A fictional think-tank scenario about European AI collapse has moved from PDF to policy table. “Europe 2031,” a speculative narrative published by the Brussels-based Arq Foundation, was read by members of the European Parliament and surfaced in track 1.5 discussions between British and German officials last week, according to The Guardian. The timing was not accidental: the piece published one day before the Trump administration blocked non-US nationals from accessing Anthropic’s Claude Fable model, an event the authors had roughly anticipated.
The scenario imagines a 2031 in which the United States monopolises roughly 70 percent of global AI compute capacity while Europe, having underinvested in datacenters, finds its economy contracting, its companies unable to adopt AI at competitive speed, and its political system absorbing the backlash. Cyber-attacks hit European firms. Populism accelerates. The euro wobbles. ASML, the Dutch lithography firm that underpins semiconductor production globally, becomes Europe’s sole remaining bargaining chip in negotiations with Washington and Beijing. It is not enough.
The scenario’s core prescription is datacenter construction, specifically in deregulated AI zones where power permitting and planning approvals move faster. Co-author Alex Petropolous frames the issue as a fixed-supply allocation problem: a finite number of datacenters will be built globally each year, and the question is where they land. Maximilian Negele, the other co-author and a former RAND researcher, describes the current situation as a “slow-moving car crash.”
That framing deserves scrutiny. Several of the US investment figures the scenario cites as evidence of American dominance have since unravelled. The $100 billion OpenAI-Nvidia deal collapsed in February. The $300 billion OpenAI-Oracle agreement is in doubt. OpenAI pulled out of the Texas datacenter project the scenario references by name. The authors acknowledge this but argue that the overall trajectory holds even if individual deals fall apart. Their pre-emptive move is to write European skepticism into the story itself: the fictional EU officials who dismiss AI as a bubble are the ones who end up worst off.
This is where advocacy scenarios and policy analysis diverge. Scenarios are not forecasts; they are pressure tools designed to make inaction feel riskier than action. The Europe 2031 authors want European policymakers to see compute underinvestment as an existential threat. That diagnosis may be directionally correct. But the prescription, build more datacenters faster, runs directly into a political problem the authors name but do not resolve: “People hate datacenters. They destroy the landscape. They support big tech. It’s a very, very unpopular policy.”
There is a second structural problem the scenario does not engage. Nicolás Casares, a Spanish MEP quoted by The Guardian, puts it plainly: if American companies build the datacenters in Europe, who controls the infrastructure and the access decisions? The Fable restriction demonstrated that a US-built model can be switched off for foreign users at the federal level. European datacenters running American AI models do not change that dependency; they deepen it by creating sunk capital without sovereignty. The relevant policy question is not only how many datacenters Europe builds, but who owns the compute stack inside them.
The Arq Foundation does not disclose its funders and describes itself as neither an advocacy NGO nor a venture-backed startup. Policymakers reading the scenario should weigh that alongside the substance.
For operators building products that serve European enterprise customers, the scenario’s real-world uptake signals something concrete: EU procurement conversations and regulatory guidance in the next 12 to 18 months will likely include harder questions about model provenance, infrastructure location, and what happens to access when US-EU relations shift. Teams designing AI-dependent products for European verticals should document their model supply chain now and begin assessing whether a European-controlled inference option is feasible, even if American frontier models remain the performance default today.
Source: The Guardian, reporting by Aisha Down, published June 20, 2026.
