Environmental Cost
The electricity and resources it takes to build and run AI
- AI draws power in two places. Training a model once, and then a little every time someone uses it afterward.
- Training spends a lot in one burst, but as usage grows, the running total from everyday use can pass what training cost.
- Computation isn't the only draw. Cooling the machines, and building and scrapping them, both take energy and materials too.
- A figure like "cost per question" varies wildly between sources, because each one counts a different slice of the picture.
- There are ways to cut it down: use a big model only when the job needs one, reach for a smaller model, and pay attention to where the electricity comes from.
Contents
1The analogy
A monthly electric bill arrives with one total on it. That single number doesn't say which appliance spent what. The fridge that never stops, the air conditioner run hard for a few days, the charger left plugged in and barely used, all of it is folded into one line.
When the bill spikes in summer, the air conditioner is the easy guess. Sometimes the real cause is several small devices left running all day instead. A single large draw and many small ones look identical once they land on the same line.
The environmental cost of AI looks like that bill. There's a share spent in one big burst, a share that piles up a little at a time with every use, and a share that barely shows up at all. Without pulling the three apart, there's no way to know what to cut.
2In detail
The lump sum and the trickle
Training a model means running an enormous number of calculations, over and over. Thousands of dedicated chips run around the clock for weeks, and that stretch draws a large, concentrated amount of power. Once training finishes, that cost doesn't come back.
Using a trained model looks tiny by comparison, one answer at a time. The problem is volume. When a service generates hundreds of millions of answers a day, that small share never stops adding up. The longer a service runs, the bigger the running total gets, and past some point it can pass what training cost.
So there's no single answer to which side is bigger. A model built and used by a handful of people spends almost everything on training. A model the whole world queries several times a day spends almost everything on running it.
It isn't just computation that draws power
Every calculation a chip runs gives off heat. Left alone, that heat would make the hardware fail, so cooling systems run alongside it around the clock. A sizable share of a data center's electricity goes into just keeping things cool.
Depending on how the cooling works, water gets involved too. Systems that circulate water to carry heat away lose some of it to evaporation. Where water is plentiful this barely matters; where it's scarce, water becomes the bottleneck before electricity does.
Then there's the cost of the hardware itself. Manufacturing dedicated chips is an energy-intensive process on its own, and a chip gets replaced every few years once it falls behind. Counting only the electricity used while running leaves this whole slice out.
Why the numbers are hard to pin down
Figures like "cost per question" show up everywhere, and they disagree by a wide margin. The first reason is scope: whether a count stops at the electricity the computation used, or reaches further to include cooling, or further still to include building the hardware, changes the answer by several times over.
The second is the conditions: a short question versus a long one, a short answer versus a long one, the size of the model, whether requests get batched together, all of these can swing the computation behind a single answer by tens of times. The third is where the electricity comes from. The same amount of electricity produces very different emissions depending on how it was generated.
The fourth is disclosure. The companies running the hardware hold the real numbers, and they don't publish all of it. Outside estimates fill the gap with assumptions, and different assumptions produce different results. Any number worth trusting has to say what it counted. Last checked: 2026-09
Where the reductions can come from
The biggest lever is usually "did this really need a large model." A simple classification job or reformatting a fixed layout is well within a small model's reach, and calling a large model out of habit can multiply the computation by tens of times. Matching model size to the job alone cuts a lot.
Builders have levers too: passing a large model's judgment down into a smaller one, computing with coarser numbers to cut the arithmetic, or caching common answers instead of recomputing them. On the infrastructure side, shifting work to times and places where the electricity is cleaner is another option.
3More precisely
Assessing environmental cost usually splits into three questions: how much electricity was used, how that electricity was generated and how much it emitted, and how much went into manufacturing and disposing of the hardware. Counting only the first two leaves the third out entirely, and counting electricity alone erases the difference between generation methods.
The analogy breaks down in one place: an electric bill is at least accurate about its total, but no such total is handed to an outside observer of AI. There's no bill to read in the first place, and the closest thing to one usually comes from an outsider's estimate rather than a real meter reading.
One more place it breaks down: better efficiency doesn't reliably shrink the total. When each answer gets cheaper to produce, it also gets used more often and in more places, and the total can rise even as the per-answer number improves. What one device saves and what the world as a whole uses are two different stories, which is why an announcement about a better per-answer figure and a report of rising overall use can both be true at once.
Last verified: 2026-09
4Try it yourself
5Common misconceptions
It's easy to think electricity is mostly spent during training, but actually the more widely a service is used, the more the running total from everyday use grows over time.
It's easy to think running a model on your own laptop has no environmental cost, but actually the cost has just moved from a distant building to your own desk.
It's easy to think there's a fixed number for what one question costs, but actually it swings by several times depending on what gets counted and under what conditions.
7One-line summary
In shortThe environmental cost of AI adds up the one-time burst of training, the running total of everyday use, and the cooling, manufacturing and disposal on top of both, and the number changes a lot depending on what gets counted.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02