what does a prompt cost the planet?
every token is a few joules on a GPU somewhere. add the chips it runs on, the months of training before it, the cooling around it — then multiply by how much you actually use it. here's an honest order-of-magnitude calculator, from sand to sentence.
since you opened this page, your selected usage would have emitted
0 mg
cached context (cheap, but there's a lot of it)
fresh input
output & thinking (expensive)
puffs = CO₂, sized by grams
the calculator
pick a use case, then tweak. every number is per request, scaled by how many you send. heavy agent use is dominated by context: each tool call re-reads the whole conversation.
advanced assumptions
±3× is the honest uncertainty: batching, hardware, model size are all secret.
Epoch AI: OpenAI's 2024 final training runs were ~10 % of its R&D compute — so up to ~×10.
Epoch: 1–10 % in early 2025; xAI ran Grok 4's RL "at pretraining scale" (100 %).
training gets split across every token ever served. popular model → small share each.
–
–
electricity
same usage, different power plug
the single biggest knob isn't the model — it's the grid. the electricity part of the exact same tokens is ~15× dirtier on a coal-heavy grid than on a nuclear/hydro one. the grey part doesn't move: chips and training were made wherever they were made.
use cases, side by side
a year of each preset. a chatbot user is a rounding error; an always-on coding agent is a long-haul flight. log scale — each gridline is 10×.
from sand to sentence
the six things your tokens carry a slice of. colours match the bar above.
zoom out: the whole industry
all data centers (not just AI) used ~415 TWh in 2024, about 1.5 % of world electricity, and the IEA projects roughly a doubling by 2030 — AI is the main driver. here is how that stacks against other things, in Mt CO₂ per year.
what to take home
- one chat message is tiny — about the energy of a few seconds of a microwave. guilt over single prompts is misplaced.
- agents are not chat. an agent re-reads a 50–150k-token context on every tool call, hundreds of times a day. that's 1,000–10,000× a casual user.
- the grid beats the model. where the GPU sits changes the answer more than which model you pick.
- training is big, but shared. a frontier run is tens of thousands of tonnes, yet split over trillions of tokens it's often a minority of your slice.
- diet and flights still dominate. even heavy agent use lands well below one transatlantic flight or a year of eating beef weekly.
- everything here is ±3×. labs don't publish per-token energy. treat numbers as orders of magnitude, not invoices.
sources & assumptions
| inference anchors | Google 2025: median Gemini text prompt 0.24 Wh, 0.03 gCO₂e (market-based), 0.26 mL water, incl. idle machines + overhead; at Google's location-based 345 g/kWh that would be ≈0.08 g (our arithmetic). Epoch AI 2025 (independent estimate): ~0.3 Wh per GPT-4o query (~500 output tokens), ~2.5 Wh at 10k input tokens, ~40 Wh at 100k. Altman 2025: 0.34 Wh average (no method given). Luccioni et al. 2023 (measured, open models): text generation ≈0.047 kWh per 1,000 queries; image generation ~60× more. |
| per-token energy | class defaults (IT energy per 1k output tokens): small 0.12 Wh, mid 0.5 Wh, frontier 1.5 Wh — chosen so a ~500-token mid-size reply ≈0.3 Wh with overhead, matching Epoch / Google / OpenAI. input tokens weighted 0.25× an output token, cached 0.025× (roughly API price ratios, which track compute; Epoch's long-input figures suggest input could be heavier). thinking tokens count as output. |
| embodied | NVIDIA HGX H100 PCF 2025 (ISO 14067, reviewed by WSP): 1,312 kgCO₂e cradle-to-gate per 8-GPU baseboard (≈164 kg/GPU), 42 % of it memory. Luccioni et al. 2023, BLOOM LCA: embodied = 22 % and idle = 29 % of 50.5 t total. here: ~400 kg per GPU slot incl. host + building share, 5-year life at 60 % use → ≈15 g per kWh of IT energy. |
| training | GPT-3: 1,287 MWh, 552 t (Patterson et al. 2021). Llama 3.1 405B: 30.84 M H100-hours, 8,930 t location-based (Meta model card). Mistral Large 2: 20.4 ktCO₂e, 281,000 m³ water for training plus 18 months of use; 1.14 g per 400-token reply (Mistral / Carbone 4 / ADEME 2025). post-training & R&D: xAI Grok 4, Epoch, OpenAI compute spend. defaults: pretraining small 2 / mid 20 / frontier 80 GWh (our guess), × (1 + post-training share) × R&D multiplier, split over lifetime tokens (1e15 / 5e14 / 3e14 output-equivalent, our guess). |
| grids | lifecycle gCO₂e/kWh, 2024, Ember via Our World in Data: world 470, US 385, EU 210, Germany 335, China 560, France 40, Sweden 35. "market-based" 125 = Google's 0.03 g ÷ 0.24 Wh. training assumed on the US grid. |
| equivalents | car: average petrol 0.16 kg/km tailpipe + 0.05 fuel supply = 0.21 kg/km; flight: long-haul economy with radiative forcing 0.20 kg/passenger-km — both UK DESNZ 2024 factors; Paris CDG → JFK 5,833 km great-circle. beef (beef herd): median ≈60, mean ≈100 kgCO₂e/kg (Poore & Nemecek 2018, OWID) → 100 g ≈ 6 kg. streaming ≈0.08 kWh/h (IEA 2020). French footprint 8.5 t/person in 2023 incl. imports (SDES). trees: the common "22 kg/yr" has no study behind it; ~10–40 kg/yr depending on species & age, ~10 kg for young planted trees (One Tree Planted). kettle (1 L, 15→100 °C) ≈0.11 kWh and phone charge ≈15 Wh from the wall: physics, not a source. |
| world | IEA Energy & AI 2025: data centers 415 TWh in 2024 (~1.5 % of world electricity), ~945 TWh by 2030; ~180 Mt CO₂ today, ~300 Mt by 2035 (base case, up to 500 Mt). aviation ≈950 Mt CO₂ in 2023 (ICCT 2025). cement ≈2.4 Gt (WEF / IEA data). road transport ≈6 Gt (derived from IEA: cars + vans 3.8 Gt, >60 % of road). France territorial 376 MtCO₂e in 2023 (Citepa Secten 2025). |
| caveats | Google, OpenAI, Mistral, Meta and NVIDIA figures are company disclosures; Epoch and Luccioni are the independent anchors. water, mining, local grid strain and rebound effects are left out. the "heavy Claude Code" preset estimates a researcher running agents all day (~800 API calls, ~60k-token context, 93 % cached) — not a measurement. |