EU AI Infrastructure and Compute Capacity Report 2026

Europe's AI compute stack in 2026 is defined by scarcity inside growth: FLAP-D colocation reached 3.6 GW live capacity (up from 1.8 GW in 2019), CBRE forecasts vacancy compressing to 6.5% by end-2026, and EuroHPC now runs 14 supercomputers, 19 AI Factories, 10 quantum systems on a EUR 7 bn budget. Yet hyperscaler depth is uneven — Azure lists 13 EU regions but only 7 support Microsoft Foundry, and Google's H100-class A3 capacity is documented in just Belgium and the Netherlands. European cloud providers hold only 15% of the local market; the top three hyperscalers hold ~70%.

The EU AI Infrastructure and Compute Capacity Report 2026 (published 2026-04-23) maps the European AI compute stack across six interacting layers — colocation capacity, hyperscaler regions, GPU and AI-platform depth, EuroHPC public compute, sovereign-cloud controls, and energy plus regulation. It is built on 80 verified public sources and 32 reproducible indicators, with provider-by-provider region maps and access-mode comparisons.

Five structural findings: (1) Capacity is expanding but tight — FLAP-D doubled since 2019, vacancy is forecast at 6.5% by end-2026, and Ireland's data centres now consume 22% of metered electricity. (2) Cloud regions exist but AI-depth is uneven — AWS P5-class accelerators are visible in Stockholm and Spain but not Frankfurt; Google H100 is limited to Belgium; Azure Foundry covers only 7 of 13 EU regions. (3) EuroHPC is becoming an industrial-access channel — Playground (2 working days) and Fast Lane (50,000 GPU-hours, 4-day approval) target SMEs and startups. (4) Sovereignty is operational, not dominant — AWS European Sovereign Cloud went GA in Brandenburg January 2026, Microsoft expanded controls February 2026, but European providers remain at 15% market share. (5) Demand is rising faster than friction is disappearing — Eurostat shows 52.7% enterprise cloud adoption and 20% enterprise AI use in 2025.

Limitations: Cloud-region counts are not harmonised across providers. GPU documentation reflects catalogued support, not real-time stock. Several capacity figures (CBRE, JLL) are scenario-quality forecasts. AI-assisted desk research, reviewed by humans, not peer-reviewed.

Six Layers of EU AI Infrastructure

EU AI infrastructure is the combined physical, cloud, accelerator, networking, and governance layers required to develop and run AI workloads inside the Union. It spans six layers: commercial colocation, hyperscaler regions, GPU and AI-platform depth, EuroHPC public compute, sovereign-cloud controls, and energy plus regulation.

EU AI Infrastructure Scoreboard

The scoreboard compiles 32 indicators across colocation capacity, hyperscaler regions, GPU and AI-platform depth, EuroHPC public compute, sovereignty offerings, energy, and regulation. Confidence: High for Eurostat, EuroHPC, EUR-Lex and provider documentation; Medium for CBRE/JLL forecasts and Synergy market estimates.

Commercial Capacity: Tight Market Inside Growth

JLL estimates FLAP-D grew from 1.8 GW (2019) to 3.6 GW (2025). CBRE projected 871 MW of European new supply in 2025 — 34% above the prior year — but expected take-up to outpace supply, with vacancy compressing from 7.8% at end-2025 to 6.5% by end-2026. That combination — high expansion with low vacancy — is the signature of a tight market, not an underused one.

Cushman & Wakefield still classifies London, Frankfurt, Dublin, Paris, Amsterdam, Milan as primary EMEA markets; Madrid, Helsinki, Berlin, Stockholm, Warsaw, Copenhagen, Brussels, Vienna, Zaragoza fall in secondary or spotlight roles. The EU-pure read: Frankfurt, Amsterdam, Paris, Dublin, Milan remain core nodes, but the growth frontier is broader — Madrid, Warsaw, the Nordics, and parts of Southern Europe are benefiting from the search for power, land, or both.

Ireland's December 2025 conditional grid-connection reopening — requiring on-site generation or batteries capable of full demand — is the clearest signal that compute expansion is now a power-system planning problem, not just a real-estate problem.

Hyperscaler Regions: Wide Footprint, Uneven AI Depth

The public-cloud footprint across EU member states is meaningful: AWS documents 6 EU regions, Azure 13, Google Cloud 11. On a simple map, the EU is no longer underserved. On a compute-depth basis, the picture is far less even.

AWS asymmetry: Stockholm exposes one of the strongest menus (P5/P5e/P5en); Spain lists P5en; Frankfurt lists P4d/P4de but not P5 in cited excerpts; Milan shows a much lighter accelerator profile (G4dn/Inf1 only). Within one provider, advanced training-class infrastructure is not uniformly present across EU regions.

Azure asymmetry: Microsoft's region list is broad, but Foundry project creation is documented in only France Central, Germany West Central, Italy North, North Europe, Spain Central, Sweden Central, West Europe. Microsoft explicitly warns that feature and model support varies by region and that quotas are regional. Azure operates a two-layer reality: wide infrastructure fabric, narrower AI-authoring footprint.

