On 21 May 2026 Arcep reported that the French datacenters it surveys drew 575,000 m3 of water in 2024, almost all of it drinking water. Counting the water behind the electricity they consume, the total comes close to 6.5 million m3.
A French datacenter draws little water itself. The water, the heat and the years it takes to get a grid connection are where compute becomes an industrial problem, and cooling engineering is the job that touches all three.
This dossier reads the water first, then the machines that are driving the change, then the choice that every new site makes between saving energy and saving water. It then reads the grid calendar and the European law on waste heat, which together decide where the next centres can be built. It ends with the jobs this creates, and with what the evidence does not show.
The visible water is the small part
Arcep, the French regulator for electronic communications, publishes an annual survey called "Pour un numérique soutenable", now in its fifth edition. The fifth edition, published on 21 May 2026, covers the year 2024. On water, it tells two stories at once.
The first story is the one usually told. The centres in the survey drew 575,000 m3 of water directly in 2024, and almost all of it was drinking water taken from the public network. That figure fell by 15% in 2024, after rising by 17% in 2022 and by 16% in 2023. A reader who stopped there would conclude that the sector had turned a corner.
The second story is larger. Producing electricity also takes water. When Arcep adds the water behind the electricity the centres consume to the water they draw on site, the total comes close to 6.5 million m3, up about 8%1 over the year. The direct draw fell while the full footprint grew.
Most of a datacenter’s water sits behind the electricity meter
- Drawn on site: 575,000 m3 m3 of water, surveyed French datacenters, 2024
- On site plus behind the electricity: close to 6.5 million m3 m3 of water, surveyed French datacenters, 2024
m3 of water, surveyed French datacenters, 2024
Source 1 Arcep
Our own arithmetic on Arcep's two figures puts the direct share at under a tenth of the total. That ratio matters for anyone who judges a site by its water meter. A centre can cut the water it draws and still raise the water its operation requires, simply by consuming more electricity. The meter on the wall is the part a town sees. The meter at the power station is the part that grows.
The same survey explains why electricity keeps growing. The centres consumed 2.7 TWh in 2024, up 12%, and their IT equipment alone consumed 1.9 TWh, up 15%. The installed capacity of the surveyed centres reached close to 600 MW, up 14%. None of these figures is alarming on a national scale. Together they show a sector whose load rises faster than its efficiency gains.
The machines are getting bigger and denser
The growth is not spread evenly. It is concentrated in new, large sites. Arcep reports that the centres that opened in 2024 averaged 20 MW of power, against 9 MW for those that opened in 2023 and 3 MW for older centres. A new site is no longer a server room. It is an industrial building with a power draw close to that of a factory.
Size brings efficiency. The usual measure is PUE, the ratio of all the energy a site consumes to the energy its IT equipment consumes. A PUE of 1.00 would mean that no energy at all goes to cooling, lighting or power conversion. Arcep reports an average of 1.30 for centres built between 2014 and 2023, against 1.52 for those built before 2014. Across all surveyed centres the average improved from 1.46 in 2023 to 1.42 in 2024.
The newest centres are far larger and use less overhead energy
| Measure | Older centres | Newer centres |
|---|---|---|
| Average power | 3 MW (older centres) | 20 MW (opened in 2024) |
| PUE by build period | 1.52 (before 2014) | 1.30 (2014 to 2023) |
| PUE, all surveyed centres | 1.46 (2023) | 1.42 (2024) |
Source 1 Arcep
The geography is concentrated too. Île-de-France holds 56% of the surveyed centres and over 70% of their capacity. Three regions together hold 90% of the centres. Any argument about water and heat in French datacenters is first an argument about the Paris region, and only then about the rest of the country.
Density is the other half of the change. Schneider Electric, launching its liquid cooling range, argues that AI chips are getting hotter and denser and that such rooms require liquid cooling. Schneider Electric, a French group, completed the purchase of a 75% controlling interest in Motivair, a company specialised in liquid cooling for high-performance computing, on 28 February 2025. It expects to acquire the remaining 25% in 2028. On 29 September 2025 it presented a liquid cooling range with Motivair: coolant distribution units, rear-door heat exchangers, cold plates and chillers for high-density AI rooms. Liquid cooling is now sold as a product line. It is no longer a laboratory option.
