The Occasional Jam No. 16

No. 16 soapbox

Tokens and T-Shirts

8 min read

I asked what my AI habit costs the planet, got an answer I liked, and then had to work out whether liking it was a reason to distrust it.

This site is mostly written by an AI, which I have already confessed to. This one is about what it costs, starting with a number that came off my own machine.

The receipt

Every session I run leaves a log on this laptop. On the morning of July 28 I added all of them up: every session, every model, every parallel copy of the project, from the first commit of the daily puzzle pipeline in early June through that morning. Fifty-one days. Six billion The unit the model reads and writes in. One token is roughly three quarters of an English word..

Six billion is a frightening number until you look at what it is made of. Ninety-five percent is the same context being read again: the project's own rules, the files under discussion, the conversation so far, re-sent on every turn because the model has no memory between turns. Strike the re-reads and about 330 million tokens remain, of which only about 24 million are words the AI actually wrote.

So before I can convert any of this into electricity, I have to decide which of those numbers is "my usage", and the first choice alone, counting the re-reads or not, moves the answer by a factor of eighteen.

Tokens in 51 days of building
6.05 billion
First commit of the pipeline to the morning I added it up.
Of that, context read again
95%
The same rules and files, re-sent on every turn.
Left after striking the re-reads
330 million
Tokens the AI wrote
24 million

Here is the whole calculation. Try both dials.

Tokens, 51 days of building
6,054,357,628
Of those, re-read from cache
5,726,641,546

Nobody publishes this number for a frontier model. Presets:

1,449 kWh about 580 kg of CO2
  • 276 cotton t-shirts
  • 1,610 miles driven
  • 15,093 hours of television
  • 1.61 months of one home's electricity
Two measured numbers, one assumption, and a verdict that swings by more than a hundredfold depending on the assumption. The presets convert Google's published Gemini figures, 0.10 Wh per prompt counting chips alone and 0.24 Wh counting the whole facility, on the assumption that a median chat prompt runs about a thousand tokens.

The token counts are measured. Everything after them is an assumption I picked, because no lab publishes watt-hours per token for a frontier model. The presets are the closest thing to solid ground: last year Google published a real measurement of a real product, and said a median text prompt to its assistant used 0.24 watt-hours, emitted 0.03 grams of carbon dioxide, and consumed about a quarter of a millilitre of water. Five drops. Nine seconds of television.

Whoever picks the boundary picks the verdict

The most useful thing in Google's report is that it published two numbers for the same prompts. Counting only the AI chips doing the thinking, a prompt costs 0.10 watt-hours. Counting the chips plus the machines sitting idle waiting for traffic, the ordinary computers around them, and the cooling for the building, the same prompt costs 0.24: two and a half times the answer. Google says the smaller number "substantially underestimates the real operational footprint", which I believe, and which also means every cheerful figure you have read might be the small one.

So there are three dials on my receipt: what counts as my usage, what a token costs, and how dirty the electricity is. Each one is a choice about where to draw a line. I set the last one at 0.4 kilograms of carbon per kilowatt-hour, roughly the US grid average, because that is the grid I am plausibly on. Google's own two figures imply 0.125, because Google buys cleaner power than I do, and swapping mine for theirs divides every t-shirt in this post by three. Keep the cost of a token inside Google's own two figures, swing the first two dials, and my seven weeks of building comes out anywhere from about six cotton t-shirts to about two hundred and eighty. That range is wide enough to hold whichever conclusion I walk in with.

The same fifty-one days, priced four ways. Nothing moved between these bars except where I chose to draw two lines.
The same fifty-one days, priced four ways. Nothing moved between these bars except where I chose to draw two lines.
ItemValue
Cache counted, whole facility276 t-shirts
Cache counted, chips only115 t-shirts
Cache excluded, whole facility15 t-shirts
Cache excluded, chips only6 t-shirts

The comparison that feels too good

Even the top of that range is small, and here is where the essay was supposed to end. One cotton t-shirt costs 2.1 kilograms of carbon dioxide over its life, which is the figure the receipt divides by, and a famous quantity of water, usually quoted at around 2,700 litres, though that figure is itself contested and I have seen credible people put it several times higher. Two hundred and eighty t-shirts is a medium-sized wardrobe. My robot cost less than a lot of people's clothes.

I noticed how good that felt, and the feeling is what made me keep writing. A conclusion that arrives already comforting deserves a second look, because comfort is a sign that I will not go looking for the error myself.

The paradox

What I just did has a name, and the name is an insult. Whataboutism: you raise a harm, I point at a bigger harm, the subject changes, nobody has to do anything. It is supposed to be a fallacy.

Except that under the moral theory most of us actually reach for, it is not a fallacy. Consequentialism judges an act by its outcomes, and if outcomes are what matter, their size is the whole question, so comparing my tokens to my t-shirts is the correct procedure. This is the part that bothers me: the move that sounds like a cheap debate trick is, on reasoning I would defend in daylight, arithmetic.

