AI Infrastructure and Scarce Grid Capacity in Europe: A Five-Part Framework from Absorption Lag to Allocation and Alternative Siting

This record contains five linked working papers examining whether Europe’s AI and data-centre infrastructure may be scaling faster than the electricity systems required to support it, what follows when scarce grid capacity must be allocated, and where compute load might be sited so that it competes less directly for constrained terrestrial grid headroom.

The first paper, “Imported Compute, Delayed Grids: The AI Infrastructure Absorption Lag in Europe,” defines the concept of Infrastructure Absorption Lag: the time gap between the physical arrival and installation of high-load AI/data-centre infrastructure and the availability of grid, generation, substation, cooling, permitting, and local connection capacity required to operate it. The paper proposes a Eurostat/Comext-based Data-Centre Import Pressure Basket as an early-warning proxy for infrastructure scaling, while explicitly noting that trade data cannot directly measure electricity demand.

The second paper, “AI Infrastructure Intensity and Grid Catch-Up Capacity in Europe: Five 1–5 Year Implementation Models,” extends the concept into a technical stress-testing framework. It models five implementation pathways: Baseline Continuation, Accelerated AI Build-Out, Compressed One-Year Shock, Grid-Constrained Deployment, and Policy Catch-Up / Coordinated Absorption. The paper compares each pathway against illustrative grid-support expansion using two metrics: the Grid Absorption Ratio and the Practical Catch-Up Period.

The third paper, “Detecting Europe’s AI Infrastructure Surge: A Eurostat/Comext Import-Basket Analysis, 2024–2026,” provides the first empirical evidence layer for the framework. Using Eurostat/Comext extra-EU27 import data, it examines a seven-code core compute basket covering servers, storage, processors, memories, other integrated circuits, networking equipment, and ADP parts/accessories. The analysis finds that the basket moved from broadly flat monthly imports through 2024 and early 2025 to a marked acceleration from the second half of 2025. The increase is driven principally by processing units / servers, while reported imports are heavily concentrated in customs-clearing states, especially the Netherlands. The paper stresses that import geography is not deployment geography, and that the findings are preliminary and import-side only.

The fourth paper, “Pole Position for Power: AI Compute and the Allocation of Scarce Grid Capacity in Europe,” turns from detection and modelling to allocation. It asks who gets priority when new connection headroom, reinforcement budgets, and clean generation are scarce. The paper’s central correction is that existing essential supply — homes, hospitals, and schools — is protected by design; the real contest is over the next available megawatt of new grid capacity and clean generation. It examines how scarcity is already being allocated through national frameworks such as Ireland’s large-energy-user self-generation requirement and the Netherlands’ social-prioritisation framework, and argues that Europe lacks a harmonised rule for deciding who gets scarce capacity when AI compute, EV charging, heat pumps, housing, industry, and public-service upgrades all converge on the same constrained grid.

The fifth paper, “Siting Compute Off the Contested Grid: A Taxonomy of Alternative Hosts for Europe’s AI Load,” turns from diagnosis and allocation to siting. It asks where compute load could go so that it competes less directly for scarce terrestrial grid capacity. The paper develops a taxonomy of alternative hosts, including repurposed thermal power stations, former coal-mine workings, urban industrial shells such as gasometers, fixed offshore or seabed installations, and mobile compute vessels. Each option is assessed against a single organising test: does locating compute there ease the contested-grid problem, or merely relocate it? The paper argues that there is no single solution, but a portfolio matched to workload: power-rich, latency-tolerant sites for training; well-connected urban shells for inference and edge compute; and mobile maritime vessels only for narrow sovereign or defence-adjacent use cases where mobility and jurisdictional separation justify the overhead. The paper explicitly distinguishes grounded, emerging, and speculative options.

