You ask ChatGPT a question and feel a little guilty about it?
You're not alone. The news is full of thirsty data centers, new nuclear deals, and AI campuses that draw as much power as whole cities. At the same time, you keep hearing that a single prompt uses almost nothing.
Both are true. That's exactly what makes the question so confusing.
In this article, I'll show you how bad AI really is for the environment, where the damage comes from, how much water AI uses, and what your own use adds up to. The full data series with a live counter are on my AI energy consumption data page.
The short version:
- Yes, AI has a real environmental impact through electricity, cooling water, and chip manufacturing. AI-focused data centers used about 155 TWh of electricity in 2025.
- A single text prompt is small. Published figures range from 0.24 to 0.34 Wh, which OpenAI CEO Sam Altman compares to running an oven for a little over one second.
- Water matters too. Data centers worldwide consumed about 560 billion liters in 2023, and most of it went into generating their electricity, not into on-site cooling.
- Images, video, and reasoning agents need far more. Video can take hundreds to thousands of times the energy of a short text prompt, and the IEA projects data center emissions to double to about 350 million tonnes by 2035.
1. Is AI Really That Bad for the Environment?
Yes. But the damage comes from scale. Your individual question to ChatGPT barely registers.
AI-focused data centers used about 155 TWh of electricity in 2025. That's just under a third of all data center electricity worldwide. By 2030, the International Energy Agency (IEA) expects about 465 TWh for AI data centers, three times as much:
gradually.aiGermany's data centers, for scale, used 21.3 TWh in 2025, according to Bitkom. Measured against all the electricity the world uses, that's still modest. All data centers together accounted for about 1.5% of global electricity in 2024. For 2030, the IEA expects about 3%.
My take:
AI isn't a climate catastrophe on its own, but it isn't harmless either. It's a fast-growing power consumer, and that growth is the problem. How much harm the electricity does depends mostly on where it comes from.
2. Why Is AI Bad for the Environment?
AI affects the environment in four places. Electricity makes the headlines. Water, carbon, and hardware follow directly from it.
2.1 Electricity
Every prompt, every training run, and every agent task runs in a data center. According to the IEA, a typical AI-focused data center uses as much electricity as 100,000 households.
You can see how fast demand grows at the big tech companies. Google and Microsoft have more than doubled their electricity use since 2021, and Meta almost has:
gradually.aiThese are total company figures, and AI is only part of them. Amazon publishes no total. The direction is still clear. Water withdrawals and nuclear deals of the big companies are on the data page.
2.2 Carbon Emissions
Electricity and water are only the inputs. For the climate, emissions are what count.
In an IEA projection, data center emissions double to about 350 million tonnes by 2035. That would be about 2% of global power sector emissions.
You can already see the trend at the companies. Microsoft's total emissions rose 25% in fiscal 2025, and the company names data center expansion as one of the drivers. Amazon's absolute emissions grew 16% in 2025.
Google reports a 2% drop in market-based emissions from its own operations in 2025. Its supply chain emissions, however, rose 25%.
2.3 Hardware and Training
Before a chip computes anything, it has to be built. Hardware manufacturing accounted for about 47 billion liters of all data center water consumption in 2023.
Then there's training large models. The Stanford AI Index 2026 estimates 5,184 tonnes of CO2e for training GPT-4 alone.
3. How Much Water Does AI Use?
Less than the headlines suggest per prompt, more than you'd think in total.
3.1 Water per Prompt
Google reports 0.26 milliliters of direct cooling water for a median Gemini Apps text prompt. That's a few drops, and Google's figures are not independently verified.
For ChatGPT, Sam Altman gives about 0.32 milliliters per average query, roughly one fifteenth of a teaspoon. He doesn't say whether that includes the water used to generate the electricity. The data page compares both figures.
3.2 Data Centers Worldwide
All data centers worldwide consumed about 560 billion liters of water in 2023 (about 148 billion gallons), according to the IEA. Only a quarter of that went into on-site cooling:
Two thirds went into supplying the energy and generating the electricity that data centers draw. By 2030, the IEA expects 1,200 billion liters a year, or about 317 billion gallons. The full curve is on the data page.
3.3 A Single Data Center
The IEA estimates that an average 100 MW data center in the US uses about 2 million liters of water a day. That's roughly 530,000 gallons, as much as 6,500 households, and most of it is indirect.
Globally, that's not much. At a single site, it can be a lot, especially where water is already scarce. That's why the location of a new data center matters as much as its size.
3.4 Withdrawal vs. Consumption
Watch out for this distinction when you read water numbers. Withdrawal means all the water taken in, consumption only the part that isn't returned, for example because it evaporates.
