and controlling whether you keep using it. Once a government demonstrates it is willing to pull a frontier model offline worldwide over a national security claim it will not fully substantiate, every other government watching has a rational reason to ask what happens if the same call gets made about a tool their own economy depends on.

Wanting an answer to that question is not paranoia. So yes, it is now effectively mandatory for any serious country to attempt some version of sovereign AI capability. Not because sovereignty is fashionable, but because the alternative is permanent dependence on a decision-making process you cannot see and cannot appeal.

Emma was not wrong to exist for this reason. Where the project went wrong sits somewhere else entirely.

The shock that should not have been a shock

Here is the part I find genuinely difficult to say without sounding harsher than I mean to be. The Fable 5 suspension landed as a jolt. Commentators who had spent time treating AI policy as a distant American argument suddenly discovered that owning the technology, not merely licensing access to it, was the entire point.

That reaction was sincere. It was also, in my view, something close to embarrassing, because nothing about the structural risk was new or hidden. It was sitting in plain sight for anyone willing to look at the shape of the industry rather than wait for an announcement to do the looking for them.

A government does not need a single dramatic event to understand that AI has become geopolitical. It needs to notice that the same handful of companies supply the underlying compute, the same handful of governments hold leverage over those companies, and the competitive dynamic that has shaped trade policy, export controls, and industrial strategy throughout the post-war era was always going to extend to the technology everyone agrees will reshape the economy fastest. That observation does not require inside information.

It requires watching the world, which is a different and far less impressive achievement than reacting correctly once the world hands you a headline. The pattern was visible well before Fable 5. Frontier AI development has concentrated inside a handful of American and Chinese companies since the current generation of large models emerged, each backed by governments that have shown, repeatedly and across unrelated industries, a willingness to use export controls, procurement rules, and licensing leverage to keep strategic technology inside trusted hands.

Anyone who had watched how Washington treated semiconductor exports to China, or how Beijing treated foreign cloud providers operating on its soil, already had the entire template for what eventually happened to Fable 5. The model changed. The underlying behaviour of the two governments capable of building it did not.

A few days before Fable 5 went dark, the United States quietly told its own AI evaluation agency to stop publishing comparisons between American and Chinese models, and separately required OpenAI to route GPT-5.6's early access through customer-by-customer government approval. None of that is a story about a single company behaving badly. It is a pattern, applied consistently, regardless of which lab reaches the capability tier that triggers the response.

A government willing to gate access to its own companies' best products, on national security grounds it controls the definition of, was never going to behave differently toward a smaller country's hypothetical future model. The lesson was available well before Emma, well before Fable 5, to anyone reading the shape of the industry rather than waiting for a single dramatic episode to supply the conclusion for them.

I raise this not to relitigate the shock, but because the shock itself explains a great deal about why Emma launched the way it did. A government and a public that arrive at a strategic conclusion through sudden alarm, rather than through patient observation, tend to reward the appearance of having an answer over the substance of one. Egomnia's chief executive Matteo Achilli called artificial intelligence critical infrastructure for a nation's economic, cultural, and democratic future, language that borrows directly from the urgency the Fable 5 episode created.

The urgency was earned. The model attached to it was not built to carry that weight, and a market primed to want a sovereignty answer immediately was always going to accept whatever arrived first, regardless of whether it could deliver.

There is a more uncomfortable version of this same point, and I will state it plainly rather than hedge around it. The people most enthusiastic about Emma in its first hours, the ones sharing the launch manifesto rather than the screenshots, were largely reacting to the word sovereignty itself, not to anything the model had actually demonstrated. That is precisely the failure mode a sudden, alarm-driven conversion to a cause tends to produce.

When the realisation that something matters arrives all at once, the appetite is for a symbol that confirms the realisation was correct, not for the slower, harder evaluation of whether the specific thing being offered can do what the symbol promises. Italy wanted proof that it had understood the moment. Emma was the first thing that looked like proof, and looking like proof and being proof are not the same achievement, a distinction the following five days made impossible to ignore.

The same instinct, applied to money instead of models

Sovereignty as a working principle is not confined to chatbots. It is currently reshaping how capital moves through European crypto markets, in a story that has nothing to do with Italy and everything to do with the same underlying logic Emma was reaching for.

Binance, the largest exchange, applied for a Markets in Crypto-Assets (MiCA) licence through Greece, and by most accounts the application had cleared its formal review process without objection. According to reporting that neither the European Central Bank (ECB) nor Greek authorities have confirmed, ECB President Christine Lagarde signalled directly to Greece's Prime Minister that Binance was not welcome in Europe, a message that reportedly reversed a more favourable position the country's own Finance Minister had held.

