브루겔 싱크탱크는 AI 반도체 기가팩토리에 할당된 EU 자금을 데이터센터 허가 및 전력망 연결 인프라 가속화로 재정향할 것을 제안했다.

The European Union should redirect funding for artificial intelligence gigafactories toward grants that reward faster permitting and grid connections for data centers, according to researchers Bertin Martens and Tillman Schenk of Bruegel. Their report argues that the primary constraint on expanding Europe's computing capacity is not capital availability but the time required to bring facilities online. Public funding should therefore target the institutional bottlenecks that delay projects, they contend.

**Current capacity gap**

The EU will account for approximately 5% of global AI computing capacity in 2026 and 5.6% by 2031, though these figures have a significant limitation: they track projects of at least 10 megawatts and exclude AI capacity in the existing cloud centers of major US providers in the EU, for which complete data are unavailable. By 2026, the EU is projected to have approximately 2 gigawatts of AI capacity, compared with about 35 GW in the United States and 5 GW in China. Even under scenarios of rapid growth, projects under consideration would bring European capacity to almost 21 GW in 2031—but because infrastructure is expanding globally, the EU's share would remain at only 5.6%.

**Private capital is not the constraint**

The case against large-scale construction subsidies begins with the projects already planned. In the inventory analyzed by Martens and Schenk, 76 of 101 projects are fully privately financed and account for 84% of planned capacity. The five gigafactories included in the report's scenario would add approximately 750 MW, representing nearly 4% of Europe's projected 2031 capacity. "Private capital is not the decisive constraint on expanding Europe's AI computing capacity," the authors write.

The European Commission justifies support for gigafactories partly by citing the need to ensure access for researchers, small and medium-sized enterprises, and the public sector. InvestAI, launched in February 2025, aims to mobilize €200 billion for AI, including a €20 billion European fund for gigafactories. Martens and Schenk challenge this rationale, proposing instead that researchers and public services could receive subsidies to rent capacity from private operators, while retaining an exception for strategic activities.

**Permitting delays carry substantial costs**

The economic case for their proposal centers on the cost of delays. A Carnegie financial model for a hypothetical 100 MW center shows that a one-year delay in starting operations reduces lifetime value by more than 5.5%—a steeper penalty than even doubling energy prices, which reduces value by 4.5%. In practice, projects require approvals from multiple authorities governing land, environment, construction, and electricity. The Commission's CADA proposal seeks to coordinate these procedures through data-center acceleration zones and one-stop information points, with a permitting deadline of no more than 12 months from submission of a complete application. However, the aggregated basic permit excludes grid-connection permits, which Martens and Schenk see as a remaining bottleneck.

The researchers propose that regions compete for grants based on permitting speed, with larger sums for faster approvals. This mechanism would redirect gigafactory funding to institutional coordination, while grid investments should be prepared in advance. The solution reflects a fundamental disagreement with the Commission: whether access to computing should be purchased through infrastructure subsidies or through grants enabling its rental.

**Sovereignty and transparency concerns**

Europe's case for locating computing capacity domestically rests on digital sovereignty—installing infrastructure in the EU gives authorities greater regulatory options. However, this does not eliminate dependencies; access to models could be restricted regardless of equipment location. The Commission's CADA proposal addresses this through four levels of cloud-sovereignty assurance with progressively stricter requirements on data location, provider independence, and control.

Martens and Schenk also recommend a voluntary label for AI services processed in the EU, supported by external audits, to prevent services partly processed outside the Union from being marketed as European. They call for greater transparency about chips and capacity already installed in Europe to establish a clearer baseline for assessing the gap between European demand and available infrastructure.

**Implementation timeline**

The Bruegel report is dated September 2026. CADA was proposed by the Commission in June and requires adoption by the European Parliament and the Council; its measures are not yet binding. The Commission estimates that construction of the first facility will begin in 2027, though this timeline remains preliminary and announced projects do not guarantee computing power available to users.