How much energy does an AI conversation really use?

I built the calculator above to help you get a grasp for how much energy AI uses, and to give you something concrete to share with others. For example, you can show how much energy a person or a company uses on a daily basis, and use that when you talk to your colleagues about how to bring AI into their daily work.

You can work out roughly how much your AI use costs, and compare it with mundane things like a car trip or charging your phone, and with big things like a whole industry. On this page I explain how it works. The story itself lives in three blogs, depending on the scale you care about:

• You: how much CO₂, water and energy does my AI use, and how does it compare with my car?
• Your business (coming soon): what does AI cost a company, and how do you use it responsibly?
• The world (coming soon): how bad is AI for the environment, really?

How it works

The calculator has three steps: pick what you do with AI, pick what you want to compare it with and press play.

What you do. There are six example cases. A short question is one small request with a follow-up, like polishing an email. A long question is a bigger task, and you can play three of them: five questions with twenty web searches (closer to how I actually use it), building a feature and hunting a bug with a coding agent. Both tiles also show the average of the 50 short and 25 long questions we measured for What is my AI footprint?, with what the example costs in brackets. For programmers there's maxing out a 5-hour Claude Code window and the real session that built this tool. And then generating five images and training a whole model.

What you compare it with. By default it's a trip by car, from one city to another. You can change the cities, the car, the number of people in it, or switch to a bus, train, plane or ship. Or compare it with electricity you know: charging your phone or laptop, a video call or a ride on an e-bike. Or with food: a 100 g portion of beef or chicken. For training a model there are whole industries, like the world's livestock farming. And next to all of that runs a water meter, in showers, baths or dishwasher runs.

Play. The conversation plays out on the left and the numbers count up as it goes. You can speed it up, pause it, or skip to the end.

Every example button on this site opens the calculator in a new tab with that example loaded. Your own settings get a "Loaded the shared settings" message with an undo, so you never lose them.

From a chat to a car trip

Under the hood every scenario goes through the same chain. All steps, with the formulas and the numbers filled in, are on the Settings & calculation tab.

Tokens. A model reads and writes text in tokens, roughly three-quarters of a word. The scenario decides how many; why a long chat needs so many more is below. Reading a token costs much less than writing one, and re-reading a token from the cache costs less still. The calculator weighs them that way.

Electricity of the chips. Tokens times the energy per token of the model. This is the most uncertain number in the whole chain; more on that below.

Datacenter overhead. Cooling, lights and power losses come on top. The datacenter's PUE says how much: 1.1 means 10% extra.

CO₂e and water. The same electricity can be clean or dirty, depending on the grid where the datacenter is. Water is what the datacenter evaporates for cooling, plus what the power plants behind it use.

How often. One conversation, or fifty a day, or a 500-person company for a year.

Compared with. For a car, the calculator works out how far that car drives on the same CO₂e, and shows that part of the trip on the map. For electric cars and electricity it uses the same kilowatt-hours instead.

Why long chats cost so much more

Here's the thing that surprises most people. A language model has no memory between messages. Each time you send something, the app sends the entire conversation till this point back to the model, plus a long hidden instruction text (the system prompt). The model reads all of it again before it writes a single word.

That's why the long research chat uses over 100 times as many tokens as the short question, even though it is only five short questions. Every web search result stays in the conversation, so the last answer re-reads all 20 searches before it.

A quick nuance: not every token costs the same. Reading a token (input) takes much less energy than writing one (output), and re-reading a token the provider has cached is cheaper still. So 100 times the tokens doesn't mean 100 times the energy. More on the cache in a moment.

It also makes the answers worse. Models get less reliable as the context grows, even on simple tasks¹. So the chat you've kept open for weeks is both the most expensive and the least sharp one you have.

The cache: a hidden discount you can throw away

Providers soften the re-reading with prompt caching: they keep the start of your conversation in memory on the server, so reading it again costs much less. Without a cache, my long research chat would use about four times as much energy. For coding agents, which re-read huge amounts of context, it's closer to eight times.

