Sovereignty is not a market share: Europe, artificial intelligence and the jurisdictional question

On 23 October 2026, an Agence France-Presse dispatch posed a question that has become commonplace in European debate: what exactly does the call for a “sovereign artificial intelligence”¹ cover? The answer it assembled is representative of the way the subject is discussed. We find there a definition, namely independence from foreign jurisdictions and foreign governmental influence², a series of ratios describing Europe’s position in the value chain, and a conclusion of caution: total sovereignty remains out of reach, it is better to reason in degrees and to abandon the idea of a fortress.

Nothing in this reasoning is wrong. We would argue nonetheless that it measures the wrong variable, and that the most instructive episode of the past year, the interruption in June of European access to two American models, is explained by none of the ratios customarily cited.

Let us begin with those ratios, for they describe a serious reality. In 2024, China accounted for roughly 60 % of the mined output of rare earths for magnets and 91 % of their refining³. The Union produces less than 10 % of the world’s semiconductors and none of the most advanced chips intended for artificial intelligence, which are designed and manufactured by a small number of American and Taiwanese firms⁴. It hosts only 5 % of global data centre capacity, against some 75 % in the United States and 15 % in China⁵. Three American providers share close to 70 % of the European cloud computing market⁶. A few exceptions exist, ASML in the Netherlands in lithography equipment, Mistral in France and Black Forest in Germany for models, but they do not alter the order of magnitude.

These figures are accurate and their publication is useful. Let us observe, however, what they measure: production capacity and market shares. They answer the question “what do we manufacture?”. They do not answer this one, which is nevertheless the sovereign question par excellence: what can be taken away from us, and who can decide it?

The year 2026 supplied on this point a demonstration that the ratios did not foreshadow. On 12 June, the US Department of Commerce sent Anthropic an is-informed letter, the act by which the administration informs a company that a licence will henceforth be required of it. The requirement covered any export, re-export or transfer of two models released three days earlier, Mythos 5 and Fable 5, to any foreign person, including the company’s foreign employees working on American soil⁷. Unable to distinguish its users with that degree of granularity, the company disabled both models for its entire global customer base. The measure was justified by the risk that these capabilities would be exploited by foreign military intelligence services⁸.

A point of legal qualification deserves full attention here, for it carries what is essential. The letter treats remote access to a model, through an application programming interface, as a making-available subject to export control, relying on section 734.13 of the applicable American regulation. Yet that same provision had served, in three earlier advisory opinions of the same administration, to ground the opposite conclusion: remote access to hosted software, without any transfer of technology or of source code, fell outside the scope of the regulation⁹. We are therefore not faced with the application of a known rule, but with a reversal of interpretation, effected by way of an individual letter, on a basis whose text has not been made public and whose reasoning the recipient company has contested¹⁰. What crosses the border, in this construction, is neither the model nor its code: it is access to a capability.

It is here that reasoning by ratios reveals its limit. The June interruption bears no relation to the 5 % of European data centre capacity. Nor would it have been averted had Europe held a quarter of that capacity, for the constraint followed neither the hardware nor its location: it followed the model’s jurisdictional attachment. A European data centre hosting an American model remains exposed to the American decision. The decisive variable is therefore neither ownership nor location, but the jurisdiction to which the technology in use belongs and, correlatively, the possibility of changing it.

The vocabulary of “degrees of sovereignty”, which the analysts cited by the dispatch rightly set against the illusion of complete independence, is well founded in its intention. But what ought to be graduated is not the share of production: it is the time and the cost of switching. A dependency from which we can exit within a few weeks, at the price of a bearable degradation of service, constitutes an ordinary commercial relationship. A dependency from which we cannot exit within a useful period constitutes a subordination, whatever the quality of the supplier and however cordial diplomatic relations may be at the moment the contract is signed. The operative criterion is not autonomy but reversibility.

