This piece first appeared in German. The English version is a rewrite rather than a line-by-line translation, and the German original stays online in the archive: KI-Rechenzentren: Was sie Betreiber, Nutzer und Umwelt wirklich kosten.

One five-second AI-generated video clip can take around 950 watt-hours, which is what an LED lamp uses burning for four days. A high-quality generated image costs around 1.2 watt-hours, and one ChatGPT message roughly 0.3, according to estimates by MIT Technology Review.

We talk about the cloud as though it sits above things, light and abstract, somewhere between the sky and the internet. It is millions of physical servers standing in enormous halls: cold, loud and hungry for power. Every stored chat, every photo, every AI query is processed there, and loading data into the cloud means storing it on another company's computer.

What the infrastructure consumes

The International Energy Agency put data center electricity use at 415 terawatt-hours in 2024, around 1.5 percent of global consumption, in its Energy and AI report. That figure has grown by about 12 percent a year since 2017, more than four times faster than total electricity consumption.

The agency projects it will more than double to around 945 TWh by 2030, slightly more than Japan's entire current consumption. AI is the central driver: its share of data center electricity has been running at 5 to 15 percent and could reach 35 to 50 percent by 2030.

In the United States alone, data centers consumed 176 TWh in 2023, 4.4 percent of national electricity use, and the Lawrence Berkeley National Laboratory expects 325 to 580 TWh by 2028. Germany's Deutsche Energie-Agentur puts national data center consumption at around 20 terawatt-hours in 2023, about 4 percent of the country's electricity demand.

The resource nobody debates

The energy argument is familiar, and the water argument is not.

Data centers cool their servers with large volumes of fresh water, and most of it evaporates. Around 80 percent of the water withdrawn leaves as vapor and never returns to the local cycle.

Li and colleagues estimate in their updated study that one technology group's own data centers evaporated more than 23 billion liters of fresh water in 2023. Almost 80 percent of that was drinking water. Training GPT-3 in Microsoft's US facilities evaporated an estimated 700,000 liters, and that was a single model from 2020. Current models are considerably larger.

Microsoft's own sustainability report for 2023 recorded a 34 percent rise in water consumption from 2021 to 2022, a jump outside researchers link to the expansion of AI workloads.

What makes these numbers pointed is location. Many of the largest clusters sit in regions already short of water. Google's data center in Council Bluffs, Iowa, used about a billion gallons in 2024 alone, the most of any of its sites. In Texas, the Houston Advanced Research Center puts data center water use at around 25 billion gallons a year and expects 29 to 161 billion by 2030.

The cloud is an oligopoly

It is not a decentralized network. According to Synergy Research Group, Amazon Web Services at 30 percent, Microsoft Azure at 20 and Google Cloud at 13 held 63 percent of global cloud infrastructure between them in the second quarter of 2025. That market turned over almost 99 billion dollars in the quarter. The next largest provider sits in the low single digits.

Concentration like that has consequences beyond market power. Training an AI model means running it on those providers' servers, and the training data sits there. The model weights sit there. Every input to ChatGPT, every prompt to Gemini, every query to Claude is processed there and in many cases stored.

For a start-up it means no AI development at meaningful scale without access to one of the three. Governments find critical digital infrastructure in private and often foreign hands. And for users, the content of their conversations, frequently highly personal, sits on servers reachable in principle by staff during maintenance, by subcontractors, by authorities, and in the worst case by attackers.

Many providers also use those inputs to train the next model, which turns users into unpaid data suppliers. At OpenAI, model training on user data is switched on by default for free and paying individual accounts, and anyone who objects has to opt out. Anthropic changed its consumer terms in 2025, and users of its Free, Pro and Max plans now decide whether their chats may be used for training. Business, enterprise and API customers are exempt by default at both.

The earlier version of this bargain was about targeted advertising. Today it is about something more valuable: the training material for the next generation of models.

The EU AI Act now imposes transparency obligations. Physical reality lags the regulation, and the question of who owns the data in our conversations with AI stays legally unsettled and technically hard to enforce while the infrastructure sits in so few hands.

What it adds up to for one person

The figures break down. A heavy user running fifteen text queries, one generated image and one short AI video a day consumes roughly the daily electricity of a refrigerator. Add the fresh water that disappears into the atmosphere as vapor on the way.

Scale that across the 900 million weekly active ChatGPT users OpenAI reported in February 2026, plus the users of Gemini, Claude, Copilot and local models, and the consumption reaches the order of a small country.

None of it is visible to the person typing. No interface says that this query cost 0.3 watt-hours of electricity and a mouthful of drinking water. That is the business model, and no oversight.

