France produced 95.2% of its electricity from low-carbon sources in 2025, but its cables and transformation substations date from before the smartphone era. This advance in the energy mix masks local connection constraints that data centers are intensifying, especially on transport networks and in certain geographic areas. Ireland faced strong growth in data center demand and restricted and conditioned new connections, particularly in constrained areas around Dublin, without imposing a general moratorium. France still has a choice.

The Essentials

  • France produced 521.1 TWh of low-carbon electricity in 2025, representing 95.2% of its mix, but distribution remains an infrastructure bottleneck.
  • Data center electrical consumption is expected to quadruple by 2035 according to ADEME projections, putting strain on networks designed for different demand.
  • In Ireland, data centers accounted for 35% of national electricity growth, saturating networks before the energy mix became an issue.
  • RTE is investing in the transport network, but regional distribution networks are modernizing at an insufficient pace relative to accelerating digital demand.
  • For the next decade, the challenge is to deliver electricity where data centers want to establish themselves, and to define the conditions for doing so.

521 TWh of Clean Power Is Useless If Cables Saturate

2025 was a good year for French electricity. The nuclear fleet recovered its pre-2022 availability levels, hydropower benefited from adequate rainfall, and solar continued its upward trajectory. According to RTE, 521.1 TWh of low-carbon electricity, nuclear and renewable, was produced in France in 2025. The figure is impressive. It has been cited in European debates as proof that decarbonization and supply security can coexist.

But a clean energy mix guarantees nothing as long as the electricity produced cannot be delivered to the point of consumption. And that is precisely where France is beginning to feel the strain.

Data centers consume continuously, twenty-four hours a day, seven days a week, with very high power peaks across reduced land areas. A large data center can require 100 to 300 MW on a single site, as much as a medium-sized city. This type of concentrated consumption requires connections to the high-voltage transport network, not the standard distribution network.

High-voltage substations capable of handling these loads are rare, their construction takes several years, and their geographic location does not always coincide with the areas where data center operators wish to establish themselves.

The Irish Lesson

Ireland offers the most documented case study. Dublin became Western Europe’s data center hub within a decade, attracted by tax incentives, Atlantic submarine cables, and good connectivity. The result: in 2024, data centers represented more than 20% of the country’s total electricity consumption, a level Ireland had never anticipated. EirGrid, the Irish network operator, warned of rapid data center growth and capacity constraints in the Dublin region; between 2023 and 2024, data centers accounted for approximately 48% of the net increase in measured electricity consumption. The Irish regulator, the CRU, introduced stricter connection criteria; this policy is part of a broader government framework, but the connection restrictions are formally the responsibility of the CRU.

Irish connection restrictions respond particularly to physical network constraints in high-concentration areas, without constituting a general moratorium. The signal is therefore transferable to any country, including those producing irreproachable electricity from a climate perspective.

France is not Ireland. Its transport network, managed by RTE, is one of Europe’s most interconnected, with 105,000 km of high and very-high-voltage lines. The distribution network has historically been designed for widely dispersed demand, but it is undergoing reinforcement and modernization; certain areas may nonetheless require substantial work to accommodate very high concentrated loads. The interface between the transport network and the distribution network can become a bottleneck for data center projects, particularly for intermediate-sized installations that do not justify a direct connection to very-high-voltage lines.

Digital Demand Exceeds Historical Forecasts

In its baseline scenario, ADEME estimates the possible multiplication by 3.7 of consumption induced by French digital uses by 2035, with a large portion potentially located outside France. This figure deserves to be unpacked, because it covers very different realities.

Part of this growth is structural: the digitization of the economy, data storage, streaming video, cloud services. This part was foreseeable and has been integrated into RTE scenarios for several years. What was less foreseeable was the acceleration linked to generative artificial intelligence. Training large language models and real-time inference are extremely energy-intensive processes. Exact data are not publicly available; an IEA estimate places GPT-4 training around 42 GWh, not several hundred GWh.

The GPU clusters running these models operate at full capacity continuously.

This demand does not follow the traditional geography of industry. It seeks affordable land, available network capacity, and low-carbon electricity to satisfy the environmental commitments of major technology companies. France ticks two of three boxes: land exists outside major metropolitan areas, and electricity is decarbonized. Available network capacity is the constraining variable.

Enedis has undertaken an ambitious investment plan for the coming years, but connection lead times remain long for a large industrial site, in a context where data center operators plan deployment cycles of eighteen months. This time lag is one of the most concrete obstacles to hosting these installations on French territory. A complementary analysis of these infrastructure dynamics can be found in the article The Factory as Battery, which discusses industrial flexibility as a network balancing lever.

RTE Invests, But the Timeline Is Tight

RTE published in 2025 a network development plan providing for substantial investments over the next ten years, responding in particular to rising digital demand and electrification of uses. These investments cover the reinforcement of existing lines, the construction of new transformation substations, and the improvement of regional interconnections.

The program is real. But it faces two structural obstacles. The first is regulatory: in France, the construction of a high-voltage line or transformation substation requires public inquiry and environmental authorization procedures that extend over several years. The second is industrial: the electrical engineering sector lacks qualified labor, and the order books of specialized companies are saturated at the European scale, in a context where all developed countries are simultaneously investing in their networks.

