Five Percent of Capacity Is All European AI

Europe accounts for just five percent of global artificial-intelligence compute capacity, and its planned projects will barely change this share.

In a policy brief of the Brussels-based Bruegel think tank, entitled How can Europe address its pressing artificial-intelligence compute infrastructure shortfall?, senior fellow Bertin Martens and analyst Tillman Schenk look for the chokepoint. In 2026, only two gigawatts of compute capacity, equating to five percent of the global total, is located in the European Union. The USA has nearly 18 times as much and China, more than twice as much. By 2031, global capacity is projected to increase more than eight-fold and the EU’s share to grow only slightly to 5.6 percent. Europe accounts for barely any significant AI models and has just one major model developer: Mistral.

Private capital is not the crucial constraint here. Speed-to-operation and grid access are more significant limitations. The proposed Cloud and AI Development Act (CADA) addresses the permitting issue but leaves grid access out of its scope. The authors suggests redirecting gigafactory subsidies to competitive grants, so that regions quicker to issue permits get more. Otherwise the data centers will queue up to the substation.

According to classical trade rules, Europe could more efficiently rent foreign compute. Geopolitics breaks this formula: the chain can be cut off by a sovereign decision in a matter of days. On top of it, a digital product is permanently connected to its producer. So computing in a foreign jurisdiction is not just lease but a political risk.

The picture is even duller as regards AI models. Europe accounts for almost none of the world’s more than 100 notable AI models. The chips are made by U.S. firms – primarily NVIDIA, with China’s Huawei making inroads.

The CADA aims to accelerate facility development but imposes sovereignty criteria that may scare away those with chips altogether. This leads to a choice: either foreign capacity on our land, or our slogans and no capacity. The current project mix chooses the latter.

Compounding this self-deception is a specific data gap. The figure of two gigawatts excludes U.S. cloud giants’ capacity already installed at European cloud data centers. There are no comprehensive datasets on these, only scattered press releases. So part of ‘domestic’ infrastructure is de facto foreign – as regards both chips and disconnection authority. The CADA barely touches on this layer. As a result, sovereignty reporting accounts for walls but not whose processor is inside and who can cut off traffic.

Private capital will not jump down this hole ahead of the grid. An unconnected data center is a warehouse with empty racks. So long as the law leaves the grid beyond its scope, the permitting competition will expedite the paperwork but not system launch. This makes a critical difference for sovereignty, as a permit can be provided in a matter of months but cable and grid power cannot.

The very logic of gigafactories is disputable as well. Money is promised to big sites, while the chokepoint proper is not box size but how soon the box gets the permit, cable, and chip. Unless tied to launch timing and to the share of processors really working in the European Union, the grant feeds an announcement. By 2031, the number of such announcements will suffice for reporting, but not for rising above five percent of global capacity. The difference between a report and sovereignty lies precisely in this gap.

Even today’s entire portfolio of projects, including gigafactories, falls short of taking the EU out of the second row by 2031. The Union’s tax revenues are not enough for the leap. Unless permitting and grid connection are accelerated, sovereignty will remain a slogan with five percent of global capacity actually hosted.

Original publication: https://www.bruegel.org/policy-brief/how-can-europe-address-its-pressing-ai-compute-infrastructure-shortfall