What Actually Happens When You Ask AI a Question
Type a question. Any question. Press enter, and start counting.
In about a second, your words will turn into light. They may cross an ocean. They'll be chopped into pieces, pushed through chips with hundreds of billions of switches, and sent back to you one fragment at a time.
The strangest part? When you pressed enter, the answer didn't exist yet.
Last time, we pretended every AI on Earth switched off for a day. Today's question comes right after that one: what exactly did we switch off? Where does it live? And what does it cost every time you ask it something? Let's follow one question, "why is the sky blue?", all the way there and back.
Leg 1: leaving your hand
First, your question has to leave your phone.
The phone turns the letters into numbers, packs them into small envelopes called packets, and encrypts them so nobody on the way can read them. Then it sends them into the air as radio waves, to your WiFi router or the nearest cell tower.
From there the packets drop into a cable, and very soon into glass: a fibre thinner than a hair, carrying your question as flashes of light.
Light in glass is a little slower than light in space. It moves at about two thirds of the speed of light, roughly 200,000 km every second. That sounds instant, but the internet is big. Every 1,000 km still costs about five thousandths of a second.
Leg 2: under the ocean
If the computer that answers you is on another continent, your question goes under the sea.
Almost all data that crosses oceans travels through cables lying on the sea floor, not satellites. As of 2026 there are 607 submarine cable systems in service and another 102 planned, according to TeleGeography.
Most are about as thick as a garden hose, with a few glass hairs inside carrying the world's calls, payments, videos and now your question.
Leg 3: the building with no windows
The cable comes ashore and runs inland to a data centre: a windowless building, often the size of several football pitches, full of computers stacked in tall racks.
It drinks electricity and needs cooling, with air and in many places with water. Almost every watt that goes into those computers comes out again as heat, so a big part of the building isn't computers at all. It's fans, chillers, pumps and pipes. The newest AI racks run so hot that air isn't enough: Nvidia's GB200 NVL72 rack uses direct liquid cooling for about 120 kW of heat in a single rack.
Your packets arrive, get decrypted, and a load balancer picks a machine that's free right now. Among thousands, one of them gets your question.
How much power do buildings like this use? The International Energy Agency says data centres accounted for about 1.5% of the world's electricity in 2025, around 485 TWh. AI-focused data centres grew by about 50% in that one year, and the agency expects data centres overall to roughly double to 950 TWh by 2030.
The chip
Inside that machine, your question meets the real engine: chips like Nvidia's Blackwell, which packs 208 billion transistors, tiny switches flicking on and off very fast.
For a big model, one chip isn't enough. Several work together, sharing one enormous file. That file is the model: hundreds of billions of numbers, learned during training, that together capture how words tend to follow each other.
Why can't your phone just answer?
Why send your words around the world at all? Size. A big open model like Llama 3.1 405B has 405 billion numbers, roughly 800 GB of data. A flagship phone has about 8 to 16 GB of working memory; Samsung's Galaxy S25 Ultra, for example, comes with 12 GB. It's like trying to fit a library into a backpack.
Phones do run small models now, for autocorrect and short replies. But the big, clever answers still live in that building.
Leg 4: chopped into tokens
The model can't read letters. So your question is first cut into pieces called tokens: "why", " is", " the", " sky", " blue", "?".
In English, a token is about four characters, or three quarters of a word, on average. Many other languages need more tokens for the same sentence.
Then every token becomes a long list of numbers, and that grid of numbers is what flows through the chips.
Attention: every word looks at every other word
Now the model does something it repeats dozens of times, layer after layer. Every piece of your question looks at every other piece and asks: how much do you matter to me?
"Blue" pays a lot of attention to "sky". "Why" tells the model you want a reason, not a list.
This trick is called attention. In 2017, a team at Google built a whole model out of it, the Transformer, and almost every chatbot you've used is built on that idea.
One piece at a time
Out comes one token: the model's best guess for what should come next. "Sunlight."
That token is added to everything before it, and the whole thing runs again to guess the next one. " is". Then " made". And so on.
That's why the answer seems to type itself out. It isn't being typed. It's being written, one piece at a time, while it travels to you.
For a fast model, the first piece usually reaches you in about a second. Models that "think" first can take much longer. After that, tokens stream back at dozens or even hundreds per second, across the same cables and through the same air, into your hand.
What does one question cost?
The companies have started to publish numbers, and they land in a fairly narrow range of about 0.24 to 0.34 watt-hours for a typical text question:
- Google measured 0.24 Wh and about 0.26 mL of water for a median Gemini text prompt in 2025. The water figure covers cooling inside the data centre, and the study excludes external networks, your device and training.
- OpenAI's Sam Altman wrote that an average ChatGPT query uses about 0.34 Wh and roughly one fifteenth of a teaspoon of water, without publishing a method.
- Epoch AI, an independent research group, estimated roughly 0.3 Wh for a typical GPT-4o query.
That's about the same as an oven running for a second, or a TV for a few seconds. Tiny.
But long questions, uploaded files and reasoning models can cost many times more. And there are a lot of these questions: ChatGPT alone was getting about 2.5 billion prompts a day in 2025.
Tiny, times billions, is not tiny anymore. That's why power companies now plan around buildings like the one we just visited. It's not "AI is destroying the planet", and it's not "it's nothing". It's a real, measurable load that grows with every question.
Back in your hand
The first words reached you in about a second, and the rest followed right behind.
In those few seconds, your question became light, crossed the ground and maybe the sea, entered a building that drinks electricity, was cut into pieces, passed through hundreds of billions of switches, and was written back to you word by word.
So "AI" isn't a little voice inside your phone. It's one of the biggest machines humans have ever built, spread across the planet, and it starts working every time you press enter.
That's what we switched off last time. Now you know where it lives.
Missed the first episode? Read What If Every AI Shut Off for 24 Hours? Next time, we slow that one second right down and watch a model think, one word at a time. When we're not making these, PretendMind builds AI tools for real companies. Say hello.