AI Energy Use
What AI costs in electricity — a real and growing number, surrounded by the most consistently misreported figures in the field.
Reviewed July 15, 2026Stable
When not to use it
- As a reason not to use AI for a task it's good at. The comparison that matters is against the alternative, which also has a footprint.
- With a number you can't source to a measurement. Almost every circulating figure for frontier models is extrapolation from guesses.
- Focusing on training when you have users. Inference dominates the lifetime footprint of anything with real traffic.
Reach for something else instead
- Smaller models cut energy and cost together, which is why it's the lever that actually gets pulled.
- Region selection is the biggest single factor and it's a dropdown — 5–10× on carbon per the measurement.
- Caching and batching reduce the multiplier on the part that dominates.
Read more on the blog
- How much energy does AI use? Training vs inferenceThe energy cost of AI is discussed constantly and measured badly. The per-query figures in circulation are stale and vary by an order of magnitude, the balance has shifted from training to inference, and the constraint that actually bites is not generating electricity but delivering it to a particular building.
- The water bottle was per 10 to 50 responsesThe most repeated environmental claim about AI dropped a qualifier from the study it cites. The concern it created is still justified, for different reasons than the number suggested.
- The binding constraint is a transformer, not a chipCapital is available and chips are shipping. The thing stopping data centres from opening is a waiting list held by utilities, and a piece of equipment on a five-year lead time.
Further reading
- Strubell et al. (2019), Energy and Policy Considerations for Deep Learning in NLP — the paper that raised the issue; the famous figure describes an architecture search, not a normal training run.
- Patterson et al. (2021), Carbon Emissions and Large Neural Network Training — recalculation with real datacentre data; published estimates overstated by up to 88×, and the four-Ms framework.
- Luccioni et al. (2023), Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model — the most complete public lifecycle accounting, including embodied and inference emissions.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Repeating the five-cars figure. It described an exhaustive architecture search, and a later recalculation with real datacentre data found such estimates overstated by up to 88×.
- Optimising the model when the grid is the variable. Datacentre location and hardware dominate the model architecture by a wide margin.
- Quoting per-query numbers for closed models. Nobody outside the lab can measure them, so the figure is someone's extrapolation and usually someone with a position.
At a glance
FieldSafety & Ethics
Dominates lifetime footprintinference, not training
Biggest levergrid region (5–10×)
Most-cited figureoverstated up to 88×
DifficultyBeginner
Where this sits
A destination. 5 concepts lead here, and nothing in the corpus depends on it.
3Levelsteps in
5Needs firstconcepts
0Opens upnothing further
1Areastays here
Computed from the prerequisite graph, not assigned. How this works