Glossary
The key terms to understand the numbers on this site, and AI in general.
Understanding AI
- Generative AI
- A family of AI that produces content (text, image, audio, video) from a prompt. Large language models (ChatGPT, Claude, Gemini) are its most common form.
- LLM (large language model)
- A neural network trained on massive text corpora to predict the next token. It powers chatbots; its footprint mostly comes from inference, on every prompt.
- Prompt
- A request sent to an AI model. Site reference: an average text prompt ≈ 0.3 Wh, 0.3 mL of water and 0.2 g of CO₂.
- Token
- The smallest unit of text a model handles (≈ ¾ of a word in English, often less in French). Models read, generate and “bill” by the token: the more tokens, the more inference consumes.
- Inference
- The stage of using an already-trained model: each prompt triggers one inference. That's what we quantify here, not training.
- Training
- The model's initial learning phase, very energy-intensive, but a one-off, then amortized over billions of requests. The site quantifies inference, not training.
- Hallucination
- When a model generates false but plausible information. It's the main reason human verification remains essential.
- Context window
- How much text a model can take into account at once (a conversation's “memory”). The fuller it is, the costlier the inference.
- Parameters
- The internal “settings” learned during training, counted in billions. As a rule, the more parameters, the more capable a model is… and the more power-hungry.
- Multimodal
- A model able to handle several data types (text, image, audio, video) within a single request.
- AI agent
- An AI that chains several steps autonomously to complete a task (so, several inferences). Useful, but it multiplies consumption compared with a single prompt.
- Reasoning model
- A model that “thinks” (generates intermediate steps) before answering. Far costlier than a text prompt: ~15 to 40 Wh per request, i.e. 50 to 100× a standard prompt.
- Image / video generation
- Producing media costs far more than text: an image ≈ 6–12 Wh (20–40× a text prompt), a video ≈ 1,000× (~940 Wh for a ~5 s clip).
Energy
- Watt-hour (Wh)
- A unit of energy: running a 1 W device for 1 h. An LED bulb uses ~10 Wh per hour.
- Kilowatt-hour (kWh)
- 1,000 Wh, the unit on your electricity bill. A space heater uses ~2 kWh in an hour.
- Data center
- The building that houses the servers where AI runs. It uses electricity to compute and water to cool down.
- GPU
- The graphics processor: the component that actually runs AI models. Powerful and power-hungry, it dominates consumption at inference.
- PUE (Power Usage Effectiveness)
- The ratio of a data center's total energy to the energy actually useful to the servers. ~1.1 (very efficient) to 1.5+: the rest goes mostly to cooling.
- TWh (terawatt-hour)
- 1 billion kWh. It's the scale of national or global consumption, where AI “weighs” at a global level.
Water
- Virtual water
- The water used to produce a good (farming, livestock, manufacturing), not just tap water. A burger thus “contains” thousands of liters of virtual water.
- Direct water (cooling)
- The water evaporated on site to cool a data center's servers. It's the “official” figure reported by Google (~0.26 mL/prompt) or OpenAI (~0.32 mL).
- WUE (Water Usage Effectiveness)
- Liters of water used per kWh consumed by a data center. Used to estimate a prompt's “direct” water.
- Water stress
- When water demand exceeds the locally available resource. A key issue around some data centers built in dry regions.
Climate & CO₂
- CO₂e (CO₂ equivalent)
- All greenhouse gases expressed as their CO₂ impact, so they can be added up.
- Radiative forcing
- The extra warming caused by flights at altitude (contrails, other gases). DEFRA and ADEME include it (×~1.9); ICAO counts CO₂ only, hence flight figures that vary twofold.
- Grid carbon intensity
- The grams of CO₂ emitted per kWh of electricity. Highly variable by country (low with nuclear/renewables, high with coal); ~480 g/kWh as a global average.
- Life-cycle assessment (LCA)
- A method that counts “cradle-to-grave” impact: manufacturing, use and end of life. It explains why an object's footprint exceeds its use-phase consumption alone.
- Scopes 1 / 2 / 3
- The emission perimeters: direct (1), from purchased electricity (2), and everything else in the value chain (3).
Reading the numbers
- Order of magnitude
- A number's approximate “size” (×10, ×100, ×1,000). On uncertain data, it's more honest and more useful than false precision.
- Median & range
- When sources diverge, we keep a credible central value (median) and show the spread (range), never a single, falsely precise figure.
A term missing? Every value cited is sourced on the Sources page.