Europe's strict AI rules can be a credible tool to shield domestic firms — but only in a narrow zone of foreign advantage; if US AI dominance is too large or rules bite evenly, regulation risks cutting Europe off from critical technologies.

Main Finding

Stringent AI regulation (as exemplified by the EU AI Act) can be a rational instrument for asserting EU digital sovereignty only under specific conditions: (1) the regulation must impose relatively larger efficiency costs on foreign (US) AI providers than on domestic firms; and (2) the US technological advantage in the EU market must be substantial but not so large that foreign firms dominate even after regulation. Outside that intermediate range, stringent rules either unnecessarily reduce technological capability available in the EU or fail to shift market share toward EU firms.

Key Points

• Research question: When is strict AI regulation a rational strategy for the EU to pursue digital sovereignty, given technological disparities and geopolitical relations with the US?
• Conceptual trade-off: Regulators weigh (a) domestic market share for home AI firms (a proxy for digital sovereignty) against (b) algorithmic performance available to EU users (which may come from foreign providers).
• Model structure: two-stage game
  • Stage 1: EU regulator chooses regulatory stringency θ ∈ {High, Low}.
  • Stage 2: Representative EU consumer allocates unit attention/time between a domestic and a US AI firm; firms’ post-regulation algorithmic efficiencies determine market shares.
• Key model parameters and interpretations:
  • τEU, τUS: pre-regulation algorithmic efficiencies of EU and US firms (τUS normalized by τEU used as relative advantage).
  • γEU, γUS ∈ (0,1): multiplicative post-regulation efficiency factors (how much regulation reduces each side’s efficiency).
  • s ∈ (0,1): regulator’s weight on market share / sovereignty relative to technological performance.
  • σ: elasticity of substitution between the two AI offerings (controls product differentiation).
• Central analytical insight: High regulation helps EU sovereignty when it causes a favorable redistribution of market share toward domestic firms (i.e., it is relatively more binding on foreign providers) and when foreign advantage τUS is large enough that leaving rules lax would yield domination by foreign firms, but not so large that even regulated foreign firms outcompete domestic ones.
• Strategic sensitivity:
  • If γUS is much smaller (regulation hurts US firms more), stringent regulation is more attractive.
  • If τUS is either very low (no need for protection) or very high (protection insufficient), stringent regulation is unlikely to be optimal.
  • The regulator’s valuation s matters: higher preference for sovereignty broadens the parameter range where strict rules are chosen.
  • Transatlantic cooperation (or the lack of it) and access to foreign cloud/compute can alter τUS and γ parameters, changing the incentives.
• Application: The authors map these results to the EU AI Act’s general-purpose AI provisions and show the Act’s success in strengthening sovereignty hinges on its redistributive effects and on the evolving US–EU technological and geopolitical balance.

Data & Methods

• Purely theoretical, analytical study — no empirical dataset.
• Methods:
  • Constructed a two-stage, dynamic game with perfect information.
  • Consumer choice modeled as allocation of scarce attention/time between two differentiated AI services; monopolistic-competition-style utility (adapted Dixit–Stiglitz framework) captures market power and near-zero marginal costs of AI.
  • Post-regulation algorithmic efficiencies modeled multiplicatively (αEU(θ)=γEU·τEU, αUS(θ)=γUS·τUS for high stringency; equal to τ’s under low stringency).
  • EU regulator’s payoff is a convex combination of domestic market share (sovereignty) and domestic algorithmic efficiency (innovation capacity), parameterized by s; tie-breaking lexicographically favors stringent regulation (reflecting unmodelled consumer-protection priorities).
  • Solution concept: subgame-perfect Nash equilibrium found by backward induction; results formalized in lemmas and propositions (e.g., closed-form expression for EU market share given θ).
• Limitations of method explicitly acknowledged by authors:
  • Simplified two-firm (EU vs US) framework and representative consumer.
  • Static framework — no explicit dynamic innovation or investment path dependence.
  • Abstracted away enforcement costs, multi-country strategic interactions beyond the EU–US dyad, and heterogeneity among firms.

Implications for AI Economics

• Regulatory strategy is endogenous to the technological balance and geopolitical context: using regulation to pursue digital sovereignty is not unambiguously welfare-improving and can act as de facto protectionism when calibrated to redistribute market power.
• The “Brussels Effect” is weaker in AI than in data protection: because AI depends on concentrated compute, data, and cross-border platforms, the EU’s market power will not automatically convert into global standard-setting unless regulation meaningfully shifts market shares and foreign providers face asymmetric compliance costs.
• Policy complementarities are essential: regulation alone may harm domestic technological capabilities (by reducing available performance); to avoid adverse outcomes, regulation should be paired with industrial policy (investment in compute infrastructure, semiconductors, cloud sovereignty, talent) to raise τEU and lower dependence on foreign tech.
• Calibration matters: well-targeted rules that disproportionately constrain foreign entrants (or that are easier for domestic firms to comply with) can achieve sovereignty goals withoutas large a loss in technological performance. Blanket, indiscriminate stringency risks technological isolation and consumer-welfare losses.
• Geopolitical coordination reduces the trade-off: transatlantic regulatory cooperation can expand the parameter region where strict standards both protect sovereignty and preserve technological access, by harmonizing compliance costs and enabling mutually beneficial data/compute arrangements.
• For market structure and competition analyses: AI regulation can act as a barrier to entry or as a leveling tool depending on asymmetries in compliance capacity — resulting impacts on market concentration, innovation incentives, and consumer surplus should be modeled endogenously in future empirical work.
• Research agenda suggested:
  • Empirical calibration of τ and γ parameters (e.g., measuring how specific provisions in the EU AI Act affect foreign vs domestic model performance).
  • Dynamic models allowing investment responses by firms (R&D, data acquisition, relocation of compute).
  • Multi-jurisdiction models including China and other major markets to study global standard competition and spillovers.
  • Welfare analyses that incorporate consumer surplus changes, enforcement costs, and political-economy coalitions.

Limitations and caveats to bear in mind: the conclusions derive from an abstract two-player model with simplifying assumptions (representative consumer, no dynamic innovation), so empirical validation and richer multi-agent modeling are needed before applying the results for concrete policy prescriptions.