Saving energy and saving water pull against each other
Every new site faces a trade-off that the headline figures hide. Outside air can cool a building, and evaporating water into that air cools it further. This is adiabatic cooling, one form of free cooling. Arcep notes that it improves energy efficiency but uses more water. Close to 20%1 of the surveyed centres use cooling systems that consume water, and a quarter of the centres opened in the last four years do.
That is the tension. The measure the industry reports most, PUE, rewards the choice that uses more water. Arcep's own note on adiabatic free cooling means that a site can report a better PUE and a higher water draw at the same time. Neither figure is false. Each one answers a different question.
The French state has taken a side where water is scarce. In its November 2025 guide to siting datacenters, the Direction générale des entreprises advises that, in water-stressed zones, dry-air cooling should be preferred over cooling towers or adiabatic systems. The advice is not a ban. It is a signal to developers and to the local authorities that approve their permits about which question the state wants answered first.
Operators publish their own figures under a separate measure, WUE: litres of water per kWh of IT energy. OVHcloud publishes WUE for each site under ISO/IEC 30134-9. For its 2025 fiscal year it reports 0.20 at Gravelines, 0.29 at Roubaix and 0.39 at Strasbourg, with 0.34 for the group.
One operator, one method, nearly double the water from site to site
- Gravelines: 0.20 litres of water per kWh of IT energy (WUE), FY2025
- Roubaix: 0.29 litres of water per kWh of IT energy (WUE), FY2025
- Strasbourg: 0.39 litres of water per kWh of IT energy (WUE), FY2025
litres of water per kWh of IT energy (WUE), FY2025
Source 4 OVHcloud
The spread is the lesson. One operator, using one method, reports nearly twice as much water per kWh at one French site as at another. The document does not explain the gap, and this dossier does not guess at it. A single national figure cannot tell a town what a new site in its territory will actually draw. The WUE figures are on-site water only, and they cannot be added to or compared with Arcep's direct and indirect total.
The grid sets the calendar
Water decides how a site is cooled. Electricity decides whether it is built at all. The DGE guide is plain about the timetable. A grid connection usually takes 2 to 3 years, and can take up to 6 to 9 years. Typical projects now range from 50 to 250 MW, and above 40 MW a site must connect to the high-voltage transmission network rather than the local one.
The state has tried to shorten the wait. In May 2025 the energy regulator, the CRE, approved a pre-reservation procedure, and a fast track of 3 to 4 years now exists for sites above 400 MW. The DGE guide also lists 63 sites that it judges favourable for datacenters. These are real measures. They do not change the order of events: the grid comes first, and a building that can be put up in less time waits for it.
A grid connection takes longer than the building it feeds
| Step | What it takes |
|---|---|
| Grid connection, usual case | 2 to 3 years |
| Grid connection, constrained case | up to 6 to 9 years |
| Fast track, sites above 400 MW | 3 to 4 years |
| Annual public reporting | from 500 kW of IT power |
| Waste heat must be used | above 1 MW, unless shown not feasible |
Sources 2 Direction générale des Entreprises · 3 EUR-Lex (Publications Office of the European Union)
For a project developer this means the cooling design can be fixed years before the site opens, on assumptions about chips and loads that will have changed by then. A room designed for air in the year it was permitted may need liquid cooling in the year it is commissioned. That gap is where cooling engineering becomes a strategic job rather than a maintenance one.
The law now counts the heat
The heat that leaves a datacenter used to be treated as waste. European law now treats it as a resource that must be accounted for. Directive (EU) 2023/1791 on energy efficiency, published in the Official Journal on 20 September 2023, does this in two steps.
The first step is disclosure. Article 12(1) provides that by 15 May 2024 and every year thereafter, Member States shall require owners and operators of data centres with a power demand of the installed IT of at least 500 kW to make set information publicly available. Commission Delegated Regulation (EU) 2024/1364 of 14 March 2024 set the first reporting date at 15 September 2024, then 15 May 2025 and every year after. Its recital expects data centres to reach 3.2% of EU electricity demand by 2030.
