How much energy and water does one AI query actually use — and how does that compare to things you already do every day? This is a genuinely contested question: credible public estimates disagree by 10–30× depending on methodology, so this tool shows ranges with sources, not single truths — here’s what’s actually known, and what isn’t. Hover (or keyboard-focus) any bar or dot for the exact figures and the source behind them.
Figures last checked October 27, 2026. What changed and why is in the changelog.
Estimate basis (applies to all range-based charts below):
Put an AI task on a familiar scale
Pick an AI task. Its estimated energy and water use (colored bar, with the whisker showing the low–high range across published estimates) is placed on the same axis as everyday activities (gray bars). Note the log scale — each gridline is 10× the previous one.
Show this task as:
Low-confidence figure: the 5-second estimates come from open video models measured on research hardware (a July 2026 preprint). The high end is an estimate for one 12-second Sora 2 Pro clip, inferred from generation time. No commercial provider has disclosed a per-video figure — treat this as a rough order of magnitude only.
Selected AI task — energySelected AI task — waterSelected AI task — carbonEveryday comparison (context)Low–high range of estimates
Show:
Energy per task
Water per task
Water figures for AI mix on-site cooling and (for some sources) lifecycle/indirect water — see each bar’s source note.
Carbon per task
Why this says 0.09 g when Google says 0.03 g — market-based vs location-based carbon
AI task vs. doing it yourself
Your laptop counts on both sides. The AI version is the query in the data center plus the minutes you spend at your laptop writing the prompt, reading the answer, and checking it. Doing it yourself is the minutes you would spend at your laptop instead. Set the minutes to match a task of yours. Shared log axis.
AI version, minutes at your laptop (prompt, read, check):
Doing it yourself, minutes:
Your laptop’s electricity:
Show:
AI version: query + your laptop timeDoing it yourself: your laptop time
Read this chart with four caveats
The laptop runs regardless. Counting your laptop on both sides makes the comparison fair, but that energy is marginal, not avoided: if you would be at your computer anyway, the minutes you save go into other work, not off the meter. What really differs between the two paths is the AI query itself and the number of minutes each one takes.
Substitution cuts both ways. One AI answer might replace several manual searches (favoring AI) — or AI’s near-zero cost per query might create queries that would never otherwise happen at all (the rebound effect, disfavoring AI). Both are real, and this comparison cannot tell you which dominates for any given use.
Some things are left out of both sides. An external monitor would add energy to both bars in proportion to the minutes. Sending a few kilobytes of text over the network adds very little. Manufacturing your laptop is a large share of its lifetime footprint, but you own it either way. Building the data center’s chips and servers is counted only in the typical chatbot query’s carbon figure.
Where it happens matters. The water and carbon views count the water and carbon behind the electricity, but not where it is drawn: a data center’s demand is concentrated in one community; your laptop’s is spread across the grid. Local grid strain and water stress in places like Phoenix — the subject of the WSJ piece this course read — are siting questions that per-query math cannot answer.
When the agent costs more than the person
In July 2026 METR proposed the expenditure horizon — the dollar amount at which an AI agent’s progress on a task equals what a human researcher achieves for the same spend. Above that point the human is more cost-effective. METR measured it on a hard ML-engineering task. Putting the same comparison on an energy axis gives a different answer than the dollar one.
Show:
AI agentHuman researcherLow–high range of estimates
Both bars are the electricity used at the same $3,300 of spend — the point where METR measured Claude Opus-4.8’s progress matching a human researcher’s. Same money, same result, very different electricity.
Expenditure horizon, Claude Opus-4.8
$3,300
The dollar crossover. Below it the agent wins on cost; above it the human does. GPT-5.5 crossed at $2,300.
The agent’s electricity at that same spend
≈15×
Central estimate. The plausible range spans roughly 3× to 75×, driven mostly by one unpublished number: how hard the 32 GPUs were actually working.
Models that made no validated progress at all
2 of 4
GPT-5 and Opus-4.1 achieved a ~$0 expenditure horizon after re-validation. Their energy per unit of progress is not favorable — it is undefined.
Read this chart with five caveats
A dollar of compute is not a dollar of labor. The two axes diverge because a dollar spent on GPU time buys on the order of 15× more electricity than a dollar spent on a person’s time. Cost parity and energy parity are different thresholds, and a metric built for one does not answer the other.