The provider carries the power and water bill, and not out of generosity. Every input, every prompt, every conversation produces data that helps train the next generation, refining language patterns, improving accuracy, extending the competitive lead. The user pays with data for convenience, and anyone paying with data never sees an invoice.

The costs are invisible on purpose

The cloud is infrastructure, not myth, and like any infrastructure it has costs that somebody carries. Here two parties carry them: the environment in electricity and water, and the user in data. That both stay invisible is the precondition for the business model working.

While three companies control over 60 percent of global cloud infrastructure and simultaneously determine access to the most capable AI systems, the question of digital sovereignty stays open, whatever regulatory frameworks get adopted. What has to change is visibility: of the ecological cost per query, and of the value of the data handed over in return.

Every time a chatbot answers, what answers is not "the AI". It is a physical system of servers, cables and cooling water. An oligopoly operates and pays for it, with infrastructure costs in the billions, and evidently still rates that price below the value of our data.

Sources

Grouped by the section they support.

Opening

  • O'Donnell and Crownhart (2025), We did the math on AI's energy footprint. Here's the story you haven't heard., MIT Technology Review, 20 May 2025, technologyreview.com. Supports about 3.4 million joules for a five-second video, about 4,402 joules for a higher-quality image and an estimated 0.3 watt-hours per ChatGPT message. The LED comparison is the author's own conversion.

What the infrastructure consumes

  • International Energy Agency (2025), Energy and AI, executive summary, iea.org. Supports the 415 TWh and 1.5 percent in 2024, growth of around 12 percent a year since 2017, more than four times faster than total consumption, and around 945 TWh by 2030, slightly more than Japan uses today.
  • Carbon Brief (2025), AI: Five charts that put data-centre energy use – and emissions – into context, 15 September 2025, carbonbrief.org, citing a report prepared for the IEA. Supports AI's share of 5 to 15 percent of data center power and a possible 35 to 50 percent by 2030.
  • U.S. Department of Energy (2024), DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers, 20 December 2024, energy.gov, on the Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report. Supports 176 TWh and 4.4 percent in 2023 and 325 to 580 TWh by 2028.
  • Deutsche Energie-Agentur (2025), Nachhaltige Rechenzentren, dena.de. Supports around 20 TWh in 2023, about 4 percent of German electricity demand.

The resource nobody debates

  • Li, Yang, Islam and Ren (2025), Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models, Communications of the ACM, arXiv:2304.03271. Supports the roughly 80 percent of withdrawn water that evaporates, the more than 23 billion liters one company's own data centers evaporated in 2023 with nearly 80 percent potable water, the 700,000 liters for training GPT-3, and the link between the 34 percent rise in another company's water consumption and AI.
  • Moss (2023), Microsoft's water consumption jumps 34 percent amid AI boom, DatacenterDynamics, 12 September 2023, datacenterdynamics.com. Supports the 6.4 million cubic meters in 2022, 34 percent more than in 2021, as stated in Microsoft's 2023 sustainability report, and the absence of an AI attribution by Microsoft itself.
  • McCauley and Scanlan (2025), Data centers consume massive amounts of water – companies rarely tell the public exactly how much, The Conversation, republished by The Current, 26 August 2025, thecurrentga.org. Supports the 1 billion gallons at Council Bluffs in 2024, as reported by Google.
  • Houston Advanced Research Center (2026), Texas Data Center Boom Could Consume Up to 161 Billion Gallons of Water Annually by 2030, 21 January 2026, harcresearch.org. Supports the 25 billion gallons today and 29 to 161 billion by 2030.

The cloud is an oligopoly

  • Synergy Research Group (2025), Q2 Cloud Market Nears $100 Billion Milestone, 31 July 2025, srgresearch.com. Supports the shares of 30, 20 and 13 percent, 63 percent combined, and quarterly revenues of 98.8 billion dollars.
  • OpenAI, How your data is used to improve model performance, OpenAI Help Center, help.openai.com. Supports training on content from individual accounts, the opt-out, and the default exclusion of business, enterprise and API use.
  • Anthropic (2025), Updates to Consumer Terms and Privacy Policy, 28 August 2025, anthropic.com. Supports the user's choice on Free, Pro and Max plans and the exclusion of Claude for Work and the API.
  • The consequences for start-ups, governments and users are the author's own analysis.

What it adds up to for one person

  • O'Donnell and Crownhart (2025), MIT Technology Review. Supports the per-query figures behind the daily estimate; the refrigerator comparison is the author's own calculation.
  • Malik (2026), ChatGPT reaches 900M weekly active users, TechCrunch, 27 February 2026, techcrunch.com. Supports the 900 million weekly active users.

The costs are invisible on purpose

  • Synergy Research Group (2025). Supports the more than 60 percent held by three companies. The conclusion is the author's own.
The link has been copied!