Europe’s difficult choices on AI

Mario Draghi argues Europe needs AI-driven growth to fund its social model plus defence and energy demands, while keeping sovereignty and values — and that it need not choose among them. The euro area–US productivity gap is estimated to have widened from $9/hour in 2018 to $21 in 2025 (cross-country level comparisons differ). ECB scenarios say fast AI adoption could add 0.3–0.4pp a year to TFP growth, which has been roughly zero since 2022. Europe controls little of the AI value chain: frontier labs cannot match US/Chinese finances and leading-edge chip production lags, creating sovereignty risk. Data is framed as Europe’s remaining sovereign zone, with a data economy that could exceed €800bn (>5% of GDP) by 2030. Caveats on capacity: EU hosts under 5% of world AI compute vs ~75% US; ordinary data-centre supply ~3GW short now and ~14GW short by 2030. Sovereign cloud can come from US operators in Europe, but sensitive uses may need European control under the Commission’s proposed tiered Cloud and AI Development Act; Amazon charges ~15% more for sovereign cloud in Germany than standard. Permitting/grid delays: data centres take ~42 months in Germany vs ~24 in the US; electricity can cost twice China’s in some regions, though Sweden is only ~10% above China. Core obstacle is fragmented demand — investors need bankable offtake while companies buy ~a year ahead; Nscale therefore sells most capacity to US buyers. ASML, Capgemini and Amadeus have committed to multiyear purchases of Mistral European Compute Units aimed at underwriting 1GW by 2030; Draghi wants larger consortiums, standardised capacity, competition among providers and progressive taxation so AI growth need not create a new tech oligarchy.