Impact Zone mark

Before you dive in

Leave your name and email and we'll let you straight through — no spam, just a note if we add new tools like this one.

Continue →
Impact Zone mark impactzone.ai
WATER · CARBON · AI

What AI Actually Costs

Published water and carbon numbers for the same AI prompt disagree by up to 2,000×. Almost none of that gap is anyone lying — it's where you draw the line around the system. This walkthrough shows you the lines, so you can read the next headline yourself.

1

One AI request is genuinely tiny — a fraction of a gram of CO₂, a few drops of water.

2

But total AI use is growing faster than efficiency improves, so the total keeps climbing anyway.

3

Where AI runs matters more than how much — the same workload can cost 1,000× more depending on local water stress.

4

"AI is terrifying" and "AI is nothing" headlines usually use the same trick: a favourable way to draw the line, presented as a fact.

THE BIG PICTURE

Improved Efficiency vs. Increased Usage

Here's the paradox: each AI request is getting cheaper to run. But the world is making so many more requests that the total keeps climbing anyway.

The good news: AI got more efficient

33× less energy
per request, in one year, for one major AI product
Google, self-reported T2
One company reporting on its own product, in an unusually good year — not a guarantee every AI product matches this.

The catch: total usage is still climbing

485 → ~945 TWh
≈ 45 million → 88 million homes' worth of electricity a year
All data centres worldwide, 2025 → 2030 (projected) T1
Source: International Energy Agency (IEA). This is every data centre on Earth — streaming, cloud storage, corporate computing, and AI, all added together.
Important distinction: "data centres" is not the same thing as "AI." AI is only a slice of that data-centre number — roughly 5–15% today, projected to grow to somewhere between a third and a half by 2030 T1. Most of a data centre's electricity today goes to things that have nothing to do with AI: video streaming, cloud storage, email, corporate computing.
Source: a report prepared for the IEA — not an IEA-authored figure itself, though the IEA commissioned and published it.
Why doesn't efficiency alone fix this? Because when something gets cheaper to use, people tend to use more of it, not less — researchers call this a rebound effect (Luccioni, Strubell & Crawford, 2025). And even the experts don't fully agree on the total: the IEA says 485 TWh (2025); another respected research body, UNU-INWEH, says 448 TWh — 8% apart, on a number that in principle should just be a matter of counting.
FOUR INTERACTIVE TOOLS

Explore the Tools

Each tool lets you move a real dial and watch a real number change, using nothing but the same underlying, sourced facts. Pick one, or go through them in order with Next.

Boundary Machine

What do we include when we measure?

Watch one AI request's cost swing by up to 2,000× as you change what counts — same facts, different lens.

Where Is the Water From?

Water basin scarcity matters

The same amount of water means something very different drawn from a drought-stricken basin than a rainy one.

Ways to Count a Human

People still eat when AI does the work

The viral "1,000× cleaner than a human" claim depends on one accounting choice — watch it collapse to zero.

Levers Ranked

Which changes actually have impact

Your own prompt count barely moves the needle. See — ranked by size — what actually does.

TOOL 1 OF 4

The Boundary Machine

What do we include when we measure?

There's no single "right answer" for what one AI prompt costs — it depends entirely on what you choose to count. Move the three dials below and watch a real number swing by up to 2,000×, using nothing but the same underlying facts.

