What’s the Environmental Impact of Your AI Use?
by Sustainability Tracker 09/11/2026
Guide, Thought Starters
by Sustainability Tracker 09/11/2026
Guide, Thought Starters
This article was contributed by Sustainability Tracker.
Last Updated: 09 November 2026
Contributors: Anne Kehlhofer – Director, ESG, Xero; Sustainability Tracker; James Tilbury – Partner, Linden Sustainability; Sydney Straver – Director, &BLOOM Sustainability & ESG
Quick answer: A typical short text prompt uses roughly a quarter to a third of a watt-hour of electricity, a few drops of water and a fraction of a gram of CO₂e. The bigger issue is total demand. Global data centre electricity use is on track to almost double between 2025 and 2030, and as each prompt gets cheaper and more efficient, people tend to use more AI overall. Your own footprint climbs steeply with how you use it: reasoning modes, long documents, image and video generation, and AI agents can each multiply energy use by hundreds of times or more. Where the electricity comes from matters as much as the prompt.
What this guide covers: the footprint of using AI. AI may also help cut emissions elsewhere, for example by improving how grids, buildings and industry use energy. The Paris-based International Energy Agency (IEA), of which Australia is a member, estimates that existing AI applications could cut about 1,400 Mt of CO₂ in 2035 if widely adopted. That’s about three times total data centre emissions in its high-growth scenario. The IEA also says there’s currently no momentum to ensure that adoption happens, so we’ve left these potential savings out of the numbers here.
How to read the numbers: A watt-hour (Wh) is the energy a 1-watt device uses in an hour, so a 10-watt LED bulb uses 10 Wh an hour. A kilowatt-hour (kWh) is 1,000 Wh, the unit on your power bill. A token is a chunk of text, roughly three-quarters of a word. CO₂e (carbon dioxide equivalent) combines all greenhouse gases into one figure.
Three providers have published energy figures for a standard text prompt.
Independent researchers land in the same range. Epoch AI estimated a GPT-4o query at roughly 0.3 watt-hours, well below the older, widely cited 3 watt-hour figure. The University of Rhode Island’s benchmarking study put a short GPT-4o query at 0.42 watt-hours.
These figures describe the lightest kind of AI use. Google’s figure covers text prompts only, excluding image and video generation. They also cover only the energy used to answer a prompt. None includes the energy used to train the model or to build the data centres it runs in, which Google and Microsoft both name as major drivers of their emissions growth. Only Google’s carbon figure includes a share of the emissions from manufacturing its chips and servers. What you ask for changes the answer dramatically.
Small per-prompt figures can also hide the bigger trend. As AI gets cheaper and more efficient, people use more of it, so total energy use keeps rising even as each prompt gets lighter. We cover this in more detail below.
| Task | Estimated energy | Source | Evidence type |
| Short text prompt | 0.16 to 0.60 Wh | Microsoft, Google, OpenAI, Epoch AI, URI | Provider figures and independent estimates |
| Text prompt with a 10,000-token document pasted in | About 2.4 Wh | Epoch AI | Independent estimate (one source) |
| 1024 × 1024 image (open-source model) | About 0.63 Wh | MIT Technology Review | Lab measurement, open-source model |
| Long prompt to a heavy reasoning model | Over 29 Wh | URI study | Independent benchmark |
| One coding-agent prompt | About 150 Wh (range 60 to 290) | Zeke Hausfather | One person’s logged use |
| Five-second video (open-source model) | About 944 Wh | MIT Technology Review | Lab measurement, open-source model |
Reasoning or “thinking” modes generate a hidden chain of working before they answer. Hugging Face’s AI Energy Score project found reasoning models used on average 30 times more energy than models without reasoning. The same model used between 150 and 700 times more energy with reasoning switched on. The researchers traced most of that to output length: with reasoning enabled, models produced 300 to 800 times more tokens.
The URI study found a similar pattern at the top end. The most energy-intensive models exceeded 29 watt-hours (about two smartphone charges) per long prompt, over 65 times the most efficient systems.
Every word you paste in has to be processed. Epoch AI estimates a query with 10,000 input tokens and a 500-token answer uses around 2.4 watt-hours (about 15 minutes of a 10-watt LED bulb), and 100,000 input tokens pushes that to around 40 watt-hours (roughly two to three full smartphone charges). A long-running conversation grows the same way, because the model re-reads the history each turn.
MIT Technology Review worked with University of Michigan researchers to measure open-source models directly. A 1024 × 1024 image from Stable Diffusion 3 Medium used an estimated 2,282 joules (about 0.63 Wh, or four minutes of an LED bulb) . A five-second video used about 3.4 million joules, roughly the energy of running a microwave for over an hour.
Agents plan, call tools, check results and loop, and each step sends the full context back through the model. Climate scientist Zeke Hausfather logged eight weeks of his own coding-agent use: 1,138 prompts triggered more than 14,000 model calls and an estimated 170 kWh of data centre electricity (enough to drive a typical electric car about 1,000 km, a little further than Melbourne to Sydney). That works out to around 150 watt-hours per prompt (a little more than boiling a full kettle), roughly 600 times a median chat prompt, with an uncertainty range of 60 to 290 watt-hours. He notes other researchers found coding agents use roughly 1,000 times the tokens of an ordinary chatbot interaction. This is one person’s logged use, so treat it as a single case. A separate lab study from KAIST in South Korea, presented at an IEEE computer architecture conference in 2026, points the same way: AI agents used up to 136 times more energy per request than a chatbot handling a similar request, mainly because they made many more calls to the model.