Google asymmetry: EU regions span Belgium, Netherlands, Germany, France, Finland, Poland, Spain, Italy, Sweden. But the GPU-locations page shows Belgium and especially the Netherlands carrying the richest accelerator mixes. St. Ghislain (europe-west1) is documented with limited-capacity H100-class A3 availability. Google has also brought Cloud Run GPUs (L4) to GA in Belgium, enabling serverless inference with ~5-second cold starts for models like Gemma 3 4B. Top-tier accelerator choice still clusters in a small number of EU markets.

Country Compute Snapshots (DE, FR, NL, IE, ES, SE, FI, IT, PL, GR, LU, DK)

Provider footprints aggregate poorly to country-level capacity. The EU's actual compute geography is the intersection of hyperscaler regions, AI Factory hosts, FLAP-D / secondary colocation depth, and energy availability. Below is a compact country snapshot for the most strategically relevant member states.

Pattern: Power and AI Factory presence — not legacy IT scale — increasingly determine which member states attract AI-ready capacity in 2026. The Nordics, Iberia, and selected CEE countries are the structural beneficiaries of FLAP-D power saturation.

EuroHPC, AI Factories & Gigafactories

EuroHPC is the EU's most direct answer to commercial compute concentration. 2026 headline figures: 14 supercomputers, 19 AI Factories, 10 quantum computers, 58 R&I projects, 38 participating states, EUR 7 bn 2021-2027 budget.

Industrial-access design. The Playground route targets European SMEs and startups — access within 2 working days for short periods. Fast Lane targets HPC-familiar users — up to 50,000 GPU hours for up to 3 months, approval within 4 working days. This is a meaningful institutional attempt to reduce friction between public compute and practical AI development.

AI Gigafactory pipeline. EuroHPC defines gigafactories as facilities for trillion-parameter models. The 2025 consultation produced 76 expressions of interest spanning 16 member states and 60 sites. The result demonstrates that demand for sovereign-scale AI compute is broad and geographically distributed across the Union — even if formal calls are still pending.

Existing AI upgrades. Public procurement calls for AI upgrades to LEONARDO (September 2024) and MareNostrum 5 (July 2025) confirm the institutional shift toward AI-optimised public compute, not only classical HPC.

Energy & Power: The New Binding Constraint

Energy is now the binding constraint on EU AI compute — not chips, not capital, not regulation. The clearest single signal: Ireland's data centres consumed 22% of metered electricity in 2024 (CSO Ireland), the highest national share in the EU. EirGrid's grid-connection moratorium was lifted in December 2025 only with the explicit condition that new sites must run on-site generation or batteries capable of full demand.

Strategic consequence. AI capacity expansion in 2026-2028 will follow power, not the other way around. Nordic baseload (hydro, wind, nuclear), Iberian solar, and selected CEE markets benefit. Frankfurt and Dublin remain critical but capacity-rationed. The next procurement criterion after "is the region available?" is "is the region power-elastic?".

Sovereign Cloud: Operational, Not Dominant

Sovereignty options are increasing, but mostly inside or adjacent to existing hyperscaler ecosystems rather than replacing them. The 15% European-provider share is the most concise indicator of the structural problem.

Legal Stack: AI Act, Data Act, NIS2, DORA, EED

AI infrastructure procurement is no longer a pure technology decision. It is also a regulatory, energy, and concentration-risk decision. See also our EU AI Act Implementation Tracker 2026 for the compliance-side detail.

Citation-Ready Evidence and Research Questions

This section packages the report into citation-ready evidence blocks for infrastructure strategists, procurement teams, journalists, policy analysts, and research reuse. The key interpretive frame is: the EU is not infrastructure-poor; it is infrastructure-uneven, with power-constrained AI-ready capacity and advanced GPU depth as the binding constraints.

Research questions and direct answers

Decision rule for 2026-2028: treat AI infrastructure as a portfolio problem. Separate inference, fine-tuning, frontier training, regulated workloads, and public research/SME access; each has different optimal regions, providers, legal constraints, and power dependencies.

Recommendations by Audience

For policymakers

• Accelerate publication and usability of the EU data-centre reporting repository created under the Energy Efficiency Directive — public infrastructure planning still operates with incomplete official statistics.
• Treat AI-ready power access as a cross-border competitiveness issue, not merely a local permitting matter. Ireland's case is a leading indicator, not an outlier.
• Use EuroHPC and AI Factories as a permanent SME-access layer — extend Playground and Fast Lane funding beyond 2027.

For infrastructure providers

• Publish better region-level transparency on queueing, quotas, and accelerator-type availability. Region footprints without depth indicators create false comparability.
• Document sustainability metrics, water dependencies, and fallback options consistently as EU reporting norms tighten.
• Treat the Foundry / Vertex AI / Bedrock regional gap as a competitive issue: regional AI-platform parity is the next procurement criterion.

For strategic planners and regulated buyers

• Use a dual-track compute strategy: commercial cloud for flexible scale and managed services, plus public-EU compute channels (EuroHPC, sovereign cloud) where sovereignty, resilience, or training access matter.
• Distinguish in procurement between inference, fine-tuning, classical HPC, and frontier training — not "GPU access" as a single category. That segmentation matches the actual fragmentation visible in provider catalogues.
• Map NIS2 + DORA + Data Act exposure against your cloud architecture before mid-2026 — the obligations stack will not loosen.

Frequently Asked Questions

Version History