What followed in this episode confirms that reading, provided it is described with caution. Access was restored by virtue of private negotiations, modifications to technical safeguards and tests conducted by the administration; the sources differ on the exact duration of the interruption¹¹. On 10 September, the European Union Agency for Cybersecurity (ENISA) obtained access to Mythos, more than three months after the announcement according to which the Union would have access to it¹². We insist on the nuance, for it is the opposite of a detail: what the Union obtained in September was not the restoration of a right, it was the granting of a negotiated faculty, conceded to a particular agency, whose terms, notably the visibility granted in return over European systems, remained to be settled. A dependency formalised in this manner is not a dependency resolved; it is a dependency administered. And the interpretation adopted in June has not been disavowed: it remains available.

A formula heard in this debate deserves to be discussed rather than taken up: “It is by creating interdependencies that sovereign power is created”¹³. The idea is classical and, under ordinary conditions, correct: interdependence deters, since it makes the cost of rupture reciprocal. It must still be symmetrical. In June 2026, interdependence did indeed exist between the American industry and its European customers; it produced no power on the European side, because rupture was costly for one party and bearable for the other. Interdependence protects only the party that disposes of a capacity for retaliation or of a fallback solution. Failing that, it is not an equilibrium: it is a bond.

We do not draw from this analysis a plea for autarky, which the analysts cited are right to describe as neither attainable nor desirable. We draw from it a displacement of the criterion of assessment. The European industrial programmes, covering critical minerals, semiconductors and computing capacity, remain necessary, and they act on the only scale on which they can act, which is that of the decade. They do not address, by contrast, the legal exposure, which is immediate and which is played out on another terrain: that of portability clauses, of documentation permitting a genuine migration, of the verified existence of an alternative, and of the capacity of an administration to continue functioning when a foreign legal order shifts. A policy that measured its progress by production shares alone might, in ten years, have reduced part of its industrial lag without having diminished its vulnerability, if the systems then in use remain attached to a jurisdiction that can suspend them.

Sovereignty, so understood, is not a rate. It is a structure of decision. The pertinent question is not what Europe manufactures, but who can say no to it, within what period it can do without, and what it consents to concede so that the answer remains negotiable. Framed in these terms, it calls for answers less spectacular than the announcement of an industrial champion, and probably more accessible: organising an exit is quicker than building a factory. That is also why it is a question which it is possible to address straight away.

AFP, « Appels à une IA européenne souveraine : de quoi parle-t-on ? », 23 October 2026, distributed notably by Boursorama and FashionNetwork. ↩
Definition reported by the dispatch cited above and attributed to Dario Maisto, an analyst at Forrester. ↩
International Energy Agency, Rare Earth Elements, summary: 60 % of the mined output of rare earths for magnets and 91 % of refining in 2024. The dispatch cited gives 60 % and 90 % for rare earths as a whole; we retain here the IEA’s perimeter, which is more precise. ↩
Dispatch cited above. No source is indicated there for these proportions. ↩
Andrea Renda, director of research at the Centre for European Policy Studies, remarks reported by the dispatch cited above. ↩
Dispatch cited above; no source is indicated there. ↩
On the act of 12 June 2026 and its scope, see Mayer Brown and Greenberg Traurig. ↩
On the justification advanced and the European reactions, see Le Temps and Euronews, June 2026. ↩
On the basis relied upon, section 734.13 of the Export Administration Regulations, and its contradiction with three earlier advisory opinions of the same department, see Mayer Brown and CSIS. For a discussion of the qualification, see also The Regulatory Review. ↩
The letter has not been made public and its full reasoning is not known; the company stated that it had not received detailed written reasons. See CSIS. ↩
The accounts differ: some mention nineteen days of interruption, others several weeks. We do not settle the question, the duration not being decisive for the argument. ↩
Bloomberg, 10 September 2026; see also France 24. ↩
Remarks by Sylvain Duranton (Boston Consulting Group), reported by the dispatch cited above. ↩