This is one task, and an unusually GPU-hungry one. Progress on the NanoGPT speedrun requires running training experiments, so both sides are dominated by H100 time rather than by thinking. The ratio measures experimental efficiency per unit of insight, not “AI versus humans” in general.
The direction flips for ordinary work. In the chart directly above, a typical AI query plus your own time at the laptop comes in below twenty minutes of doing the task by hand; only a long reasoning query comes in above it. Task type decides the answer. Nothing here says AI is generally the higher-energy option, and nothing in the chart above says it is generally the lower one.
METR calls its own harness “likely inefficient.” The agents had continuous reserved access to the GPUs, which invites wasteful experimentation. A better-run agent would land lower. This is a first-generation setup and should be expected to improve.
The largest uncertainty was never published. METR reported the hardware (4 nodes, 32 H100s) and the time limit (5 days) but not how hard those GPUs were actually working. That single assumption moves the agent estimate by about 3×, which is more than every other assumption combined.
Right-sizing: your choices matter
For a simple query, a frontier model can use anywhere from about 3× to 100× the energy of a small model that would answer it just as well — roughly 3–5× between adjacent model classes under production serving, rising toward 100× when a reasoning model works a long prompt. Model choice is still the largest per-query lever a user controls.
Small modelFrontier model
Efficiency is improving fast
33×
more energy-efficient per median prompt (and 44× lower carbon), May 2024 to May 2025.
Source: Google, 2025
…but don’t read that stat alone (Jevons paradox)
But over the same period (calendar 2024–2025) Google’s own totals kept climbing: total electricity use rose 27% in 2024 and a further 37% in 2025, and water consumption reached 10.9 billion gallons in 2025, up 34%. Its operational emissions fell 2%, but its supply-chain emissions rose 25%. A textbook Jevons paradox: far more efficient per query, yet the total footprint still grows because usage grew faster than efficiency improved.
Zooming out: the aggregate numbers
Per-query numbers are small; the aggregate footprint is not. It is real, it is growing fast, and it is concentrated in particular places. These are the headline figures, stated plainly.
Data centers: how much of it is AI?
Data centers are a real local concern: their electricity, water, land, and noise fall on the communities around them. But “data center” does not mean “AI.” Most data-center electricity today still runs everything else online, such as email, streaming, cloud storage, websites, banking, and business software. AI is the smaller part, and the fastest-growing one.
World data-center electricity in 2025: 485 TWh
Two credible estimates of the AI share. They differ because they draw the line in different places. Both are shown against IEA’s 485 TWh total.
IEA: whole facilities built mainly for AI
AI ~155 TWh
everything else ~330 TWh
Chip-based estimate: AI chips only
AI 46–82 TWh (dark to light); everything else ~400 TWh
Either way, AI was well under half of data-center electricity in 2025: roughly a tenth to a third. An independent 2026 estimate lands in between, at about 20% of 448 TWh (Aczel et al., 2026).
Today
AI is about 1/10 to 1/3 of data-center electricity
The rest is the ordinary internet and business computing.
Because the growth is AI, and new AI facilities are large and built in particular places. A small share of world electricity can still be a large share of one region’s water or grid. In the Phoenix area, for example, on-site cooling water for data centers was about 385 million gallons a year, and Ceres projects more than 3.7 billion as planned facilities are built (Ceres, 2025; some water-policy experts question the study’s assumptions). A fair summary: most of today’s data-center footprint is not AI, but most of its growth is. When a headline says “data centers,” ask whether it means all of them or the new AI buildout.
Where the growth is actually coming from
It is volume, not per-query cost
Energy per prompt is falling quickly. Total tokens processed is rising far more quickly, and increasingly from programmatic and agentic traffic rather than people typing.
How much of electricity demand growth is data centers?
The answer depends entirely on which system you are looking at — which is why this question produces so many contradictory headlines.
The short version
Capacity is not consumption
…and it is wrong
Why, in order of size
The transferable lesson
What this tool does not measure
Why it still matters
Why do the public numbers disagree so much?
Estimates for the same task differ by 10–30× in the public record. That isn’t (mostly) anyone lying — it’s six methodological choices that compound. Expand each one.
What we genuinely don’t know
Epistemic humility is the point of this section, not a footnote. Anyone quoting precise figures for the items below is estimating, not reporting.