intelligenza artificiale

AI sovereignty: the ability to walk away keeps you safe, says WEF – Francesca Zanolla

On June 12, a US government directive pushed Anthropic to switch off its two most powerful models worldwide; five months on, the World Economic Forum is using that episode as a stress test for companies that run on someone else’s artificial intelligence. Francesca Zanolla, AI Competitiveness Manager at the WEF’s Centre for AI Excellence, discussed it in an interview with Digitalic and the DigitMondo podcast. Her starting point is an observation about the C-suite: “Almost none of the CEOs we work with ask us to focus on end-to-end independence across the stack.” What they ask for is control: over their data, over their suppliers, and over their ability to keep operating whatever regulatory or market shift comes along. The practical upshot, as the Forum sees it, is that sovereignty is measured by your ability to walk away from a vendor.

Francesca Zanolla: “What the Anthropic case taught us”

We have reconstructed the events on Digitalic: on June 9 Anthropic launched Claude Fable 5 and Mythos 5; three days later the US Department of Commerce, through an export control directive, ordered it to suspend access to both models for any foreign national, inside and outside the United States. The restriction was so broad that the company chose to switch them off for everyone; the order was withdrawn on June 30, and Fable 5 came back online on July 1. As for what it meant for European companies, we had already asked at the time whether Washington could pull the plug on corporate AI.

For Zanolla, the episode gave concrete shape to a risk that had been discussed for some time. “It made people aware that export controls don’t apply only to chips, not only to physical components, but can extend all the way to models, to every cloud-based AI service,” she explains; cutting off those services, she adds, “can happen immediately and on a massive scale.”

The most useful part for anyone running IT systems is who held up. From its conversations with the companies it works with, the Forum has drawn a clear picture: organizations that had built model-agnostic layers were able to shift workloads to other models and keep running; those tied to a single provider, with no alternatives in place, “took the hit.” What made the difference was the layer separating the model from the applications: those who could swap out the engine switched tracks; those who had wired it straight into their processes went down with it.

A definition that is still missing

The WEF’s work on the subject began almost two years ago with an absence. “The absence of a shared definition is itself part of the problem,” says Zanolla. The first attempt to fill the gap is the report Rethinking AI Sovereignty, published with Bain & Company on January 20, 2026, during the Annual Meeting in Davos. It defines AI sovereignty as the ability to shape, deploy and govern AI ecosystems in line with one’s own values, while retaining strategic and operational control, flexibility and resilience.

The report acknowledges that the United States and China are present in every link of the AI value chain, with a full-stack approach few others can afford; for most countries, Zanolla argues, copying it “is not only unthinkable, it isn’t even desirable.” The formula the document settles on is strategic interdependence: targeted local investment combined with trusted international partnerships.

That first strand, aimed at national economies, has since been joined by a second one, again with Bain, focused on the individual company; according to Zanolla, it is the more relevant of the two from a business standpoint today.

Follow the rule or redesign the architecture

Zanolla names no names, but she offers two cases that illustrate two possible responses to sovereignty constraints. The first involves a global carmaker operating in China, where local rules require vehicle usage data, including images of the surroundings captured by the cars, to stay in the country. Its answer was a hybrid deployment: fleet data on local infrastructure, the rest of the business run as a global operator. The rule is followed to the letter.

In the second case, the constraint forced a rethink of the whole system. A global pharmaceutical group needed to work with patient data held by several institutions; by law, that data could be neither moved nor pooled. “The company decided to bring the model to the data,” Zanolla recounts, turning to federated learning: each institution trained the model locally, sharing only what the model had learned. The data stayed where it was; the model learned from all of it.

For Zanolla, these examples show that a new balance can be struck without disrupting operations, provided companies invest in innovation and in a very deep understanding of the technology. Similar cases exist among smaller companies too, though for them the costs can weigh far more heavily.

The exit test

Strategic interdependence can sound like a more elegant word for dependence. Zanolla answers with a sharp distinction: the difference lies in the ability to choose, because “dependence essentially means you can’t leave.” Being strategically interdependent means forging targeted, deliberate partnerships, securing specific guarantees and keeping a credible way out, like the one that allowed some companies to ride out the June episode.

The test, as she frames it, is a practical one: what are the contractual exit rights; how does data return work; has workload migration been approved in advance; who keeps the code, the documentation and the critical components if the relationship breaks down. A dependency is strategic, she concludes, if it meets two criteria: the company is aware of it, and the exit is designed from day one, rather than waiting for a crisis to ask how much it depends on others, and on whom.

Frank Karlitschek, founder of Nextcloud, had reached the same conclusion from a very different vantage point, that of a European open source provider: in his DigitMondo episode, which we drew on when reporting on Christian Fu Müller joining Nextcloud, sovereignty meant the ability to change your mind. When the organization that convenes the world’s largest corporations in Davos and a company founded to offer alternatives to their platforms put the same criterion at the center, reversibility becomes a requirement to be written into contracts.

Three pillars and a diagnosis

The framework developed for companies rests on three interdependent pillars: architectural resilience, operational management, and governance and compliance. Architecture is the foundation, because architectural choices are the hardest to reverse; operational management is about keeping control of vendor relationships, so that no single supplier gets to dictate the terms; governance ensures the company remains accountable across the markets where it operates, and it is the pillar that cuts across all the others.

