Three American companies today host more than 70% of the computing data of European businesses [5]. Several French ministries entrust their sensitive systems to actors subject to American law. AI is becoming the central infrastructure for all economic and public decision-making. The trade-off posed for 2027 is clear: regain control of the machines on which the State’s decisions run, or continue deciding on infrastructures it does not control.

Sixty-one billion market, fifteen percent European

The European cloud market reached 61 billion euros in 2024 [10]. The share of European providers fell from 29% in 2017 to 15% in 2022 [11]. Amazon, Microsoft and Google together account for 70% of the regional market. In France, the infrastructure cloud market represented 9 billion euros in 2024. More than 70% went to American actors [10].

The investment imbalance is massive. The four American hyperscalers are preparing to deploy between 635 and 665 billion dollars in AI in 2026. The previous year, they had spent 381 billion [10].

In global venture capital, the United States concentrates 89% of AI operations by value. Europe captures 5% [17]. French Tech raised 7.2 billion euros in 2025: a real rebound, but nowhere near American scales [17].

This imbalance shows in usage patterns. In 2024, only 23% of French companies with ten or more employees displayed high or very high digital intensity. In Scandinavian countries, this rate exceeds 55% [8]. In 2025, 42% of Danish companies used AI, 37.8% of Finnish companies, 35% of Swedish companies [8]. France sits in Europe’s middle ground.

Two clarifications limit a catastrophic reading. On public online usage, France stands out favorably: in 2025, 90.8% of French internet users used public online services, against 75.0% on average in the Union [21]. The OECD ranks France among the top five European countries for digital transformation of the public sector [9]. But vulnerability lodges itself between the performance of public interfaces and control of infrastructure.

Dependence Has a Legal Price

Digital technology radically changes relationships between nations. It enables real-time control of a growing share of actors. Through the appropriation of the datasphere and intermediation platforms, new forces emerge that force public authorities to reposition themselves. This is the central thesis of Stéphane Grumbach in The Empire of Algorithms (Armand Colin, 2025) [14].

This appropriation has a concrete legal translation. The Cloud Act authorizes American authorities to access data hosted by American companies, wherever they are located in the world. Data stored in Europe by American subsidiaries can be seized in case of investigation, without going through international judicial cooperation procedures.

Certain ministries use non-European computing solutions for sensitive data. The Court of Auditors established this in its report of October 31, 2025 on the civil information systems of the State [5]. The health data platform has been hosted for more than five years with an American actor. The public procurement platform Place calls in part on a Canadian group, with no firm migration timeline. No ministry has a complete map of its sensitive data [5].

The institutional diagnosis was already established in July 2024. The Court of Auditors noted that DINUM’s capacity to drive an inter-ministerial policy had been altered by frequent leadership changes since 2018 [6]. The results remained “mixed.” Above all, the Court found an absence of consolidation of the State’s digital spending and its projections [6]. This budgetary blind spot makes any coherent investment strategy impossible.

AI as a Tipping Point

The debate on digital sovereignty has long focused on data and cloud. AI shifts it to more radical ground: the decision models themselves. The models, training data, graphics chips and cloud are the property of large American groups. France has only one actor of global scale in this field: Mistral.

AI is a general-purpose technology. It will modify all production processes and the way work is done, in all sectors, not just tech. This is what Axelle Arquié, economist at CEPII, documents in “The Double Shock of AI: Employment and Taxation,” published in 2026 [13].

The fiscal shock she describes completes the infrastructure shock. “White-collar” workers pay more taxes and social contributions than other worker categories. A 20% shock on this part of the labor market could produce a very violent scissor effect for public finances [13].

If AI productivity gains are captured by American actors on American platforms, the French tax base erodes doubly. First through the substitution of jobs taxed in France. Then through the location of profits outside territory [13].

In March 2026, Coface and the Observatory of Threatened and Emerging Jobs published the most granular mapping to date of French employment exposure to AI [12]. In 2026, 3.8% of jobs are vulnerable to generative AI. In a two- to five-year horizon, this rate rises to 16.3%, or about 5 million jobs [12]. Infrastructure dependence amplifies this risk. If adaptation to AI happens on tools France does not control, the conditions for this adaptation are set elsewhere: speed, access, cost, sectoral priority.

