---
title: "How Much Electricity Does One AI Prompt Use? 5 Surprising Facts"
url: "https://learnaitodayonline.com/how-much-electricity-does-one-ai-prompt-use/"
description: "How much electricity does one AI prompt use? About 0.3 watt-hours for a typical ChatGPT query — but the real number can vary by 100 times or more."
author: "Robert Waithaka"
published: "2026-08-23"
last_reviewed: "2026-08-22"
categories: ["AI News"]
tags: ["level-beginner"]
site: "Learn Artificial Intelligence"
approx_tokens: 3129
---

# How Much Electricity Does One AI Prompt Use? 5 Surprising Facts

How much electricity does one AI prompt use? Depending on who you ask, the answer is 3 watt-hours, 0.3 watt-hours or 0.24 watt-hours — a ten-fold gap between the most-cited estimates, before you even consider long documents or image generation. The truth, as with AI water consumption, is that there is no single number. But once you understand what each estimate is measuring, the picture is clear: a single prompt is cheap in energy terms, and the total is enormous. Here is what the research actually says, in plain English.

## How Much Electricity Does One AI Prompt Use?

The number you will see most often — roughly 3 watt-hours per ChatGPT query, about ten times a Google search — comes from a 2023 estimate by Alex de Vries. It was built on early assumptions: a GPT-3.5-sized model, older A100 chips, and an unusually long average query of around 1,500 words. Researchers at Epoch AI revisited the math in early 2025 with newer hardware and more realistic usage, and put a typical GPT-4o query at about 0.3 watt-hours — ten times lower. Google’s own measurement of a median Gemini text prompt is lower still: 0.24 watt-hours, with 0.03 grams of CO2 and about five drops of water.

To make that tangible, 0.3 watt-hours is less electricity than an LED lightbulb or a laptop uses in a few minutes — about nine seconds of television. The average US household uses roughly 28,000 watt-hours a day, so one prompt is a rounding error on your daily bill.

![One prompt versus one household day versus one training run](https://cdn.sanity.io/images/gfihpee1/production/96c5aa1fbea591433f5674adc1153becf45503cc-1140x720.png)

_Figure: the scale gap — a typical prompt is 0.3 watt-hours, an average US household uses ~28,000 watt-hours a day, and training GPT-4 consumed ~50 gigawatt-hours. Sources: Epoch AI; EIA via Epoch AI; MIT Technology Review._

Why do estimates disagree by ten times or more? The variables are model size, chip generation, token counts, batching and what gets counted. Some figures cover only the GPU; others include cooling, networking and data-centre overhead. None is wrong — they are measuring different things.

| Estimate | Watt-hours | Who measured it | What's included |
| --- | --- | --- | --- |
| “3 Wh per ChatGPT query” | ~3 | Alex de Vries (2023) | GPT-3.5-era model, A100 chips, long ~1,500-word average query |
| “0.3 Wh” | ~0.3 | Epoch AI (2025) | Typical GPT-4o query on H100s, batching and overhead included |
| “0.24 Wh” | 0.24 | Google | Median Gemini text prompt, full data-centre measurement |
| Long input (10,000 tokens) | ~2.5 | Epoch AI | Input processing dominates when documents are attached |
| Very long input (100,000 tokens) | ~40 | Epoch AI | Roughly 200 pages of context |
| BLOOM-176B (measured) | ~4 | Luccioni et al. | 2022 research deployment without batching |

_Table: per-prompt electricity estimates, decoded. The differences come from model size, token counts, chip generation, batching and accounting boundaries._

![How much electricity does one AI prompt use — published estimates](https://cdn.sanity.io/images/gfihpee1/production/2cd83860c2292751d7e4d38d842010ac0393b655-1709x770.png)

_Figure: published estimates of electricity per AI prompt, in watt-hours (log scale). Sources: Google via EcoFlow; Epoch AI; Alex de Vries (2023); Luccioni et al. via Epoch AI._

## How Much Electricity Does ChatGPT Use Per Day?

Multiply small numbers by billions and they stop being small. ChatGPT is estimated to handle around a billion messages a day, which works out to roughly 12.5 megawatts of continuous inference load — comparable to the power used to train GPT-4o. Industry estimates put total daily AI inference demand above 850 megawatt-hours. And inference, not training, is the real energy story: an estimated 80–90% of AI’s computing power goes to running models for users, not building them.

