The power bill
behind AI.
A free, neutral explainer of the electricity and infrastructure modern AI runs on: what it uses, where the constraints are, and how the power is being built.
The energy questions
What AI really costs to run
Expand any topic for a plain, factual answer, then read the deep dive. A neutral overview, not advocacy.
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.
- 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.
A neutral, factual overview for general understanding. Energy figures are often estimates with varying methodologies: verify specifics against primary sources before acting.
Power is the new constraint.
The story of AI energy is shifting from chips to electricity. For a while the scarce input was accelerators; increasingly it is megawatts: getting enough firm, affordable, ideally clean power to the right place, fast enough.
Two phases drive the footprint: training a model once (an intense burst) and serving it forever (smaller per query, but constant and at huge scale). For popular models, lifetime inference energy can exceed training.
Efficiency keeps improving per query, but total demand still climbs as adoption grows. That is why grid interconnection, cooling, and clean-power supply now shape where and how fast AI can be built.
Keep exploring
Related AI tools
Sizing AI infrastructure?
Plan compute and power together.
We help organizations plan AI infrastructure against real workloads, including the power, cooling, and siting realities, then build and run the systems on top.