On markets with network effects, the danger is dynamic lock-in. Jean Tirole demonstrated this in his work on regulating digital markets [20]: once installed, this lock-in makes any subsequent entry structurally impossible. The appeal of hyperscalers rests on functional richness, economies of scale, and ecosystem dynamics. These mechanisms create self-reinforcing dependencies.

What Europe Has Put on the Table, and What Still Needs to Come

The European institutional response took shape in two phases. The first, regulatory, produced GDPR, the AI Act (fully applicable since August 2026) and the Data Act. The SREN law of May 2024 gave Arcep a role as regulator of cloud service providers [1]. Arcep launched a consultation on interoperability and portability and recommends to the Commission that it designate major cloud providers as access controllers under the DMA [3].

The second phase, industrial, is more recent and more ambitious. The European Commission adopted on June 3, 2026 its proposed regulation on cloud and AI development, called CADA [10]. The objective: at least triple the Union’s data center capacity within five to seven years. Fully meet the needs of European businesses and administrations by 2035. CADA is part of a technological sovereignty package, alongside the Chips Act 2.0 and a European open source strategy.

But the regulatory strategy carries a risk that Carl Benedikt Frey documents in How Progress Ends (Princeton, 2025) [19]. Institutions can block progress as surely as the absence of rules. Regulation that stops at compliance without creating conditions for a credible industrial alternative can freeze dependence by legitimating it.

Microsoft extended in June 2025 its range of sovereign solutions for Europe, relying on its Blue partnerships in France and Delos Cloud in Germany. This strategy aims to neutralize the regulatory argument before it translates into losses of public contracts. As long as the parent company remains subject to American law, the technical guarantees of isolation remain to be examined with caution.

In 2025, 83% of cloud and software spending by European businesses went to American actors [22]. The mechanism is self-reinforcing: dependence generates financial flows that consolidate the dominant position of hyperscalers. Investing more in AI without diversifying infrastructure providers increases the constraint.

The State Regains Ownership of Its Models and Sensitive Data, or It Renounces Decision-Making

The trade-off is simple to state, difficult to assume. The State ceases to be tenant of its own decisions. The sensitive data of administrations, whether health, taxation, intelligence or public procurement, is hosted on infrastructures over which public authority holds legal and technical control, without exception drawn from cost or availability of supply. The Court of Auditors established that no ministry today has a complete map of its sensitive data [5]. This blind spot is a de facto delegation to actors subject to American law.

The lock-in of markets with network effects is undone through the capacity to exit, not through contractual negotiation. Tirole demonstrated this, and it holds here with particular precision [20]. The trade-off transfers the competence of qualifying infrastructures—that is, the decision of knowing which data can go where—from each ministry to a single inter-ministerial authority. This authority would have a veto right over public digital markets. Solutions like “sovereign cloud” backed by American parent companies would lose their eligibility for sensitive-level data.

This migration is the first phase. The horizon is European, and it is built through public procurement: CADA must become a tool of critical mass for qualified offerings. Interoperability and reversibility complete the system, by attacking the economy of proprietary lock-in. The third stage is training: a national plan for open source technologies, backed by the European strategy, removes public servants from proprietary ecosystems.

The main trade-off concerns time and money. Migrating the health data platform to a qualified SecNumCloud environment has a real investment cost. Migrating the public procurement platform Place and ministerial messaging likewise: rewrites, tests, skills development, higher market prices in the short term. This cost falls on ministerial budgets and project timelines.

What this trade-off closes is the ease of contracting with the most mature and fastest-to-deploy actors [5]. French and European providers gain a guaranteed domestic market and must demonstrate in short timelines that they can handle the load. The alternative is a double shock on employment and the tax base whose effects are already being calculated in 2026 [13]. Investment in ownership of public infrastructure remains less costly than the bill of dependence, provided one begins before the network effects of AI make exit structurally impossible.

Sources

[1] Arcep, “Annual Report 2025: Volume 3,” July 2025, https://www.arcep.fr/uploads/tx_gspublication/ARCEP-RA2025-TOME_3-Norme_A.pdf (accessed 26/08/2026).

[2] Arcep, “Public Consultation on Interoperability and Portability of Cloud Services,” July 2025, https://www.arcep.fr/actualites/actualites-et-communiques/detail/n/cloud-170625.html (accessed 26/08/2026).

