France’s transmission operator had reserved nearly 18 gigawatts of grid connection capacity for about 80 data-center projects by May 2026, up from roughly 5 GW for 40 projects at the end of 2024. The rapid growth of this RTE project queue shows the scale of developers’ plans. It does not measure electricity being consumed now.

That distinction matters as France decides where new centers can connect and how to account for their environmental effects. A Senate committee report adopted on September 16 calls for faster connections to AI facilities alongside stronger environmental measurement. The two aims depend on knowing which projects become operating sites and how much power those sites actually draw.

The big change

  • What changed: France’s reserved transmission capacity for data-center projects has risen from about 5 GW at the end of 2024 to nearly 18 GW in May 2026. The grid must plan for requests whose eventual use is uncertain.
  • Why it matters: Reserving a connection can shape grid planning years before a building is fully equipped. A resident or policymaker assessing an infrastructure proposal needs its expected operating demand and local effects, not a national reservation total presented as present consumption.
  • What to watch: RTE’s project queue, commissioning and measured load will show how much capacity becomes demand. France’s existing reporting rules and the Senate’s proposed AI disclosure standard address a related gap: facility electricity data alone cannot identify the share used by AI.

Three different numbers

A gigawatt measures power: how much a connection could draw at a given moment. A terawatt-hour measures energy used over time. Turning a connection rating into annual electricity consumption would require assumptions about which projects open, when they open and how heavily they operate. RTE says its 18 GW figure is capacity reserved for projects. It expects some announced projects not to proceed and others to increase consumption gradually.

RTE estimates that existing French data centers use about 10 TWh a year, around 2% of the country’s annual electricity consumption. It says only eight centers are connected directly to its high-voltage network, with a combined connection rating of 800 MW; most existing sites connect to distribution networks. These figures cover data centers as a sector, including cloud, storage and other computing, rather than AI alone.

There is also uncertainty in the baseline. A government statistical study published in July identified 471 larger data centers that used 4.3 TWh in 2024. Allowing for smaller sites, it estimated 4.3–6 TWh for metropolitan France that year. RTE’s roughly 10 TWh is a more current sector estimate, while the government study is based on identified 2024 consumption and an allowance for smaller sites. The published figures have different dates and methods, and the material reviewed here does not reconcile them. Neither is a meter reading for the 18 GW queue.

RTE offers a narrower observation about how new connections develop. Centers connected to its network over the past two to three years currently use, on average, about 20% of the power they requested, it says. The operator reports that developers expect a gradual build-out over 10–15 years. That 20% describes a subset of recently connected centers at a particular stage; applying it to all pending projects would manufacture a national load estimate.

The planning question behind the queue

The Senate report recommends prioritizing AI data-center connections and changing RTE’s order from “first arrived, first served” toward “first ready, first served.” Its stated aim is to remove projects that occupy queue positions without proceeding. That recommendation recognizes a planning problem inside the headline capacity figure: a reserved place is still several steps from an operating facility. It is a proposal, not an adopted connection rule.

RTE’s 2030–2035 outlook puts data-center electricity use at 15–20 TWh in 2030 and 23–28 TWh in 2035 under a rapid-decarbonization trajectory. The operator explicitly assumes that some announced projects will not materialize and that operating sites will ramp up over time. ADEME’s January scenario study asks a different question: how electricity use attributable to French digital activity could evolve across facilities in France and abroad. Its business-as-usual scenario reaches 3.7 times the current level by 2035, with nearly two-thirds of that scenario’s consumption abroad. These are modeled paths with different boundaries, not observed future demand or figures to add together.

Measuring the environmental effect

Connection capacity alone gives no site’s annual electricity use, cooling-water intake, emissions or mix of AI and other workloads. France’s data-center reporting rules require operators with installed IT power at or above 500 kW to report measures including total energy, water input and reused waste heat. The ministry says individual submissions to the European platform may be confidential. Facility reporting helps track operations, but it does not by itself assign each facility’s footprint to AI services.

The Senate’s 16-recommendation report says environmental data for AI remain incomplete. It proposes that France’s telecom regulator define environmental data that language-model suppliers would publish, and calls for a common European method for assessing large models. These are recommendations, not disclosures already available to compare models. The report also seeks quicker connections and better environmental performance at data-center sites, placing both demands on the same expansion.

In a September 27 opinion column, Le Monde journalist Stéphane Foucart argues that attention to hypothetical AI catastrophes distracts from environmental and social costs. His argument is a perspective on why to scrutinize these costs; the queue, consumption estimates and reporting rules above come from the cited official records. BIG CHANGE did not meter a data center or calculate an AI-only share. The useful next comparison is between reserved capacity, completed connections and reported use as those records develop.

Sources & further reading