Editor’s note: This column is part of CEPR’s “Europe 2050: Geometries of peace, power, and prosperity” initiative, which adopts a long-term perspective to explore what Europe should aspire to become by 2050.

On 12 June 2026, the US Commerce Department issued an export control directive ordering Anthropic to suspend access to its most powerful AI models for all non-US persons. For the first time, Washington used unilateral administrative authority to shut down a frontier AI model globally – without prior consultation with allies, without any process by which foreign governments could contest the order, and without a specified basis that any European institution could invoke (US Commerce Department 2026, Anthropic 2026). The precedent is not about those specific models. It is about the authority itself: any frontier AI system that European institutions, governments, or defence contractors depend on can be switched off by US administrative decision, without notice, without appeal.

This is not an AI regulation story. It is a sovereignty story. And it was entirely predictable. Unlike energy or semiconductors, AI dependency has no gradual adjustment mechanism: it can be switched off globally in hours, as 12 June demonstrated.

The choice Europe made

In 2024, the EU deliberately excluded military and national security AI from its landmark AI regulation. Member states – led by France – insisted that defence decisions must remain with national capitals, and this position was reflected in Article 2(3) of the regulation (European Parliament and Council 2024). The consequence: Europe built no instrument to govern its own military AI, and none to govern anyone else’s. When Washington acts in this space, European institutions are observers, not parties.

Washington drew the logical consequence. In January 2026, the US Department of Defense dropped the requirement that military AI remain under human control, replacing it with “any lawful use, free from usage policy constraints” (US Department of Defense 2026). The 12 June directive on commercial AI models extends the same logic: unilateral authority, no consultation, no European recourse. These are not isolated decisions. They are a consistent architecture – and Europe is outside it. The practical meaning is simple: when Washington makes decisions about AI systems that operate on European territory or affect European citizens, there is no European institution with standing to contest or delay them.

The price

Three US companies supply 65% of the European cloud services market (Marcelin 2025) – a share that has grown substantially as European providers have seen their market position decline significantly over the past decade. This is the infrastructure through which European institutions, governments, and defence contractors increasingly process sensitive data. European defence ministries and intelligence services run on infrastructure they do not control and cannot independently audit. Every year of delay raises the cost of building an alternative and narrows the space for any independent European position.

The cost of closing the military AI gap is now on record. The Kiel Institute’s Sparta 2.0 estimates €500 billion over a decade – €50 billion per year – to close Europe’s ten critical military capability gaps (Schularick et al. 2026). The two most directly relevant: a sovereign military command software system comparable to what US forces now deploy would cost €10-20 billion over three to four years; scaled autonomous systems production would cost over €30 billion over three to five years (Schularick et al. 2026). High-technology defence investment of this kind generates multiplier effects up to €1.50 per euro – compounding strategic investment, not pure expenditure (Schularick et al. 2026).

The financing architecture exists but distributes unevenly. ReArm Europe offers €800 billion combining national fiscal flexibility with EU-backed loans for joint procurement (European Commission 2025). But the states most exposed to AI dependency – those carrying large legacy debts – have the least room to invest unilaterally. Without a decision on joint financing for AI military investment, the money will not reach where the capability gap is largest.

The political price must be named. France drove the exclusion of military AI from the EU’s 2024 regulation (European Parliament and Council 2024), insisting on defence as a national prerogative. The same France is Europe’s largest defence industrial power and the country with the most to gain from a European military AI programme. Any move to govern allied military AI on European territory requires France’s agreement – and that agreement will not come without an industrial architecture that gives France the leading role. The governance question and the industrial question are the same negotiation. Treating them separately produces neither.

What can be done

Ten to fifteen years separate today’s dependency from a European capability sufficient to negotiate as a peer (Schularick et al. 2026). During that period, Europe will lack a veto over allied AI decisions affecting its territory and will pay a structural premium on digital infrastructure that constrains sovereign action. Three steps are available now – not as a substitute for the investment, but as the minimum action in the absence of industrial capacity.

First, European institutions should formally establish that Washington’s January 2026 removal of human-control requirements from military AI constitutes a significant change in the security environment for European institutions (US Department of Defense 2026). A formal political assessment – achievable before the end of 2026 – creates the basis for subsequent action and establishes that Europe is aware of the gap before the window for pre-operational negotiation closes.

Second, the European Commission should open consultations with Washington on notification: a commitment that the US will inform European partners before AI-assisted operations affect European networks or European-flagged assets. Human oversight of US military AI is not currently negotiable – Washington’s doctrine explicitly rejects it. Notification is. Europe’s leverage is its digital market, not its military capacity. The 12 June precedent makes the stakes concrete: Europe needs a process by which it can formally respond to unilateral decisions affecting European citizens, not learn of them from press reports.

Third, Europe’s sovereign cloud investment programme – the Cloud and AI Development Act, projected for adoption in late 2027 (Marcelin 2025) – should be accelerated and explicitly tied to these consultations. Market access is Europe’s most concrete leverage over US AI providers. It should be deployed as part of a governance negotiation, not spent separately on unrelated priorities.

None of these steps closes the gap. They buy time while the investment matures. For the full empirical documentation of Europe’s AI governance gap, see Perugini (2026).

The 2050 answer

A Europe with credible AI military capacity, cloud sovereignty, and defence industrial scale does not need to negotiate notification rights with Washington. It negotiates as a peer. That Europe does not exist in 2026. Building it is the €500 billion project: deployed as a coherent European programme, financed through instruments that reflect where fiscal capacity actually exists, and negotiated with France as an industrial partner rather than a veto problem. That is the project for 2050.

The short-term answer is harder: for the next decade, Europe will operate under a gap it cannot close by political means alone. The steps above buy time. Building industrial capacity closes the gap.

Conflict of interest: This piece was drafted using Anthropic Claude. On 12 June 2026, the US government issued an export control directive suspending global access to Anthropic’s frontier models – the same type of unilateral authority this piece identifies as the governance problem.

References

Anthropic (2026), “Public statement on government export control directive”, 12 June 2026. anthropic.com

European Commission (2025), “ReArm Europe Plan/Readiness 2030”, 19 March 2025. 

European Parliament and Council (2024), Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence (AI Act). 

Marcelin, T (2025), “Cloud and AI Development Act”, EPRS Briefing PE 779251. European Parliamentary Research Service. europarl.europa.eu

Perugini, M, (2026), “The Governance Gap: How the United States Dismantled AI Oversight and What It Means for Europe”, Zenodo, 21 May. 

Schularick, M, N Lange, R Obermann and T Enders (2026), “Achieving European Defence Autonomy: A Roadmap for Overcoming Critical Dependencies (Sparta 2.0)”, Kiel Institute for the World Economy, May. 

US Commerce Department (2026), “Export control directive on Anthropic AI models”, 12 June 2026.

US Department of Defense (2026), “Artificial Intelligence Strategy for the Department of War”, Memorandum, 9 January.