The Secret Journey of Your AI Question: From Phone to Power Grid and Back
Ask an AI chatbot a question and the answer lands in two or three seconds. In those seconds, your words take a round trip most people never think about. Here is the whole journey, second by second.
Second 0: You press send
Your phone does almost none of the work. It simply wraps your question into data and pushes it into the internet through your WiFi or mobile network. Your phone is the postbox, not the brain.
Second 0 to 1: The question races to a data centre
Your words travel through cables, often under oceans, to a data centre: a warehouse packed with thousands of computers stacked in racks. There are roughly 11,000 such buildings worldwide, and your question is routed to whichever one runs your chatbot, quite possibly in another country. It can cross half the planet in well under a second, because light in a fibre optic cable covers about 200,000 kilometres every second.
Second 1 to 3: The chips do the thinking
Inside, your question lands on GPUs, special AI chips built to do millions of small calculations at once. Around 8 in 10 of them are made by one company, NVIDIA. The chips run your words through a giant neural network that predicts the best response word by word, thousands of calculations for every single word it writes.
This is the expensive moment. A single top AI chip draws around 700 watts, nearly ten times your ceiling fan, and the building runs hundreds of thousands of them at once. That is why one large data centre can use as much electricity as a small city.
The two silent partners: the grid and the water pipe
While the chips think, two utilities keep them alive. The first is the electricity grid, the national network of power stations, pylons and cables. The International Energy Agency projects that data centres will double their electricity use from 485 terawatt hours in 2025 to about 950 by 2030, more than all of Japan uses today, and grids are struggling: in the UK, some projects wait up to 15 years just for a connection.
The second is water. The chips’ heat has to go somewhere, and most data centres get rid of it the way your body does, by evaporating water. A large facility can consume up to 19 million litres a day, as much as a small town. When you count the water evaporated at the power plant too, one 100 word answer has been estimated to cost up to a bottle of water, though AI firms say a typical prompt uses far less, and the truth depends on the model and answer length.
Second 3: The answer flies back
The finished text retraces the same route, across the cables, back through your network, onto your screen. Total round trip: a few seconds, one small sip of a city’s worth of electricity and water, all invisible to you.
Next time the answer appears instantly, remember what actually happened: your words visited a warehouse of chips fed by a power grid and cooled by a water supply, possibly on another continent, and came back before you finished your sip of tea.