Santiago.

🔋 Where does the energy AI consumes come from?

October 8, 2026 · 7 min read

Illustration of an exhausted young man in bed at dawn, looking at his phone while threads of energy flow from his chest to an AI agent and a nearly empty red battery floats above his head

When people talk about how much energy artificial intelligence uses, they almost always mean two things: what it costs to train a model and what it uses every time we run it. But if you ask me where the energy AI consumes comes from, I'd say it comes from three places. The first one gets a lot of attention. The other two, almost none.

🏭 1. The one everyone measures: data centers

According to the International Energy Agency (IEA), data centers used around 415 TWh in 2024, roughly 1.5% of the world's electricity, and its projections point to around 945 TWh by 2030. 🔌 A single query, on the other hand, uses very little: Google estimates that a typical text prompt to Gemini takes about 0.24 Wh, roughly one second of a microwave.

The takeaway seems clear: each query is tiny; what matters is the total, and the fact that it never switches off.

🏠 2. The one nobody adds up: our devices

AI doesn't only live in data centers. It also lives in the phone that dies before the day is over, in the laptop left on all night so an agent can keep working, in the charger that never gets unplugged, in the router that never rests and in the screen lighting up our face at two in the morning. 📱💻

Each device uses little. But multiplied by hundreds of millions of people, in every home and office in the world, it stops being little.

This isn't a baseless idea. A study by Malmodin and colleagues, published in Telecommunications Policy (2024), estimated that in 2020 the whole digital sector used about 4% of the world's electricity: around 223 TWh for data centers and around 250 TWh for networks. And according to summaries of the same work, user devices (phones, computers, tablets) used even more than either of those. That data predates the AI boom, and nobody has yet measured how much of what our devices use today is because of it. That's exactly why it's worth naming: it's energy spread across millions of nightstands that never makes the headlines.

🧠 3. The one we don't even feel: ours

And then there's the third. It doesn't come from any power outlet. It comes from us.

🌙 Just one more minute

There's a scene that keeps happening more and more. Someone wakes up before dawn, still half asleep, and before getting out of bed they already have their phone in hand. They check what the AI agent did overnight, give it one more instruction and promise themselves they'll go right back to sleep.

That promise is almost never kept.

I know because, many times, that person has been me. For weeks I've been working long hours with AI: I get up very early to keep going and I go to bed very late. And I'm not an isolated case:

  • 🌐 Well-known people in the industry talk about it. According to Axios, Garry Tan, CEO of Y Combinator, said he spent 19 hours straight working with agents and didn't fall asleep until 5 a.m. Andrej Karpathy, the person who coined the term vibe coding, has even talked about a kind of "AI psychosis". And the CTO of a tech company compared these tools to slot machines.
  • ⏳ Companies see it too. In 2025, Anthropic added weekly limits to its subscription plans because some users had its coding tool running "continuously in the background, 24/7".
  • 📈 There are more of us every day. According to Pew Research, the share of U.S. adults who use ChatGPT went from 18% in 2023 to 44% in 2026, and OpenAI reports around 900 million weekly users.

🎰 Why it's so hard to stop

It isn't a lack of willpower. There are very human mechanisms at work:

  • ⚡ Instant reward. You ask for something and within minutes you see it working. Our brain reacts strongly to anticipating a reward, especially when the outcome is uncertain: "let's see if it comes out perfect this time". That's why the slot-machine comparison keeps coming up. It's an analogy, not a diagnosis, but it makes sense.
  • 🚀 The illusion of the impossible. Projects that sat in a drawer for years now get built in weeks. That's real, and it's wonderful. But when everything seems possible, every idea becomes urgent.
  • 💸 "Use it or lose it". Many AI subscriptions have limits that reset over time. If you don't use them, you feel you're wasting what you already paid for. A classic Harvard study on gyms found something similar: people who pay up front go a lot right after paying, pushed by that very feeling. Applying it to AI is my own observation, not a study, but I recognize it in myself.
  • 📱 No boundaries. The agent keeps working while you have dinner and pings you in your pocket. Work and rest live in the same device.

🧟 The tiredness you don't feel

This is what worries me most. I go through the day sleepy, in a kind of zombie state, and yet I don't feel tired. Every so often I feel I'm moving forward, making progress, getting things done. The satisfaction of seeing what AI does gives back what it takes from me. Or so it seems.

It's like an energy loan: you feel you have plenty, but you're borrowing it from your sleep.

The data suggests that feeling can fool us:

  • 🧪 METR (2025). In a trial with 16 experienced developers, those using AI took 19% longer, yet believed they had been 20% faster. It's a small study, with tools from early 2025. The valuable part isn't that AI slows us down, but that feeling fast and being fast aren't always the same thing.
  • 🏢 Berkeley Haas (Harvard Business Review, 2026). An eight-month study at a tech company found that AI didn't reduce work, it intensified it: more tasks, faster pace, fewer breaks.
  • 🌫️ BCG (2026). A survey of almost 1,500 workers described what they called AI brain fry: mental fog and slower decisions in about 14% of respondents.

🍔 Whose fault is it?

I think the responsibility is shared. It's a bit like fast food: nobody forces you to buy it, but it's designed so that saying no is really hard. Tech companies know very well how our attention works, and that gives them a responsibility.

But that doesn't take ours away. We're still the ones who decide when to open the phone and when to put it down, even if it's hard.

🌱 We're still human

I haven't tried anything to slow down yet, and I'd rather be honest about that. But I want to start: setting disconnection hours without buying into the illusion that if I stop, I lose. Because that's exactly what keeps you from stopping.

If a friend told me they were in the same place, I'd tell them this: at some point some things have to be paused, because we're still human. We can't set aside our daily life and the things we love to chase, so aggressively, a dream that may be taking many other things down with it. Sleep among them. 😴

In the end, everything we build, an app, a project, a company, we build looking for well-being. And well-being isn't being glued to a screen all day. Well-being is living: enjoying what that investment can give us, with the people we love and the things that make us happy.

So yes, let's measure the watts in data centers. Let's add up the ones in our homes too. And let's start measuring the rest.

⚡ May the biggest energy AI consumes never be our own life energy.

📚 Sources

✍️ By Santiago Perea Restrepo

← Back