Count the whole lifecycle, not one convenient number.

Change workload, scale, hardware, location, utilization, and useful output. Every estimate shows its assumptions and uncertainty.

Electricity12.4 MWh

Operational carbon4.2 tCO₂e

Cooling water18.6 kL

Per useful task12 Wh

Training, inference, chip fabrication, cooling, and useful output belong in the same conversation. Precision without matched boundaries is only false confidence.

Compare ways to complete the task

Approach	Energy/task	Privacy	Best when

Transparent formulas

energy = chips × power × hours × utilization × PUE

carbon = kWh × grid intensity

per useful task = total ÷ successful outputs

Embodied hardware is allocated by the share of lifetime compute used in this scenario.

What drives the answer?

Should this task use a model?

Rules are stable and repeatable Data is highly private Open-ended synthesis is needed Must work offline

Start with a smaller model or deterministic program, then measure whether complexity adds useful outcomes.

Recommendations are engineering prompts, not moral verdicts. Social benefit and harm require evidence outside this calculator.

Uncertainty range

Current high estimate is 2.9× the low estimate. Grid intensity and utilization dominate this scenario.

Knowledge check

Why can per-request comparisons mislead?

Impact scenarios: multipliers, useful-output denominator and boundaries

Your browser does not support video.

Read the explanation

This impact calculator multiplies eight accelerators by assumed point forty-two kilowatts, seven hundred twenty hours, fifty-five percent utilization and PUE one point twenty-five. It then applies a chat workload multiplier point seven, a central uncertainty multiplier one and a scale factor point seventy-three, yielding eight hundred forty-nine point eight nine five two kilowatt-hours. Changing only the workload to local uses point zero eight instead of point seven and gives ninety-seven point one three zero eight eight. Bars use point six pixels per simulated kilowatt-hour. These multipliers are authored assumptions, not measurements of a model, real hardware, a data center or an on-device deployment. The entered grid intensity is three hundred forty grams per kilowatt-hour, so the chat scenario gives about two hundred eighty-eight point nine six four kilograms of carbon-equivalent. A fixed water coefficient of one point five liters per kilowatt-hour gives about twelve hundred seventy-four point eight four three liters. The bars explicitly use different kilogram and liter scales. The displayed labels round or switch units at thresholds; they are not additional precision or measurement. Useful tasks are defined as hours times sixty times useful-output percentage, thirty thousand two hundred forty at the default seventy percent. Dividing energy in watt-hours by that denominator gives about twenty-eight point one watt-hours per assumed useful task. Unchecking Inference for an inference workload multiplies simulated energy by point two; the local scenario drops from about ninety-seven point one three to nineteen point four three kilowatt-hours. Unchecking Training affects only training through a point twenty-five multiplier. Other lifecycle flags, including fabrication, server build and cooling, change display labels but do not add or remove modeled energy terms. Bars compare the two local outputs at five pixels per kilowatt-hour. The sensitivity numbers are fixed labels rather than computed derivatives. The recommendation rule also checks creative, repeatable and privacy flags in a fixed order. This program cannot demonstrate a complete lifecycle inventory or compare real tasks with matched quality.