How bad is using AI for the environment?

As a data scientist and AI engineer, there are few questions I get as often as this one: how much energy am I "wasting" when I use AI? How much CO₂ am I emitting, and how much water am I taking from the locals every time I ask a chatbot something?

Honestly, it's both a lot and not that much... Yes, AI uses a lot of energy. But in the world we live in, it gets overshadowed by other things we do every day, like driving a car or eating meat.

In this blog I'll put things into perspective and help you answer the question: how bad is it when I use AI, really? All numbers come from the AI energy calculator I built, and every example has a button that opens it in a new tab, so you can play with it yourself.

We will mainly focus on CO2, energy and water aspects of running AI since these are the main concerns about the use of AI.

This blog is about you. If you're wondering what AI costs your company, see How to use AI responsibly as a business (coming soon), and for the planet as a whole, How bad is AI for the environment? (coming soon)

How many tokens do I use?

Before we dive into the costs of AI, we first have to make sure we understand what using AI actually means. AI models read and write text in small chunks called tokens, roughly three-quarters of a word each, and tokens are what cost energy. So here are a few examples to give you a feel for how many tokens you burn through with your tasks.

In this blog I'll stick to text and leave images out, since text is what most of us use day to day.

Small tasks (~1,000 tokens)

On the low end of the spectrum you have the mundane tasks: rewriting an email, answering a few questions, asking for a recipe or generating a personal letter for your best friends birthday... These are the tasks a lot of people use AI for every day, and to some they already feel like a huge CO₂ burner.

Click through the tabs to get a feel for some of these use cases.

Large tasks (~1,000,000 tokens)

Then there are the bigger tasks: digging into a research question, building a simple website or adding a feature to your product. These can easily outcompete the small tasks by three orders of magnitude.

In these kinds of tasks the context tends to grow and grow. The AI researches a lot and drags all that information into the conversation.

This is where we run into the cache. When an answer is generated, the provider stores the conversation so far, so it doesn't have to recalculate all of it for the next message. It's a bit more complicated but an extensive explanation is outside the scope of this blog, but you can read up on it in the cache section of the calculator blog.

From a thousand to a million

As you can see, the number of tokens you burn after sending a prompt to an LLM varies enormously. I hope this helps you get a grasp of how much you tend to use for your work.

The cost of your AI usage

So now that we're on the same page: how much does all this cost? And how does it compare with other things we do?

These values are best guesses. Precise numbers on the energy use of AI aren't published for every model on every server, so we have to make some assumptions. Values also vary a lot on different factors, so the values we take could be off by a factor of two on either side. You can check the assumptions in more detail in the settings of the calculator.

How much energy and CO₂ does a single AI chat cost?

Have a short chat with a chatbot, like polishing an email, and you've used just under a watt-hour (that's the average of the 50 short chats above). But those are just numbers, and what do numbers mean anyway, right? In CO₂ terms that's the equivalent to driving a petrol car about a metre and a half. You'd need around 2,600 of those short chats to drive a short commute of 4 kilometres, and around 16,000 for a 25-kilometre one.

So yes, AI uses energy. But your car is a lot worse.

When I compare with "your commute" in this blog, I use two of them: a short commute of about 4 kilometres each way, a quick drive across town and common in most countries, and a medium commute of about 25 kilometres, the average Dutch car commute, roughly Rotterdam to Den Haag¹. A long commute of about 70 kilometres, Amsterdam to Rotterdam, is much rarer: only about one in ten commutes is 50 kilometres or longer.

So using AI isn't bad at all, right?! The car is barely leaving your parking spot at this point. Let's now compare the same chat with something else: charging your phone⁴. A short chat equates to about 5% of a phone charge. But have a long chat, the kind where the AI searches the web for you, builds a small website or fixes a bug, and you're now at a whopping 200%... That's two full charges, for about 37 Wh (the average of the 25 long chats above). Suddenly it feels different, doesn't it?

Nothing changed except the comparison. Next to a car, AI looks harmless; next to a phone, it looks huge. Most headlines about AI's footprint are really a choice of comparison, so it pays to look at more than one.

And how much water does a chat cost?

Data centres evaporate water to keep their chips cool, and the power plants behind them evaporate even more. A short chat costs about 1.6 millilitres, a third of a teaspoon. A long chat is about 70 millilitres, a small glass and 0.1% of a shower. Most of that isn't the datacenter at all: in the calculator's default (AWS in Virginia) about 96% is used at the power plants.

