The debate regarding technological sovereignty is frequently misframed, as if it consisted of possessing 'European' artificial intelligence models in opposition to American or Chinese ones. That is merely the surface. The fundamental question is different: who controls the entire stack upon which everything else rests. Not merely the models, but the chips, the data centres, the cloud, the training data, the deployment tools, the distribution channels, and the recommendation systems. Sovereignty is not contested at the visible apex, but within the foundations that almost no one observes.

And the foundations are in very few hands. The Stanford AI Index estimates that almost 90% of the relevant AI models of 2024 emerged from industry, not from universities or the public sector; and that in that year the United States produced forty of those models, China fifteen, and Europe three. In the cloud, the concentration is even more distinct: the European Parliament itself notes that three American companies account for approximately 65% of the European cloud services market, and on a global scale those same three account for around two-thirds of the infrastructure. Computing—the processing capacity that trains and operates AI—is, according to specialised literature, the most concentrated bottleneck of all: detectable, quantifiable, and controllable by a few. The European gap, therefore, is not merely regulatory. It is one of material capacity.

It is appropriate to be fair regarding what is being done, without purchasing the optimistic narrative. Europe has reacted: an action plan for AI, public computing factories and 'gigafactories', a Data Act in force since September 2025 that seeks to facilitate the switching of cloud providers, a cloud sovereignty framework that decomposes the word into evaluable criteria. It is positive that it does so. Yet none of those initiatives is equivalent to sovereignty as of yet: they are infrastructures in deployment, contracting frameworks, strategies in progress. The case of Gaia-X, the European federated cloud project, is instructive: in order to respond to the dependence upon large providers it ended up incorporating them, and from that emerged a sovereignty that was weaker and more pragmatic than promised. Declaring an AI to be 'European' is insufficient; what counts is reducing real dependence—being able to migrate, audit, execute, train, and safeguard data without being trapped by an external provider, jurisdiction, or business model.

Thus far, the issue is one of infrastructure and industrial policy. Yet there exists a layer that concerns me more closely, the cultural, which is rarely named in these debates. If those who command attention—the platforms, the models, the recommendation algorithms—are in few hands, then those few hands also condition which languages, which works, which memories, and which forms of creation circulate, are recommended, or remain buried. Recent literature formulates this with prudence, and it is appropriate to treat it as such: algorithmic systems have become agents of cultural mediation; they rank content and produce visibility under logics of prediction and attention; metrics—views, 'likes', followers—function as a new form of capital, displacing part of the authority from credentials towards algorithmic performance. I do not assert that algorithms systematically bury all minority creation: that would be a law that the sources do not support. I assert something more measured and more unsettling: that there exists a real risk of invisibilisation and homogenisation, and that its form depends upon who designs the architecture.

For that is the key, and it provides the title for this: architecture is not neutral. It is unnecessary for anyone to censor for some things to be seen and others not. It suffices that the infrastructure—technical, economic, algorithmic—quietly favours certain formats, certain tempos, certain signals of success over others. Technological sovereignty, viewed from the perspective of culture, is also sovereignty of visibility: the capacity for that which is created in a language, on a margin, on a small scale, not to depend entirely upon systems designed elsewhere and with a different logic.

I conclude where I usually do, without despondency. The first step, when faced with an architecture that presents itself as air, is to perceive it. To name the three hands that hold the cloud, to distinguish infrastructure in deployment from achieved sovereignty, not to confuse possessing an open model with controlling the stack that makes it function. That which is seen can be discussed, regulated, diversified. That which is assumed to be neutral, cannot. And in culture, as in almost everything, freedom commences by knowing upon which foundations one is standing.

On open conversation

This text continues the series regarding digital and cultural infrastructures that I have been publishing in this notebook: the material and energy cost of AI, the geographical distribution of computing power, and now architecture as a layer that decides without being seen. The texts share a fundamental thesis: in contemporary systems, the decisive factor is not individual conduct but infrastructural design, and that which presents itself as neutral is almost never so. If anyone wishes to intervene from the perspectives of AI governance, the political economy of technology, digital sovereignty, or cultural practice, this notebook remains open.

Sources

Stanford HAI. AI Index Report 2025.

European Parliamentary Research Service. Briefing on the Cloud and AI Development Act. 2025.

Synergy Research Group / Statista. Global cloud infrastructure market shares (Q1 2026).

Sastry, G.; Heim, L.; Belfield, H. et al. Computing Power and the Governance of Artificial Intelligence. arXiv:2402.08797, 2024.

European Commission. AI Continent Action Plan (2025); Cloud Sovereignty Framework (2025); Data Act (applicable from 12 September 2025).

Baur, A. Analysis of Gaia-X and European digital sovereignty. Information, Communication & Society, 2025/2026. DOI: 10.1080/1369118X.2025.2516545.

Sánchez-Vera, F. (2025); Valiati, L. and Möller, C. (2025); Wong, K. (2025), regarding algorithmic mediation and cultural visibility.