What does it actually cost the planet when you press send on an AI prompt — and where does that cost land?

Introduction

Most coverage of AI's environmental cost reports aggregate figures — terawatt-hours per year, billions of litres of water — that are too large to feel. This project starts at the unit of one prompt and traces the cost upstream into electricity, water, and atmosphere, then re-aggregates deliberately. The intended reader uses generative AI tools regularly, has read a headline about its environmental footprint, and wants to know whether the concern is justified.

One prompt

What follows is the per-prompt cost of running a frontier large language model — the GPT-class systems from OpenAI, Anthropic, Meta, and others. Pick a model, pick a prompt length, click Send: energy in watt-hours, water in millilitres, and carbon in grams of CO₂ tick up live, with everyday equivalences for each.

Models do not all cost the same — and the next section ranks them.

Comparing models

Same data, two cuts. The dashboard below ranks frontier models by per-query energy within each query-length tier, where length is measured in tokens (the small chunks of text a model processes — roughly a short word or punctuation each). At the long tier (1500-token prompts, the kind of reasoning task a chatbot handles many times per day), per-query energy varies 184× across deployed models: from Llama 4 Maverick at 0.27 Wh up to o3-pro at 49.93 Wh. The ranking is stable at the extremes; the middle of the pack reshuffles as queries lengthen.

The first dashboard ranks models within a tier; the next traces how each individual model scales from short to long prompts. Where the ranking answers which models are most expensive, the line traces answer which models are most prompt-sensitive — the two questions don't always have the same answer.

The shape of each line is the diagnostic: a flat trace means the model is roughly insensitive to prompt length; a steep trace means cost grows fast with context.

Where the cost lands

Hyperscalers — the cloud providers running these models — route inference traffic (the model-running work triggered each time someone presses send) dynamically across regions and rarely disclose which physical facility served any single request. The map shows where the major cloud regions are located, not where any one query went. Toggle between two layers: grid carbon intensity (grams of CO₂ per kilowatt-hour generated, the lower the cleaner) from Ember 2025, and Aqueduct's baseline water-stress score on its 0–5 scale. Hong Kong and Singapore lack water-stress scores under Aqueduct's hydrological-basin methodology.

Below is the same query traced end-to-end. The Sankey starts with one user prompt — a typical long-form coding question sent to GPT-4o — and follows it through the cellular network, the cloud data centre, the regional electricity grid, and finally to the carbon emissions released into the atmosphere. Each transition shows what the prompt picks up along the way.

The derived figure (~0.73 g) sits 3.8% below Jegham's published 0.755 g because the integration uses Virginia's actual 2025 grid intensity (327 gCO₂/kWh from Ember) where Jegham assumed a static 0.34.

One to a billion

The per-query cost is small. The aggregate cost depends on how many prompts get sent. The dashboard below takes one selected model — defaulted to GPT-4o, the model that has served most consumer ChatGPT traffic — and projects its energy, water, and carbon footprint to one billion, fifty billion, and one hundred billion daily prompts. Use the dropdown to swap in a different model and see how the multiplier changes.

The everyday equivalences — household-years of electricity, Olympic swimming pools of water, transatlantic flights of CO₂ — aren't the analytical claim. They are a way to make the magnitudes legible.

Training scale

The project's analytical focus is inference, but training compute is the necessary historical context. Training the largest models has scaled on a sustained log-linear trajectory in FLOP (floating-point operations, the unit of raw compute work). Only fifteen recent frontier-model releases publish verified training compute; the chart shows those.

Disclosure is uneven across the industry, and what cannot be measured cannot be charted.

Two stories from the same data

Identical data, opposite framing, opposite conclusions. The two charts that follow are built from the same per-query inference numbers; the framing decisions diverge completely. The first reads honestly — full sample, log axis, neutral palette, attribution at title-level contrast. The second weaponises every legitimate-looking technique a non-pedant reader is unlikely to catch: cohort cuts, axis truncation, alarmist colour, and a geographic frame that imports a worst-case region's stress context onto data from regions that aren't that region.

Now the same data, framed differently:

Six techniques the second chart uses

1. Sample restriction. The chart shows only the 31 of 62 long-tier models with water consumption above 25 mL — the heavy half. The 31 efficient models that would drag the average down are excluded by construction.
2. Axis truncation. The x-axis starts at 25 mL, not zero. Differences within the heavy tail get visually magnified, and every value looks big.
3. Geographic recontextualisation. Per-query water values come from globally distributed Jegham benchmarks. They are framed as if they apply at Qatar Central, the global worst-case water-stress region (Aqueduct score 5.0). The numbers are not altered — but the x-axis label imports Qatar's stress context onto data from regions that are not Qatar.
4. Aggregate framing. Per-query values are projected to one billion daily users multiplied by 365 days, producing tens of billions of litres per year, expressed in Olympic swimming pools.
5. Palette and typography. Deep red primary, bold heavy title weights, newsroom-graphic register. The visual establishes urgency before any number is processed.
6. Source de-emphasis. Full attribution and methodological caveats sit in the footer at low contrast, easy to skim past.

The lesson is methodological. Neither chart is "the truth" in any complete sense; both are readings of the same underlying data. The reader's defence is recognising the moves — knowing what a cohort cut, an axis truncation, or a swap of geographic context look like — well enough to spot them before the conclusion lands.

Findings

1. Per-query inference energy varies 184× across currently deployed frontier models.

   At the long tier (1500-token prompts), Llama 4 Maverick uses 0.27 Wh per query; o3-pro uses 49.93 Wh. Same prompt, two orders of magnitude in cost.

2. Aggregate consumption is dominated by the largest reasoning models.

   Provider routing decisions and user-side model choice both have real environmental leverage — one o3-pro long-tier prompt is roughly equivalent to 184 Llama 4 Maverick prompts at the same length.

3. Geographic distribution of the cost is highly uneven.

   Virginia's 2025 grid intensity is 327 gCO₂/kWh against Washington's 124 — a 2.6× ratio at identical workload. Arizona's water-stress score, 4.39 on Aqueduct's 0–5 scale, is the highest in the United States.

4. Per-query efficiency is improving fast; aggregate demand is growing faster.

   Both statements are true. The reconciliation is a Jevons-paradox argument the project's report develops further.

Methodology

Five public datasets feed the analysis. The Jegham per-query benchmark (How Hungry is AI?) gives per-query energy, water, and carbon for 66 frontier models across three query-length tiers. Epoch AI provides a 3,509-system database with training compute, parameters, and disclosed power draw. Ember publishes 2025 electricity grid carbon intensity at country and US-state granularity. The IM3 Open Source Data Center Atlas geocodes 1,479 US data-centre facilities. The WRI Aqueduct 4.0 publishes country and province-level baseline water stress on a 0–5 scale. All five sit under permissive licences (CC-BY 4.0 or ODbL).

Three known limitations bound the analytical claims. First, inference numbers are estimated, not measured: proprietary providers (OpenAI, Anthropic, Google) do not publish per-query energy directly, and Jegham's confidence intervals are wide. Second, water consumption is sometimes reported as on-site cooling water and sometimes as off-site thermoelectric water for grid generation; the headline figure can change by an order of magnitude depending on which the source uses. Third, geographic attribution is coarse — the map shows where cloud regions are, not where any specific query went.

We publish the full code, data preparation notebooks, and Tableau workbook sources at github.com/IIxoskeletonII/Thirsty-Machines.