AI energy · Topic
How much energy AI uses
AI's energy footprint comes from two phases: training a model once (very intensive, but one-off) and serving it to users (less per query, but constant and at huge scale). At population scale, inference increasingly dominates total energy use.
Key facts
- Training a frontier model consumes a large amount of electricity over weeks of running tens of thousands of accelerators.
- A single AI query uses a small but non-trivial amount of energy; multiplied by billions of queries, inference becomes the larger long-run draw.
- Energy use scales with model size, usage volume, and how efficiently the hardware and software are run.
- Reported figures vary widely and are often estimates: methodology matters, so treat single numbers with caution.
Takeaway
Inference at scale, not one-off training, is the structural driver of AI's growing energy demand.