Agile InsightsBriefing
A server unit, a glass of water by a steel tap, and a balance with two empty pans on a dark workbench, labelled Power, Water and The Trade-off.

Episode 5 · Primer · About 6 minutes

Are data centres evil?

AI's power, water and carbon footprint is a trade-off an executive can manage, not a verdict.

#leadership

Download the MP3 · Linda, an AI voice, reads a script written from the sources below.

In this episode

01How big the electricity demand isThe International Energy Agency estimates data centres used about one and a half per cent of the world's electricity in 2024, and expects that to more than double by 2030, driven mainly by AI. Locally it can be very large.
02Training is large, each request is tinyTraining is a big one off cost, while Google reports its median text prompt uses about 0.24 watt hours. That is one company's method, for text only. Images, video and agents cost more, and millions of requests add up.
03Why water and carbon figures need careOperators rarely disclose enough about water for independent checking, so treat any single number with care. On carbon, matching annual electricity use with renewable purchases is not the same as running on clean power every hour.
04The Australian pictureThe market operator estimates data centres used 2.8 per cent of the main grid in 2025 to 26, and around thirteen per cent a decade later in its central scenario. Government expectations are policy, not law.
05What an organisation can doPut environmental cost into every AI business case and claim only benefits you will track. Use the smallest model that does the job, and ask suppliers what they report per service and how it is verified.
Sources and links (8)
IEA: Energy and AI · IEA executive summary
The International Energy Agency's global estimates and scenarios for data centre electricity.
AEMO 2026 Electricity Statement of Opportunities · AEMO report (PDF)
The market operator's August 2026 data centre forecasts for the National Electricity Market.
Expectations of data centres and AI infrastructure developers · Department of Industry page
The Australian Government's five expectations, announced 23 March 2026.
National AI Plan · Department of Industry page
The plan the data centre expectations sit under.
Measuring the environmental impact of AI inference · Google Cloud blog
Google's method and figures for energy, carbon and water per prompt.
Carbon Emissions and Large Neural Network Training · arXiv paper
Peer-reviewed style estimate of training energy and carbon, including GPT-3.
Making AI Less Thirsty · arXiv paper
Research estimating the water footprint of training and running AI models.
Microsoft: a milestone in our journey to carbon negative · Microsoft blog
Microsoft's February 2026 update on matching its electricity with renewables.

Written from public sources and checked on 10 October 2026. Company and research figures are estimates made with their own methods. Check the current documents before relying on any figure.

How this maps to what we do

Environmental impact is becoming part of AI decisions: where models run, what they cost in energy and water, and what you can say about it.

Standard or rule
What an organisation needs
Where we help
Greener AI commitments
Environmental cost counted in AI business cases and reported with the same discipline as other costs.
Value Measurement & Board ReportingAI Strategy Build

These are Agile Insights AI Strategy & Governance offerings.

Want to know more?

Sign up for our podcast at agile-insights.com.au.

Read the transcript

Linda Are data centres evil? No. But they are not free either. AI is making data centres one of the fastest growing users of electricity, and that cost is something an executive can measure, manage and put in a business case. By the end of this episode you will know what the evidence says, where it is weak, and what your organisation can do. I'm Linda, and this is the position as at October 2026.

Linda First, what is it? The environmental footprint of AI is the energy, water and carbon needed to build and run AI systems. It is not one standard or law. Governments publish expectations, suppliers publish sustainability reports, and your organisation decides how much to count. Our practice treats it as a governance topic, reviewed alongside cost, risk and security. We call that commitment Greener AI.

Linda Why does it matter? Boards hear two stories. One says AI is wrecking the climate. The other says it is a rounding error. Both are partly true, and neither helps you decide. Picture a team rolling an assistant out to five thousand staff. Nobody asks what it draws from the grid, and the sustainability report never mentions it. That is the gap. It matters to chief information officers, sustainability and risk leads, procurement and finance, usually at reporting time, in tenders, and in board papers.

