A single AI text prompt uses roughly 0.2–0.3 watt-hours of electricity — far less than the alarming numbers that circulated in 2023, but the fleet-level total is what’s exploding. Per-query water use is a fraction of a millilitre; the real story is data-center demand more than doubling by 2030.
The big picture · energy & water per query
- ⚡~0.24 Wh per prompt. Google’s Aug 2025 technical report puts a median Gemini text prompt at 0.24 Wh. Sam Altman cited ~0.34 Wh per ChatGPT query (June 2025); Epoch AI estimates ~0.3 Wh for a GPT-4o query.
- 💧Water is tiny per query, big at scale. Google reports ~0.26 mL per prompt; UC Riverside research shows the totals add up across billions of queries and during training.
- 📈The macro number that matters. The IEA puts data-center electricity at ~415 TWh in 2024, rising to ~945 TWh by 2030 — roughly a doubling, with AI the main driver.
- ⚔️“Methodology wars.” Per-query estimates vary 10×+ depending on model size, data-center efficiency (PUE), grid mix, and whether one-time training is counted. Treat any single figure with caution.
(Google, Aug 2025)
by 2030 (IEA)
(Google, Aug 2025)
Per-query energy estimates compared
The headline figures cluster around 0.2–0.3 Wh — close to a Google web search — but an older, widely-circulated estimate of 2.9 Wh per query (from 2023–24) is roughly 10× higher. The gap is the “methodology war” in a single chart.
Sources: Google technical report (Aug 2025); Epoch AI; Sam Altman, “The Gentle Singularity” (June 2025); older estimate widely cited 2023–24. Mixed 🟢 Active
Per-query figures, sourced
| Metric | Value | Source & date | Confidence |
|---|---|---|---|
| Energy, median Gemini text prompt | 0.24 Wh | Google technical report, Aug 2025 | Vendor 🟢 |
| Water, median Gemini text prompt | 0.26 mL | Google technical report, Aug 2025 | Vendor 🟢 |
| Energy, ChatGPT query | ~0.34 Wh | Sam Altman, June 2025 | Vendor 🟢 |
| Energy, GPT-4o query (est.) | ~0.3 Wh | Epoch AI | Medium 🟢 |
| Data-center electricity, 2024 → 2030 | ~415 → ~945 TWh | IEA, 2025 | High 🟢 |
| GPT-3 training emissions (one-time) | ~552 tonnes CO₂ | Patterson et al., 2021 | Medium ⚫ |
Why the estimates disagree so much
- Model size & routing. A small distilled model costs a fraction of a frontier model; “per query” hides which model answered.
- Boundary choices. Some figures count only the GPU; others add cooling, networking, idle capacity, and data-center overhead (PUE).
- Training vs inference. Training is a large one-time cost; amortizing it across queries changes the per-query number dramatically.
- Grid & location. A query on a coal-heavy grid emits far more CO₂ than the same query on hydro — even at identical energy use.
Data lineage
FAQ
How much energy does one AI prompt use?
Current vendor and research estimates cluster around 0.2–0.3 watt-hours for a typical text prompt — Google reports 0.24 Wh, Sam Altman cited ~0.34 Wh, and Epoch AI estimates ~0.3 Wh. That’s comparable to a Google search and far below older 2.9 Wh estimates.
If each query is small, why the concern?
Scale. The IEA projects data-center electricity roughly doubling to ~945 TWh by 2030, with AI the main driver. Billions of queries plus training and idle capacity add up even when each query is cheap.