How much electricity does one AI query use?

About 0.3 watt-hours for a median text query. Google reports 0.24 Wh for a median Gemini prompt; OpenAI cites 0.34 Wh for ChatGPT; independent benchmarks land around 0.42 Wh for a larger model. All are in the same ballpark as a single Google search and far below an hour of video streaming (~77 Wh).

How much water does AI use per query?

Roughly 0.3 millilitres of on-site cooling water per median query, per first-party figures from Google (0.26 mL) and OpenAI (~0.32 mL). Independent estimates that also count the water used to generate the electricity run several times higher. Either way it is a tiny fraction of the water embedded in everyday items like a cup of coffee (~130 L) or a burger (~2,400 L).

Is AI worse for the environment than streaming video?

Per use, no. One hour of HD streaming uses roughly the electricity of a few hundred AI queries. The activities we treat as completely normal — streaming, driving, eating beef, watering a lawn or a golf course — generally have far larger per-use footprints than a single AI query.

So AI has no environmental impact?

That's not the claim. The per-query cost is tiny, but global query volume is enormous and growing, data centres concentrate water and power demand in specific (sometimes stressed) locations, and training is energy-intensive. The point is perspective: judge AI against the other things we accept, and focus concern on scale and siting rather than the single query.

AI & the environment

Is AI really an environmental disaster?

The headlines say AI is boiling the oceans and draining the reservoirs. The per-use numbers tell a quieter story. One AI query uses about 0.30 Wh of electricity and roughly 0.30 mL of water — less than a Google search, and a rounding error next to a burger, a coffee, or a cotton t-shirt. Here are the receipts.

Electricity per query

0.30 Wh

≈ 18 sec of a 60 W bulb

Water per query

0.30 mL

on-site cooling water

Queries = 1 almond

≈ 14k

same water footprint

Queries = 1 hr streaming

≈ 257

same electricity

Figures as of July 2026. Every number on this page links to a source below. This is an argument about perspective, not a claim that AI's footprint is zero — see the honest version.

The unit

Put one query in perspective

The number that matters is per use. Here's what a single AI query actually costs, and what it takes to notice.

Electricity

A median text query runs on about 0.30 Wh — Google reports 0.24 Wh for Gemini, OpenAI puts ChatGPT at 0.34 Wh. That's roughly 18 seconds of an old 60-watt bulb, or a few minutes of an LED. You'd run 257 queries on the electricity of a single hour of HD streaming.

Range: 0.24–0.34 Wh (first-party); ~0.42 Wh in independent tests.

Water

The cooling water for one query is about 0.30 mL — a few drops. To match the water embedded in one almond you'd need about 14k queries. To match one beef burger, 8M.

Range: ~0.3 mL on-site (first-party); several mL if you also count water used to generate the electricity.

The water story

Water: one AI query vs. everyday life

Virtual-water footprints for things nobody protests. The AI bar is on the same axis — you just can't see it.

One AI query is 0.30 mL. The coffee bar next to it is about 433k times larger; the burger, about 8M times. At this scale the AI bar is thinner than the line around it.

Each item is measured per its own unit (per cup, per shirt, per query); food/product figures are total virtual water. Bars are not a single shared serving size — they're the real footprint of one of each thing.

The electricity story

Electricity: an AI query barely registers

A single query costs about the same as a Google search — and disappears next to a bulb, a microwave, or an hour of Netflix.

One hour of HD streaming uses the electricity of roughly 257 AI queries. Leaving a single old 60-watt bulb on for an hour equals about 200.

The comparison

Things we don't think twice about

Each of these is a normal, accepted part of daily life. Each dwarfs an AI query — here's by how much.

One beef burger

2.4k L

≈ 8M AI queries to equal its water

≈ 660 US gallons; beef is ~15,400 L/kg (Water Footprint Network).

Source

One cotton t-shirt

2.5k L

≈ 8.3M AI queries to equal its water

Virtual water to grow and process the cotton (~2,500 L).

Source

One hour of HD video streaming

77 Wh

≈ 257 AI queries to equal its electricity

≈ 0.077 kWh/hr (IEA); device + network + data centre combined. HD estimates range to ~0.12 kWh.

Source Source

One cup of coffee

130 L

≈ 433k AI queries to equal its water

Virtual water to grow, process and ship the beans (~130–140 L per cup).

Source Source

A handful of almonds (1 oz, ~23 nuts)

97 L

≈ 323k AI queries to equal its water

~4.2 L (1.1 gal) per almond kernel; water-footprint accounting puts it as high as ~12 L.

Source Source

A 10-minute shower

75 L

≈ 250k AI queries to equal its water

≈ 20 US gallons at an efficient 2 gpm showerhead (EPA WaterSense).

Source

Zoom out: one golf course

A single golf course drinks about 1.18M L of water — every day.

That one course, for one day, uses the same water footprint as roughly 3.93B AI queries — billions. The United States has over 16,000 of them. Nobody is writing headlines about the reservoirs golf is boiling.

Average US 18-hole course; regional use ranges from under 100,000 to over 1,000,000 gallons per day (USGA).

The honest version

Where the concern is actually valid

'In perspective' is not 'nothing to see here.' The reframe only earns trust if it's fair — so here's the other side, straight.

The problem is scale, not the single query

Per query the cost is tiny, but the world runs billions of queries a day and the total is climbing fast. Data centres already use on the order of 1.5% of global electricity, and AI is the fastest-growing slice. 'Small × enormous volume' is still a real number.

Water is local, and concentrated

A third of a millilitre averaged globally hides the fact that data centres cluster in specific places — some of them already water-stressed. The aggregate looks fine while a particular town's aquifer does not. Averages don't drink from your reservoir.

Whose number are you using?

First-party figures (Google, OpenAI) count on-site cooling. Independent studies that also count the water and energy used to GENERATE the electricity land several times higher. We used mid-range, on-site figures and flagged the ranges — but the honest answer is 'it depends what you count.'

Training is real, but amortized

Training a frontier model is genuinely energy-hungry — but it happens once and is then spread across billions of queries. Per-query training cost rounds to a rounding error; the headline training numbers are describing a one-time capital cost, not your chat.

Efficiency is improving fast

Google reported the energy per median Gemini prompt fell ~33× in a single year. Hardware, model design, and cooling are all improving. Today's per-query number is a moving target — downward.

Questions

The quick answers

Receipts

Sources

Every figure on this page traces to one of these. Figures as of July 2026; verify current data before quoting.

Where sources disagree we used a defensible mid-range value and noted the spread. Water footprints for food and products are "virtual water" — the total embedded in production — and vary by region and method. AI per-query figures are first-party, on-site numbers; see the honest version for what that does and doesn't include.