AI and the environment

Whatever else this technology is, it is buildings. Not a cloud and not a metaphor — warehouses of machines, substations, transmission lines, cooling plant, sited on real land with real neighbors. They are hot, they draw a great deal of electricity, and they are going up quickly. To a lot of people living nearby, the decision looks as though it was already made by the time they heard about it.

What makes this a serious objection rather than a mood is that the cost is invisible exactly where the choice gets made. A car burns fuel you buy. A data center burns power somebody else buys, in a building you will never see, on a grid you did not pick. Your screen shows a blinking cursor and nothing else. The only way to know what a thing costs is to be told, and on this you are mostly not told.

A parent asking about it is asking the ordinary question. A new industry arrives with growth first and accounting later, the benefits are spread thin and far away, and the costs — the water, the transmission lines, the plant that does not close on schedule — land on whoever happens to live near it. That has happened often enough to be a pattern rather than a suspicion. If you care what the world looks like when your child is grown, wanting to know what is being built and what it needs is not obstruction. It is the question that is supposed to be asked before the concrete goes down.

Training a large model is a single concentrated industrial event. Great banks of processors run flat out for weeks or months in one place, doing nothing useful for anyone yet, to produce a finished model. That is a large amount of electricity spent up front, before a single person has asked the thing a question. It is the part usually described as the environmental cost of AI, and it is real.

Then there is running it, which is what happens every time anyone uses it. One answer is a very small slice of what training cost. That is the honest shape of it, and it is why a per-question figure sounds so slight next to the training number.

But a model that many people use is answering constantly, all day, everywhere, for as long as it stays in service — and across the deployed life of a widely used model, serving it adds up to more than training it did. This is the distinction worth carrying away. The upfront cost is the one that gets quoted, because it is dramatic and finite. The ongoing one is small per use, unbounded, and grows with adoption. The aggregate is where the environmental question sits.

A data center uses the electricity mix of its local grid, which changes by place, hour, and season. The same work can therefore have very different effects in two locations.

Cooling can use water, or use air and more electricity. The tradeoff matters most locally: drawing water beside a large river is different from drawing it in a dry region with a stressed supply.

Growing demand can lead utilities to build new generation and transmission. In some places, it has also kept fossil plants open longer or delayed their retirement.

If this cost matters to you, using these tools less is a coherent response rather than a symbolic one, and it does not require you to first settle a number nobody has. It helps to know that the load is not spread evenly. Generating images and video is considerably more work for the hardware than a short piece of text, and re-running the same request over and over because the answer came back nearly right is where a lot of ordinary waste sits. If you are going to use less of something, it is worth knowing which part is heavy.

It is also consistent to hold both positions at once — that you want your child to understand this technology, and that you do not want your household to be a heavy user of it. Understanding how something works has never required consuming much of it. Most of what is worth knowing about how these systems are trained and where they fail can be learned from a handful of sessions and a clear explanation, and a family that stops there has not missed the point.

Whether efficiency gains outrun demand growth is the question the whole thing turns on, and it will be settled by what happens rather than by anyone's argument about it. Both directions have serious people behind them, and neither has the evidence to close it. Anyone telling you now that this is a bump efficiency will absorb, or that it is a step change that swamps everything else, is forecasting — and in most cases the forecast was chosen to fit a conclusion already held.

What this costs where you live is knowable in principle and mostly unavailable in practice. You could ask how much water a nearby facility draws, how much of it comes back, what your utility is building to serve it, and who is paying for that. In a great many places you will not get an answer, because nobody is obliged to give one. That is not a law of nature. It is a reporting rule that was never written, and rules like that get written when enough people ask.

Whether disclosure ever becomes mandatory is a political question rather than a technical one, and until it is answered, every argument about this — including this page — is being conducted with the load-bearing numbers missing, held by the parties being argued about. A parent who takes all of that in and decides to use these tools sparingly, or not at all, has read the situation rather than flinched from it. Deciding without adequate numbers is the only kind of deciding on offer here, and choosing restraint in the face of a cost nobody will quantify is a defensible reading of it, not the absence of one.