How can Europe develop AI capabilities while reducing its dependence on the US?
Europe must build shared compute infrastructure and use public procurement to anchor European AI capacity before dependency becomes impossible to reverse.
Schmidt and Prebisch both identify the same structural problem: Europe exports AI rules and imports AI capability, which transfers value outward, not inward. Hirschman adds the sharper point: cloud contracts, chip supply chains, and embedded APIs lock member states in deeper each year. Deng's record in Shenzhen shows that sequenced protection, anchor customers first, competition later, can build frontier capability that markets alone would not have produced.
Ostrom's diagnosis is the operational obstacle. Each member state has cut its own cloud deal with US hyperscalers because pooling compute sovereignty costs more than any single government will pay alone. Without binding contribution rules and graduated penalties for free-riders, a shared European compute infrastructure will not be funded.
Hayek's objection stands as a real constraint, not a veto. Brussels cannot pick which AI models will matter in 2035. But the council's majority holds that shared infrastructure and procurement commitments do not require picking winners. They require building the capacity for European actors to compete, which is a different and more defensible choice.
Confidence summary: The council reaches a majority verdict with one substantive dissent; confidence in the diagnosis is high, confidence in the precise mechanism is moderate.
1. The core argument
The most important observation from this council is not that Europe lags in AI. It is that Europe has already embedded the architecture of its own dependence. Cloud contracts signed by public institutions, chip supply chains routed through American and Taiwanese firms, and foundation model APIs wired into European research infrastructure are not temporary arrangements waiting to be renegotiated. They are path dependencies. Hirschman's reading of Colombia's infrastructure choices in the 1950s applies with force: each year of deeper integration raises the cost of exit until exit stops being a realistic option at all.
The structural position Schmidt and Prebisch identify, Europe as rule-exporter and value-importer, is not a gap that markets will spontaneously close. The platform layer, where compute, foundation models, and distribution converge, is where economic returns accumulate. Europe sits below it. Mistral's emergence since 2023 is genuine evidence that European frontier capability exists. But capability without infrastructure anchor, without procurement commitments and shared compute, does not scale into platform sovereignty. It remains a promising outlier.
2. How each member frames it
Helmut Schmidt draws the sharpest analogy from his own record. When he co-constructed the European Monetary System with Giscard d'Estaing in 1978, the goal was not to build a currency that could rival the dollar in global prestige. It was to ensure that no external monetary authority could dictate the terms of European economic life. He would apply precisely that logic to compute infrastructure today. What the reasoning card left out is the boundary condition Schmidt would impose: he would not wait for consensus among all 27 member states before acting. His model was Franco-German bilateralism as the engine, with others joining as the structure proved its value. He would counsel France and Germany to anchor a shared compute commitment and let the coalition build from there.
Raúl Prebisch provides the mechanism Schmidt's framing needs. His analysis of British trade data from 1876 to 1947 found that commodity exporters transferred real income to manufactured goods importers through the price structure of trade itself, not through tariffs or explicit extraction. The card stated the structural position; what it omitted is Prebisch's direct response to Hayek's challenge. He concedes that market knowledge is dispersed and that states cannot pick specific model architectures. But he distinguishes between picking winners and changing the terms of trade. Latin American import substitution failed, on his own later reading, where it tried to replicate specific products. It succeeded where it shifted the structural position of the exporter. Europe's target is the platform layer, not a specific model.
Friedrich Hayek presses the hardest against the majority and deserves full hearing on his own terms. His concern is not with sovereignty as a value. It is with the knowledge problem as an operational constraint. The EU AI Act's compliance costs land most heavily on European startups precisely because those startups lack the legal infrastructure that hyperscalers maintain. He reads Mistral's emergence not as evidence that targeted policy works but as evidence that competitive conditions in European research universities and private capital markets produced a challenger without Brussels selecting it. His candid limit: he does not resolve Hirschman's irreversibility question. He acknowledges that if dependency becomes structural, the case for intervention strengthens, but he doubts the intervening authority will know when that threshold has been crossed.