📅 The core figures on this page are from Google's own May 2025 data — the most recent methodology of its kind publicly available. Given AI's documented year-over-year efficiency trend, treat them as likely conservative (i.e. today's real numbers are probably somewhat lower), not as current-day measurements.
Choose your settings, then check the impact at the bottom of the page.
1 · What are we measuring?
Same technology, three different-sized questions — one message you send, one person's whole year of use, or one single AI product (like Gemini or ChatGPT) totaled across everyone who uses that one product.
2 · Energy: how much of the process counts?
Answering your prompt takes more than just the chip that does the "thinking" — like a car needs more than just its engine. Widen the ring to include the rest of what's really involved. Click a greyed-out ring to see why it's locked.
3 · Water: how much of the process counts?
Cooling the computer is only step one. Making the electricity that computer runs on takes water too, at the power plant — most companies don't count that part. Meta has since disclosed information suggesting this off-site share may be even larger than the working estimate this tool uses — worth treating these off-site figures as a floor, not a ceiling.
4 · How do we count clean-energy purchases?
Market-based counts the clean-energy contracts a company has bought, wherever in the world that clean power was generated. Location-based counts what actually came out of the power plug at that specific location. Same electricity — two honest ways to report it.
Numbers below are translated into everyday comparisons — a petrol car's tailpipe emissions, an irrigated almond's water footprint, and Google's own "running a TV" comparison — so you can feel the size, not just read the unit.
Energy T2
0.24 Wh
Carbon T2
0.03 gCO₂e
Water T2
0.26 mL
Primary source · T1

Li, Yang, Islam & Ren, "Making AI Less Thirsty," arXiv 2023. GPT-3 in Microsoft US data centres: ~10–50 mL per response.

Newspaper calc · T3 for the number

Washington Post × UC Riverside, 2024: 519 mL for one 100-word GPT-4 email, one region, one season.

Same lab, on-site only

~2.2 mL per request for an average US data centre — the like-for-like comparator.

Headline as circulated

"A bottle of water per email." No boundary stated.

What it got compared against

Google's 0.26 mL, on-site only — presented as a ~2,000× contradiction. Mismatched boundaries.

Verdict, both directions

519 mL is a defensible calculation, not a typical one. Google's own comparator (vs "45–50 mL") was also boundary-mismatched. Both sides compared unfavourably in their own favour.

Primary source · T1

Tomlinson, Black, Patterson & Torrance, Scientific Reports 2024. Peer-reviewed, real.

The allocation mechanism

Attributional (full life share) vs. marginal/consequential (what actually changes). Same study supports both readings.

Functional-equivalence check

Applies to writing & illustrating — exactly two of the tasks this briefing rates "invalid as stated."

Headline as circulated

"AI is 1,000× cleaner than a human doing the same task." Allocation choice dropped.

Verdict

Legitimate under attributional allocation. Collapses to ≈0 under marginal allocation for the common case: net-new work nobody would have paid a human to do.

TOOL 2 OF 4

Where Is the Water From?

Water basin scarcity matters

A litre of water evaporated in a rainy region and a litre evaporated in a drought-stricken one are not the same event. Moving the same AI workload to a different location can swing its real-world water impact by up to 1,000× — with the volume of water used staying exactly the same.

How stressed is the local water supply? T1
1.0
This is a scarcity score researchers call AWARE. 0.1 = a water-rich region · 1.0 = the world average · 100 = an already-dangerously-overdrawn basin
Important nuance: that "1.0 = world average" is an area-weighted average — every patch of the planet counted equally, including vast, barely-used basins. Where water actually gets consumed (a consumption-weighted average) runs considerably higher, because human and industrial activity clusters disproportionately in already-stressed regions. Don't read 1.0 as "what a typical user experiences" — most real-world water use already happens somewhat above it.
≥ 1.0 — at or above the world average, water-cooled infrastructure deserves real scrutiny.
≥ 10 — ten times more scarce than the world average; this app's own halfway marker toward the AWARE cap, not an official published threshold. Risk escalates sharply from here toward 100.
AWARE is calculated per river basin, per month — not per country — so there's no single official "UK score" or "Kenya score" to quote precisely. That's not a gap in this app; it's the whole point of this tool. What follows is general, well-documented water-stress knowledge for orientation, not an official published AWARE decimal.
🇬🇧
UK

Close to the world average nationally — but South East England, which supplies London, runs noticeably more stressed than the rest of the country.

🇯🇵
Japan

Generally water-abundant thanks to heavy monsoon rainfall, though dense industrial regions draw down local basins more than the national picture suggests.

🇰🇪
Kenya

Split in two: the wetter highlands sit near the world average, while the arid north and east rank among the most water-stressed regions anywhere.

