Measurement guide
What can we actually measure?
Electricity and water use are measurable. Connecting them to one AI request requires a defined system, an observation period, and an allocation method. AI Footprint Lab currently works with published evidence and information you enter; it has no connection to a provider’s meters or routing records.
| Question | What can be observed? | What else is needed? |
|---|---|---|
| How much electricity? | A device, server or facility meter records energy during a time interval. GPU telemetry covers supported hardware components. | Time, workload, idle baseline, hardware, task count and meter boundary. Shared servers require allocation; a GPU reading excludes unmetered components. |
| Which model or provider? | A local model and hardware can be specified. Providers can instrument their own serving systems. | The actual model/version, routing, batching, caching and serving location. Selecting a brand in a menu would not supply those observations. |
| How much carbon? | Electricity use can be observed; emissions factors describe electricity-generation accounting. | Matching geography and time, a stated location-based or market-based method, and separate embodied emissions. Energy × a factor is an estimate. |
| How much water? | Facilities can meter water entering and leaving. Consumption is distinct from withdrawal. | A matching time period and allocation to the AI workload. Cooling water, upstream electricity water and manufacturing water have different boundaries. |
| What happens to the watershed? | Streamflow, water levels, quality and ecological conditions can be monitored. | Location, season, source, competing demand and ecological interpretation. A global litre total cannot establish a local consequence. |
A practical classroom investigation
- Choose a local workload and record exactly which equipment the meter covers. Keep any battery charge level and background activity as consistent as possible.
- Record cumulative energy before and after a repeatable group of tasks. Use enough tasks and time to exceed the meter’s resolution.
- Repeat the run and record the spread. Measure an idle interval of the same duration separately. A baseline comparison supports interpretation; it does not automatically isolate causation.
- Use the lab below to calculate the interval energy and an allocated per-task average. Describe the result as specific to that run.
- Discuss what would change on different hardware, with longer outputs, or when using a remote server.
A student can learn from provided observations without installing a model or creating an account. The worked example below is explicitly fictional.
Work with electricity readings
Enter cumulative kWh readings from a plug-in electricity meter before and after a local workload. AI Footprint Lab calculates the difference; it does not connect to or verify the meter. Keep a separate record of the date, hardware, model/version, task, output length, meter resolution, and repeated runs.
Using a cloud chatbot? A meter on your laptop covers the laptop only. It does not measure the remote model. To observe inference locally, run a model on the metered computer.
Enter both readings, duration and completed task count to calculate. Leave the grid factor blank when it is unknown.
Readings and descriptions stay in this open page until you clear, reload, or leave it. They are not put into share URLs or the scenario calculator’s report. Download this observation record separately to preserve its inputs and limits.
Explore models and places
Compare published model energy tests or read a water-use case study from The Dalles, Oregon. The scenario calculator offers published U.S. regional electricity factors. Each keeps its own system boundary and source period.
Still outside the calculations: automatic provider routing, live facility telemetry, comprehensive watershed impacts, and compatible life-cycle inventories for manufacture, training and end of life.
Methods to consult
Hardware energy readings — NVIDIA documents power and energy telemetry for supported GPUs. This is a component boundary, not automatically a whole-system reading.
Software measurement and estimation — CodeCarbon uses supported hardware interfaces and estimation fallbacks. Inspect which method was used for each component.
Production serving and water allocation — Google’s study combines serving measurements with allocation methods and annual conversion factors. Its per-prompt water result is calculated, not a prompt-level water-meter reading.