Technology
Estimate AI inference emissions from assumptions you enter yourself — calls, energy per call, grid carbon intensity, and facility overhead (PUE). Nothing here is a measured footprint for any specific model.
Results update as you type. No calculate button. All calculations run locally in your browser.
The energy-per-call and grid-intensity values are example inputs, not measured values for any model or region. Replace them with your own measured or provider-supplied figures before using the result.
—
—
| Item | Value |
|---|
Facility overhead is the difference between total energy and inference energy. If PUE is 1.0, overhead is zero.
Bars are scaled to the larger energy figure. They show how much PUE adds, not a comparison with any other system.
| Scenario | Value |
|---|
Each row reuses your energy-per-call, PUE, and grid inputs and varies only the call count, so the figures scale linearly.
Energy (kWh) = calls × Wh per call × PUE ÷ 1,000. Emissions (kg CO₂e) = energy × grid grams per kWh ÷ 1,000. Per-call emissions (g CO₂e) = Wh per call × PUE × grid grams per kWh ÷ 1,000.
Calls 100,000, energy 0.5 Wh per call, PUE 1.2, grid 450 gCO₂e/kWh. Energy = 100,000 × 0.5 × 1.2 ÷ 1,000 = 60 kWh. Emissions = 60 × 450 ÷ 1,000 = 27 kg CO₂e. Per call: 0.5 × 1.2 × 450 ÷ 1,000 = 0.27 g CO₂e; per 1,000 calls: 270 g CO₂e. These inputs are examples only — replace the energy-per-call and grid figures with your own before relying on the result.
This is a scenario model, not a carbon accounting result. It covers only the inference energy you describe: training, model development, hardware manufacturing, networking, cooling beyond PUE, and embodied emissions are excluded. Energy per call varies by model, hardware, batching, and prompt length, and grid intensity varies by place and time — both are user inputs here precisely because no single published value fits every case. No live emissions data or certified accounting is provided.
Energy is calls times Wh per call times PUE, divided by 1,000 to get kWh; emissions are energy times grid grams per kWh, divided by 1,000 to get kilograms of CO₂e.
Because they are assumptions, not constants. A different model, GPU, or region gives different values, so the page treats them as inputs you must supply and labels the defaults as examples only.
Power usage effectiveness: total facility energy divided by IT energy. A PUE of 1.2 means 20% overhead for cooling and other facility loads on top of inference energy; 1.0 means no overhead is counted.
No. It covers only the inference calls in your scenario. Training, fine-tuning, and hardware manufacturing are separate and much larger in some cases, and are not modeled here.
Each row in the volume table reuses the same per-call figure, so doubling calls doubles the estimate. Real systems may batch or cache, which would break that linearity — another reason to replace the example per-call figure with measured data.
No. It is a planning estimate from your own assumptions. Formal reporting needs measured energy data, a documented grid factor for the right region and period, and the reporting framework your organisation follows.
Results are estimates only and do not constitute professional advice.