Training GPT-4 consumed about 50 gigawatt-hours — enough to power San Francisco for three days — but it happens once. Inference happens constantly, forever. For context, ChatGPT’s estimated 2.5 billion daily prompts carry a water footprint roughly equal to the basic annual domestic needs of 500,000 people in Sub-Saharan Africa — a reminder that small per-prompt numbers still add up. And for a model like BLOOM, researchers found it took 590 million uses to match the energy cost of training. For hugely popular models, usage can overtake training in a matter of weeks.

| Metric | Figure | Source |
| --- | --- | --- |
| ChatGPT messages per day | ~1 billion | Epoch AI |
| ChatGPT inference load | ~12.5 MW continuous | Epoch AI |
| Total daily AI inference demand | >850 MWh | IEEE Spectrum estimate |
| Share of AI computing used for inference | 80–90% | UN University; MIT Technology Review |
| GPT-4 training energy | ~50 GWh (≈ 3 days of San Francisco) | MIT Technology Review |
| Uses of BLOOM needed to match training | ~590 million | Hugging Face (Luccioni et al.) |
| Water footprint of ChatGPT's ~2.5B daily prompts | ≈ annual domestic needs of 500,000 people | UN University |

_Table: AI's energy use at a glance — inference dominates because it never stops._

![Inference vs training share of AI energy use](https://cdn.sanity.io/images/gfihpee1/production/43d5cd49b8948f1b09e838c71768402192fb63ca-1186x608.png)

_Figure: 80–90% of AI computing power goes to inference (running models for users), not training. Sources: UN University; MIT Technology Review._

## Why Reasoning Models Changed the Maths

Reasoning models — o1, o3, Deep Research — think before they answer, generating long chains of thought. Epoch AI found o-series models produce around 2.5 times as many tokens as GPT-4o, and studies cited in the research suggest reasoning prompts can emit up to 50 times more CO2 than simple ones.

Input length matters just as much: a query with a 10,000-token document attached costs roughly 2.5 watt-hours, and a 100,000-token upload (about 200 pages) approaches 40 watt-hours. Agentic AI pushes this further: Gartner estimates agents can consume five to thirty times more tokens per task than a standard chatbot, because they read screens, call tools and repeat the loop. The cheapest prompt is a short question with a short answer.

## AI Energy Use Compared to Everyday Things

Fair comparisons help. MIT Technology Review’s measurements on open models found Llama 3.1 8B used about 114 joules (0.03 watt-hours) per response — a tenth of a second of microwave time. The giant 405B model used about 6,700 joules (1.9 watt-hours) — eight seconds of microwave. Generating a 1024×1024 image with Stable Diffusion 3 Medium used roughly 2,300–4,400 joules (0.6–1.2 watt-hours) — five and a half seconds of microwave.

Older 2023 research found image generation could cost as much as a full phone charge; the newer numbers are much lower. A five-second AI video, by contrast, can use as much electricity as running a microwave for over an hour. The honest ranking: short text < long text < image < video. Misleading comparisons usually pick one extreme — a giant model doing a hard task — and present it as “the” cost.

| Task | Energy | Everyday equivalent |
| --- | --- | --- |
| Llama 3.1 8B text response | ~114 J (0.03 Wh) | ~0.1 seconds of microwave |
| Llama 3.1 405B text response | ~6,706 J (1.9 Wh) | ~8 seconds of microwave |
| Stable Diffusion 3 image (1024×1024, 25 steps) | ~2,282 J (0.6 Wh) | ≈ half the 50-step image energy |
| Same image, 50 steps | ~4,402 J (1.2 Wh) | ~5.5 seconds of microwave |
| Median Gemini text prompt | 0.24 Wh | Less than 9 seconds of TV |
| 5-second AI-generated video | ≈ 1+ hour of microwave | MIT Technology Review estimate |