How much depends on what the grid runs on. Coal and nuclear plants evaporate water in their cooling towers, wind and solar use next to none, and hydro dams lose water from their reservoirs (how much of that to blame on electricity is disputed, because reservoirs also serve farms and drinking water). So a clean grid isn't always a dry one: the calculator counts about 15 litres per kWh for hydro-heavy Norway and under 1 litre for gas-heavy Singapore.

You'll find scarier numbers online, like "a bottle of water per conversation". Estimates per prompt run from a quarter of a millilitre (Google and OpenAI, counting only their own buildings) to around 15 millilitres (researchers who also count the power plants)⁵. The bottle came from an early study on GPT-3-era hardware, and its authors now reportedly give a much lower figure for today's models.

So for you, it's tiny. The problem is where it happens. A datacenter takes all its water from one place, often drinking water, and most of it on the hottest, driest days. In July 2022, while OpenAI was finishing GPT-4, Microsoft's datacenters in West Des Moines, Iowa, used about 43 million litres, 6% of all water in the local district⁶. Microsoft's datacenters in Middenmeer used 84 million litres of drinking water in 2021, four to seven times what the municipality had told residents⁷. That wasn't for AI yet, but AI runs in the same kind of buildings, and it's the main reason so many more are being built. And clean water is getting scarcer: about half the world's people face severe water shortages for part of the year⁸, and even in the rainy Netherlands, Vitens has had to turn away new business connections because there isn't enough drinking water⁹. A thirsty new datacenter doesn't help.

The bigger picture on this topic will be in How bad is AI for the environment? (coming soon)

How much energy does AI use per day for me?

Before we can truly describe your personal impact, we have to establish how much of an AI junkie you truly are. So I created three personas, each one a working day of AI:

• The light user polishes an email from time to time and asks some mundane questions: 10 short chats a day.
• The intense user uses AI as a personal assistant for many tasks during the day: 50 long chats.
• The power user maxes out their $200-a-month subscription from time to time: two 5-hour Claude Code windows¹⁰.

I'm leaving out the power² users here, the ones who max out multiple accounts all day, every day. These people are truly nuts and belong to the top 0.01% anyway.

These are personas, not averages. Nobody publishes how many tokens a typical office worker uses. The closest there is: ChatGPT users send about 3 to 4 messages a day on average, and most of that isn't for work¹¹. The three users here are there to make the numbers concrete, from very light to as heavy as it gets.

Tokens are what the model reads and writes; what counts for the climate is the CO₂e. So here are the three users next to activities that might be part of your daily routine: charging your phone⁴ or laptop¹², a commute by car¹³, and 100 grams of chicken or beef¹⁴.

The AI, the phone and the laptop all run on electricity, so where that electricity comes from matters a lot¹⁵. Switch the grid and watch those bars move; the car and the meat stay where they are.

The grid makes a big difference. On a low-carbon grid, like Sweden's, even the power user stays below a portion of chicken; on a coal-heavy one the intense user creeps up to almost a full portion. From here on I use the mixed grid, like Virginia, the world's biggest data centre hub.

A year of AI next to a year of driving

One day is one thing. What about a whole year?

Take the intense user's day, 50 long enquiries, and repeat it every working day: 250 days. That's about 320 kWh and 135 kilograms of CO₂e, the same as driving about 600 kilometres. The car commuter from Rotterdam to Den Haag drives about 11,000 kilometres a year just to get to work and back, and emits about 2.5 tonnes doing it. Someone with a short 4 km commute drives about 2,000 kilometres a year and emits about 450 kilograms.

So a whole year of heavy chatting is about 5% of a year of medium commuting, and about 30% of a short one.

Programmers who max out a 5-hour Claude Code window every working day end up higher: this equates to 950 kWh and 400 kilograms a year, around 1,800 kilometres of driving. That's a real number: about 16% of the medium commute, and almost as much as the whole short commute. Still, the medium commute is six times bigger.

What am I contributing?

So my AI use is small next to my car. But "small next to my car" isn't the same as "nothing". Let's zoom out once more.

If everybody on earth, all 8.2 billion of us¹⁶, used AI like the intense user does all year, the world would need about 2,600 TWh a year for it. That's more than six times what all data centres in the world use today, AI and everything else together¹⁷. Of course nobody chats like that, let alone everybody. But it shows why AI's footprint is a question of scale: tiny for one person, huge when billions of people do it every day.

And for me personally? An average Dutch person is responsible for roughly 8 tonnes of greenhouse gas a year (the country's 144 million tonnes¹⁸ divided over 18 million people¹⁹). My 135 kilograms of chatting is under 2% of that. A 25 km car commute is about 30%.

