AI's electricity bill is a design problem
Most of the energy behind an AI product is decided by choices your own team makes.
In April, the International Energy Agency reported that data centre electricity use rose by about 17% in 2025. It expects data centre consumption to roughly double by 2030, with AI-focused facilities growing fastest. Its outlook a year earlier put that 2030 figure at more than 945 terawatt-hours a year.
Those numbers are usually discussed as a grid problem or a climate problem. For a company deploying AI, they are also a design problem. Every request has an energy cost and a price, and both depend mostly on choices your own team makes.
Where the waste comes from
- Using a very large general model for a narrow task a small model handles just as well.
- Sending the same long context with every request instead of caching it.
- Generating three drafts when the workflow only ever uses one.
- Running overnight batch jobs on expensive real-time infrastructure.
None of these appears on a bill as "energy". They show up as a cloud bill that grows faster than usage.
What efficiency-first looks like
The IEA makes a point that is easy to miss: there are energy efficiency standards for equipment, but none for AI models. Until there are, efficiency is up to the people building the system.
In practice, that means choosing the smallest model that meets the quality bar, fine-tuning it on the task instead of padding prompts with instructions, caching aggressively, and running work that isn't urgent in batches, in cloud regions with cleaner grids where you can.
It also means measuring. If you can't say what a thousand requests cost you in money, you can't say what they cost in energy either.
The good news
Efficient AI and cheap AI are mostly the same thing. The work that lowers your inference bill also lowers the energy behind it, which makes this one of the few sustainability decisions that pays for itself.