--- title: The Environmental Cost of AI source: https://steadman.ai/newsletters/david/ai-environmental-cost.html published: 2026-07-25 summary: The full workings behind the essay "Burned and earned" — what a prompt, a task and a day of AI cost in energy, carbon and water, every number sourced and checkable. --- # The Environmental Cost of AI *The workings behind the essay, every number sourced.* This page accompanies the essay [Burned and earned](https://steadman.ai/newsletters/david/archive.html#email-2026-07-25) (25th July 2026). The essay makes the case: the cost of using AI is the footprint of the machine minus the footprint of whatever would otherwise have done the work. Here are the full workings behind every number in it, the model, the ranges and the sources, so you can follow the arithmetic or argue with it. ## What a run costs ### Generation 1: Chat, a prompt *A question, or "synthesise this document".* * Money: under a penny * Energy: 0.24 Wh (nine seconds of television) * Carbon and water: 0.03 g CO2e, 0.26 ml (five drops) * Human-energy floor: about 12 seconds The one measured anchor: [Google's published figure](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference) for a median text prompt, covering chips, host machines, idle capacity and cooling. Google also reports that this figure fell 33-fold in a single year. ### Generation 2: Agents, a task *Research a topic, build a model, make a deck.* * Money: about £1 * Energy: ~65 Wh (range 25–250) * Carbon and water: ~8 g CO2e, ~70 ml (a few sips) * Human-energy floor: about 50 minutes By energy, a task is a couple of hundred prompts. Not hundreds of questions: one long, growing conversation the model re-reads on every step. ### Generation 3: All-day AI, a day *Triage the inbox, watch the channels, run a few tasks.* * Money: a few pounds * Energy: ~0.25 kWh (the kettle, boiled twice) * Carbon and water: ~30 g CO2e, ~270 ml (a mug) * Human-energy floor: about 3.3 hours Range 0.1–2.5 kWh: a light day sits near the bottom, a heavy multi-agent power user above 1 kWh. Even the heavy day is a fraction of a tumble-dryer cycle. ## The workings Every number above, with its inputs. Sourced figures link out. Anything called an estimate is a working assumption, not a measured value, because no vendor yet publishes per-run energy for agents or day-long assistants. The point of the model is that you don't have to take the estimates on trust: they are built up from token counts you can change. ### 1. The one measured anchor A median Gemini Apps text prompt uses **0.24 Wh** of energy, **0.03 g** CO2e and **0.26 ml** of water, on [Google's published methodology](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference). That covers data-centre serving (chips, host machine, idle capacity, cooling) but excludes model training, networks and your own device. Everything else here is anchored to this one number. Google also reports the energy of that prompt fell 33-fold over the twelve months to May 2025. From it, two intensities the whole grid uses: **0.125 g CO2e per Wh** (0.03 ÷ 0.24) and **1.08 ml water per Wh** (0.26 ÷ 0.24). Applying Google's own ratios keeps every row on one energy mix. ### 2. The model A small, reusable Python model estimates tokens, cost and energy for each generation, built bottom-up from realistic activities. Its logic: * **One primitive:** an inference call with input and output tokens. Everything else is a composition of calls. * **The agent loop:** language models are stateless, so on every step of a task the whole running context is re-sent as input. Context grows by what the model writes plus any tool result it reads. This is why a task costs hundreds of prompts: it isn't hundreds of questions, it's one long, growing conversation the model re-reads on every turn. * **Calibration:** energy is set at 0.5 joules per input token and 2.0 per output token (decode is dearer than prefill), chosen so the 150-in / 400-out median prompt lands exactly on Google's 0.24 Wh. Cost uses representative mid-tier frontier pricing, $3 and $15 per million input and output tokens. * **Two simplifications, pulling opposite ways.** The model ignores prompt caching, so the agent-tier figures lean high. It also treats energy per token as constant however long the context, when the longest calls are dearer per token, which leans low. The ranges below absorb both. Running it on a typical run of each generation: | Type | Tokens (in / out) | Cost | Energy | CO2e | Water | Human-hours (floor) | |---|---|---|---|---|---|---| | Prompt | 150 / 400 | $0.006 | 0.24 Wh | 0.03 g | 0.26 ml | 12 s | | Task (research, code, deck) | 371,000 / 22,000 | $1.45 | 64 Wh | 8 g | 69 ml | 51 min | | Day (five tasks + housekeeping) | 1,426,000 / 94,000 | $5.69 | 250 Wh | 31 g | 271 ml | 3.3 h | A task lands at about 260 times a prompt, a day at about 1,000 times, which is the independent check on the "hundreds or thousands of times" figure: it falls out of the token counting, it isn't borrowed. ### 3. Ranges, not false precision The two larger figures are estimates, so the model was run across a spread of activity, not a single point: * **A task:** 25 to 250 Wh depending on how many searches, code runs and revisions it does; ~65 Wh typical. * **A day:** 0.1 to 2.5 kWh. A light day (one task plus housekeeping) is about 0.1 kWh; a typical five-task day about 0.25 kWh; a busy eleven-task day about 0.59 kWh; a heavy multi-agent power user north of 1 kWh. The top of the range is that power user running parallel loops, not one assistant ticking over. The orders of magnitude hold. The decimals don't. ### 4. Human-hours: the floor A person at a laptop draws about 75 watts. Dividing each run's energy by 75 W gives the human desk-time that draws the same power: a prompt is about 12 seconds, a task about 51 minutes, a day about 3.3 hours. Lighting and heating in the room would raise the human figure and shorten the machine's break-even, so 75 W is the cautious choice: it makes the human look cheaper than they are. ### 5. Human-work: the