_Table: measured AI energy use versus everyday things. Values are from MIT Technology Review (2025) and Epoch AI/Google figures._

![Model size and task type drive energy per response](https://cdn.sanity.io/images/gfihpee1/production/6e52278262e2779b772918aa46ea64eadfa57148-1250x747.png)

_Figure: measured energy per response for open models — a 405B-parameter model uses roughly 60 times more energy than an 8B model. Source: MIT Technology Review (May 2025)._

## Does Using a Smaller Model Help?

Massively. Model size is the single biggest lever: 50 times more parameters meant roughly 60 times more energy in MIT’s Llama tests. A fine-tuned specialist model used for classification consumed around 30 times less energy than a general generative model doing the same job, in Hugging Face’s research. Efficiency is improving fast too — Google reports cutting energy per prompt by more than 30 times and carbon by 40% in a single year, and DeepSeek’s architecture activates only 37 billion of its 671 billion parameters per token. Prompt design matters as well: in testing, simple prompts used nine times less energy than complex creative ones. Choosing the right model for the task beats every other user decision.

## Running AI Locally: More or Less Energy?

It depends — on your hardware, your model and your usage. Buying a data-centre GPU (an H100 costs roughly $25,000–30,000) only pays off against cloud renting at about $2 per hour after 17–20 months of near-constant use, and you still pay the electricity, cooling and maintenance bills. A consumer rig draws far less than a data-centre cluster, but it lacks the batching efficiency of hyperscale inference, where one GPU serves many users at once.

For occasional use, cloud is almost certainly the lower-energy option: you share infrastructure that is already running. For heavy, repetitive workloads, a small quantized model on your own hardware can beat a big cloud model on energy per task — and it keeps your data local. Our [Ollama guide](https://learnaitodayonline.com/ollama-download-install-run-local-ai-models/) walks through getting started. Either way, running locally doesn’t make AI energy-free; it moves the meter to your wall socket.

## How to Read the Next AI Energy Headline

When you see a claim about AI and electricity, ask three questions: which model, which task, and what’s counted? A 3-watt-hour figure and a 0.24-watt-hour figure can both be true. The useful questions are about the aggregate — data centres already use more than 4% of US electricity and could reach 12% by 2028 — and whether the growth is justified by what AI actually delivers.

If you want the mechanics behind the meter, our explainer on [what an LLM is](https://learnaitodayonline.com/what-is-an-llm/) covers the basics, and our companion piece on [how much water AI uses](https://learnaitodayonline.com/how-much-water-does-ai-use/) completes the environmental picture. For the practical side — trying smaller models and seeing the difference — the [ChatGPT vs Claude vs Gemini comparison](https://learnaitodayonline.com/chatgpt-vs-claude-vs-gemini/) and the [AI glossary](https://learnaitodayonline.com/ai-glossary-essential-terms/) are good next stops.

---

## Sources

- [How much energy does ChatGPT use? (Epoch AI, 2025)](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use)
- [We did the math on AI’s energy footprint (MIT Technology Review, 2025)](https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/)
- [Making an image with generative AI uses as much energy as charging your phone (MIT Technology Review, 2023)](https://www.technologyreview.com/2023/12/01/1084189/making-an-image-with-generative-ai-uses-as-much-energy-as-charging-your-phone/)
- [Making AI Less “Thirsty” (arXiv, 2025)](https://arxiv.org/abs/2304.03271)
- [2024 United States Data Center Energy Usage Report (Lawrence Berkeley National Laboratory)](https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf)
- [Data Centers and Water Consumption (EESI, 2025)](https://www.eesi.org/articles/view/data-centers-and-water-consumption)

_Last reviewed: 22 August 2026. Energy figures move quickly; this article will be re-checked quarterly._