Which uses more electricity: training AI, or people using it?
Using it, in the little data there is. Training is a big one-off; use (called inference) repeats with every question. Google said machine learning was 10% to 15% of its energy, about three fifths of it for use and two fifths for training. AWS estimated inference at 80 to 90% of machine-learning cloud demand. Researchers measuring open models found use overtakes training after hundreds of millions to tens of billions of answers: 200 to 590 million for Hugging Face’s BLOOMz models.
The bigger pictureUse adds up fast: ChatGPT alone got about 2.6 billion messages a day by July 2025, by Hum’s sums from OpenAI’s count. All the world’s data centres together used about 485 TWh in 2025.
Why the claim misleads
Training is paid once per model, though each new model is trained again. Use keeps going for as long as people keep asking.
What nobody knows yet
- No company has published a training versus use split since chatbots took off; Google’s is from before 2022.
- OpenAI, Google and Anthropic don’t publish how much energy training their newest models took.
- Splits are counted different ways: Meta put about one-third of its machine-learning carbon on use in a 2021 paper, counting data and storage alongside training.
Sources (7)
- Google Research blog (David Patterson, February 15, 2022; Google's own machine learning, measured by Google) Company says so
“ML training and inference are only 10%–15% of Google’s total energy use for each of the last three years, each year split ⅗ for inference and ⅖ for training”
Quote checked on the page, 25 Oct 2026 - Luccioni, Jernite & Strubell, Power Hungry Processing (FAccT 2024), citing AWS Company says so
“According to AWS, the largest global cloud provider, inference is estimated to make up 80 to 90% of total ML cloud computing demand”
Quote checked on the page, 25 Oct 2026 - Luccioni, Jernite & Strubell, Power Hungry Processing (FAccT 2024), citing Meta Company says so
“a 2021 publication by Meta attributed approximately one-third of their internal end-to-end ML carbon footprint to model inference, with the remainder produced by data management, storage, and training”
Quote checked on the page, 25 Oct 2026 - Luccioni, Jernite & Strubell, Power Hungry Processing (FAccT 2024; Hugging Face, open BLOOMz models, measured) Measured
“from around 200 million inferences for the smallest model, BLOOMz-560M, to over 590 million inferences for the biggest model, BLOOMz-7B.”
Quote checked on the page, 25 Oct 2026 - Morrison et al., Allen Institute for AI (ICLR 2025; open OLMo models, measured) Measured
“We find that for most models tested, the number of inferences required to outweigh training costs is in the hundreds of millions to tens of billions, except for the most over-trained models.”
Quote checked on the page, 25 Oct 2026 - NBER Working Paper 34255 (OpenAI and Harvard), How People Use ChatGPT Measured
“By July 2025, 18 billion messages were being sent each week by 700 million users, representing around 10% of the global adult population.”
Quote checked on the page, 18 Oct 2026 - IEA, Key Questions on Energy and AI (2026) Estimated
“electricity consumption from data centres roughly doubling from 485 TWh in 2025 to 950 TWh in 2030”
Quote checked on the page, 3 Oct 2026
Hum’s sums
- 18 billion messages a week × 52 ÷ 365 ≈ 2.6 billion a day
Cite this
Google Research blog (David Patterson, February 15, 2022; Google's own machine learning, measured by Google), as quoted in Hum, “Which uses more electricity: training AI, or people using it?”, https://humearth.org/q/training-vs-using (quote checked 25 Oct 2026).
Cite the original source first; Hum is where you found it.
What do Wh, kWh and TWh mean?
Wh is a watt-hour: a 1-watt light left on for an hour. 1 kWh = 1,000 Wh, and 1 TWh = 1 billion kWh. Hum compares each number with something everyday so you don’t need the units.
Send this answer
Where it’s been measured, using AI takes more energy than training it: about three fifths of Google’s machine-learning energy went on use (Google, 2022).