This double bottleneck, administrative and industrial, means that even an investment decided today will take time to produce effects on the ground. Enedis proposes, in coordination with RTE, an indicative mapping of network capacities; RTE also develops anticipated connections and identified sites with the State for certain very large consumers. This is a pragmatic approach, but it assumes operators play along, which they do not always do spontaneously, preferring to establish themselves where land is cheapest rather than where the network is most ready.

Risks of Unmanaged Localization

The Irish experience illustrates that the concentration of data center projects in Greater Dublin can quickly create capacity constraints and justifies coordination of localization with network operators. In France, Île-de-France already concentrates a very significant share of the existing fleet, through proximity to corporate headquarters and long-distance fiber optic cables. This geographic concentration increases connection constraints in certain areas; the existence of truly unused capacity elsewhere must be established zone by zone.

Geographic dispersal of data centers would be technically possible. Large cloud applications tolerate latencies of a few tens of milliseconds, which allows localization several hundred kilometers from end users. The applications most sensitive to latency—high-frequency trading, some industrial applications—constitute a minority of the market. The rest of the sector could establish itself in regions where the network is less loaded, provided it finds adequate digital connectivity infrastructure.

This is where coordinated public policy can make a difference. Several European countries have begun conditioning aid for data center establishment on their localization in zones with strong network availability. Sweden and the Netherlands have developed public maps of priority connection areas. France has the tools to do the same; RTE already publishes data on network reception capacities by zone, but the translation of this data into territorial planning policy remains embryonic. This tension between productive sovereignty and dependence on private investors is at the heart of the debate on industrial decarbonization that other major economies are also traversing.

How Far Will Demand Go, and Who Captures the Value

According to ADEME’s baseline scenario, consumption induced by French digital uses could be multiplied by 3.7 by 2035, with a large portion potentially located abroad; RTE estimates that current industrial projects awaiting connection, notably data centers, represent demand equivalent to nearly three times current consumption of the entire French industrial sector.

First scenario: France chooses not to manage the situation and lets the market decide locations. Data centers establish themselves where conditions are met, essentially in Île-de-France and a few connected regional metropolitan areas. The network saturates quickly in these zones, RTE is forced to invest urgently to serve unanticipated demand concentrations, and connection costs explode. Operators, who negotiated network access rates before saturation, capture most of the value. Local authorities and taxpayers absorb network reinforcement costs without necessarily benefiting from tax revenues, often optimized at the European scale.

Second scenario: ADEME’s scenario 2 frames locations in consultation with territories and prioritizes uses according to their social, environmental, or health utility; it does not contain, in the examined source, specific conditions for local employment and network cost contribution. In this case, data centers disperse more widely, RTE investments are better anticipated and better distributed, and peripheral regions capture part of the economic benefits. This scenario assumes coordination between RTE, Enedis, local authorities, and the State that exists in institutional frameworks, but whose specific adaptation to very large data centers may need reinforcement.

Third scenario: computational demand restructures around less centralized architectures. Researchers and actors like Lê Nguyên Hoang (Tournesol) advocate for pluralistic AI model governance, which potentially implies different distribution of computational loads. If tomorrow’s AI models are smaller, more distributed, trained on federated data rather than centralized, network pressure would be reduced. This scenario remains hypothetical; current trends point in the opposite direction, toward ever-larger models, but it depends in part on industrial policy and regulatory choices that Europe still has the capacity to influence.

In all three cases, one signal should be closely monitored: connection lead times. If data center connection requests begin generating lead times exceeding five years outside Île-de-France, the bottleneck becomes structural and difficult to resolve within the decade’s timeframe. Conversely, if RTE succeeds in reducing these lead times through procedure simplification and proactive spatial planning, France could transform its energy mix advantage into real competitive advantage for hosting European digital infrastructure.

Digital Sobriety as an Adjustment Variable

One element is often missing from this debate: computational demand is not entirely exogenous. A significant portion of data center consumption is linked to uses whose energy efficiency can be improved. Major operators, Google, Microsoft, Amazon, have made considerable progress on PUE (Power Usage Effectiveness), the indicator measuring what share of a data center’s total energy is actually devoted to computation rather than cooling. The best centers now achieve a PUE close to 1.1, compared to 2 or higher for aging centers.

But these efficiency gains have been absorbed by growth in demand. This is the classic rebound effect. ADEME has documented this phenomenon for digital technology in France: the efficiency gains per unit of computation are real, but the multiplication of uses far outweighs them. This does not argue for curbing uses, but for reflecting on the incentives influencing operators’ technological choices: a well-constructed carbon price signal, network connection pricing that truly reflects infrastructure costs, and minimum energy efficiency standards for new data centers could help slow demand growth without prohibiting any use. Environmental data and their lucid reading show that these efficiency levers exist and can be activated without catastrophism.

France possesses a rare asset in Europe: near-entirely decarbonized electricity production, a network operator capable of long-term planning, and a nuclear industry providing stable baseload power. These strengths can contribute to French territory’s attractiveness for data centers seeking to reduce their carbon emissions. This advantage can translate into voluntarist industrial policy, guided localization, incentive pricing, simplified connection procedures, or dissipate through regulatory inaction. Ireland showed what the second option costs.

Sources

  1. RTE – Annual Review 2025: French Electricity Generation
  2. ADEME – Assessment of Data Center Energy Consumption in France (2026)
  3. EirGrid – Annual Renewable Energy Report 2024 (Ireland, data center data and moratorium)
  4. Enedis – Distribution Network Development Plan 2024-2028