How much energy does ChatGPT use?

How much energy does one ChatGPT query use?

A single typical ChatGPT-style text query uses on the order of 0.3 watt-hours of electricity — very roughly enough to power a 10-watt LED bulb for under two minutes. In 2025 Google measured its median Gemini text prompt at 0.24 Wh, and Epoch AI estimates a typical ChatGPT query at ~0.3 Wh. Small models use far less; reasoning, long answers, and image generation use much more. The honest range still spans about 10×.

An important correction: a widely repeated earlier estimate put a typical query at around 3 Wh. With production measurements now available, that figure is considered roughly 10× too high for a typical short query. We've recentred this page — and the calculator — on the lower, measured numbers.

These are estimates and measurements, not a meter on your own session. For models that don't publish data, researchers infer energy from chip power draw, model size, and typical response lengths; Google now publishes measured figures directly. Oviedo et al. 2026 (Joule) report a frontier median of 0.31 Wh, with reasoning-heavy queries running much higher (~3.9 Wh). We headline a median and show a range rather than a false-precision single number.

What actually happens when you send an AI prompt?

When you hit send, your prompt travels to a datacenter where specialized GPUs run the model's billions of parameters to predict your answer one token at a time. The electricity cost is dominated by these chips, plus cooling and power overhead. This is inference — running the model, not training it.

Training a large model is a separate, enormous one-time cost. But it is amortized across billions of subsequent queries, so the per-prompt share of training is tiny. This page — and the calculator — focus on inference, the cost you actually trigger each time you press send.

Why does the length of the answer matter?

Output length is the biggest lever on a single query's energy. Models generate text autoregressively — producing one token, feeding it back in, then producing the next. Each generated token requires another full pass through the network, so writing a long essay costs far more than reading a long prompt.

This is why a chatbot summarizing a long document (large input, short output) is cheaper than one drafting a long report (short input, large output). Input tokens are processed in parallel and are comparatively cheap; output tokens are the expensive part. The calculator models this with an affine formula: a fixed per-query overhead plus a marginal cost per token, weighted heavily toward generated tokens.

How much water does ChatGPT use?

AI uses water two ways: on-site evaporative cooling at the datacenter (Scope 1, ~1.15 mL/Wh), and water embedded in generating the electricity it consumes (Scope 2, ~3 mL/Wh). Combining both gives roughly 4 mL of water consumed per watt-hour, so a typical ~0.3 Wh query uses about a millilitre — a tiny fraction of a small cup. This follows Li et al., Making AI Less Thirsty (arXiv:2304.03271), counting water consumed (evaporated) rather than withdrawn. The total ranges widely by region — from ~0.5 mL/Wh for efficient hyperscalers up to ~9 mL/Wh where cooling is water-intensive, as in hot, dry climates.

The split matters because it depends on where the datacenter sits and how its grid is powered. A facility on a hydro-heavy grid has very different off-site water than one drawing from gas plants. As with energy, treat these as ranges, not precise per-prompt readings.

How does this compare to everyday things?

Honestly, a single text prompt is small — at ~0.3 Wh it's a minute or two of a small LED bulb, well under 2% of a smartphone charge, and about a millilitre of water. The environmental story is not one prompt; it is scale: heavy daily use across hundreds of millions of people, and especially image and video generation, which can cost many times more than a text reply.

So the takeaway is balanced. Don't agonize over a single question — but recognize that the totals add up, and that generating images is a meaningfully heavier operation than asking a question. Carbon depends on your grid; the calculator defaults to the global average (~440 g CO₂e/kWh, IEA Electricity 2026 / Ember 2026) and has a region toggle so you can switch to the US preset (~348 g CO₂e/kWh, EPA eGRID2023).

Want the numbers for your own usage?

Plug in your model, typical answer length, and how many prompts you run, and see energy, water, and carbon with an honest uncertainty range.