🇺🇸
Texas

Swings from comparatively wet in the east to severely over-drawn in the west — the Permian Basin and Rio Grande valley are among the most stressed basins in the US.

🇦🇺
Australia

The world's driest inhabited continent overall. The Murray–Darling Basin, its agricultural heartland, is chronically over-allocated, while the wetter coasts sit far lower.

How much is being used — prompts per day
100
Actual water used, one year (from the 0.26 mL/prompt baseline)
—
Real-world impact, once you factor in how scarce that water already is T1 method
—
"Litre-equivalents" means: if this same water had been drawn from a basin at the world average scarcity score (1.0) instead, it would carry the same real-world impact as this many litres. That's what multiplying by the scarcity score actually does — it re-expresses the same water in terms everyone already understands.
The top bar never moves when you drag the scarcity slider — the amount of water used is identical. The bottom bar is what actually matters, and it can swing 1,000× wider on siting alone. No AI company publishes exactly where each request was served, so nobody outside these companies can calculate this bottom number for a real product today — this is a demonstration of the lever, not a measurement of any specific AI service.
Cooling trade-off
Water burden
Energy / carbon burden
Illustrative, not measured — no multiplier for this trade-off is sourced in the briefing. Air cooling saves water and costs energy; evaporative cooling saves energy and costs water. There is no free configuration.
T1 Ireland: data centres are now 23% of national metered electricity (2025), up from 5% in 2015 — small globally, decisive locally. (CSO Ireland, 7 Jul 2026)
UNVERIFIED "~2/3 of data centres built since 2022 are sited in water-stressed regions" (ELI fact sheet, Jan 2026) — chain of custody incomplete. Flagged, never used as a number in this app.
TOOL 3 OF 4

Ways to Count a Human

People still eat when AI does the work

You've probably seen the claim "AI is 1,000× cleaner than a human doing the same task." It comes from real, peer-reviewed research — but the honest answer flips to roughly zero depending on a single accounting choice. See both answers for yourself below.

How much AI is being used — prompts per day
100
This is the footprint of one person using AI heavily, every day, for a full year.
Electricity, one year T2
8.8 kWh
Carbon, one year T2
1.1 kgCO₂e
Water, one year T2
9.5 L
Which accounting method for the comparison below?
Market-based and location-based are the same two honest choices from the Boundary Machine. Independent full audit uses a competitor's (Mistral AI's) own third-party-audited numbers, which count more of the full picture.
≈ equivalent to driving
1.5 km
≈ days of household water use
0.06 days
Watch the driving distance change with no change to how much AI was used — just which honest accounting method you picked. That's the same lesson as the Boundary Machine, wearing different clothes. (The water figure draws on a separate set of studies than the driving figure, so its source names differ — read the small caption under each number for its real source.)
Pick a task to compare against a human doing it
How valid is it to compare a human doing the task you selected (above) to AI doing it?
TOOL 4 OF 4

Levers Ranked

Which changes have impact
Choose a role to see how people in those roles can have the most influence on the impact of AI
⚠ These are rough scales, not precise measurements — think "roughly 10× bigger," not "exactly 13.2%."
The data below is not interactive, but shows where AI impact can be shifted.
LeverMagnitudeTierNote
VALUE LENS · TOOL 1 OF 4

What's the Real Alternative?

Before this tool tells you what AI was worth, it needs to know what would have happened without it. That answer changes everything downstream — and for most real AI use, the honest answer is "nothing would have happened," which is exactly the case this tool refuses to guess a number for.

