How much energy does a ChatGPT query use?

Nobody knows precisely, and the honest answer is a range. A widely cited 2025 analysis from Epoch AI put a typical short GPT-4o query at roughly 0.3 Wh, OpenAI's Sam Altman later stated an average of about 0.34 Wh, and Google reported a median of 0.24 Wh per Gemini text prompt. Older 2023-2024 estimates landed near 3 Wh. That's a 10x spread, which is exactly why this calculator shows a low-to-high band instead of pretending to know one number.

Does AI use more energy than a Google search?

Probably a bit more per query, but they're closer than headlines suggest. Google's long-standing figure for a classic search is about 0.3 Wh — essentially the same as Epoch AI's estimate for a short text chatbot query. Longer AI responses and reasoning models use more, and image generation uses far more. The old "one AI query equals ten Google searches" claim came from the ~3 Wh estimates that newer analyses (and the providers' own published figures) have since revised sharply downward.

How much energy does AI image generation use?

Much more than text — roughly an order of magnitude. The landmark Hugging Face study by Sasha Luccioni and colleagues measured about 2.9 Wh per image on a large open image model, with the least-efficient model tested burning 11.5 Wh — about one full smartphone charge per picture. Newer, smaller image models come in lower, which is why this calculator uses a 1-3 Wh band per image. Video generation runs far higher still.

Is using AI bad for the environment?

One query is tiny — a short text chat is a fraction of a phone charge, and even a year of heavy use is well under 1% of a typical home's electricity. The concern is aggregate scale: billions of queries a day, plus the far larger energy of training models and building data centers, add up to real grid and water demand. This tool deliberately shows numbers instead of moralizing: individual impact is small, the collective total is where the debate genuinely matters.

How accurate is this calculator?

It's an honest range, not a measurement — and we'd rather say so. The per-query figures come from public 2025-2026 estimates that disagree by roughly 10x, because per-query energy depends on the model, the hardware generation, data-center efficiency, and whether you count idle capacity. We carry a low and a high bound through every calculation and show the sources for each. Treat the result as an informed ballpark for curiosity, not an audited carbon report.

AI Energy Calculator — What Does Your Chatbot Cost?

Enter how many AI text conversations you have per day (and, optionally, images you generate per week). You'll get an honest low-to-high energy range for the year, translated into phone charges, toast, EV miles, and a share of your home's electricity — every figure sourced, because per-query AI energy is genuinely disputed.

How the AI energy calculator works

This tool answers a simple question — "what does my AI habit cost in electricity?" — without pretending to a precision that doesn't exist. It multiplies your daily text conversations by a per-query energy band of 0.3 to 3 watt-hours, adds 1 to 3 Wh for each image you generate, and scales to a year. The catch is that the per-query figure is one of the most disputed numbers in tech right now: recent rigorous estimates for a short ChatGPT-class text query sit around 0.3 Wh (Epoch AI, February 2025 — a figure OpenAI's Sam Altman later echoed at 0.34 Wh, and Google roughly matched with a reported 0.24 Wh median per Gemini text prompt), while the older, widely cited figure was ~3 Wh — now generally considered too high, but not indefensible for long, reasoning-heavy responses. Rather than pick a side, we carry both bounds all the way through, so your result is an honest band, translated into things you can picture: phone charges, LED-bulb hours, EV miles, toast, and a share of your home's annual electricity.

Why so uncertain? Because AI companies rarely publish per-query energy, and the honest inputs — model size, hardware generation, GPU utilization, data-center overhead, and how long the answers are — are mostly hidden or highly variable. A short chat is genuinely tiny. An image is several times bigger. Video is far bigger still — enough that we deliberately left it out rather than guess. Training a model dwarfs all of it, but that's a one-time cost spread across billions of queries, so this calculator covers your usage (inference), not training.

The formula

yearly kWh = (chats/day × per-chat Wh + images/week ÷ 7 × per-image Wh) × 365 ÷ 1000

Computed twice — once with the low figures (0.3 Wh per chat, 1 Wh per image) and once with the high (3 Wh per chat, 3 Wh per image). The equivalents use fixed, sourced constants: a full smartphone charge ≈ 0.012 kWh (the EPA-derived figure used in the Luccioni study's phone-charge comparisons), a 10 W LED bulb per hour, an EV at 0.28 kWh/mile (real-world fleet average), a slice of toast at 0.033 kWh (a ~1,000 W toaster for about two minutes), and an average US household's ~10,500 kWh per year of electricity (EIA).

