The figure that should stop us in our tracks
A single conversation with ChatGPT, of 5 to 50 requests depending on length, consumes around 500 millilitres of water. That is the equivalent of a small bottle of water, evaporated somewhere in a data centre in Nevada, Iowa or Virginia. Across hundreds of millions of daily users, we are talking about millions of litres of water evaporated every day, to produce poems, emails and essays.
This calculation comes from a study by the University of California, Riverside published in 2023, and confirmed since by several independent publications. Training the GPT-3 model alone in Microsoft's data centres in the United States directly evaporated 700,000 litres of fresh water. For training GPT-4 and GPT-5, we are talking about several million litres. And that is just training; the recurring cost of each use is added on top.
This article is not an ecological manifesto. It is a factual account of a subject that the French tech press almost never addresses, but which is becoming in 2026 one of the industry's biggest points of tension. As we documented in our article on the AI bubble, the economic boom has a physical flip side. Here it is.
The real scale of electricity consumption
Let us start with the official figures. According to the IEA (International Energy Agency, the international energy agency that advises countries on energy policy), global data centre electricity consumption rose from 415 TWh in 2024 to around 460-490 TWh in 2025. To give a sense of scale, 1 TWh (terawatt-hour, that is one thousand billion watts for one hour) represents the annual electricity consumption of around 100,000 European households. The 2030 projection is 945 TWh, a doubling in six years.
In the United States, the picture is even starker. Data centres now account for 4.4% of national electricity consumption, against 1.9% in 2018. And according to projections by American institutions, that figure could reach 12% by 2028. That means that by then, more than a tenth of all electricity produced in the United States would go to powering computer servers.
This growth is almost entirely driven by AI. "Classic" data centres (which host your emails, Netflix, your personal cloud) have been stable for 10 years thanks to efficiency gains. AI data centres consume 5 to 20 times more electricity per square metre, because of the GPUs (Graphics Processing Units, graphics processors manufactured notably by Nvidia, which are the chips used massively to train and run AI models) that generate enormous heat.
A ChatGPT query consumes between 0.3 and 3 Wh (watt-hours), i.e. 3 to 10 times more than a conventional Google search. Across billions of daily queries, the gap compounds. If Google Search shifts predominantly to its AI Mode (the AI version of search announced at Google I/O 2026, which we picked apart in our full article on the keynote), the electricity consumption of the internet could literally increase tenfold. No country currently has the electrical capacity to meet that scenario.
The water problem: what nobody measures
Electricity is now well documented. Water, far less so. Yet data centres are enormous consumers of it.
In the United States, the direct water consumption of data centres rose from 21.2 billion litres in 2014 to 66 billion in 2023. And that is only the water used to cool the servers. If you add the water consumed indirectly to produce the electricity that powers those same data centres, you reach 800 billion litres a year. For reference, that is the equivalent of roughly 320,000 Olympic swimming pools.
In concrete terms, a medium-sized data centre consumes up to 110 million gallons of water a year for cooling (416 million litres), equivalent to the annual water consumption of around 1,000 American households. The largest data centres can drink up to 5 million gallons a day (19 million litres), equivalent to the consumption of a town of 10,000 to 50,000 inhabitants. And there are currently more than 5,400 data centres in the United States alone.
The problem is that many of these facilities are built in regions already under water stress: Arizona, Nevada, west Texas, parts of California. While states ration water for farmers and households, they grant it generously to buildings that evaporate millions of litres a day to cool the graphics cards running AI models.
| Metric | 2026 figure |
|---|---|
| Water for 5 to 50 ChatGPT queries | 500 ml |
| Water to train GPT-3 | 700,000 litres |
| Electricity per ChatGPT query vs Google | 3 to 10 times more |
| Data centres' share of US electricity 2026 / 2028 | 4.4% / projected 12% |
| Microsoft water consumption, fiscal 2022 | 6.4 million cubic metres |
| Projected AI carbon footprint 2025 | 32.6 to 79.7 million tonnes of CO2 |
The revolt of local communities
Here is the angle the media often overlook. Data centres are not welcomed with open arms by the communities where they are built. On the contrary, since 2024 a wave of opposition has been gathering pace.
According to Data Center Watch (a monitoring body for data centre projects in the United States), between March and June 2025, citizen opposition led to $98 billion of projects being blocked or delayed. Heatmap Pro, which analyses US municipal public records, has counted at least 25 projects cancelled in 2025 in response to local objections.
Concrete cases are not hard to find. In Memphis, Tennessee, protesters opposed xAI's plan (Elon Musk's artificial intelligence company, creator of the Grok chatbot) to use gas turbines to power a data centre. In Indianapolis, after months of intense citizen opposition, a lawyer representing Google announced to the city council in September 2025 that the project was being withdrawn. Similar cases are multiplying in Virginia, Georgia and Oregon.