Pick a task category, then answer the one question below. Your answer is what gates everything else in the Value lens — and you can always come back and change it.
Task category
What would have happened if you hadn't used AI for this?
✓ A real comparison applies here
A genuinely displaced purchased service. This is the cleanest case for a value calculation — there's a real dollar figure (what you would have paid) to compare AI's cost against.
Next: "The Full Bill" will ask you to choose between attributional and marginal allocation, the same honest fork the footprint side already uses for human comparisons — full substitution value looks very different depending on which you pick.
✓ A real comparison applies here
Headcount that scales with task volume is one of the conditions the source briefing treats as a legitimate substitution claim — the counterfactual is real because staffing would genuinely have moved.
Next: "The Full Bill" will use your organization's fully-loaded cost per hire (wage × the FTE multiplier for your region), not just a bare wage.
↷ Different question, still a real one
This isn't "AI instead of hiring someone" — it's "AI saved my own time." That's a legitimate value driver, but it doesn't price against a wage table; it prices against what your own hour is worth, which has no single right answer.
Next: head to "What's Your Time Worth?" instead of "The Full Bill" — it walks through the three defensible ways to value a solo operator's hour.
✗ No human baseline exists
Net-new work — nobody was ever going to pay a person to do this. The source briefing's own finding is that most real AI use falls exactly here. A comparison to a human is not conservative or generous in this case; it's meaningless, because that human was never going to do it.
Per how this tool is scoped: this shows as "value estimate not available yet," not a guessed number. The valuation question for genuinely net-new AI output — what it's worth with no human price to compare against — is real and still open; this tool won't pretend otherwise to give you a tidier answer.

Where this gate comes from

Options A and B are two of the three legitimate conditions quoted directly from the source briefing's §6, Step 5 — the same test that governs when a footprint-side "AI vs. a human" comparison is meaningful. This tool reuses it rather than inventing a second version, so a comparison that's disqualified on one side of the app is disqualified on both. T1
The briefing's third condition — a displaced physical activity or trip — is left out of the buttons above on purpose. It's real (the briefing's own example is a journey made solely to perform a task now done remotely), but it's the one the briefing itself calls "rare," and none of this taxonomy's twelve task categories genuinely involve a trip AI is standing in for. Rather than offer a choice that would almost never be the honest answer for any real task here, it's omitted from this gate — noted here so the source's full test stays visible, not silently dropped.
The net-new finding (D) — "most AI use is net-new work with no human baseline" — is the briefing's own central finding in that section, not an assumption made for this tool.
The reroute case (D) follows the research plan's own distinction between substitution value and the value of a user's own time (Part 1, item 8) — treating "I'd have done it myself" as the same question as "hire someone" would misprice it, since a solo founder's hour isn't a market wage.
VALUE LENS · TOOL 2 OF 4

The Full Bill

Where "What's the Real Alternative?" said a comparison is meaningful, this is the actual math — the AI-assisted path's real cost against what the displaced human alternative would genuinely have cost, shown as visible arithmetic rather than one polished number.

What's sourced and what's yours, made explicit throughout: the FTE multiplier is real, cited data (StatCan / BLS national accounts). Your wage rate, hours, and AI tool cost are your own inputs — this tool has no standing wage table to silently plug in, and won't pretend otherwise.
Your task
Coming from "What's the Real Alternative?" with A or B selected — a real hire or contract, or headcount that scales with volume.
Task category
Region drives real multiplier
The human alternative
What the displaced hire or contract would genuinely have looked like.
Hours it would take a human your input
Their hourly rate, base wage your input Not auto-filled — plug in your own wage-benchmark lookup for this task category and region.
The AI-assisted path
What it actually took you, including your own review time.
Your review/editing hours your input
Your own hourly value your input
AI tool cost, allocated to this task your input Current pricing changes fast — this isn't pre-filled from a price list that could already be stale.
Domain risk tier shapes a modeled estimate

