I build AI products and I want to work on climate. That's an uncomfortable combination if you don't look at the numbers honestly. So here's how I think about AI's energy use, without the hype in either direction.
The big picture
AI runs in data centres, and data centre electricity demand is growing quickly, driven heavily by AI. The International Energy Agency's 2025 Energy and AI report estimated data centres used around 1.5% of global electricity in 2024 and projected that demand could roughly double by 2030.
At the global level, that's still a small share. Locally it can be huge. A cluster of data centres can strain a regional grid, raise prices, and delay coal plant retirements.
Training vs inference
- Training a frontier model uses a lot of energy once.
- Inference, answering queries, uses less per call but happens constantly.
For popular models, inference dominates over time. That means product decisions (how often you call a model, which model, how long the prompts are) matter more than most developers realise.
Per-query numbers: handle with care
You'll see claims like "one AI query uses X times a Google search." Published estimates vary a lot, depending on:
- Model size and architecture
- Hardware generation
- Prompt and output length
- Data centre efficiency and cooling
- Whether the estimate includes idle capacity and networking
I don't quote a single number without its assumptions, and you shouldn't either.
Beyond electricity: water and hardware
- Water is used for cooling, sometimes in water-stressed regions.
- Embodied emissions come from manufacturing GPUs and building data centres.
- Grid mix matters most. The same query on a hydro-powered grid and a coal-powered grid has very different emissions.
The benefit side
AI can reduce emissions or harm elsewhere:
- Grid balancing and renewable forecasting
- Building and industrial energy optimisation
- Flood early warning and climate risk
- Deforestation monitoring
The honest framing: AI is a tool with a footprint. Whether it's worth it depends on what it does.
What developers can do
- Use the smallest model that works. Route easy tasks to small models.
- Cache. Don't recompute identical answers. Prompt caching helps too.
- Cut tokens. Shorter prompts, shorter outputs where possible.
- Batch non-urgent work.
- Choose regions with cleaner grids when you can.
- Measure. Tools like CodeCarbon estimate emissions for your own workloads.
My Electricity Cost calculator won't measure a data centre, but it's handy for estimating your own GPU workstation's energy cost.