Now run that arithmetic on everybody. Each person's share of any large collective harm is negligible, because that is what "collective" means. If negligible means innocent, then several billion innocent people produce a catastrophe that nobody committed. The philosopher Derek Parfit's answer is that the error is treating imperceptible as zero. Honest moral arithmetic has to add up the effects too small to feel.

Parfit's thought experiment

A thousand torturers, each of whom turns a dial one notch too small for any victim to feel. Every torturer can say, truthfully, that their own contribution makes no detectable difference to anyone. Together they cause agony.

The argument that my share is negligible and the argument that the aggregate is catastrophic are the same calculation run with different denominators. I got to choose the denominator. I chose the one that flattered me, twice, once by counting tokens the way that suited me and once by comparing myself to clothing rather than to the person who did not use any of this at all.

A comparison that changes what you do is analysis. A comparison that changes only how you feel is anaesthesia.

If I have a fixed budget of effort or money for reducing harm, knowing that my wardrobe outweighs my robot tells me where to spend it. You can usually tell which kind of comparison you are holding by asking whether anything about tomorrow is different now that you know.

Who taught me to count like this

The arithmetic never asks why I am auditing my personal consumption, item by item, at all. The habit has an author. In 2004 BP, then in the middle of a two hundred million dollar rebrand toward "Beyond Petroleum", worked with the advertising agency Ogilvy and put out a personal carbon footprint calculator, inviting ordinary people to total up their commuting and their groceries and their holidays. Before that campaign, almost nobody searched for the phrase. The agency won a major industry award for it in 2007.

One of the largest oil companies in the world taught the public a unit of moral accounting, and the unit it chose was the individual.

I do not think this makes personal accounting fraudulent, and I notice the irony of a man who just published a slider saying so. But measuring myself by the item makes me the moral agent and turns the firm into scenery. It stays quiet about the decisions with leverage: where a data centre gets built, which grid it draws from, whether its water comes from a river with any to spare, and what any of them must disclose.

This post has the same asymmetry. I could count my tokens to the digit, because the log is on my laptop. I could not measure one litre of the water, because that number belongs to companies under no obligation to share it, and the best figure available is one company's, about its own product, checked by nobody else. The information a consequentialist needs is held by the parties it would judge. That is a choice about disclosure, and disclosure is business ethics.

The shopping trap

The deeper cost of the footprint frame is what it does to the question. It converts a political problem into a shopping problem, and shopping problems are answered by buying differently, which is the one response the seller can live with. Consumer ethics of this kind tends to collapse into a question about identity, "what kind of person am I", and arithmetic cannot answer that, which is why the arithmetic never satisfies.

The political philosopher Iris Marion Young argued that for harms produced by large structures, responsibility does not work by tracing a causal thread back to a culprit. It comes from participating in the structure at all, and it is forward-looking and shared: what you owe is a contribution to changing how the thing works. By that account, what I am responsible for is less my token count than what I say about disclosure, what I am willing to pay for, and what I am prepared to have counted.

This argument has an evil twin: "it is all structural, so my choices are irrelevant". That is the negligibility argument again in political clothes, and it ends in the same comfortable place. Taking part in the structure is a reason to act on it, and no excuse not to.

Where this leaves me

I am not going to stop. Saying so plainly seems better than performing an agony I do not feel. The number really is small. Someone who wants to be furious about AI and the environment should be furious about training runs, and about data centres sited in places that cannot spare the water, and not about a man in the margins of his week building a word game.

But the smallness licenses almost nothing. It does not survive averaging, because the harms here have a geography a global figure erases: a litre drawn in a wet place is not a litre drawn from a failing river, and my slider cannot tell them apart. If disclosure ever showed me which one I was doing, it would change my behaviour in a way two hundred and eighty hypothetical t-shirts never will. The missing number is the scandal, and it is nobody's personal failing.

I ran the numbers, found them small, and caught myself enjoying it. The enjoyment is the part worth reporting.

One last disclosure, since this post is built on them. This essay was drafted by the thing it is accounting for, against rules I wrote, and edited by me.

Sources

  1. Google Cloud: measuring the environmental impact of AI inference Both published figures, 0.10 watt-hours counting the chips and 0.24 counting the facility around them. The presets on the receipt are converted from these.
  2. World Resources Institute: the apparel industry's environmental impact in 6 graphics Where the t-shirt's carbon and water figures come from, contested as the water one is.
  3. Grist: Footprint fantasy The history of BP's 2004 calculator, the Ogilvy campaign around it, and the award in 2007.
  4. Derek Parfit, Reasons and Persons (1984) The harmless torturers, and the case against treating imperceptible as zero.
  5. Iris Marion Young, Responsibility for Justice (2011) Responsibility for structural harm as forward-looking and shared, rather than traced back to a culprit.
  6. My own session logs, 7 June to 28 July 2026 Every token count in this post. They are on this laptop and nowhere else, which is the asymmetry the post is about.