AI leaders describe very fast growth in demand from agents: programs that chain requests continuously for users and businesses. Behind the statements, a structural shift is visible: inference is taking a growing share of compute demand. That could reshape France's data center map.

DOES YOUR SITUATION LOOK LIKE ONE OF THESE?

A small 2 MW regional AI data center: what building to look for, and what power to check before visiting?We want to convert an 8,000 m² logistics warehouse into a data center. Slab, clear height, power: will it work?The primary substation 4 km away shows capacity, but Enedis mentions reinforcement. Is 12 MW at MV (HTA) realistic?
Xavier WatrelosXavier Watrelos
HV grid connection & industrial projects
Anis MokdedAnis Mokded
Design, commissioning & operationsAll situations →

1. TRAINING AND INFERENCE HAVE DIFFERENT NEEDS

Training runs in campaigns, in very large campuses, and tolerates distance from users. Inference is persistent and benefits from proximity to users, fibre and enterprise customers.

Projections cited by Hitachi Energy (from McKinsey, global) Total data center demand: about 82 GW in 2025, 219 GW in 2030
Inference: about 21 GW in 2025, 93 GW in 2030
Training: about 23 GW in 2025, 62 GW in 2030
Typical sizes: training campuses from 200 MW to multi-GW, inference hubs of 20 to 100 MW, edge nodes below 10 MW

2. WHAT IT COULD MEAN IN FRANCE

Large campuses will still be needed. But if inference dominates, a growing share of projects could be 10 to 50 MW around major metros: Paris region, Lyon, Marseille, Lille, Toulouse, Bordeaux, Nantes. This is a hypothesis, not a certainty.

3. THE BOTTLENECK SHIFTS

Grid connection. Up to a few tens of megawatts, a data center usually connects to Enedis' distribution grid (MV). In dense areas, substations are often already in demand: part of the bottleneck moves from transmission to distribution. See 20 to 30 MW data centers and the 40 MW threshold and MV and HV grid connection.

Land and buildings. Near cities, land is scarce: converting warehouses, brownfields or industrial buildings becomes a serious option.

Local acceptance. More projects closer to residents mean more questions on noise, water and power. See data center water use in France.

4. THE TAKEAWAY

AI agents do not remove the power problem: they move it to more numerous, smaller sites closer to cities. The questions stay the same at a different scale: which capacity at which substation, by when, in which building, with what local reception. See also AI factory in France.

Sources: Hitachi Energy, "AI data centers 2030-2035: the power shift" (June 2026), citing McKinsey projections; Techniques de l'Ingénieur on lengthening connection delays (March 2026). This article is an analysis: the projections cited belong to their authors and may change. General information.

FREQUENTLY ASKED QUESTIONS

What is the difference between training and inference for data centers?

Training runs in campaigns in very large campuses and tolerates distance. Inference is persistent and benefits from proximity to users, favouring smaller sites near metros.

How does a 10 to 50 MW data center connect in France?

Usually to Enedis' distribution grid at medium voltage, up to a few tens of megawatts. Above that, connection falls to RTE's transmission grid.