The central finding across the record is that Europe’s AI infrastructure risk is not determined only by the total scale of future data-centre electricity demand. It is also determined by the rate, sequencing, location, siting model, and allocation of infrastructure arrival. AI infrastructure can be imported and installed in months, while the electricity infrastructure required to operate it is planned and built over years. This creates a timing mismatch that conventional long-range demand projections may understate.

The modelling results show that identical five-year deployment totals can produce very different grid-stress outcomes depending on timing. In the Accelerated Build-Out and Compressed One-Year Shock models, both reach 9 GW of additional IT load over five years, but the compressed-shock model front-loads 5 GW into Year 1, producing much higher early absorption stress. This demonstrates that the most hazardous scenario is not necessarily the largest total deployment, but the one in which several years of infrastructure arrive before support capacity can catch up.

Scenario arithmetic shows how quickly fixed IT load becomes material electricity demand. At a PUE of 1.3, 1 GW of new IT load implies roughly 11.4 TWh/year, while 5 GW implies roughly 56.9 TWh/year. These figures are conversion illustrations, not measured estimates, but they show why gigawatt-scale additions can become significant for regional energy systems.

The empirical Comext analysis adds a first evidence layer. It shows that EU27 extra-EU imports of the core compute basket rose from broadly stable conditions in 2024 and early 2025 to a late-2025 acceleration, with year-on-year growth remaining elevated into early 2026. The strongest product-level signal is in processing units / servers, suggesting that the acceleration is concentrated in compute-relevant categories rather than spread evenly across generic electrical equipment. The analysis does not claim to measure installed capacity, electricity demand, or grid stress.

The papers also find that the problem is fundamentally local, nodal, and jurisdictional, not simply EU-wide. Data-centre loads tend to cluster around fibre routes, existing grid nodes, cool climates, favourable jurisdictions, available land, and inherited infrastructure. This means national or EU-level electricity totals may appear manageable even while specific regions, substations, or connection zones experience severe bottlenecks. It also creates a regulatory perimeter problem: capacity may be physically integrated with European demand while sitting outside, or only partially inside, emerging EU reporting, permitting, and allocation safeguards.

The Policy Catch-Up model shows that high AI infrastructure growth is not automatically unmanageable. In the modelling frame, the largest five-year deployment can become the best-managed outcome if grid planning, connection reform, renewables, storage, demand response, cooling, waste-heat reuse, and alternative siting are accelerated alongside deployment. Scale alone is not the decisive problem; unmatched rate of arrival, unmanaged allocation, and poor siting are.

This record therefore establishes a five-part sequence: conceptual framework, stress-test modelling, first empirical import evidence, governance/allocation analysis, and siting taxonomy. The quantitative schedules, grid-support tracks, and absorption-ratio thresholds are illustrative modelling assumptions, not empirical forecasts. The Comext findings are preliminary, value-based, and import-side only. The siting taxonomy is qualitative and exploratory, distinguishing between commercially deploying options, emerging configurations, and speculative concepts. Further work will require quantity-bearing extracts, CN8-level disaggregation, origin analysis, validation against EU data-centre reporting, ENTSO-E load data, transmission-system operator connection data, comparative analysis of national connection-priority frameworks, and technical assessment of alternative hosts such as repurposed power stations, mine-water systems, urban gasometer heat buffers, fixed offshore compute, and maritime platforms.

Note: This is an independent working-paper release. The analysis is conceptual, methodological, exploratory, and policy-analytic. It is supported by illustrative modelling scenarios, preliminary import-side evidence, comparative examples of national grid-allocation frameworks, and a qualitative taxonomy of alternative siting options. It does not constitute energy-market, investment, legal, engineering, or policy advice.

This record should be read alongside the related Zenodo record:
https://zenodo.org/records/20408140
which provides broader context for the energy-infrastructure and AI/digitalisation framework.

Core proposition:
A 2030 electricity problem becomes a 2026 grid-and-allocation problem if the infrastructure arrives early — and a siting problem if all of it is allowed to land on the same contested grid.

Licence: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).