You can't switch the cache on yourself (with the exception of some API calls), but you can easily throw it away:

• Waiting too long. The cache expires. On the Claude API it lives 5 minutes after the last use, OpenAI's typically 5 to 30 minutes²; Claude Code on a subscription keeps it for an hour³. Come back to a long thread the next day and the whole history gets re-read at full price.
• Switching models halfway. Each model has its own cache, so switching recomputes the whole conversation³. Claude Code is one of the few tools that tells you: when I switched from Fable to Opus in the middle of a session, it warned me the full history would be re-read on my next message. Other apps don't tell you, but the same thing happens.

So: stick to one model per chat, and start a new chat for a new question.

The knobs that matter most

You can change almost everything, but four knobs make most of the difference.

• The model: a small model needs more than ten times less energy per token than a frontier model. The default is Claude Opus 5.5, which I think is a reasonable benchmark for how people work today.
• The datacenter: the same chat emits about twenty times less CO₂ in Stockholm than in Virginia, because the Swedish grid runs on hydro, wind and nuclear. The electricity barely changes; where it comes from does.
• The cache: switch it off and a long chat uses about four times as much energy, a coding agent about eight times.
• The scale: one chat is nothing. Every chat of every employee for a year is something.

What each of these means for you, your company and the world is in the three blogs linked at the top.

Make it your own

Maybe you don't agree with my numbers. That's fine; on the Settings & calculation tab you can change all of them.

• Pick another paper. Every number is a dot on a strip: one dot per paper, measurement or company report. The highlighted one is the default. Click another dot to use that one instead, shift-click a few to use their average, or type in your own value.
• Read the source. Double-click a dot and the paper opens in a new tab, so you can check it yourself. The Sources tab lists everything with the exact quote or table used.
• Add your own. Add a model, a datacenter with its own PUE and grid, a vehicle or a destination. Handy if you want to use this for a talk at your company.
• Carpool. By default you're alone in the car. Add people and the car's CO₂ gets split between them, which makes AI look worse.
• Share. The share button gives you a link with all your settings in it.

To be transparent about my own process: I picked the default sources together with AI. I asked it which ones it thought were the most trustworthy, and most of the many papers behind the calculator were analysed with AI's help rather than read cover to cover by me. That's exactly why every alternative is one click away.

Why every number here is uncertain

Hardly any AI company publishes energy per token. What we do have are independent measurements, all done differently:

• Lab measurements of open models on bare GPUs, often at maximum efficiency⁴
• Estimates from outside, based on API speed and assumed hardware⁵
• A few company disclosures, like Google's median Gemini prompt of 0.24 Wh⁶ or OpenAI's 0.34 Wh per ChatGPT query⁷

These differ by more than a factor of 10 for similar models. For open models like DeepSeek, Llama or Mistral there are papers and measurements to go on. For closed models like Claude Opus or GPT-6 Astra there aren't, so their numbers are assumptions based on their likely size. That's why the calculator shows every data point as a dot instead of one number.

So treat the results as an order of magnitude: likely within a factor of two or three, not to the decimal. Even when a number is ten times too low, the conclusions in the blogs mostly still hold.

What's left out entirely: your own device and the network (usually small next to the datacenter), and making the hardware. I left the hardware out of the comparisons because it's a one-off cost up front that makes comparing tricky, but it does count. On an average grid it adds about 10%; on a very clean grid it can be more than the electricity itself. You can switch it on in the settings.

Read more

• How much CO₂, water and energy does my AI use? A single question, a day at the office, coding agents and a whole year, next to your commute and your phone.
• How to use AI responsibly as a business (coming soon) What AI costs a company in money and CO₂, and the levers that make it greener.
• How bad is AI for the environment? (coming soon) Data centres next to transport and industry, training the models and why the pinch is local.