Training or use: where does AI's energy really go?
60%: that's the share of Google's AI energy that went into running models day to day (inference) over several recent years, against 40% for training them. Intuition usually says the opposite: that training a model is the big-ticket item. That's true for a moment, at the scale of a single model. It stops being true the moment that model ships and gets used millions of times.
At Google, over several consecutive years, 60% of AI energy went into running models day to day (inference), against 40% into training them. Training a large model is dramatic and concentrated (up to 1,287 MWh for GPT-3 in one run), but it's the billions of queries that follow, added up over months or years, that end up weighing more.
Figures updated on June 30, 2026
How much energy does training a large model use?
1,287 megawatt-hours (MWh) of electricity and 552 tonnes of CO2e: that's what training GPT-3 cost in 2020, in a single run, according to the reference study by David Patterson and coauthors (Google, 2021). For a newer, bigger model the order of magnitude climbs further: training Mistral Large 2 emitted 20,400 tonnes of CO2e and used 281,000 cubic meters of water for cooling, according to the life-cycle assessment Mistral AI published with ADEME and Carbone 4 in 2025, the first of its kind made public and peer-reviewed for a large language model.
And day-to-day use, inference, how much does that weigh?
0.24 Wh: that's the energy used by a median text prompt sent to Gemini, according to the environmental report Google published in 2025 (0.03 g of CO2e, 0.26 mL of water), a measurement that covers inference only, not training. At Mistral, a 400-token exchange (prompt plus reply) costs 1.14 g of CO2e and 45 mL of water. Taken alone, a single prompt is negligible next to the hundreds of tonnes of CO2e from training. It's multiplying that changes the picture.
So which weighs more in the end, training or use?
60% of the AI energy used in Google's data centers, each year over 2019 to 2021, went to inference, against 40% to training, according to the same research team (Patterson et al., 2022). The logic tracks: a model trains once, but can be queried billions of times over months or years once it ships.
How many prompts does it take for inference to overtake training's cost?
About 17.9 billion prompts (400-token exchanges): that's what it would take, using Mistral's own published life-cycle figures, for inference's cumulative CO2e to match the 20,400 tonnes emitted by training Mistral Large 2. That number is illustrative, we calculated it ourselves from two figures Mistral published, but it shows the scale: a widely-adopted model can clear that bar within months. ChatGPT alone gets about 18 billion messages a week (NBER/OpenAI, 2025): at that scale, training quickly becomes a minor share of the total.
Could this flip as models keep getting bigger?
+30% a year: that's the projected growth in electricity use by servers dedicated to AI, driven mainly by inference, according to the International Energy Agency (IEA, Energy and AI report). Training compute for frontier models is also growing fast, about 4 to 5 times a year per Epoch AI, but it stays concentrated in time. At Amazon, as early as 2019, inference already made up 80 to 90% of compute costs for machine learning systems at scale, a structural imbalance that has only grown since.
Related equivalents
Sources
- Patterson et al. (Google) : répartition entraînement/inférence de l'énergie IA, 2019-2021 (arXiv 2204.05149)
- Patterson et al. : émissions carbone des grands réseaux de neurones, GPT-3 (arXiv 2104.10350)
- Mistral AI : analyse de cycle de vie de Le Chat, avec l'ADEME (juillet 2025)
- IEA : Energy and AI
- Google Cloud : impact environnemental de l'inférence IA
- NBER / OpenAI : « How People Use ChatGPT » (Chatterji et al., 2025) : 700 M d'utilisateurs hebdomadaires, ~18 Md de messages/semaine, part professionnelle 47 % → 27 % (juin 2024 → juin 2025)
Frequently asked
Does a single AI prompt use more energy than training the model?+
No, never on its own: a prompt costs a fraction of a gram of CO2e (0.03 to 1.14 g across the studies here), against hundreds or tens of thousands of tonnes for a model's full training run. It's only the sum of billions of prompts that eventually overtakes that total.
Why do people often say training is the bigger polluter?+
Because training is concentrated into a few weeks and disclosed all at once (one study, one striking number), while inference is diffuse: billions of tiny requests spread over months, rarely added up in public discussion.
Does the 60/40 training-inference split hold for every AI model?+
No, it's the average Google measured on its own infrastructure between 2019 and 2021: it shifts with a model's popularity, lifespan and provider. OpenAI, Anthropic and Meta haven't published an equally detailed public breakdown; Mistral is one of the few to publish a full assessment (training and inference) for a single model.