The costs not on the invoice

A query feels weightless. Nothing about the interface suggests otherwise, and the price is often zero.

It is not weightless. This lesson is about what sits behind it — stated carefully, because the topic attracts both dismissal and exaggerated numbers, and neither helps you decide anything.

Energy and water

Two distinct costs.

Training a large model consumes a great deal of electricity, once, in concentrated fashion. Published figures for large models run to hundreds of megawatt-hours upward.

Inference — running the model, every time anyone asks anything — is small per query and enormous in aggregate. At present scale, inference has overtaken training as the dominant ongoing cost for widely-used systems.

Water is the one people miss. Datacentres use it for cooling, and unlike electricity it is drawn locally. A datacentre in a water-stressed region is a local issue regardless of the grid’s carbon intensity.

Two honest caveats. Per-query figures circulating online vary by orders of magnitude and many trace to poor sources — E7’s rules apply to this topic as much as any other. And efficiency per query has improved substantially, while total consumption has risen, because usage grew faster. Both are true, and quoting only one is how this subject usually gets argued.

The people in the pipeline

Less visible and, for many, the sharper issue.

Models are shaped by large amounts of human work: labelling data, ranking outputs, writing demonstration responses, and reviewing material to train safety filters.

That last category deserves naming. Making a model refuse to produce the worst things requires people to look at the worst things — violence, abuse, child sexual abuse material. This work is frequently outsourced to lower-income countries, sometimes at very low pay, with documented psychological harm and inadequate support. It has been the subject of investigative reporting and litigation.

The comfort of a well-behaved assistant rests on that work. It is worth knowing whose.

Neither panic nor dismissal

Two unhelpful responses.

“It’s catastrophic, don’t use it.” Not proportionate, and it ignores genuine value — including uses that reduce other consumption.

“It’s trivial, a video stream uses more.” Sometimes true per query, and it skips aggregate growth entirely, plus water and labour, which are not comparable to streaming at all.

The honest position: real costs, unevenly distributed, borne substantially by people who see none of the benefit. That does not mean stop. It means choose deliberately.

Proportionate use

What that looks like in practice:

Match the tool to the task. The largest model is not required to fix a sentence. Smaller and cheaper models handle most routine work, and choosing one is the single most effective thing an individual can do.

Do not loop pointlessly. Twenty regenerations because you have not said what you want is E6·L3’s problem and this one.

Prefer once to repeatedly. Save the result. Reuse the prompt.

Weigh it against the alternative. Sometimes the assistant replaces a flight, a print run, or three people’s afternoon. Sometimes it replaces thinking for ten seconds. These are not the same trade.

Small, undramatic, and it accumulates — the aggregate is exactly what the numbers say matters.

Try it now (5 minutes)

Look at your recent use. Find one task you routinely give a large model that a smaller one would handle.

Switch it. That is proportionality — not abstinence, just not defaulting to the biggest thing available for everything.

Check your understanding

1. Which is currently the dominant ongoing energy cost for widely-used models? Training Inference — running the model for every query Storing the training data

2. Why does water deserve separate mention from electricity? It costs more It is drawn locally, so a datacentre in a water-stressed region is a local issue It is used during training only

3. The safety filters that make an assistant well-behaved depend on: Automated detection alone People reviewing harmful material, often outsourced, with documented psychological harm Rules written by the model provider's lawyers

Recap

Training and especially inference consume significant energy, cooling water is a local resource, and safety filters rest on human review work that has caused documented harm. Reject both panic and dismissal: the costs are real and unevenly borne. Choose proportionately — match the model to the task, avoid pointless looping, and weigh each use against what it actually replaces.