What It Does

There are a lot of confident, uncited claims about how much energy AI uses. This tool replaces the hand-waving with a transparent calculation where you can see and change every assumption. Two modes:

Inference - energy, CO2, and water for a single query and at daily volume. The math is simply the GPU’s power draw times how long it spends generating your tokens (output_tokens / throughput), scaled up by datacenter overhead (PUE).
Training run - the one-off cost of training a model, using the standard 6 × params × tokens FLOP estimate divided by achievable GPU throughput to get GPU-hours, then energy and CO2.

Carbon comes from the grid’s carbon intensity (gCO2/kWh), which varies enormously by region - the same query emits ten times more CO2 in India or China than in France or Norway. Water comes from a datacenter cooling coefficient.

Relatable Equivalences

Raw watt-hours are hard to feel, so every result is also expressed as something concrete: EV miles driven, phone charges, web searches, and for training runs, US homes powered for a year. That’s what makes the number land - and what makes it shareable.

Everything Is Editable

The presets - GPU power draw, serving throughput, grid carbon intensity, PUE - are representative published figures with a verified date. None of them is hidden. Change any of them to match your setup and the whole calculation updates live.

Pairs naturally with the Reasoning Cost Calculator (dollars instead of joules), the Inference Latency Estimator (where the throughput number comes from), and the EV Charging Cost Calculator if the miles equivalence makes you curious.

Limitations

Order of magnitude, not measurement. It models GPU electricity only. It excludes networking and storage energy, the embodied carbon of manufacturing the hardware, idle capacity, and cooling beyond the PUE multiplier.
Throughput is the biggest lever and the least universal number - it depends heavily on the model, batching, and serving stack. The default is representative; set it to your measured value for a real answer.
Grid intensity is an annual average. Real marginal carbon shifts hour by hour with the generation mix.