So the honest answer is: you're contributing something, but not much. If you want to make a difference, your car, your flights, eating meat and your heating are where the tonnes are. And if you work at a company, the bigger lever is there; more on that in How to use AI responsibly as a business (coming soon). For the world as a whole, see How bad is AI for the environment? (coming soon)

So, should you feel guilty?

For a short chat: no. A single chat is tiny next to a car trip, a shower or a steak. But "tiny" times billions per day is not tiny, and agents that work for hours use thousands of times more than a chat. The useful question isn't "Is it bad if I use AI for this task?" but "How can I use AI responsibly in order to reap the benefits while having a minor impact on our environment?" Those choices can make a 100-fold difference.

How to use AI sustainably: 6 ways to shrink your footprint

There are a few things you can do to keep your personal CO₂ footprint low while still enjoying AI.

1. Do some straightforward tasks yourself

Maybe you don't need AI to change the colour from blue to green on your website? A small edit you can make in ten seconds doesn't need a model that re-reads your whole project first.

2. Start a new chat instead of continuing an old one

A long chat history increases the cost of creating new text, because every new message re-reads the whole conversation. The cache helps a lot here, but even the cache isn't a magic bullet that cancels it out completely. Of course, in some cases the chat history is vital/important in order to answer the question correctly, so decide for yourself when a new chat is the best decision.

3. Use a smaller model when possible

Fable (Anthropic) and Astra (OpenAI) probably use significantly more energy than their smaller counterparts, Opus and Sol respectively. Start small, and only upgrade when the answer isn't good enough. That being said, sometimes small models are so ineffective that they actually need more energy to complete the task at hand.

4. Turn off extended thinking (reasoning mode)

Most models let you choose how hard they think: low, medium, high, and on some models extra high or max. All that thinking is text too, thousands of hidden words the model writes before you see an answer, and writing is the expensive part. In Microsoft's estimates a reasoning-style query writes about 15 times more tokens and uses about 13 times more energy than a typical one²⁰.

And more thinking isn't always better. A study of 53 models on basic maths found zero accuracy gain from low to medium to high effort on GPT-5 and OpenAI's o-series²¹. In a test of seven models on short coding tasks, low effort solved every task, and high never solved one that low missed, while costing up to 10 times as much²². Anthropic's own researchers even found that letting models reason longer can make them worse: they get distracted by irrelevant details and talk themselves out of the right answer²³.

So medium is a good default. It costs less energy than high or extra high, and sometimes it's actually better. Save the max setting for the really hard problems, and for a quick question go even lower.

5. Think about the cache

This one is a bit trickier, and it's covered in more detail in the AI energy calculator blog. The short version: the provider keeps your conversation in a cache for a few minutes, which makes re-reading it much cheaper. Once that cache has expired, the whole conversation has to be read again at full price. So picking up yesterday's chat costs way more energy than chatting in one sitting.

6. Choose a provider with greener servers

This one is also a bit tricky, since most providers don't tell you which servers your chat runs on, or what powers them. And it matters a lot: the same chat on a Swedish grid emits about 20 times less CO₂ than in Virginia, where much of the American cloud sits¹⁵.

Here's what we know about the big ones:

• Google (Gemini) is the most open. About 65% of the electricity in its data centres was carbon-free in 2025, measured hour by hour, and it publishes an energy figure per prompt²⁴.
• Mistral runs mostly in France, where the grid is largely nuclear, and is the only one that has published a full life cycle analysis of its model: about 1.14 g CO₂ per answer²⁵.
• Anthropic (Claude) and OpenAI (ChatGPT) publish no emissions data at all²⁶. And some of their newest capacity runs on gas: Anthropic rents xAI's Colossus 1 in Memphis²⁷, and at least three of OpenAI's Stargate sites get on-site gas plants²⁸.
• xAI (Grok) is the clear worst. Its data centres around Memphis run on dozens of gas turbines, many of them without a permit, and the temporary ones only have to be gone by July 2027²⁹.

Be careful with "100% renewable" claims too. Most of the time that means the company buys as much green energy over a year as it uses. Your chat at 10 at night can still run on gas.

So what can you actually do? If you only use the chat apps, not much beyond picking Gemini or Le Chat over the others. If you use AI through an API or a cloud platform, you can often pick the region: a data centre in Sweden or Finland beats one in Virginia by a mile. And open models can run on GPU clouds in Iceland and the Nordics that run on hydro and geothermal power.

Want to try your own day? Open the calculator, pick your own cities, car and model, and see where you end up.