Withdrawals are far larger than consumption. A rough reading of a chart in the IEA report puts data center withdrawals at about 5,200 billion liters in 2023. On the company level, Google's data centers consumed about 39.8 billion liters in 2025, or 10.5 billion gallons, according to its 2026 environmental report.
4. Is ChatGPT Worse for the Environment Than Google Search?
Not much for a simple text question. Based on published figures, both are in the same ballpark.
You'll often read that a ChatGPT prompt uses ten times the electricity of a Google search. That only works if you assume about 3 Wh per ChatGPT prompt. Epoch AI considers that figure about ten times too high, at least for a typical GPT-4o query with a short input. For very long inputs, Epoch gives much higher figures.
More recent figures sit well below 3 Wh:
gradually.aiGoogle's 0.24 Wh apply to the median Gemini Apps text prompt and include idle machines, CPU, memory, and data center overhead. They are not independently verified. The 0.34 Wh for ChatGPT come from OpenAI CEO Sam Altman as an average, without a published method.
That said:
The Google search figure comes from a 2009 blog post. Against figures from 2025, that's a rough comparison, not a measurement.
It gets more expensive when ChatGPT thinks longer or works on its own. For a reasoning agent task of four to six model calls in a row, an IEA model estimate gives about 50 Wh, and that's GPU electricity only. The IEA calls this an order of magnitude, not a precise figure.
By my math, that's about 150 times Altman's average. The two figures come from different methods and aren't directly comparable. The direction is clear anyway.
You can try any number of prompts in the calculator on the data page.
5. How Bad Are AI Images and AI Video?
Much worse than text.
According to the IEA, a high-resolution image can take about 10 times the energy of a short text prompt. For video, it's hundreds to thousands of times, depending on length and resolution. The chart uses a log scale, so each step is ten times the previous one:
The IEA gives factors instead of fixed watt-hours, because consumption depends so much on length and resolution.
A video you regenerate five times because the first attempt didn't land can cost as much energy as hundreds or thousands of text questions.
6. How Much Does Your Own AI Use Add Up To?
Very little while you chat. More once you use images, video, or agents.
Sam Altman compares the electricity of an average ChatGPT query to an oven running for a little over one second.
Let's scale that up.
At 100 prompts a day, Altman's figure adds up to about 12 kWh a year. An average US home used 10,791 kWh of electricity in 2022, according to the US Energy Information Administration:
gradually.aiIn short:
Normal chatting keeps you well under 1% of an average home's electricity.
Agents are a different story. Ten reasoning agent tasks a day add up to about 180 kWh a year under the IEA's model estimate, and that's just the GPUs.
I've been working with Claude Code since mid-2025, more than 800 hours by now. If you use AI that intensively, the agent numbers are the ones to watch.
7. Can AI Help the Environment?
Yes, the potential is big. It hasn't been realized yet, though.
The IEA estimates that well-proven AI applications could save more than 13 exajoules of energy by 2035 if barriers to wider adoption are overcome. That would equal 3% of global final energy consumption.
Savings could come from optimized industrial processes, smarter building management, and better route planning in transport.
However:
The IEA attaches conditions. These applications would need to spread widely over the next ten years. And energy saved can lower costs and create new demand elsewhere, a rebound effect.
The models themselves are getting more efficient too. Google reports that the energy for a median Gemini Apps text prompt fell by a factor of 33 between May 2024 and May 2025. That figure is not independently verified.
The IEA also expects cheaper, better AI to attract more use. On balance, AI's total electricity use is set to keep rising in the near term.
So AI can help the environment. Whether it does gets decided mostly in factories, power grids, and buildings.
8. What Can You Do?
You don't have to avoid AI to be environmentally conscious. These five habits make the biggest difference:
- Skip reasoning models and agents for simple questions. A reasoning agent task sits in the order of 50 Wh in the IEA's model estimate, an average ChatGPT query at 0.34 Wh according to Sam Altman.
- Generate images and video on purpose. Write a sharper prompt instead of rerolling ten times.
- Use smaller models where they're good enough. According to the IEA, smaller specialized models need far less energy for narrowly defined tasks.
- Don't paste documents you don't need into the chat. Long inputs with documents or images increase the energy per prompt, the IEA notes.
- Favor providers that publish numbers. Google reports electricity, water, and CO2e per prompt, and Mistral published a lifecycle analysis with the French environment agency ADEME.
And yes:
The bigger lever sits with the operators. Where they get their power and where they build their data centers matters more than any habit in your chat window.
All electricity and water figures, with a live counter and a value register, are on my AI energy consumption data page.