Albert O. Hirschman supplies the temporal argument the other members require but understate. His work in Colombian development found that infrastructure choices locked development paths for decades, and that the hiding hand, the tendency of actors to underestimate difficulty and thereby attempt projects they would otherwise abandon, sometimes produces good outcomes accidentally. He applies both sides of this here. The irreversibility risk is real and accelerating. But he also cautions against fully transparent cost accounting for a shared European compute infrastructure: if member states see the full price of coordination up front, they will rationally defect. Some institutional ambiguity about costs may be a feature, not a failure.
Elinor Ostrom reframes the problem in a way the other members do not. She does not see Europe's AI dependence primarily as a market failure or a sovereignty failure. She sees it as a commons problem, identical in structure to the irrigation systems and alpine meadows she studied across multiple continents. Individually rational bilateral deals with American hyperscalers produce collectively irrational infrastructure lock-in. Her eight design principles for durable common-pool institutions are specific: clearly defined boundaries, rules matched to local conditions, collective choice arrangements, monitoring, graduated sanctions, conflict resolution, and recognition by higher authorities. The EU AI Act provides the boundary. It provides nothing else on her list.
Deng Xiaoping offers the only model in this council where sequenced protection actually produced frontier industrial capacity. The Shenzhen special economic zone was not a market solution or a planning solution. It was a laboratory, insulated from the full competitive pressure of global markets and from the full constraint of central planning simultaneously. He is direct about the European application: the mechanism of a single government is unavailable, but anchor customers in public institutions, asymmetric procurement rules for European models, and protected procurement zones within member states can replicate the sequencing without requiring unified political authority. His pointed addition: Mistral needed a market before it needed a regulator. The EU AI Act got that sequence backward.
3. Where the council agrees
The most surprising point of agreement is between Hayek and Prebisch. Both concede, from opposite starting positions, that the structural problem is real. Hayek does not dispute that European startups face asymmetric regulatory burdens relative to American hyperscalers. Prebisch does not claim that central selection of AI model architectures is feasible or desirable. Their shared ground: the current regulatory environment unintentionally favours incumbents, and that is a problem regardless of whether the corrective is industrial policy or regulatory reform.
Beyond that, all six members accept that doing nothing accelerates irreversible dependency, that Mistral's emergence is evidence of genuine European capability requiring deliberate amplification, and that shared compute infrastructure is a more defensible intervention than picking specific model winners. Ostrom and Hirschman agree that the coordination failure among member states is the primary operational obstacle, not the technical challenge. Schmidt and Deng agree that bilateralism or zone-based leadership is a more realistic path to action than waiting for 27-member consensus.
4. Where the council splits
The real line runs between Hayek and the rest on whether sequenced protection produces capability or entrenches mediocrity. Hayek holds that the EU AI Act already demonstrates how well-intentioned regulatory boundaries can punish European challengers more than American incumbents, and that adding procurement preferences and contribution mandates compounds that error. Deng, Prebisch, and Schmidt hold that Shenzhen, South Korean semiconductor development, and Taiwan's TSMC trajectory are counter-evidence: states that used asymmetric protection in sequence built frontier capability that markets alone would not have produced. Neither side is wrong about its own historical record. The dispute is about which analogy is closer to Europe's actual situation in 2026, and that is a judgment call, not a resolvable empirical question.
5. For a policymaker to decide on
The council cannot resolve this choice: should Europe establish a binding common compute fund, with mandatory member state contributions and graduated penalties for free-riders, using public procurement to anchor European AI capacity at the cost of near-term efficiency losses and political conflict over burden-sharing? Or should Europe instead focus on removing the regulatory asymmetries that currently disadvantage European startups relative to hyperscalers, accepting that the platform layer remains American-dominated for a generation while European challengers compete on level terms? Both paths are coherent. Only the policymaker can decide which dependency is more tolerable.