What does your AI really cost the planet?
Pick a model, a workload and your volume. We estimate the electricity, water and carbon of running it, as an honest range, because the real answer swings up to 25× per query. Then neutralize it in one API call.
We build energy from the tokens up
Most calculators multiply one “0.3 Wh per query” constant, which breaks the moment a model reasons. We model the tokens, sample the uncertain coefficients, and report an honest range.
Questions, answered honestly
A median text chat with a frontier model uses roughly 0.2-0.4 Wh, about a few seconds of an LED bulb. But the real answer swings up to ~25× per query depending on output length, model size, batching and hardware, and a reasoning model can use 10-50× more. That is why we lead with a P5-P95 range, not a single figure.
Energy scales with the tokens a model generates, one at a time in the decode phase. “Thinking” models emit far more of them, some use tens of times more tokens to answer the same question, so they can cost 10-50× more per task even on identical hardware.
Input length, output length, batch size, GPU type, utilisation, the serving engine, and the data-center overhead all swing the number. Measured energy for one model class spans up to ~25× across tasks, which is exactly why we lead with a range.
Less accurate, and we say so. OpenAI, Anthropic and Google don’t publish the parameter counts or throughput the model needs, so those figures are inferred from comparable open-weight models and flagged “estimated” with a wider band. Google’s single disclosed per-prompt figure is the one closed-model anchor we can calibrate against.
No, this is inference only, which is what scales with your usage. Training is a large one-time cost amortised across billions of queries; for high-volume inference it adds a small per-query amount we may fold in later.
Reduce first (shorter outputs, caching, smaller models, cleaner regions), then neutralize the rest with high-integrity credits, ideally durable carbon removal. That’s exactly what the 1clickimpact API automates: one call captures the carbon (or plants the trees) for a given amount, so you can wire offsetting straight into your product per transaction.
Why the energy footprint of AI matters
Every AI query draws real electricity, evaporates real water in data-center cooling, and, depending on the grid, emits real CO₂. A single answer is tiny, but multiply it across the billions of prompts sent each day and the footprint of AI inference becomes a serious line item in the world’s energy budget.
The catch is that the true cost per query is genuinely uncertain: it swings up to 25× depending on how many tokens the model generates, how it is batched, which accelerator it runs on, and how clean the local grid is. A reasoning model can use 10-50× the energy of a plain chat reply. That is why this calculator leads with a P5-median-P95 range instead of a single false-precise number.
Measuring is the first step; the point is to act on it. Reduce where you can (shorter outputs, caching, smaller models, cleaner regions), then neutralize the remainder with high-integrity carbon removal. The 1clickimpact API captures the carbon or plants the trees for a given amount in a single call, so offsetting can be wired straight into your product, per transaction.