[3] European Commission, “Country Report on the Digital Decade 2024: France,” 2024, https://digital-strategy.ec.europa.eu/fr/node/13602 (accessed 26/08/2026).

[4] Eurostat / OECD, “AI use in enterprises and AI adoption by SMEs,” December 2025, https://www.oecd.org/content/dam/oecd/fr/publications/reports/2025/12/ai-adoption-by-small-and-medium-sized-enterprises_9c48eae6/1dbf96ef-fr.pdf (accessed 26/08/2026).

[5] Court of Auditors, “The Sovereignty Challenges of the State’s Civil Information Systems” (report S2025-1479), October 31, 2025, https://www.ccomptes.fr/fr/publications/les-enjeux-de-souverainete-des-systemes-dinformation-civils-de-letat (accessed 26/08/2026).

[6] Court of Auditors, “Oversight of Digital Transformation of the State by the Inter-ministerial Department of Digital Affairs,” July 10, 2024, https://www.ccomptes.fr/fr/publications/le-pilotage-de-la-transformation-numerique-de-letat-par-la-direction (accessed 26/08/2026).

[7] European Commission, “Cloud and AI Development Act (CADA): Proposed Regulation COM(2026) 502 final,” June 3, 2026, IT Social coverage, https://itsocial.fr/cloud-infrastructure-it/cloud-infrastructure-it-articles/geopolitique-de-lia-partie-4-les-fournisseurs-de-cloud-europeens-doublent-de-taille-en-5-ans-mais-reculent-en-parts-de-marche/ (accessed 26/08/2026).

[8] Synergy Research Group, cloud market share data Europe 2024, cited in IT Social, 2026, https://itsocial.fr/cloud-infrastructure-it/cloud-infrastructure-it-articles/geopolitique-de-lia-partie-4-les-fournisseurs-de-cloud-europeens-doublent-de-taille-en-5-ans-mais-reculent-en-parts-de-marche/ (accessed 26/08/2026).

[9] Coface / Observatory of Threatened and Emerging Jobs, “Exposure of French Employment to AI,” March 2026, https://anthemcreation.com/en/artificial-intelligence/ai-jobs-5-million-at-risk-france-coface-oem/ (accessed 26/08/2026).

[10] Axelle Arquié, “The Double Shock of AI: Employment and Taxation,” Political Economy, 2026/2, no. 110, https://shs.cairn.info/publications-de-axelle-arquie–111394 (accessed 26/08/2026).

[11] Stéphane Grumbach, The Empire of Algorithms: A Geopolitics of Control in the Age of the Anthropocene, Armand Colin, 2025, https://www.dunod.com/histoire-geographie-et-sciences-politiques/empire-algorithmes-une-geopolitique-du-controle-ere (accessed 26/08/2026).

[12] EY France, “Venture Capital Barometer 2025: Annual Results,” January 2026, https://www.ey.com/fr_fr/insights/fast-growing-companies/barometre-ey-du-capital-risque-les-resultats-annuels (accessed 26/08/2026).

[13] Carl Benedikt Frey, How Progress Ends: Technology, Innovation, and the Fate of Nations, Princeton University Press, 2025, https://press.princeton.edu/books/hardcover/9780691233079/how-progress-ends (accessed 26/08/2026).

[14] Jean Tirole, Work on Regulating Digital Markets, Recent Reference “European Defence,” Project Syndicate, 2026, https://www.project-syndicate.org/columnist/jean-tirole (accessed 26/08/2026).

[15] European Commission, “Country Report 2025: France,” European Semester, Working Document SWD(2025) 210, June 2025, https://economy-finance.ec.europa.eu/document/download/9f4d2810-f2df-4b16-8eb8-7189923a7622_en?filename=FR_CR_SWD_2025_210_1_EN_autre_document_travail_service_part1_v3.pdf (accessed 26/08/2026).

[16] Cigref, Data on Cloud and Software Spending by European Businesses, figure cited to Cabinet Ministers, June 2025, Studeria coverage, August 2026, https://www.studeria.fr/articles-de-blog/couts-ia-generative-budgets-dsi-souverainete-numerique (accessed 26/08/2026).