Linda Now the evidence, in four ideas. The first is scale. The International Energy Agency estimates data centres used about 415 terawatt hours of electricity in 2024, roughly one and a half per cent of the world's total. It expects that to more than double, to around 945 terawatt hours by 2030, with AI the main driver. Globally that is still small. Locally it can be very large.

Linda The second idea is training versus everyday use. Training is a big, one-off cost. A 2021 paper estimated that training GPT-3 used about 1,287 megawatt hours. That model is small by today's standards, so treat it as a scale marker. Everyday use is tiny per request. Google reports its median text prompt uses about 0.24 watt hours and 0.26 millilitres of water, around five drops, and says the energy figure fell thirty three times in a year. But that is one company's method, a median, for text only. Images, video, long documents and automated agents cost more, and millions of small requests add up.

Linda The third idea is water. Researchers at the University of California, Riverside estimated that training GPT-3 in Microsoft's American data centres could evaporate about 700,000 litres on site, and 5.4 million litres once the water behind the electricity is counted. Water is mostly a local issue, shaped by cooling design and location. Operators rarely disclose enough for independent checking, so treat any single water number with care.

Linda The fourth idea is carbon, which depends on the grid and on accounting. In February 2026 Microsoft said it had matched all its annual electricity use with renewable energy for 2025. Yet trade press reporting on its latest sustainability report says its total emissions rose about twenty five per cent in the 2025 financial year, driven by data centre construction. Both can be true. Matching annual purchases is not the same as running on clean power every hour.

Linda Where are the estimates weak? Companies measure different things, independent data is thin, and forecasts swing widely. The Australian Energy Market Operator reports that more than forty per cent of data centre projects have dropped out or slipped since 2025.

Linda Now Australia. In its August 2026 outlook, the market operator estimates data centres used about 5 terawatt hours in 2025 to 26, or 2.8 per cent of the main grid. Its central scenario reaches about 34 terawatt hours by 2035 to 36, around thirteen per cent. A high growth case reaches 52. Developers have proposed 67 gigawatts of connections, but the operator says that is not a forecast.

Linda Government has set expectations too, announced on 23 March 2026. New data centres should not push up energy prices, should secure new clean energy, pay their share of network costs, and use water efficiently with transparent reporting. Energy intensive proposals that do not align will not be prioritised in Commonwealth assessments. They do not cover small enterprise sites. Expect that pressure to reach you through supplier contracts.

Linda How does this fit the stack? Your data foundations and AI governance run on someone else's data centres, so the footprint is largely a supplier and design question. Greener AI puts it into governance: an environmental check when each project is proposed, a decision about model size, and quarterly review.

Linda Three first steps. One, put environmental cost into every AI business case, beside the benefit you expect, and claim only benefits you will track. Two, choose where and how models run: use the smallest model that does the job, cache repeat requests, switch off unused integrations, and prefer lower carbon regions where your data rules allow. Three, measure and ask. Ask suppliers what they report per service, whether figures are location based or purchase based, whether water and training are included, and how they are verified. The primary sources are in the show notes.

Linda As at October 2026, the government expectations are in force as a policy position, not a law. Forecasts will move, and supplier figures change every reporting cycle. Check the current documents before you rely on any number here.

Linda To recap: data centres are not evil, but AI is a real and growing load, and your choices about models, locations and suppliers decide your share. If you'd like to know more, sign up for our podcast at agile-insights.com.au.

More on #leadership

Episode 38 · Primer and discussionBudgeted for seats, billed for workEpisode 34 · Primer and discussionWhen the test stops telling you anything: hunger beats tenureEpisode 23 · Primer and discussionWhat FDE means at OpenAI: skills, not custom code.Episode 21 · DiscussionTrust cannot be rushed: how much authority will you give AI?Making You Think 13 · Primer and discussionHow to Lead a Quest, by Jason FoxMaking You Think 12 · Primer and discussionThe Game Changer, by Jason Fox

Every episode tagged #leadership

All of the briefings (38) · Follow Briefing in Apple Podcasts or any podcast app

Everything from Agile Insights: briefings, Making You Think and the reference series