AI energy consumption: how much electricity does AI actually use?
One AI chat prompt uses roughly 0.2 to 2 watt-hours of electricity: seconds of a microwave. US data centers as a whole used about 176 TWh in 2023 (4.4% of national consumption), and the LBNL reference projection lands between 325 and 580 TWh by 2028, with AI the main driver of growth.
Water on this site is mostly computed from energy, so this page is the foundation everything else stands on.
Per query: smaller than the headlines
The best-documented per-query figures:
| Figure | Source |
|---|---|
| 0.24 Wh per median Gemini prompt | Google official (2025) |
| 0.34 Wh per average ChatGPT query | Altman (2025) |
| ~0.3 Wh per GPT-4o query | Epoch AI estimate |
| ~0.9 Wh per 400-token GPT-4.1 exchange | Jegham et al. benchmark (2025) |
Where queries get expensive is length and mode. Reasoning models (o3-class, DeepSeek-R1) burn 5–20× a standard exchange because they generate long hidden chains of thought. Media generation is its own league: ~3 Wh per image (Luccioni et al., FAccT 2024), and anywhere from ~25 to ~1,000 Wh per 5-second video clip. The only public measurement of a current-quality model is ~940 Wh (MIT Technology Review, 2025), and Sora- and Veo-class providers disclose nothing. One short AI video clip ≈ a couple of hours of streaming at our mid value, or a dozen at the best public measurement.
The biggest lever you have: which model
Per-query energy spans two orders of magnitude across deployed models (Jegham et al. 2025 benchmarks; these are the bands the calculator runs on):
| Class | Per ~1,000 tokens | Examples |
|---|---|---|
| Efficient | ~0.1–0.6 Wh | nano tiers, Claude Haiku, Gemini Flash |
| Standard | ~0.4–1.5 Wh | GPT mini tiers, Claude Sonnet, Mistral Large, Llama 70B |
| Frontier | ~1–3.5 Wh | GPT-5-class, Claude Opus, Gemini Pro |
| Reasoning | ~3–15 Wh | o3-class, DeepSeek-R1, extended thinking |
Routing a simple question to an efficient model instead of a reasoning one changes the cost 10–50×: the largest footprint decision a user actually controls. For intuition: a full phone charge (~12 Wh) buys roughly 35–50 typical prompts at the official figures, and an average US household's daily electricity (~29 kWh, EIA) equals about 85,000 of them.
frontier class · 2 Wh per 1,000 tokens (mid benchmark)
The same prompt can be “five drops” or “a bottle of water” depending on what you count: just data-center cooling (Low), the water behind the electricity (Mid), or training and hardware too (High).
Read the methodology16.9gallons
= 63.9 liters
ChatGPT (standard), 25 prompts/day · per year, Mid scenario
Energy
12.8 kWh
Carbon
4.5 kg CO₂e
The fleet: where the real numbers live
Per-query efficiency improves every year; total consumption rises anyway, because volume grows faster. The honest framing is that AI is the marginal driver of US electricity demand growth; data centers are why US load forecasts bent upward for the first time in 15 years. What that does to water depends entirely on which grid and which cooling design serves the new load, which is the subject of our data center map.
From energy to carbon
Carbon follows energy through the grid mix, so the same prompt emits differently in different places. The two anchor figures: Google reports 0.03 gCO₂e per median Gemini prompt (Low-leaning, clean-energy procurement; 2025), while Mistral's lifecycle analysis reports 1.14 gCO₂e per 400-token response (High, full lifecycle). The ~40× gap is mostly boundary, partly grid. Either way the scale is: well under a gram for a typical prompt on the average US grid, against ~404 g for one mile in an average US gasoline car (EPA).
Energy → water, the conversion this site runs on
- Low (0.3 L/kWh): efficient operator's on-site cooling.
- Mid (5 L/kWh): + water consumed generating US-grid electricity.
- High (30 L/kWh): + training amortization and hardware manufacturing, derived from Mistral's published LCA (45 mL per 400-token response) at this site's assumed ~1.5 Wh per response.
The full derivation of each factor, with sources, is on the methodology page.