Show your working

Where this is sourced, and where it's yours

The FTE multiplier (1.4× US, 1.15× Canada) is real, cited data — US Bureau of Labor Statistics Employer Costs for Employee Compensation (March 2026) and Statistics Canada's national accounts (Table 36-10-0103-01). T1 The US/Canada gap is structural, not a rounding difference: Canada's publicly-funded healthcare removes the largest single driver of the US figure.
Review-time benchmarks exist for exactly two categories — code review (roughly 200–400 lines per session, under 500 lines/hour) and editing (copyediting 3–15 pages/hour, line editing 4–15, developmental 2.5–15, proofreading 6–17.5). Every other category shows "no professional body has published a quantitative standard for this" rather than a fabricated number, because that's genuinely the state of the evidence.
The error-cost line is a modeled estimate, stated as one — not a citation. It scales with the domain-risk tier you picked, per the research finding that hallucination-rate studies measuring "grounded and checkable" tasks (single digits to teens) and "open-domain recall with nothing to check against" (58–95%+) are measuring genuinely different things, not disagreeing about one true rate.
"Capacity freed up" vs. "Full substitution value" is the same attributional-vs-marginal fork the footprint side already uses for human comparisons, applied to dollars instead of emissions. Full substitution value assumes the fully-loaded human cost was genuinely avoidable — true if this was a real hire/contract decision, an overstatement if that person or budget existed regardless. Capacity freed up is the more conservative default, matching the augmentation-first positioning this build settled on.
VALUE LENS · TOOL 3 OF 4

What's Your Time Worth?

This is where "I would have done it myself, just slower" from the counterfactual gate lands. There's no wage table for your own hour — but there are three genuinely different, defensible ways to price it, and which one fits depends on your actual situation, not a formula.

Same lesson as everywhere else in this app: pick a different valuation method below with the same hours and watch the number move — not because anything about your work changed, but because "what is my hour worth" doesn't have one right answer.
Choose how to value your hour
Three methods from the UN's own methodology for unpaid/personal work — reused here, not built for this purpose, and the research plan says so plainly.
Your normal hourly rate your input
Hours this task took you your input
Was this during paid working hours, or personal/off time? real multiplier

Show your working

Where this comes from

The three methods — opportunity cost, replacement-specialist, replacement-generalist — are named directly in a UN methodology for valuing unpaid and personal work, and map almost exactly onto the "solo founder's hour" question this tool needed an answer for. They aren't invented for this app. T1
The business/personal time split (100% vs. 50% of rate) comes from US DOT federal travel-time valuation guidance — adapted from transportation economics, not purpose-built for AI value calculations, which is exactly why it's labeled as reused rather than presented as bespoke research. T1
VALUE LENS · TOOL 4 OF 4

Cost Meets Value

The same boundary and accounting choices you made on the Boundary Machine, carried forward — not a separate, flattened "Google measured it" number pretending those choices don't matter here too.

The Boundary Machine's whole point was that the same prompt can swing several-fold in water and carbon, depending on what you choose to count. This tool doesn't get to quietly ignore that and hand you one tidy number — it reads whatever you last set on the Boundary Machine.
Your water & carbon assumptions
Read live from the Boundary Machine — change them there and this updates automatically.
Per prompt, with your current Boundary Machine choices:
—
—
—
Put a dollar value on that water
Water price swings enormously by region — same idea as the AWARE scarcity score, now in dollars.
Region
—
Put a dollar value on that carbon
Carbon price varies almost 100× depending on which honest price you're using — social cost, market price, or offset market.
Price basis
—
What did AI create, for this prompt?
Pick a figure you've worked out on the previous two tools, or enter your own.
The three numbers, paired up
All three at the same per-prompt granularity, all from the boundary you set on the Boundary Machine.
$ of water used
—
$ of carbon released
—
$ value created
—
Explore subscription cost per prompt
A different way to sanity-check "$ value created" — what does a typical subscription actually cost, per message?
—

Water pricing is aggregator-cited, not primary government data — treat it as illustrative, not something to quote commercially without checking your own local rate. US average and California figures: consumer water-cost surveys, 2025. Toronto: City of Toronto industrial rate, 2025. Denmark: DANVA "Water in Figures" 2025 report, converted from EUR.

Carbon pricing is on firmer ground: the social cost of carbon is the US EPA's own December 2023 report ($204/tonne, 2023 dollars). EU ETS market price is live trading data, mid-2026 (~€83/tonne). Voluntary offset prices come from 2026 carbon-market trackers — nature-based and direct-air-capture credits are genuinely different products, not two estimates of the same thing.