Worked example

15 text conversations a day plus 10 images a week, projected to a year:

Low = 15 × 0.3 + (10 ÷ 7) × 1 = 5.9 Wh/day → × 365 ÷ 1,000 = 2.2 kWh/year. High = 15 × 3 + (10 ÷ 7) × 3 = 49.3 Wh/day → 18 kWh/year. So the honest answer is 2.2–18 kWh per year, midpoint about 10.1 kWh.

That range equals roughly 180–1,499 phone charges, 216–1,799 hours of a 10 W LED bulb, 7.7–64.2 miles in an EV, 66–545 slices of toast — and 0.021%–0.171% of what an average US home uses in a year. Small individually; the debate is about multiplying it by billions of users.

Why the estimates disagree by 10x

The gap between 0.3 Wh and 3 Wh isn't sloppiness — it's genuinely different accounting. The higher, older figure (traceable to a 2023 analysis by Alex de Vries, back when GPT-4-era hardware was the baseline) assumed longer responses, less efficient chips, and generous overhead. The newer, lower estimates — Epoch AI's 0.3 Wh analysis (February 2025) and Google's published 0.24 Wh median for Gemini text prompts (August 2025) — reflect newer hardware, mixture-of-experts models that activate only a fraction of their parameters, and realistic (shorter) query lengths. But the low numbers have their own caveats: company self-reports are medians (long reasoning queries cost much more than the median), and critics note they may exclude idle capacity held in reserve and some data-center overhead (the PUE multiplier). Both bounds are defensible depending on what you count, which is why this calculator refuses to pick one.

Training vs. inference: the amortization argument

Training a frontier model is enormous — estimates for a single big training run reach into the tens of gigawatt-hours, millions of times a single query. So why doesn't it swamp your personal total? Amortization: that one-time cost is spread across the billions of queries the model serves over its lifetime, adding only a fraction of a watt-hour to each. Whether to count it at all is one of the genuine disagreements between estimates. What's not disputed: your marginal query tonight doesn't re-train anything — the training electricity was spent either way — but a world that trains ever-more models on spec is a different accounting question, and an honest calculator admits that's beyond its pay grade.

One year of chatbots vs. one hot bath

Here's the perspective section, with the math shown. A heavy user — say 30 conversations a day — lands at 3.3 to 33 kWh per year (30 × 0.3 or 3 Wh × 365). Heating one bath — roughly 120 liters warmed by 25 °C — takes about 3.5 kWh. So a year of heavy chatbot use is one hot bath at the low estimate and around nine at the high — and even the high end is about 0.3% of an average home's annual electricity. If you want to move your personal needle, the levers are elsewhere: video generation (orders of magnitude more per clip than text, and the main reason per-use footprints may climb), and — entirely outside your control — where the data centers sit, since a query served from a hydro- or nuclear-heavy grid carries a fraction of the carbon of one served from a coal-heavy one. No doom, no dismissal: the numbers are small, the error bars are real, and the aggregate is a legitimate infrastructure story.

Sources & method

Every figure is embedded in the calculator with its source. Estimates are current as of July 2026 and will keep moving as models and hardware change.

Text query, 0.3 Wh (low) — Epoch AI, "How much energy does ChatGPT use?", February 2025 (typical GPT-4o query). Corroborated by OpenAI CEO Sam Altman's stated ~0.34 Wh average (June 2025) and Google's reported 0.24 Wh median per Gemini Apps text prompt (August 2025).
Text query, 3 Wh (high) — the older, widely cited estimate (Alex de Vries, Joule, 2023, ~2.9–3 Wh), now generally considered too high for typical queries but retained as an honest upper bound for long, reasoning-heavy use.
Image generation, 1–3 Wh — Sasha Luccioni et al., "Power Hungry Processing" (2023): ~2.9 Wh average per image on a large open model, up to 11.5 Wh worst case; see MIT Technology Review's coverage. Newer, smaller image models come in lower — hence the 1 Wh floor.
Smartphone charge 0.012 kWh — the EPA-derived full-charge figure used in the Luccioni study's comparisons. EV 0.28 kWh/mile — real-world average; efficient EVs run ~0.25, big trucks 0.4+. Toast 0.033 kWh/slice — ~1,000 W toaster for ~2 minutes. LED bulb — 10 W nominal.
US household ~10,500 kWh/year — US Energy Information Administration (10,791 kWh in 2022; ~863 kWh/month in 2024).

Per-query energy estimates are genuinely debated and depend on model, hardware, and data-center efficiency — figures as of July 2026. Treat these as honest ranges, not measurements.

AI Energy Calculator