I have written repeatedly on this blog about lean websites: every kilobyte counts, skip the custom fonts, do not generate images you do not need. This site practices it - system fonts, almost no JavaScript, no dynamically generated social images.
I also run AI models every single day. To write code, to research, to draft articles. Data centers hum somewhere so that I can avoid typing a function.
Both of those are true at the same time, and I have been putting off writing about the tension. So let us do the uncomfortable thing and look at the actual numbers - because the honest position is neither “AI is boiling the oceans” nor “it is basically nothing, relax.”
What a query actually costs
Start with the number people argue about most: energy per prompt.
The figure that circulated for years - about 3 Wh per query - turns out to be roughly ten times too high, according to analysis by Epoch AI. OpenAI’s own stated figure for an average query is around 0.34 Wh, and independent estimates land somewhere in the 0.3 to 3 Wh range depending on model and length. That puts a typical AI query at roughly three to ten times a web search.
Let that land properly. It is real, and it is not apocalyptic. A single prompt costs about as much energy as running an LED bulb for a couple of minutes. If your mental model was “each question I ask melts a glacier”, that model was wrong - and being wrong in the alarmed direction is not more virtuous than being wrong in the dismissive one.
Two important asterisks, because honesty cuts both ways: those figures are marginal estimates that exclude training, and they come largely from the providers themselves.
Now the number that should worry you
Per query, small. In aggregate, enormous.
Global data center electricity consumption is heading for roughly 565 TWh in 2026 - a 26% jump over the previous year. AI-optimized servers alone account for about 175 TWh of that and are on track to draw more power than conventional servers by 2027. The IEA projects data centers reaching roughly 950 TWh by 2030 - comparable to the entire electricity consumption of Japan today. Water use at hyperscalers rose 25-40% year over year, concentrated in specific regions where that matters a great deal.
So which story is true? Both - and this is the part almost every article gets wrong:
“AI uses more energy than ever” is true on absolute volume and false on per-query efficiency.
Google reduced the energy footprint of a median Gemini prompt by a factor of 33 - and its carbon footprint by 44 - within twelve months. Each query got dramatically cheaper. We just made vastly more of them. That is Jevons paradox playing out in real time: efficiency gains get eaten by volume growth.
One more shift worth knowing, because it changes who is responsible: inference now dominates. Roughly 63% of frontier-model lifecycle energy is inference, about 37% is training. The energy story is no longer mostly about giant training runs by labs. It is about everyday usage - by people like me and you.
Where AI genuinely helps sustainability
I refuse to write this as pure penance, because the ledger has a credit side, and it is not small:
- It builds leaner software. The optimizations that keep sites small - dead code removal, image pipelines, bundle analysis, replacing a dependency with a native API - are tedious, mechanical work that AI does willingly. Every kilobyte it helps remove is served millions of times.
- It finds waste at scale. Unused features, idle infrastructure, N+1 queries, expensive loops. This is exactly the kind of pattern-matching where models are strong and humans are bored.
- It automates the efficiency work nobody schedules. The pruning jobs, the audits, the “is this still needed?” review - the maintenance I wrote about, done consistently rather than never.
An hour of inference that removes 200 KB from a page served a million times a month is not an environmental cost. It is an environmental investment with an absurd return.
Where it is just indulgence
And the debit side, honestly:
- Asking a model what a
grepwould answer. Semantic search over your own codebase to find a string you could have found in 200 ms of local CPU. - Regenerating instead of thinking. Firing off five variations because it is easier than deciding what you actually want. Fast, cheap-feeling, and pure waste.
- Vanity runs. Rewriting prose eleven times to move a comma. Generating images because a slide “needs something”.
- Bad session hygiene. Sprawling context, repeated full-codebase reads, loops that redo the same work - the same patterns that show up as cost spikes when you measure per-PR spend.
Notice that every item on that list is also a productivity problem. The wasteful uses are rarely the valuable ones. That is convenient, and it is the basis of my whole framework.
My framework: efficiency per outcome
I do not track my personal watt-hours, and I am suspicious of people who claim they do. Instead I use two rules that are simple enough to actually follow:
1. Efficiency per outcome, not per action. The question is never “did I use AI?” but “did this use of AI produce something worth its cost?” A model call that prevents a wasted week of work is an efficiency win. Five regenerations to avoid making a decision are not - and they are not a good use of my time either.
2. The million-times rule. What matters most is the artifact that gets served, not the tool used once to build it. Every visitor to a page pays for its weight, forever. The model call that helped me build it was paid once. So I optimize hardest where the multiplier is: the delivered product stays radically lean, even if building it consumed real energy.
That is why this blog looks the way it does - no dynamically generated social images (an image per post, generated and served, for a preview thumbnail), system fonts instead of downloaded ones, almost no client-side JavaScript. The thing served a million times is as light as I can make it. The tool used once to write it was heavier. I think that trade is correct, and I am prepared to defend it.
The household ledger
Think about how a sensible household handles money. Not with a purity test where every purchase is either virtuous or shameful - that is exhausting and nobody sustains it. Instead: know roughly what things cost, spend deliberately on what earns its keep, cut the recurring drains that give nothing back.
The family that agonizes over a single coffee while a broken freezer runs all night in the garage has the wrong model. So does the family that never looks at the statement at all.
That is where I have landed on AI and sustainability. My prompts are the coffee: real, worth knowing about, not the thing that decides the budget. The bloated page served a million times is the freezer running in the garage - the recurring, invisible drain that actually adds up. So I keep an eye on both, spend where it multiplies, and stop pretending I am either innocent or doomed.
Look at your own statement honestly. That is more useful than any guilt - and it is the only version of this conversation that survives contact with how we actually work.