The motivation attributed to that intervention is specific. Binance dominates stablecoin liquidity across Europe, and that scale is said to complicate the ECB's own ambitions for a digital euro, with policymakers reportedly preferring that established banks, not a large foreign-controlled exchange, manage the equivalent flows.

Whether or not that account is precise in every detail, the shape of it should be unmistakable to anyone who has just read several paragraphs about Anthropic. A central bank, worried about a foreign-controlled platform holding too much influence over how money moves inside its currency zone, used the leverage available to it to keep that platform out, citing concerns it has not been required to fully substantiate in public. That is the same instinct driving Washington's posture toward Fable 5, pointed at a different layer of the stack.

One case concerns who gets to use the most capable AI model. The other concerns who gets to clear the most liquid stablecoin flows inside the eurozone. Both are sovereignty disputes, dressed in the specific vocabulary of their own domain, and both follow the same underlying rule: when a foreign actor accumulates enough structural leverage over something a government considers strategically essential, the government finds a way to act on that leverage, formal authority or not.

I find this comparison more unsettling than reassuring, and I want to say so directly. If an institution with Lagarde's standing and the ECB's actual regulatory weight is willing to intervene informally, outside its own formal competence, to protect monetary sovereignty over stablecoin flows, then the instinct Italy displayed toward Emma was not an overreaction invented out of nowhere. It was a smaller, weaker version of exactly the calculation a serious European institution was already making about a different layer of the same technology stack.

The difference is not that Italy's instinct was wrong. The difference is that Egomnia tried to express that instinct through a product launch, while the ECB expressed the identical instinct through actual regulatory leverage, and only one of those two approaches had any chance of working.

Read against that backdrop, Emma stops looking like an isolated Italian misstep and starts looking like a small, badly executed instance of a logic that the ECB, the US Commerce Department, and a 550-million-parameter Italian chatbot are all, in their own register, pursuing at the same time. The difference between Lagarde's intervention and Egomnia's product launch is the gap between an institution with real leverage to apply and a company with a manifesto but no comparable capability behind it.

Sovereign in name, foreign at the foundation

The United Kingdom offers the clearest illustration of how far the gap between sovereignty language and sovereignty substance can stretch even when a government does almost everything right.

A year after that pledge, the country can point to genuinely substantial results: up to £11 billion committed to domestic data centre capacity, 200,000 UK developers enrolled in NVIDIA's developer programme, AI certification embedded directly into government-funded apprenticeships, and a sovereign fund backing domestic AI startups built on the country's most powerful computer. None of that is theatre.

It is the kind of patient, well-funded, multi-year build that a serious sovereignty strategy actually requires, and it is precisely the category of effort Emma skipped entirely in favour of a press release.

And yet the word doing the heaviest lifting in that entire programme is "sovereign," attached to infrastructure whose most critical layer, the chips themselves, comes from a single American company. The UK's own most powerful computer runs on NVIDIA Grace Hopper superchips. The training, the certification, the startup fund, all of it sits on top of hardware the country does not manufacture and cannot replace on short notice from a domestic alternative.

Sovereignty, in this case, means controlling where the computer physically sits and who gets access to it. It does not yet mean controlling what the computer is made of.

I do not raise this to suggest the UK's effort is worthless, because it plainly is not, and the contrast with Emma is instructive precisely because the UK got the hard part right. Even a government spending billions of pounds, building real talent pipelines, and producing measurable outcomes still cannot escape dependence on the one supplier whose silicon underwrites more than 90% of the continent's AI computing capacity. If a well-resourced, methodical, multi-year national programme cannot fully close that gap, a single Italian company launching a laptop-scale chatbot with a press release was never going to close it either.

The lesson is not that sovereignty is impossible. It is that sovereignty is layered, that owning one layer while depending entirely on another is still progress, and that nobody gets to claim the word until they can say plainly which layer they actually control.

The apprenticeship pathway built into that programme makes the layered point even sharper. Embedding NVIDIA's developer credentials inside government-funded vocational training is a genuine policy achievement, since it routes a serious technical skillset through the same national apprenticeship structure that has historically trained electricians and accountants, available to school leavers and career changers without a university degree standing in the way. That is sovereignty over talent, a layer few governments bother building deliberately, and the UK built it on purpose.

But talent trained to operate someone else's hardware is still talent operating someone else's hardware. The country can produce as many certified engineers as it wants and still find itself negotiating, at the next renewal cycle, with the same single supplier holding the same pricing power it held before any of those engineers existed. Emma never reached the stage where this distinction would even apply, because Emma never built a layer worth defending in the first place.

The number Italy keeps skipping past

Here is where the irony becomes difficult to ignore.

