The sky does not float. Somebody is standing under it, and you have never seen him

the claim

The cloud is a building

Type a question to a chatbot and watch what your own senses report. The answer condenses out of nowhere, word by word, weightless. Nothing hums. Nothing warms up. The word the industry chose for where this happens — the cloud — agrees with your senses completely: somewhere up, somewhere soft, somewhere that does not have a street address.

Every part of that impression is false, and falsifiable. Your question leaves your phone as radio, reaches a mast, drops into a glass fibre, and travels — as light, in glass, at two-thirds of its top speed — to a very specific building with a fence, a car park and a water bill. Inside, in a hall longer than a football pitch, it joins a queue for a slab of silicon that is, at this moment, one of the hottest man-made objects in the room it sits in. The slab does the arithmetic you met in the other pieces — the maps, the stencils, the Oracle's scroll — at a few hundred trillion operations per second, for you, for about a second. The answer retraces the fibre. The words appear. Total round trip: often under a second, which is precisely why you have never once been forced to think about any of it.

The Greeks would have found the word “cloud” very funny. They knew what holds a sky up, because they had done the accounting: it takes a Titan, all of him, forever. This piece is that accounting for our sky. It runs on four meters — distance, heat, water, and hands — and each one gets an instrument you can push on.

meter one

Heat is the rent

Here is the single most physical fact in this whole series. A chip has no exhaust pipe, no moving parts, nowhere for the energy to go. So essentially every joule of electricity that enters a chip leaves it as heat. Not some of it. Not the wasted part. All of it. Computing does not consume electricity the way a car consumes petrol, leaving some of it as motion; computing borrows electricity on its way to becoming warmth. The arithmetic is a toll gate the energy passes through.

A single modern training chip draws about as much power as a room heater — seven hundred watts, more for the newest ones — and a training machine straps eight of them together. The building we imagined above holds tens of thousands. Which means a data centre is, in the most literal sense available, a building-sized electric heater that thinks. And a heater that must not be allowed to get hot: silicon starts misbehaving around ninety degrees, so every watt that goes in must be found, collected and carried out of the building, continuously, forever. The chip below is honest — its temperature follows from the power you feed it and the cooling you grant it, by the same physics as a pot on a stove.

meter two

The bill for one answer

So what does your question cost, when it lands on that hot chip for its second of attention? This is a question with a genuinely honest answer and a genuinely dishonest history, so let us do it properly: with a range, and with the reasons the range is wide.

A short chatbot answer costs somewhere between a fraction of a watt-hour and a few watt-hours of electricity, depending on the size of the model, the length of the answer, and how many questions the building is juggling at once. Google, in the first detailed public accounting by a company running these systems at scale — its own accounting, with its own boundary-drawing, which the sources note — measured a median of about a quarter of a watt-hour per text prompt: roughly nine seconds of a running television. Independent researchers, measuring open models on their own hardware, find numbers from several times smaller to ten times larger. The honest picture is a band, not a number, and the instrument below draws the band instead of hiding it.

A quarter of a watt-hour is genuinely small. And it is multiplied by billions of questions a day — that is the trick of the sky, every single stone of it is light — and it is only the second bill. The first bill was training: months of thousands of chips running flat out, before the model could answer anything at all. That one is paid once, in advance, and it is paid in gigawatt-hours.

meter three

The river through the building

All that heat has to go somewhere, and “somewhere”, at building scale, usually means water. The cheapest way to shed a megawatt is the way your own body does it: evaporation. Warm water is trickled through open towers on the roof; some of it becomes vapour and carries the heat into the sky — the honest, physical sky — taking roughly one to two litres of water per kilowatt-hour of computing with it. That water is not borrowed. Evaporated water leaves the local river, lake or aquifer and comes back as rain somewhere else entirely.

There are other ways. You can cool with fans and radiators alone — dry cooling — which spends almost no water but pays for it in extra electricity, and struggles on hot days, which are exactly the days the grid is already struggling. You can put the building somewhere cold and let the weather do the work. Every siting decision is this trade, made with a spreadsheet: water against watts against distance-to-users, in a place with a real river and real neighbours. The instrument below hands you the spreadsheet.

meter four

The hands in the pile

Electricity and water hold up the answering. Something else entirely holds up the knowing, and it is not made of silicon. You met the pile in Procrustes: the mountain of examples a machine learns from. Here is the fact about the pile that the word “data” is designed to make you forget: nearly every piece of it passed through human hands. The sentences were written by people. The photographs were taken by people. And an enormous share of it was labelled by people — someone sat at a screen, looked at an example, and recorded a judgement, for money, thousands of times a day.