ceiling The floor is not what the run replaces. [METR](https://metr.org/time-horizons/) measures the length of task, timed on expert humans, that an agent finishes at 50% reliability; that horizon has doubled roughly every seven months for six years, from nine seconds for a 2020 model to about 14.5 hours for Claude Opus 4.6 in early 2026. By METR's May 2026 update the frontier reading (Claude Mythos, in preview) had reached roughly 16 hours, which is the ceiling of what METR's current task suite can reliably measure: the models have outgrown the ruler. So a task drawing 51 minutes of human-equivalent energy can stand in for hours of human work. That gap, an order of magnitude, is where the environmental case lives. Caveats: the 50% reliability point is not the same as dependable delegation, the tasks are mostly software work, and the horizon is a frontier ceiling, not a typical run. ### 6. The physical comparisons * **Car:** a UK petrol car emits about 165 g CO2 per km, a diesel about 170, fleet average nearer 130, on the [DESNZ conversion factors, 2024](https://www.gov.uk/government/publications/greenhouse-gas-reporting-conversion-factors-2024). A 50-mile (80 km) round trip in a petrol car is about 13 kg. * **Van:** a diesel van is heavier per mile than a car, so a dedicated cross-city run is a few kilograms of CO2. Shared delivery rounds spread that across many parcels, so a few kilograms is the honest ceiling for a dedicated trip, not a precise figure. * **Flight:** a nonstop transatlantic round trip emits [about one tonne](https://theicct.org/wp-content/uploads/2021/06/ICCT_transatlantic-airline-ranking-2014_0.pdf) of CO2 per passenger on average, with some routes higher (ICCT, 2014 figure, CO2 only). Counting the extra warming from burning fuel at altitude would push the effective climate cost higher still, so the comparison is deliberately conservative. ### 7. The aggregate figures Data-centre electricity is projected to [roughly double](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary), from about 485 TWh in 2025 to about 950 TWh in 2030, with AI-focused consumption tripling, and the fall in energy per task described as unprecedented. [Our World in Data's July 2026 analysis](https://ourworldindata.org/how-much-energy-do-data-centers-and-artificial-intelligence-use) independently lands on the same starting point and supplies the denominator: data centres used about 485 TWh in 2025, around 1.5% of the world's electricity, with AI-focused data centres at about 155 TWh, or 0.5%. The global average hides sharp concentration: data centres draw about 5% of US electricity, more than a quarter of Virginia's and more than a fifth of Ireland's. And the carbon a run produces depends on the grid behind the data centre: the same watt-hour is several times dirtier on a coal-heavy grid than on a nuclear- or renewables-heavy one. The grid here applies Google's own reported mix throughout (0.125 g CO2e per Wh); a different operator on a different grid would shift the carbon column, in either direction, without changing the energy figures or the argument. The capital markets have stopped treating the growth as speculative: National Grid, the British utility, [is investing $1.75bn](https://www.ft.com/content/39f4842b-7389-485c-9d28-1a199607e56b) in a developer supplying power to US data centres. Both the collapse in energy per task and the doubling in total are true at once: that is the Jevons paradox, and the reason the last column of the grid, not the first, is where the essay ends. ## Four questions to ask in your organisation 1. **Did it replace real work?** A drive avoided, a contractor not briefed, an afternoon returned. If yes, the subtraction almost always lands in AI's favour, by an order of magnitude or more. 2. **If it replaced nothing, was it worth commissioning?** Price it in human-hours, the currency you already use for contractors and agencies. A pointless task is 50 minutes of human-equivalent energy spent on something nobody needed. A useful one nobody would otherwise have done can be worth every watt. 3. **Is the run matched to the job?** The one way to lose on real work is to over-power it: a task-sized agent on five minutes of work, or an all-day system watching for a need that never comes. That is waste in a clever hat. Match the run to the job and the sum falls the right way. 4. **What grid is your provider on?** The energy of a run is fixed by the work; the carbon is not. Providers publishing their energy mix, and siting on cleaner grids, differ several-fold on the same workload. It is a fair procurement question, and increasingly answerable. ## Where every number comes from * [Google's published methodology](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference) for a median text prompt (0.24 Wh, 0.03 g CO2e, 0.26 ml water): the one comprehensive measured anchor. Task and day figures are modelled up from it by counting the tokens a realistic run gets through; the full workings, assumptions and ranges are set out above. * [METR time horizons](https://metr.org/time-horizons/): the length of expert job a model finishes at 50% reliability, to May 2026. * [IEA, Energy and AI](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary): the 485 to 950 TWh projection. * [Our World in Data](https://ourworldindata.org/how-much-energy-do-data-centers-and-artificial-intelligence-use) (Hannah Ritchie, July 2026): shares of global and national electricity, and the grid-mix point. * [UK DESNZ conversion factors](https://www.gov.uk/government/publications/greenhouse-gas-reporting-conversion-factors-2024) and [ICCT](https://theicct.org/wp-content/uploads/2021/06/ICCT_transatlantic-airline-ranking-2014_0.pdf): the car and flight comparisons. This page accompanies the essay [Burned and earned](https://steadman.ai/newsletters/david/archive.html#email-2026-07-25) (25th July 2026). The essay makes the case; this page shows the workings. --- This is an artefact created for David's weekly email. [See the others here](https://steadman.ai/newsletters/david/).