Set how much you use each kind of AI and see the estimated electricity, carbon and water over a year, with everyday comparisons. Built on published 2025-26 measurements; every figure is uncertain, so ranges are shown.
Cleaner grids mean less CO₂ for the same electricity. Approximate 2025 averages; big AI companies often buy cleaner power than the local average.
Messages to ChatGPT, Claude, Gemini and similar.
A typical quick message ≈ 0.3 watt-hours, about one second of microwave use.
Pictures from tools like DALL·E, Midjourney or Stable Diffusion.
One image ≈ 1.2 watt-hours, about a phone on charge for 6 minutes. Surprisingly close to a text answer.
Clips from tools like Seedance, Veo or Kling. Video is by far the most energy-hungry AI use.
One 5-second clip ≈ 500 watt-hours, like running a microwave for half an hour. Energy grows faster than clip length.
Sessions with tools like Claude Code or Cursor agents, where the AI works through a task in many steps.
A typical session ≈ 41 watt-hours, roughly 140 quick chatbot messages.
At the usage you've set.
The calculator multiplies your usage by published per-use energy estimates, converts electricity to carbon using average grid intensity for your region, and to water using a typical data-centre figure. It covers running the AI models (inference) only; it excludes model training, hardware manufacturing and your own device's power. Last reviewed July 2026.
| Activity | Central | Range used | Basis |
|---|---|---|---|
| Chat: quick question | 0.3 | 0.15 – 1 | Google reports 0.24 Wh for a median Gemini prompt (with 0.03 g CO₂e and 0.26 ml onsite water); OpenAI reports 0.34 Wh for an average ChatGPT query; Epoch AI estimates 0.3 Wh. |
| Chat: long or complex | 2.5 | 1 – 10 | Epoch AI estimates ~2.5 Wh for a ~10,000-token prompt with typical output. |
| Chat: deep research / reasoning | 10 | 5 – 40 | Benchmarks (Jegham et al., 2025) measured 30+ Wh for long prompts to reasoning models such as o3 and DeepSeek-R1; press reports suggest GPT-5-class responses can average ~18 Wh. |
| AI image | 1.2 | 0.3 – 3 | Measured open-model studies (Hugging Face researchers; MIT Technology Review reporting) put a high-quality 1024×1024 image at roughly 0.3 to 3 Wh. |
| Video: ~5 s clip | 500 | 100 – 1,000 | MIT Technology Review reported ~0.94 kWh for a 5-second state-of-the-art clip in 2025; other 2026 estimates put a 10-second Sora clip near 1 kWh. Measured studies ("Video Killed the Energy Budget", 2025) show energy roughly quadruples when clip length doubles; smaller open models use far less. Central values here are deliberately mid-range. |
| Video: ~10 s clip | 1,000 | 300 – 3,000 | |
| Video: ~20 s clip | 3,000 | 800 – 10,000 | |
| Coding: short task | 10 | 3 – 30 | Analysis of real Claude Code session logs (Couch, 2026) found a median session of ~592,000 tokens ≈ 41 Wh, about 140× a quick chat message; a heavy multi-agent day reached ~1,300 Wh. |
| Coding: typical session | 41 | 15 – 120 | |
| Coding: long / multi-agent | 130 | 50 – 400 |
| Factor | Value |
|---|---|
| Grid carbon intensity (approx. 2025 averages) | UK ≈ 125 g; France ≈ 45 g; EU ≈ 155 g; US ≈ 350 g; India ≈ 650 g; world ≈ 420 g CO₂e per kWh (Carbon Brief; EEA; IEA Electricity 2025). Falling a few per cent per year in most regions. |
| Water | ≈ 2 litres per kWh central (0.5 – 5), covering data-centre cooling plus water consumed generating the electricity. Onsite-only figures are far lower: Google reports ~0.26 ml per prompt. |
| Everyday comparisons | Electric oven ≈ 2 kWh per hour of use; streaming video ≈ 80 Wh per hour (Carbon Trust / IEA estimates; mostly the TV itself, so it includes your device where the AI figures don't); typical UK household ≈ 2,700 kWh/year; average petrol car ≈ 170 g CO₂e/km; shower ≈ 50 litres. |
How uncertain is this? Genuinely quite uncertain. Only Google has published a detailed per-prompt methodology; most other figures are independent estimates or single company statements. Per-use numbers could be a few times higher or lower than the central values, which is why every total shows a range. Efficiency is improving rapidly (Google reports a 33× drop in energy per prompt in one year), but models are also getting bigger and usage heavier, so treat results as order-of-magnitude guidance.
What this leaves out. Training frontier models, building data centres and manufacturing chips are significant, growing energy demands at grid level even when each individual use is small. The IEA projects data-centre electricity use will roughly double from 2024 to 2030, with AI the main driver. Your per-use footprint being small does not mean AI's system-level footprint is.
Sources: