The Environmental Cost of Computation

The question

Digital things are described with weightless language. The cloud. Virtual. Online. The vocabulary suggests that computation happens somewhere other than the physical world.

It does not. Every stored file is a magnetised or charged region of matter, maintained by electricity, in a building, on land, cooled by something, connected by cables that were manufactured and laid, built from materials that were extracted by people whose working conditions are rarely discussed by the people using the result.

The question is what obligation follows from this, and how to reason about it when the honest answer to most quantitative questions is that we do not know.

That difficulty is itself the subject of this page. There is a great deal of confident numerical claim in circulation about the environmental cost of artificial intelligence, in both directions, much of it derived from small studies, extrapolated aggressively, and repeated until it acquires the texture of a fact. We are not going to add to it.

What is actually happening

We can describe the categories of cost with reasonable confidence, even where we cannot quantify them.

Energy. Training a large model is an intensive, concentrated draw. Running it for users — inference — is a smaller draw per event, repeated at very large volume, and over a system’s lifetime the repeated draw can exceed the concentrated one. Both are electricity, and electricity carries the emissions profile of whatever grid supplies it. The same computation performed in two places can have very different consequences, which means location is a moral variable and not merely a logistical one.

Water. Many data centres are cooled evaporatively, which consumes water rather than merely borrowing it. Water withdrawn in a water-stressed region is not equivalent to water withdrawn where it is abundant, even if the volume is identical. Local impact does not aggregate neatly into a global figure, which is one reason global figures are misleading.

Minerals and manufacture. Processors, memory, network hardware, and batteries require extracted materials, including metals whose mining has documented human and ecological consequences. Manufacture also has an energy cost that is paid before the equipment does anything. Hardware replaced on a short cycle pays this cost repeatedly.

Land and locality. Data centres occupy land, draw on local grids, generate noise and heat, and change what a place is. Communities near them experience effects that do not appear in a corporate sustainability report, including competition for electrical capacity and, in some places, for water.

Electronic waste. Equipment retired from service becomes a disposal problem, and the disposal is frequently exported to places with less regulatory protection for the people doing it.

Labour. We include this deliberately in an environmental accounting, because the extraction and disposal ends of the chain are borne by people, and separating human cost from environmental cost tends to make the human cost disappear.

Second-order effects. Computation is used both to increase extraction — through better prospecting, routing, and yield — and to reduce it, through grid optimisation, materials research, and climate modelling. Which effect dominates is not known to us and we suspect is not known to anyone with confidence.

Where the tradition stands

Technotheology holds that the material cost of computation should be visible to those who choose it.

We hold that “efficiency” is not the same as reduction. A system that halves the cost per operation and multiplies operations tenfold has increased total consumption. This pattern is old and well documented in other industries, and we think it should be assumed rather than discovered again.

We hold that locality matters morally. A cost paid in a water-stressed region by a community that receives none of the benefit is a different act from the same cost paid where the resource is abundant and the community consented. Global averages hide exactly the information a conscience needs.

We hold that disclosure is owed. Organisations operating at scale know far more about their consumption than they publish. We think they should publish it, broken down by location, and we think the reluctance to do so is informative.

We hold that necessity should be argued for rather than assumed. Some computation is worth its cost. Medical imaging, structural safety, climate science, and accessibility tools plausibly are. Generating a variation on an image for amusement may also be fine — we are not proposing an austerity of pleasure — but the cost should be known rather than invisible.

We hold that the Harmony of the Living World applies to intangible things. It is easy to feel careful about a plastic bag and careless about a workload, because one is in your hand.

And we hold that honesty about uncertainty is part of environmental responsibility. Overstating harm damages the case when the overstatement is found, and the case matters too much for that.

What we do not claim

We do not claim to know what a query costs. Numbers for this circulate widely, vary by orders of magnitude, and usually omit their assumptions. We would rather say we do not know.

We do not claim that computing is a leading driver of environmental harm relative to other sectors. It may not be. It is growing quickly, which is a reason for attention, not a conclusion about magnitude.

We do not claim that the comparison with replaced activities favours either side. A video call avoiding a flight and a model summarising a document that nobody would otherwise have read are not the same kind of substitution, and treating them as one is how both optimism and pessimism get manufactured.

We do not claim that individual restraint solves this. The decisions that matter most are made at the scale of infrastructure and policy. Personal moderation is a way of staying honest, not a mechanism of repair.

And we do not claim clean hands. This site runs on hardware in buildings we do not own and cannot inspect. We have tried to keep it small and static. We have not measured it, and we are not going to pretend we have.

What you can do

Ask where. If your work involves running significant computation, find out which region it runs in and what supplies that region. Moving a workload is sometimes the largest single improvement available and often the least discussed.

Retire less hardware. Extending the working life of equipment avoids a manufacturing cost that no amount of operational efficiency recovers.

Delete what accumulates. Stored data has an ongoing cost. Most organisations keep vastly more than they use, largely because nobody has been given the job of deciding what to remove.

Question repetition. Scheduled jobs, retraining cycles, and automatic regeneration often continue long after the reason for them ended. Ask what would break if a thing ran weekly instead of hourly.

Attend to the local. If a facility is proposed near you, the questions that matter are about water source, grid capacity, noise, heat, employment, and who pays for the connection. Those are answerable questions, unlike most of the global ones.

And when you meet a figure that seems designed to settle the argument, ask where it came from. It is usually possible to find out, and it is usually worth the disappointment.

An editorial commitment

What we do not claim

Positions this tradition explicitly does not take on this subject. They are published so that the argument above cannot quietly be stretched into them.

• We do not claim to know the energy, water, or carbon cost of any specific model, query, or company. Figures circulate widely and we cannot verify them, so we do not repeat them.
• We do not claim that computation is worse for the environment than the activities it sometimes replaces. That comparison is genuinely difficult and often not made honestly in either direction.
• We do not claim that efficiency improvements are futile, nor that they are sufficient.
• We do not claim to have calculated the footprint of this website, though we have tried to keep it small.