Cost/of/an/Answer
Energy · Carbon · Water — per query

The price of an answer.

A web search and an AI prompt can return the exact same fact. They don't cost the same to produce it. This page defaults to the most recent (2025) figure on both sides — a company's own disclosure where one exists, and the closest credible independent estimate where it doesn't. Every number is labeled either way, and you can switch sources with the toggles below.

AI Prompt
0.00Wh
electricity per prompt
Carbon
Water
Scroll for the full comparison
01 — The full comparison

Same question, three different meters.

Energy is only one part of the bill. Carbon depends on how the electricity was generated; water is drawn to cool the servers that answered you. Here's how a single search stacks up against a single AI prompt on each meter.

Energy — watt-hours (Wh)
Search
AI prompt
Carbon — grams CO₂e
Search
AI prompt
Water — millilitres
Search
AI prompt
02 — Why it still matters

The per-query number isn't the whole story.

The headline "AI uses 10× more energy than search" comes from a 2023 estimate that today's companies say is out of date. The current, self-reported figures above are much closer together — sometimes AI comes out lower. But three things change the real picture.

01

One prompt is rarely one message.

A search is usually a single round trip. A chat is a back-and-forth — several prompts to get to one answer, each one billed separately.

02

Complexity multiplies the cost fast.

The 0.24–0.34 Wh figures above are for short text answers. Long documents, reasoning-heavy prompts, code generation, or image and video output can run 10–100× higher per request.

03

Scale erases small differences.

Billions of prompts a day means even a fraction of a watt-hour compounds into real electricity, water, and grid demand — regardless of which side of this page is technically "cheaper."

04

Comparisons age unevenly.

Google's only official search figure is from 2009 — this page defaults to a 2025 independent re-estimate instead, so it isn't comparing 16-year-old hardware to brand-new AI infrastructure. On that fairer footing, modern search comes in well below either AI figure. The underlying reason: a lookup against an index is inherently cheaper than a full neural-network pass per output token. That gap has narrowed with more efficient AI serving, but it hasn't closed. Switch the search card to "2009 official" to see the only number Google has actually put its own name to.

Published per-query energy estimates (Wh), lowest to highest
03 — Sources

Every figure, traced back.

Where a company has published its own number, this page uses it. Where it hasn't — a modern search figure — the closest credible independent estimate is used instead, clearly labeled as such. Further independent research is cited below for context on why estimates vary.

Google · Official Blog · Jan 2009
"Powering a Google search" — 0.0003 kWh and 0.2 g CO₂ per search, self-reported by Google SVP Urs Hölzle.
Read source ↗
Wim Vanderbauwhede · Independent estimate · 2025
Re-estimates a modern search at ≈0.04 Wh and 0.02 g CO₂ by applying published hardware-efficiency and grid-carbon trends to Google's own 2009 baseline. Not company-disclosed — included because Google has issued no official update in 16 years.
Read source ↗
Google · Technical paper / Cloud Blog · Aug 2025
"Measuring the environmental impact of delivering AI at Google Scale" — 0.24 Wh, 0.03 gCO₂e, 0.26 mL per median Gemini text prompt.
Read source ↗
arXiv · Peer-reviewable preprint · Aug 2025
Elsworth, Huang, Patterson et al., "Measuring the environmental impact of delivering AI at Google scale" — full methodology behind the Gemini figures.
Read source ↗
OpenAI · Sam Altman, CEO · Jun 2025
"The Gentle Singularity" — 0.34 Wh and 0.000085 gallons (≈0.32 mL) of water per average ChatGPT query.
Read source ↗
Epoch AI · Independent research · Feb 2025
"How much energy does ChatGPT use?" — independently modelled ≈0.3 Wh for a typical GPT-4o text query, used here only for range context.
Read source ↗
Online Learning Consortium · Summary of Jegham et al. · Dec 2025
A 30-model benchmark found per-prompt energy ranging from ≈0.4 Wh (short prompts) to 33+ Wh (long, complex prompts) — cited here for the upper end of the range.
Read source ↗