How much energy does AI use?
Pick an example question and compare published estimates for the electricity used to generate one answer. The numbers vary dramatically with model, prompt length, response length and reasoning effort.
Median Gemini Apps text prompt in May 2025 using Google’s comprehensive serving methodology, including accelerator, CPU/DRAM, idle capacity and data-centre overhead. Google also reported 0.03 gCO₂e and 0.26 mL of water per median prompt.
Sam Altman stated that an average ChatGPT query uses about 0.34 Wh. The public statement does not provide a model breakdown or detailed methodology, so PAH treats it as a product-wide reference rather than a model-specific measurement.
Compare the same question across models
The independent benchmark below estimates inference energy for fixed token workloads. Choose the example closest to what you are asking.
| Model | Estimated energy | Relative use | 10 W LED equivalent | Evidence |
|---|
Model figures are from Jegham et al., How Hungry is AI? (2025, updated preprint). They are infrastructure-aware statistical estimates, not electricity-meter readings from the providers’ data centres.
Your AI-use estimate
Estimate what repeated use would mean over a year. Electricity price is editable and is used only to translate energy into a household-style cost comparison.
Why the answer can change by 100×
A “prompt” is not a unit of computational work. A small model giving a short answer can finish quickly. A reasoning model can generate many internal tokens, run for longer and use much more accelerator time. Large input documents also have to be processed before the first output token appears.
In the cited benchmark, the long workload is estimated at about 0.45 Wh for GPT‑4.1 nano and 39.22 Wh for o3 — roughly an 86× difference for the same token-size test.
The energy used by model training, chip manufacture, networking, your own device and other lifecycle impacts is not consistently included in per-query figures, so PAH does not add them into a false “precise” total.
How confident should you be in these numbers?
What about the newest AI models?
Current products move faster than environmental disclosure. As of 15 September 2026, PAH found no comparable public model-specific Wh-per-query figure for the newest GPT‑5.6 family, and the Google 0.24 Wh figure is a Gemini Apps median rather than a Gemini 3 Pro-specific measurement. We will keep those separate until a like-for-like figure is published.
Does AI training count?
Yes, but it answers a different question. This calculator estimates inference: the electricity associated with running an already-trained model to answer a request. Training a frontier model is a large one-off or periodic workload, while inference happens every time users ask questions. A full lifecycle assessment would also consider hardware manufacturing, data-centre construction and the carbon and water intensity of the electricity supply.
Sources and methodology
Google: median Gemini Apps text prompt, May 2025 — 0.24 Wh using Google’s comprehensive methodology. Google Cloud methodology and results.
ChatGPT product-wide reference: Sam Altman, “The Gentle Singularity”, June 2025 — average ChatGPT query ~0.34 Wh. Original statement.
Model-by-model estimates: Jegham, Abdelatti, Elmoubarki & Hendawi, “How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference”, arXiv:2505.09598. Research paper.
Mistral transparency note: Mistral published lifecycle and marginal inference carbon/water data for Mistral Large 2, but not a directly comparable Wh-per-response figure, so it is not placed in the energy ranking. Mistral environmental disclosure.
Sources checked: 15 September 2026. Results are estimates, not an electricity bill from your individual AI session.
How to use the AI energy calculator
Choose the model or estimate you want to explore, enter the number of prompts or queries you expect to make, and set an electricity price if you want a cost comparison. Treat the result as an estimate rather than a meter reading: published figures vary because models, hardware, data centres, prompt lengths and measurement methods differ.
Worked example
If an estimate assumes 0.3 Wh per query, 1,000 queries would correspond to about 300 Wh, or 0.3 kWh. The calculator makes that multiplication explicit so different published assumptions can be compared on the same basis.
AI energy calculator FAQs
Is there one fixed energy figure for a ChatGPT or AI query?
No. Energy use varies with the model, hardware, workload, prompt and response length, batching and data-centre efficiency. That is why this page shows estimates and sources rather than presenting one universal number.
Does the calculator include training energy?
No. Per-query estimates are intended to describe inference or use-time energy. Training a model is a separate workload and should not be added to each query without a clearly defined allocation method.
Why do published estimates disagree?
Researchers and companies may measure different hardware, system boundaries and workloads, and some estimates include more infrastructure overhead than others.