Behind the scenes
What AI consumes, and how it really works
Machines, electricity, water, tokens: we open the hood, with sourced, dated figures and telling comparisons.
The numbers to remember
≈ 0,3 Wh
per question asked to an AI assistant, i.e. an LED bulb lit for 2 minutes (Epoch AI)
1,5 %
of world electricity used by data centers in 2024, 415 TWh (IEA)
× 2,3
by 2030: 945 TWh projected, nearly 3% of world electricity (IEA)
≈ 0,5 L
of fresh water per 10 to 50 answers, cooling included (UC Riverside)
Sources: Epoch AI (2025), Google (2025), International Energy Agency (2024-2030), UC Riverside. Figures evolve fast: we update them with our weekly watch.
How does an AI learn?
An AI's life has two very different phases. Training first: for weeks or months, tens of thousands of specialized processors adjust billions of internal "settings" by analyzing huge amounts of text and images. It is the most expensive phase, done once per version. Inference next: each time you ask a question, the already-trained model computes its answer. One request costs almost nothing... but multiplied by billions of daily uses, it is what weighs the most overall.
Read the discovery version
Before answering your questions, an AI goes to school for months: thousands of computers make it read almost the whole Internet, again and again, until it can guess how sentences continue. After this big school (called training), it can answer in a second: like reciting a very well-learned lesson.
The means deployed: a giant industry
Behind every answer lies an industrial chain: specialized chips (GPUs, 700 to 1,000 watts each under load, i.e. a microwave oven), assembled by the tens of thousands in data centers, powered by reinforced grids and cooled with air or water. AI giants invest tens of billions per year in these infrastructures, design their own chips (Google, Amazon, OpenAI, Anthropic) and sign dedicated energy contracts, including nuclear and solar.
Read the discovery version
AIs don't live in your phone: they live in huge buildings full of computers, called data centers, guarded day and night. Inside, very powerful special chips get so hot that big fans and water are needed to cool them down, like a cyclist you spray after a big race.
What is a token?
The token is the base unit of every language AI: a fragment of a word, about 3-4 characters on average (an English word is 1 to 3 tokens). The model turns your question into tokens, computes thousands of probabilities for each, then generates its answer one token at a time. It is also the billing unit of professional AI: you pay per million tokens read and written. And it is the memory unit: a model's "context window" (e.g. 1 million tokens for the latest ones) is how many pieces it can hold in mind at once.
Read the discovery version
An AI doesn't read words like you: it cuts sentences into small pieces called tokens, like Lego bricks. "Fascinating" may become two bricks: "fascin" + "ating". The AI then guesses the next brick, again and again, very fast: that's how it writes its answers, brick by brick!
So, does it consume a lot or not?
At the individual scale, the impact is modest: about 0.3 Wh per question (Epoch AI; Google measures 0.24 Wh and 0.26 mL of water for a median Gemini request), far less than one minute of video streaming. At the global scale, it is another story: data centers used 415 TWh in 2024 (1.5% of world electricity) and the IEA projects 945 TWh in 2030, driven by AI. Cooling water counts too: about half a liter per 10 to 50 answers. The real issue is not your evening question, but the collective trajectory, and how fast the industry covers it with low-carbon energy.
Read the discovery version
Asking an AI one question uses about as much as a small bulb lit for two minutes: almost nothing! But since millions of people ask millions of questions every day, it all adds up, like a giant anthill. So the big AI houses build solar panels and look for tricks to consume less.
🌿 Good habits
- Reuse the same conversation rather than opening ten: the AI keeps the context and works less.
- Precise questions = short answers = fewer generated tokens.
- No need for AI for everything: a calculator or a classic search often consumes less.