THE CLOSING SUMMARY

Facts Summarized

Everything in this app answers one of two questions: what did that prompt actually cost, and what did it actually create? Both answers turn out to hinge on the same kind of decision — which honest line you draw — not on some AI product being secretly better or worse than another.

What it costs

  1. One AI request is tiny — but exactly how tiny depends entirely on which boundary you're quoting. Carbon runs from 0.03 g (Google, market-based, full serving stack, as published) up to roughly 1.14 g (a full-lifecycle boundary that includes embodied hardware and data-centre construction, not just serving energy). Water runs from 0.26 mL (Google, on-site cooling only, as published) up to roughly 45 mL (on-site + off-site, one specific study's region- and season-specific estimate — Li, Yang, Islam & Ren, T1). Neither range is one number with fuzzy edges — each endpoint is a different, real boundary, and blending them without saying so is exactly the mistake this tool exists to catch.
  2. For any one person, even heavy daily AI use adds up to about a kilogram of CO₂ a year — roughly the same as driving a few kilometres. Your own prompt count is not a meaningful climate decision.
  3. Each AI request keeps getting cheaper to run — but so many more requests are happening that total usage is still on track to roughly double by 2030. Efficiency is winning the battle and losing the war.
  4. Water is the harder problem, because it's local. The exact same AI workload can differ a thousandfold in real-world impact depending only on which water source served it — and no company publishes that location data, so nobody outside them can check. AI infrastructure like data centres should not be built in water-stressed regions if they use evaporative cooling, since that trades energy savings for water burden in exactly the places that can least afford it.
  5. Both viral claims — "a bottle of water per email" and "a thousand times cleaner than a human" — use the identical trick: pick a favourable way to draw the line, then present the result as if it were simply a fact.

What it's worth

  1. Most real AI use is net-new work — nobody was ever going to pay a human to do it. That's why so many "AI saved us $X" claims don't hold up: there's no human price to compare against, so a defensible tool says "value estimate not available" instead of guessing one.
  2. Even when a real hire or contract genuinely was avoided, the dollar figure still depends on a choice: "capacity freed up" (time redeployed, not cash in the bank) versus "full substitution value" (the human cost really disappeared). Same arithmetic, two honest readings — the same fork the footprint side already uses for market-vs-location carbon accounting.
  3. There's no wage table for your own hour. Three defensible methods — opportunity cost, specialist replacement rate, generalist replacement rate — can price the identical task at very different dollar amounts, with no single right answer among them.
  4. Water and carbon can now be priced in the same dollars as the value AI created, using your own boundary choices from the Boundary Machine — not someone else's flattened assumptions. And the price you put on that water or carbon swings just as widely as the footprint numbers do: water rates run roughly 6× apart across regions, carbon prices span nearly 100× depending on whether you use the social cost of carbon, a market price, or a voluntary offset.
One more thing worth remembering: whenever you see a headline about "data-centre electricity," that number includes Netflix, Zoom, cloud storage, email, and every corporate server on Earth — not just AI. AI's own share of it today is still only about 5–15%.

How confident should you be in each number?

T1 Peer-reviewed research, a national lab, or an international agency
T2 A company's own disclosure, with its method shown publicly
T3 Mentioned for context only — never the source of a number used here
UNVERIFIED Flagged so you can see it, but not used to calculate anything

The cost side is built from "What AI Actually Costs in Water and Carbon," compiled 29 July 2026. The value side draws on a separate research plan for pricing AI's output in dollars, plus the same water/carbon addendum used throughout the Value lens. Every figure is sourced; where this app does its own math on a number a source didn't publish directly, it's marked app's arithmetic instead of being presented as if someone else measured it.

Even the experts don't fully agree: the International Energy Agency says the world's data centres used 485 terawatt-hours of electricity in 2025 (roughly what 45 million homes use in a year). A United Nations research institute says 448 — 8% lower, on a number that in principle should just be a matter of counting. If two respected research bodies can't agree to the nearest tenth, be skeptical of anyone who states an AI footprint number as if it were precise.