Transparency · sources
Sources & references
Every number this calculator uses, grouped by topic, with its value, plausible range, and the public source it comes from. The figures here are generated from the same data the calculator runs on, so they never drift out of sync. For how they fit together, see the methodology.
Text generation energy
Energy is modelled as a fixed per-query overhead plus a marginal cost per token, with generated (output) tokens costing more than prompt (input) tokens.
- Large model — per-query overhead0.1 Wh (range 0.05 – 0.2 Wh)
Fixed per-query cost (idle machines, host CPU/DRAM, routing) for a frontier model. Google measured the TPU-only-to-full gap as ~0.10 Wh.
Google measured inference disclosure (2025) ↗ - Large model — output (decode) token0.0006 Wh/token (range 0.0002 – 0.0015 Wh/token)
Marginal energy per generated (decode) token; dominates long answers. ~3:1 costlier than input tokens. A ~500-token query lands near 0.3 Wh, matching Google measured (0.24 Wh) and Oviedo et al., Joule 2026 (arXiv:2509.20241).
Epoch AI; confirmed by Oviedo et al., Joule 2026 (frontier median 0.31 Wh, IQR 0.16-0.60) ↗ - Large model — input (prefill) token0.0002 Wh/token (range 0.00005 – 0.0005 Wh/token)
Prefill (reading the prompt) is far cheaper per token than decode; keeps summarization from being overstated.
Prefill-vs-decode ratio (~3:1) from inference profiling ↗ - Small model — per-query overhead0.05 Wh (range 0.02 – 0.15 Wh)
Fixed per-query cost for a small/efficient model (e.g. 7-13B class).
Luccioni et al., "Power Hungry Processing"; ML.ENERGY ↗ - Small model — output (decode) token0.0002 Wh/token (range 0.00005 – 0.0008 Wh/token)
Marginal decode energy per token for a small model (~0.00003-0.00006 Wh/token measured for 8B-class).
ML.ENERGY longitudinal LLM inference measurements ↗ - Small model — input (prefill) token0.00007 Wh/token (range 0.00002 – 0.0002 Wh/token)
Prefill per-token energy for a small model.
Prefill-vs-decode ratio (~3:1) ↗
Image generation energy
Image generation is priced per image rather than per token.
- Large model — per image3 Wh/image (range 1 – 11.5 Wh/image)
Per-image generation energy. Luccioni measured mean ~2.9 Wh; the 2025 diffusion-scaling paper measured 3.58 Wh for a 20B model at 1024px (SDXL-class up to ~11.5 Wh). Varies with resolution/steps.
Luccioni et al. 2311.16863; confirmed by "Energy Scaling Laws for Diffusion Models" (2025) ↗ - Small model — per image1 Wh/image (range 0.1 – 3 Wh/image)
Per-image energy for a smaller/distilled image model (modern efficient models reach <0.5 Wh).
Luccioni et al.; "The Hidden Cost of an Image" (2025) ↗
Datacenter overhead (PUE)
Compute energy is multiplied by datacenter Power Usage Effectiveness to account for cooling and facility overhead.
- Power Usage Effectiveness1.15 × (range 1.1 – 1.6 ×)
AI/hyperscale datacenter overhead multiplier. AI runs in hyperscale facilities (Google 1.09 fleet, Q4 2025; AWS ~1.15); the broader industry average is higher (~1.54, Uptime Institute 2025).
Google 1.09 (Q4 2025) / AWS ~1.15 (hyperscale); Uptime Institute 2025 (industry avg 1.54) ↗
Water
Water combines on-site datacenter cooling (WUE) and water embedded in generating the electricity used (EWIF).
- Water intensity4 mL/Wh (range 0.5 – 9 mL/Wh)
Total water consumed per Wh = Scope 1 on-site cooling (WUE ~1.15 mL/Wh, Google measured 2024 fleet) + Scope 2 off-site electricity generation (EWIF ~3.1 mL/Wh, US avg, Li et al. arXiv:2304.03271) ≈ 4.0. Consumption basis (water evaporated/lost), not withdrawal. Li et al. remains the only source decomposing on-site + off-site water per-energy; the Scope 1 term is grounded in Google's 2024 measured WUE. Reality-checks: Google measured 0.26 mL/prompt (on-site only) and Mistral's LCA ~45 mL/response (lifecycle) bound different scopes. Low 0.5: best-case hyperscaler (Microsoft WUE 0.30, zero-water cooling) on a low-water grid. High 9.0: water-stressed regions where on-site WUE alone reaches ~9 mL/Wh (e.g. Arizona summer). The paper projects rising TOTAL water volume by 2027 from demand growth, not rising per-Wh intensity.
Scope 1: Google measured fleet WUE 2024; Scope 2: Li et al. EWIF (consumption basis) ↗
Carbon — grid intensity
Energy is converted to carbon using grid intensity. The calculator defaults to the global average and offers a US preset via the grid-region toggle.
- Global average440 g CO₂e/kWh (range 50 – 700 g CO₂e/kWh)
Global-average grid carbon intensity, centred between the two newest full-year-2025 figures (IEA 435 g CO2/kWh; Ember 458 g CO2e/kWh). The default — the US grid is cleaner than the world average. Trending down ~3%/yr.
IEA Electricity 2026 (435 g, 2025) / Ember 2026 (458 g, 2025) ↗ - United States348 g CO₂e/kWh (range 50 – 700 g CO₂e/kWh)
US-average grid carbon intensity (767 lb CO2/MWh = ~348 g CO2e/kWh, eGRID2023, ~7% below eGRID2022).
US EPA eGRID2023 (national average) ↗
Everyday equivalents
Relatable comparisons are derived from public reference figures. Each is the amount of energy, water, or carbon equal to one of the everyday thing.
- Smartphone charge19 WhEPA GHG Equivalencies — ~19 Wh of wall energy per smartphone charge ↗
- US household's daily electricity29564 WhUS avg household ~29.6 kWh/day (EIA, 10,791 kWh/yr 2022) ↗
- Google search0.3 Wh
Dated figure — see source note.
Google's 2009 estimate ~0.3 Wh/search (dated; a modern search is ~7x lower) ↗ - Cup of water240 mLUS legal cup = 240 mL (FDA, 21 CFR 101.9) ↗
- One second of a shower126 mLEPA WaterSense showerhead 2.0 gpm = 7.57 L/min = ~126 mL/s ↗
- One metre driven (petrol car)0.25 g CO₂US avg passenger car ~400 g CO2/mile = ~0.25 g/m (EPA) ↗
These are estimates with genuinely wide ranges — the published figures for AI energy use span roughly 10×. Advanced mode lets you override any assumption to match a newer source or your own situation.