Someone drew the box around every pedestrian the driving pile knows about. Someone listened to hours of Amharic radio and typed what was said — at Phonetico that someone has a name and an hourly rate we set, which is part of why this chapter exists. And when a chatbot learned manners — learned to prefer the helpful answer over the sneering one — it learned from hundreds of thousands of judgements of exactly the kind you are about to make: two answers side by side, a person, and the question which is better?

The going rate for this work, on the big labelling platforms, runs from under two dollars an hour in Nairobi or Manila to fifteen or twenty for specialists. The hardest of it — filtering the worst of the internet out of the pile so the model never learns it — is work that leaves marks on the people who do it, and it was done, for the systems you use, at those rates. The instrument below will not show you that part. It will do something more useful: it will measure you, and then scale you up.

the ledger

Who can afford a sky

Add the meters up. A frontier model — the biggest kind, the kind whose name you know — is trained on tens of thousands of top-end chips running for months. The best public estimates put the bill for a single such training run in the tens to hundreds of millions of dollars, and each generation costs several times the one before. Include the buildings, the substations, the water rights and the salaries and the number climbs toward billions. There are perhaps a dozen organisations on Earth that can pay it, and none of them are universities, and none of them are countries the size of ours.

Sit with what that concentration means. Every piece in this series has ended with the choices are ours — but a choice needs hands to make it with, and if only a dozen skies can be afforded, then it is a dozen boardrooms choosing what the sky contains: which languages it speaks, which pages went into the pile, whose voice it answers in. Atlas at least stood at the edge of everyone's world. These stand rather closer to some people than others.

Now the honest counterweight, which is our own working day. You do not need a frontier sky to hold up a language. The speech recognisers this series keeps returning to — the ones learning Amharic, Tigrinya, Afaan Oromo — are trained at Phonetico on machines that fit in one ordinary room: a handful of the same hot chips, a domestic-scale electricity bill, no river required. Weeks of patience instead of megawatts. That is not a boast about frugality; it is the practical fact this whole series is built on. The ideas in these pieces are not owned by the people with the largest buildings. A small team, a corpus gathered with consent, and one room of shoulders can hold up a real sky over a real language — ours does. The concentration at the top is real, and it matters, and it is not the whole story.

the last thing

The mountain

The myth has one more scene, and the Greeks — who put their warnings inside their geography — made it the ending. Perseus, flying home with the Gorgon's head in its bag, stopped at the western edge of the world and asked Atlas for shelter. Atlas, warned by a prophecy about a son of Zeus, refused him. So Perseus, from the piece before this one, held up the one face no one can look at — and the Titan turned to stone. His beard became forests. His shoulders became ridgelines. He became the Atlas Mountains, which stand in Morocco to this day, holding up the Moroccan sky, and nobody who looks at a mountain wonders how it feels about the weight.

That is the fate of all infrastructure that works: it becomes landscape. The electricity grid was once the wonder of the age; now it is scenery you notice only in a blackout. The machines in this series are petrifying the same way, in real time, in front of you — folding into phones and search boxes and call centres, becoming the mountain range your generation will simply live among. The Greeks did tell one more story, though. Heracles, needing the golden apples, once took the sky onto his own shoulders — the only figure in the whole mythology who ever volunteered for the weight, just for an hour, just to make a trade possible. He is also the only one who ever felt how heavy it was and said so out loud.

You cannot hold the sky. But you have now done the next best thing: you have felt the weight — the milliseconds in the glass, the watts becoming warmth, the litres leaving the river, the afternoons of the people at the bench. A person who knows what the sky weighs asks better questions about it: whether this errand needed a frontier model or a small one, where the building is and whose river it drinks from, who labelled the pile and what they were paid. Those questions are the hour of holding. The gods at the feast never ask them. You are not obliged to be a god at the feast.

the glossary

The words the engineers use