The second step is use. Article 26(6) provides that Member States shall ensure that data centres with a total rated energy input exceeding 1 MW utilise the waste heat or other waste heat recovery applications, unless they can show that it is not technically or economically feasible. Article 26(7)(d) requires a cost-benefit analysis for a newly planned or substantially refurbished centre above that threshold. Some French guidance restates the heat duty as applying only to new centres. The directive's own wording on the duty is broader, and this dossier follows the directive.
The scale of the prize is not trivial. ADEME, the French agency for ecological transition, published an outlook in January 2026 on datacenter consumption in France from 2024 to 2060. It puts the net waste heat that could be recovered on French soil at 4 to 13 TWh in 2035, depending on the scenario. That heat is warm, but not hot. Using it means finding a district heating network, a greenhouse or an industrial user close enough to take it, and signing a contract that outlasts the servers.
Isère shows the argument in miniature
The debate is not only national. On 20 May 2025 ICI, the local Radio France service, reported on two centres planned in Isère, at Eybens and Villefontaine, in buildings formerly occupied by Hewlett-Packard and DXC. Data One provides the walls and the technical infrastructure. The consortium behind the project described the Eybens site as an investment of more than 800 million euros.
Kevin Polizzi, president of the Unitel group and a member of the consortium, told the reporter that the servers are cooled in closed circuits and that water use is "anecdotique". He gave a figure of 100 to 200 m3 of water, without stating the period it covers, and a starting power of 15 MW with a target of 1 GW. Alternatiba Grenoble, a climate group, contested the project in the same report. These are the consortium's own statements, reported by a journalist. They are not measurements, and this dossier does not compare them with the Arcep survey.
The Isère case is useful for another reason. The questions the reporter asked were the three questions this dossier asks: how much water, how much power, and who benefits from the heat. They are also the issues the DGE guide asks local authorities to weigh when a site is proposed.
The counter-case
The evidence has limits, and a reader should weigh them. First, the direct water drawn by the surveyed centres is about 0.02% of the fresh water drawn in France in 2021, which went mainly to agriculture. This dossier does not claim that datacenters threaten the national water supply. Its claim is narrower: that water becomes a local and industrial question as sites grow.
Second, the populations differ. Arcep surveys the operators it questions, which covers about 160 centres. ADEME's inventory found 352 active centres consuming 8.16 TWh in 2024. Once private rooms, public centres and server racks in office buildings are added, its model starts from 10 TWh. The two counts measure different populations and are printed side by side, never merged.
Third, the future is uncertain. ADEME's trend scenario multiplies the consumption of centres on French soil by 3.7 by 2035, and by 4.4 once the centres abroad that serve French uses are counted. Its other scenarios, on French soil, range from a 42% fall to a 7.5-fold1 rise by 2050. No single scenario is presented here as the forecast.
What would change the reading
Three developments would weaken the thesis. If Arcep's next edition showed indirect water falling as fast as direct water, the gap between the meter and the power station would stop growing. If grid connections in the main regions fell reliably below 2 years, the calendar would stop being the binding constraint. And if dry cooling matched adiabatic cooling on energy at high density, the trade-off between water and energy would lose its force. None of the documents read for this dossier shows any of the three.
What this means for your career
The documents point to three kinds of work. The first is thermal and facilities engineering: designing and running cooling for rooms whose density changes faster than their buildings, including liquid systems of the kind Schneider Electric and Motivair now sell as a range. The second is energy procurement and grid planning: the people who secure a connection years ahead and manage the reporting that the EU directive requires every year. The third is heat-recovery project management: finding the user for the heat, building the business case and running the cost-benefit analysis that article 26 asks for.
None of the sources read for this dossier prints a salary or a headcount for these roles, so this dossier gives none. What the documents do show is that each of the three jobs sits where a site's water, heat and power decisions are made. A candidate who can explain why a better PUE can mean a higher water draw, and why a building can wait years for its grid connection, already understands the problem these employers are hiring to solve.