Bruegel, the Brussels economic institute, put EU artificial intelligence (AI) compute capacity at 2GW in 2026, or 5% of global capacity, against 78% for the United States and 11% for China. The figures come from a policy brief by Bertin Martens and Tillman Schenk published on 10 September 20261. A gigawatt powers between 450,000 and 750,000 GPUs on Bruegel's own reckoning, depending on efficiency, so the EU's entire 2026 capacity comes to somewhere between 900,000 and 1.5 million of them. On the authors' projection the bloc reaches 20.89GW by 2031 and still holds 5.6% of world capacity, because everyone else builds too.
Martens and Schenk argue that capital is not what holds Europe back. On figures Bruegel relays from Phillips-Robins et al at the Carnegie Endowment, German data centres take 42 months from permit to operation, against 24 months in the United States23. The delay cost comes from a model rather than a measurement. Carnegie's discounted cashflow run covers an illustrative 100MW greenfield site over twelve years, with GPU rental held at $10 an hour and 70% utilisation in every country it tests. Modelled on those assumptions, a one-year slip costs more than 5.5% of lifecycle value, against 4.5% for a doubling of energy prices. A government weighing an energy subsidy against a planning reform is being told, on those modelled numbers, that the planning reform pays better, despite energy costs dominating the European debate.
That lands on Europe's flagship answer to the shortfall. EuroHPC opened its €4.12bn AI Gigafactories call on 30 July, targeting €20bn of private capital across seven sites4. Despite that headline, Bruegel calculates that the five planned Gigafactories add roughly 750MW between them, about 4% of the EU's own projected 2031 total. Bruegel also runs the experiment directly: accelerating the planned pipeline by one year raises the EU's 2031 share from 5.6% to 6.7%5. Speed buys about one percentage point, and even that leaves Europe near six per cent of world capacity while Washington and Beijing keep building.
Bruegel argues a position rather than reporting a measurement, and the position is contested inside a sovereignty debate. Its prescription runs to four items: redirect subsidies towards the fastest-permitting regions, create EU-wide data-centre acceleration zones, put grid investment into the Cloud and AI Development Act, and attract money from the US hyperscalers, the few American firms running cloud computing at global scale, through faster permits rather than block it. That last item asks Europe to buy its way out of the gap with American capital, and the evidence on delay costs comes from an American think tank paper titled "The compute coalition: how to build the future of AI in the free world"6. A Bruegel economist, Mario Mariniello, made a related argument in June: that the sovereignty package mimicked its rivals while European providers held 15% of their own cloud market.