On where to invest with a limited budget, Zanolla rejects a one-size-fits-all answer: the WEF’s message is to invest in all three pillars in a targeted way, weighting each according to the use cases you most want to scale. The starting point is a diagnosis: map existing systems, figure out where you are most exposed and concentrate your efforts there. Trying to tackle everything at once, she warns, sends costs soaring.

Digital embassies for states and companies

Among the Forum’s proposals is an idea born far from artificial intelligence. In 2017, ten years after the cyberattack that knocked out the websites of its parliament, banks, ministries and newspapers, Estonia signed an agreement with Luxembourg for the first data embassy: copies of public registries kept in a Luxembourg data center, protected by immunities similar to diplomatic ones.

The principle, Zanolla explains, is the same as for a physical embassy: a facility abroad that operates under the jurisdiction of the sending country, on the basis of an agreement between the two states. On May 11, the WEF and Bain published AI Infrastructure in the Age of Sovereignty, which proposes a common framework to make these arrangements trustworthy; until now, they have been negotiated case by case.

The concept, according to Zanolla, is becoming relevant for businesses as well: a company that needs more computing power than it can build in-house, but wants to stay under the legal control of its home country, could move part of its data abroad and keep managing it under its original jurisdiction. With physical infrastructure constrained by many factors and concentrated in a handful of economies, it is one possible answer for mid-sized countries.

For SMEs, the real risk is not thinking about it

Much of Italy’s economy is made up of companies with a few hundred employees and IT budgets nowhere near those of large groups. Asked whether digital sovereignty risks becoming a luxury for them, Zanolla starts from an asymmetry: a vulnerability that hits a large company can be mitigated one way or another; if it hits a small one, “it can have a very, very serious impact on its survival.” That is why, far from being a luxury, failing to think through the consequences becomes “a huge limitation on operational capacity.”

I also put to her a question prompted by the work of Philippe Aghion, winner of the 2025 Nobel Prize in Economics, on distance from the technological frontier: if the resources needed to be sovereign are so great that they erode competitiveness, does chasing independence still make sense? Zanolla’s answer is proportionality: decide which data, decisions and areas you want to control, and for everything else rely on trusted partners. The opposite risk, wanting maximum control everywhere, leads to unsustainable costs with no benefit in the medium term; that is why, she says, it makes more sense to reason by principles than by compliance checklists.

There is, however, an obstacle further upstream. According to Zanolla, very few technology leaders today have a clear enough view of their own spending to make decisions on that basis; mapping where the money goes is the first step, understanding where you are exposed is the second, and neither is a trivial exercise.

Who sets the rules between agents

The last theme looks ahead to the next Annual Meeting. Today the debate centers on how to govern an AI agent inside one’s own organization; far less attention, Zanolla observes, goes to how one company’s agents will work with another’s across national borders, each bound by different rules. “Who sets the rules when two different agents close a transaction? Who is accountable when something goes wrong?” she asks. Agents are being built faster than the rules that govern them, and the Forum expects this to become a particularly hot area in the coming months.

In June, drastic as it was, the chain of decisions was easy to read: a government signs a letter, a company decides how to comply. An agreement between a Chinese agent and a Spanish one exchanging information, with no traditional contract behind it, has no obvious jurisdiction yet. The WEF’s exit test assumes you know whom you depend on; with agents, the first challenge will be noticing.


AI sovereignty: the ability to walk away keeps you safe, says WEF – Francesca Zanolla - Ultima modifica: 2026-11-17T07:03:29+00:00 da Francesco Marino

Giornalista esperto di tecnologia, da oltre 20 anni si occupa di innovazione, mondo digitale, hardware, software e social. È stato direttore editoriale della rivista scientifica Newton e ha lavorato per 11 anni al Gruppo Sole 24 Ore. È il fondatore e direttore responsabile di Digitalic

Recent Posts

Sovranità dell’AI, il WEF: si salva chi sa andarsene – Francesca Zanolla

Il 12 giugno una direttiva americana ha spinto Anthropic a spegnere in tutto il mondo…

41 minuti ago

Palantir, l’Europa cerca un’alternativa sovrana per l’intelligenza artificiale militare

Palantir: Francia, Germania e Paesi Bassi cercano un'alternativa europea per l'AI militare. Arcadia, Helsing e…

22 ore ago

AI UNTOLD: al MEET di Milano il Gruppo E dà voce all’AI che le aziende non dicono

AI UNTOLD, l’evento del Gruppo E al MEET di Milano: da Leonardo alla governance degli…

4 giorni ago

Dimissioni Jensen Huang: chi ci guadagna davvero

Dimissioni Jensen Huang: lo scandalo Nvidia nasconde la partita USA-Cina sui chip. Chi ci guadagna,…

4 giorni ago

“Riscrivere il potere”: il libro di Francesca Moriani e Francesco Frugiuele che smonta la gerarchia dall’interno

Riscrivere il potere, il libro di Francesca Moriani e Francesco Frugiuele, presentato a Milano: sei…

6 giorni ago

Cisco premia Collabora: è Innovation Partner of the Year per il South EMEA

Al Cisco Partner Summit di Vancouver, Collabora ha ricevuto il riconoscimento per l’innovazione nell’area South…

6 giorni ago

Digitalic © MMedia Srl

Via Italia 50, 20900 Monza (MB) - C.F. e Partita IVA: 03339380135

Reg. Trib. Milano n. 409 del 21/7/2011 - ROC n. 21424 del 3/8/2011