How much electricity does a prompt use? What about water? And what happens when a data centre moves in down the road?
Those questions belong at different scales. Keeping them separate makes it easier to understand the costs, challenge a headline and decide where our effort counts.
These notes follow the slide deck prepared by Nadia Ellis for 12 November 2026.
Start with three scales
- Your prompt
- The resources used for one task. A short text answer, an image and a long video involve different amounts of work. The model, answer length and task all affect the result.
- The planet
- The combined demand from data centres, the electricity that supplies them and the equipment they rely on. Check whether a figure covers all data centres or the portion attributed to AI.
- One place
- The effects on a particular grid, catchment and community. Local power supply, water availability, noise, jobs and consent conditions belong in this conversation.
Nobody lives in the global average. A small share of world electricity can still be a substantial demand in one place.
A prompt is not a standard unit
The deck compares estimates for text, images and video. Treat each figure as an estimate for a particular task, model and set of conditions. Company-reported figures also need to be labelled as such.
A comparison with boiling the jug can help us picture an unfamiliar unit. It needs its assumptions alongside it: how much water, which appliance and what is included in the AI estimate.
Efficiency per task and total demand are separate questions. More efficient tools can be used for more tasks. We need to look at both.
Water headlines need a boundary
Before sharing a large water number, ask what it counts:
- On-site cooling: water used at the data centre itself.
- Electricity generation: water associated with producing the power it uses.
- Equipment: water involved in making chips and other hardware.
Then check the year, location and method. Water withdrawn and water consumed measure different things. A consented maximum is a ceiling on permitted use, rather than a measurement of daily use.
The hairdryer example in the slides is a constructed scenario. Its assumptions are part of the example, so keep them attached to the headline.
Watch: data centre water, what gets measured
Follow the water through cooling, electricity generation and hardware manufacturing. Press Play to watch, or use Back and Next to move through the seven steps.
The local deal deserves scrutiny
The deck puts Southland alongside examples from Memphis, Santiago, Hamina and Falun. The questions to bring to any proposal are practical: where will the power come from, what happens to the water, can heat be reused, and can the facility reduce demand when the grid is under pressure?
A fair local deal answers four questions:
- Who decides? Do mana whenua and neighbours receive information early enough to influence the decision?
- Who benefits? Which jobs, services and financial benefits are demonstrated, and which remain promises?
- At what cost? What happens to water, power, noise and landscape? Who pays for the clean-up if the site closes?
- Who checks? What gets metered and reported? Which commitments become enforceable conditions?
Judge each use, and keep two ledgers
The session also looks at AI applications in contrail forecasting, leak detection, irrigation, energy systems and scientific research. For each example, ask what has been measured, who measured it and whether anyone is using it at scale.
One ledger records environmental harm reduced. The other records value we may judge worth the resources used, such as medical research. A medical breakthrough does not take a tonne of carbon dioxide out of the air.
Judge each application on its own evidence. A successful use elsewhere does not settle the case for a particular project.
Where our effort counts
- Match the tool to the task. Think about what you need, especially before generating repeated images or long videos.
- Ask what a number includes. Which task, which water, which year and compared with what?
- Ask suppliers for evidence. Look for measured use, clear reporting boundaries and independently checked results.
- Stay involved locally. Ask the people building and approving infrastructure who benefits, who bears the cost and who checks the promises.
If it is coming here, let’s be the place that does it well.