Europe's record supercomputer expansion tells a totally different story: 35 new systems across 23 countries, reaching more than 3 million researchers once fully operational. Inside that expansion, Italy's IT4LIA is the largest single deployment in the entire announcement, built through a partnership between the EuroHPC Joint Undertaking, Italy's national supercomputing consortium CINECA, the Ministry of University and Research, and the Italian Cybersecurity Agency, delivering more than 8,000 graphics processing units (GPUs) and a combined 82 exaflops of AI training performance.

CINECA's own language around the project describes it as strengthening the country's technological autonomy, the same phrase Egomnia reached for to describe a model roughly fifteen thousand times smaller in parameter count than the frontier systems IT4LIA is built to train.

Italy did not need Emma to have a credible claim to taking AI sovereignty seriously. It already has one, funded at industrial scale, built through the same European coordination mechanism the UK's own infrastructure relies on, and operating at a tier where the conversation about national capability actually means something. That programme will not produce a viral chatbot any time soon, because exaflop-scale training infrastructure is not designed to.

It is designed to let researchers, companies, and public institutions train and run models that can do something IT4LIA's own scale makes obvious: that the difference between a sovereignty claim and a sovereignty capability is measured in dollars and engineering hours, not in how confidently a company writes its launch manifesto.

This is the detail that makes Emma's failure sting in a specific way for anyone who actually looked. Italy was not absent from the serious version of this race. It picked the wrong project to put its name on publicly, while the project that should have carried that name was already running, mostly uncovered, one floor down. But this is what happens when you chase press releases.

Achilli's manifesto language about sovereignty as critical national infrastructure would have described IT4LIA accurately. Applied to Emma-5, it described a press release.

Even the giants cannot agree on how to play this

It would be a comfortable ending to say that the small countries are confused and the large ones have it figured out. They do not, and the clearest evidence sits in what happened, or rather did not happen, when the two countries that actually dominate frontier AI sat down together.

The Beijing summit in May 2026 put AI governance formally on the agenda, with Nvidia's Jensen Huang travelling on the presidential plane. Stanford University's own AI Index had measured the performance gap between American and Chinese frontier models at 2.7%, down from a range of 17.5 to 31.6 percentage points just three years earlier. The two countries best positioned to write a shared framework for how frontier AI gets evaluated before reaching the public sat across from each other with that gap nearly closed, and left with no agreement at all.

The reason is structural, not personal. The American system, run through the Center for AI Standards and Innovation (CAISI), gives federal evaluators voluntary, non-binding access to models from five major labs, with no legal mechanism to act on what those evaluations find and no power to stop a company that decides to walk away from the arrangement.

China's Cyberspace Administration runs the opposite system: mandatory government review before public release, carrying real legal force, with roughly 350 large language models already filed under it as of last year. Two governments cannot sign a binding bilateral agreement when one side's domestic commitment is voluntary and the other's is compulsory. There was nothing to reconcile in Beijing, because the two systems were never built to be compatible in the first place.

Notice what this means for the entire argument this piece has been building. Sovereignty is not a contest with a finish line where the winners compare notes and agree on the rules. Sovereignty is a live, ongoing structural conflict even between the two actors with the most resources, the most leverage, and the clearest incentive to find common ground.

If Washington and Beijing cannot agree on how to govern the technology they each lead in producing, no smaller country should expect the rules of this race to be handed down from above, settled, or stable. Italy, the UK, every EU member state building inside the EuroHPC programme, every government watching Lagarde's intervention in Greece, all of them are playing a version of a game the two largest players have not finished negotiating between themselves.

That is, in the end, the most useful way to read Emma's five days online and its quiet suspension. Not as a national embarrassment to laugh off or defend, but as the smallest, most exposed entry point into a contest playing out simultaneously at every scale: between nations over who controls frontier models, between a central bank and an exchange over who clears digital euros, between a sovereign fund and a single chip supplier over what "sovereign" can honestly mean, and between the two countries with the resources to set the terms for everyone else, who still could not agree on terms for themselves.

Italy is not wrong to want a seat at that table. Emma simply proved that wanting a seat and building a chair are not, yet, the same project, and the country's actual answer to that problem, sitting inside IT4LIA's exaflops rather than inside a viral screenshot, deserved the manifesto far more than the chatbot that borrowed it.

This sector sits at the centre of fintech more directly than most outside it realise, since the same sovereignty contest now running through AI policy is the one already reshaping who controls the infrastructure behind digital payments, stablecoins, and the next generation of financial services built on top of both.

Editor's note

Every piece published on The Bright Minded goes through careful verification, but mistakes can happen. Readers who spot an error or have additional information can write to rosalia@thebrightminded.com.