The main reason for opposition is not ideological. It is concrete: these facilities drive up residential electricity prices, drain local groundwater, generate noise and heat, and create very few local jobs. A large data centre typically employs 50 to 150 people for tens of billions of dollars of investment. The investment-to-jobs ratio is one of the worst in the entire economy.
The tariff scandal: who really pays?
The point that is becoming a major political issue in 2026: residential electricity bills are rising because of AI data centres. Several studies in the United States have shown that in regions where data centres are being built en masse (Virginia, Ohio, Texas), ordinary households are seeing their bills climb by 8 to 15% a year, without using any more electricity.
Why? Because data centre operators negotiate preferential rates with power companies, and infrastructure costs (new high-voltage lines, additional power plants) are spread across all consumers. In practice, households are subsidising the infrastructure that Google, Microsoft, Meta and OpenAI benefit from.
This situation creates an uncomfortable political equation: American tech companies, which are collectively worth more than $10 trillion on the stock market, are making the poorest households pay for their energy infrastructure costs. It is a massive and silent transfer of wealth, which no legislation currently regulates.
The Big Tech companies (Microsoft, Google, Amazon, Meta) had all made commitments to carbon neutrality by 2030. None of them is on track to meet them. Worse, several have rowed back on their targets in 2025-2026 in order to power their AI ambitions. Microsoft admitted in late 2024 that its emissions had risen by 30% compared with 2020, because of AI. Google acknowledged a 48% increase over the same period. Promising carbon neutrality by 2030 while building 50 new data centres hungry for fossil energy: the contradiction is no longer tenable.
The technological solutions that are emerging
Honestly, the industry has not stood still in the face of these problems. Several technical avenues are emerging and deserve to be named.
Liquid cooling. Legacy data centres use air cooling (industrial air conditioning), which is very electricity-hungry but uses little direct water. New "direct-to-chip" or "immersion cooling" systems (cooling by direct contact with the chips or by immersion in a non-conductive liquid) reduce direct water consumption by 70 to 90% and significantly improve PUE (Power Usage Effectiveness, an indicator of data centre energy efficiency: a PUE of 1.0 would be perfect, the global average is around 1.5).
Nuclear power. In 2024 Microsoft signed a landmark deal with Constellation Energy to restart the Three Mile Island nuclear plant (famous for the 1979 accident) in order to power its data centres. Google and Amazon are following with investments in SMRs (Small Modular Reactors, smaller modular nuclear reactors that are more compact and quicker to deploy than conventional plants). Nuclear could absorb a significant share of AI demand by 2030, with no CO2 emissions.
Hyperlocalisation and edge computing. Edge computing (running computations as close to the user as possible rather than in a distant, centralised data centre) could significantly reduce the need for massive cooling. Advances in compact models (such as Gemini 3.5 Flash, which we analysed in our comparison) make this approach technically viable.
More efficient models. Part of AI consumption is down to the inefficiency of current algorithms. In 2025 DeepSeek demonstrated that it was possible to train a model equivalent to GPT-4 for roughly 1/10th of the compute cost. If this efficiency becomes widespread, AI consumption could grow more slowly than expected. As we explored in our article on Chinese models, the American embargo on advanced chips has paradoxically forced China to innovate on efficiency.
What this changes for you
You are not responsible for the consumption of Microsoft's data centres. But a few practical actions remain within your reach.
Choose the right tools for the right tasks. Using OpenAI's GPT-5 for a simple search is around 10 times more energy-intensive than using classic Google Search. If you just need a factual piece of information, stick with traditional search.
Favour efficient models. Gemini 3.5 Flash or Claude Haiku consume significantly less than Claude Opus or GPT-5. For 80% of uses, they are more than sufficient.
Avoid over-generation. As we documented in our article on Hell Grind, the first fully AI film shown at Cannes, the team generated 16,181 clips to keep 253. Every generation consumes energy. Be precise in your prompts to reduce repeated attempts.
Find out about your providers. If you are in France, OVH (the leading French cloud host) has a markedly more decarbonised energy mix than AWS or Azure in some American regions. For businesses, choosing a low-carbon cloud is now a concrete argument.
The question we need to ask collectively
AI will probably improve productivity, speed up medical research and democratise certain professions. As we explored in our article on the medicine of the future and our analysis of AI discovering new knowledge, the benefits are real. But the political question that really arises in 2026 is this: what proportion of our water and our electricity are we collectively prepared to allocate to AI, and who should pay the price?
This question has no simple answer. It is, however, becoming a major political issue in several American states, and soon in Europe. The environmental regulation of AI will probably be one of the great legislative projects of 2027-2030, alongside the AI Act that we documented in our article on the limits of regulation.
In the meantime, just remember this: every time you ask ChatGPT to write you a poem, around 100 millilitres of water evaporate somewhere. That is neither serious nor trivial. It is simply a physical reality that nobody was measuring three years ago, and which is gradually becoming impossible to ignore.