Estimates range widely because they measure different things. All four figures below cover using a model (inference). None includes training it. Training is a large, one-off cost, like the emissions from constructing an office building before anyone works in it. UC Riverside researchers estimated that training GPT-3 in Microsoft’s US data centres used about 5.4 million litres of water (about two Olympic swimming pools) .
The UC Riverside researchers found the result depends on when and where a model runs. Climate, cooling technology and the local power mix all change the answer.
Location matters most for water. The same few millilitres mean far more where water is scarce, and a 2025 Bloomberg analysis found about two-thirds of new data centres built or in development since 2022 are in areas already under high water stress. Per-prompt figures can’t show that, so it’s worth asking providers where their data centres are and how they’re cooled.
Carbon depends on two things: how much electricity a prompt uses, and how that electricity is generated. The same prompt can produce very different emissions depending on where it runs and how emissions are counted.
Market-based figures reflect the clean energy a company has purchased. Location-based figures reflect the actual mix of the local grid, and Google’s technical paper describes them as excluding a company’s clean energy procurement. Both are legitimate. They answer different questions.
Google reports 0.03 grams of CO₂e per median Gemini text prompt, calculated with its 2024 fleet-wide carbon intensity. It doesn’t publish a location-based equivalent. Independent analysts at Greenly estimate that on a location-based approach, the figure would be closer to 0.09 grams, about three times higher.
To show how much, we applied Google’s 0.24 Wh figure to Australia’s official grid factors. For 2025–26, the Clean Energy Regulator sets Victoria’s factor at 0.78 kg CO₂e per kWh and NSW’s at 0.64. At those rates, the same prompt would produce about 0.19 grams in Victoria and 0.15 grams in NSW. That’s illustrative only: Gemini prompts aren’t necessarily processed in Australia.
The URI study found the most resource-efficient models stay below 0.3 grams of CO₂e per query.
| Daily chat user | Heavy coding-agent user | |
| Usage | 20 short prompts a day for a year (7,300 prompts) | Hausfather’s logged use over eight weeks |
| Energy | About 1.75 kWh | About 170 kWh |
| CO₂e at Google’s market-based rate | About 0.22 kg | Not applicable |
| CO₂e if powered by Victoria’s grid | About 1.4 kg | About 133 kg |
The heavy agent user consumed roughly 97 times more electricity in eight weeks than the daily chat user does in a year. For most people, a year of casual chatting has a small footprint. Heavy, automated use is where the numbers become significant.
Per prompt, often yes. Overall, the gains are being outpaced by growth in use. Economists call this the rebound effect, or Jevons paradox: when something becomes cheaper and more efficient, people tend to use more of it.
Google says energy per median Gemini text prompt fell 33 times over the 12 months to May 2025.
The picture across models is less tidy. When Hugging Face compared newer models against similar-sized models from nine months earlier, 9 of 15 used the same or more energy.
Total demand keeps rising regardless. Google’s 2026 Environmental Report shows its electricity use rose 37% in a single year, even as per-prompt efficiency improved. The IEA projects global data centre electricity use roughly doubling from 485 TWh in 2025 to 945 TWh in 2030 (2030’s figure is more than three times Australia’s total annual electricity use), with AI-focused data centre consumption tripling.
Australia’s grid is feeling the growth directly. In its 2026 Electricity Statement of Opportunities, AEMO forecasts data centre electricity consumption in the National Electricity Market rising from about 5 TWh in 2025–26 to 34 TWh by 2035–36, lifting their share of grid electricity from around 3% to about 13%. New South Wales accounts for 66% of current data centre consumption, and Victoria 30%.
Buying clean energy cuts emissions most clearly when it adds new renewable supply, and how other purchases should count is still debated. When data centre demand grows faster than new renewables are built, other generators, including fossil fuel plants, make up the difference.
Not every AI prompt sent from Australia is processed here, and providers generally don’t tell users which region serves their requests. Growth in AI use shows up in Australian electricity planning either way.
Energy, carbon and water are only part of AI’s footprint. AI also has social impacts, and for many organisations they matter as much as emissions. Data centres can raise local electricity prices and affect air quality, as the legal case over gas turbines near Memphis shows. Further up the supply chain, AI relies on people who label data and moderate content, often for low pay. AI is also changing jobs and the skills people need, and its outputs can carry bias that affects how people are treated. These issues deserve their own attention, and we cover each one, with sources, in our AI platform comparison.
The biggest levers sit with organisations and providers. Many people don’t choose their model at all, because AI comes built into the tools they already use. Organisations can set AI standards in procurement and ask vendors for emissions data based on their actual usage. Providers decide which models run by default, how their data centres are powered and cooled, and what they disclose.
Several gaps remain as at November 2026:
We’ll update this guide as providers publish more data.
Published figures for a typical text prompt are 0.24 Wh (Google, median Gemini Apps prompt) and about 0.34 Wh (OpenAI’s CEO, average ChatGPT query). Microsoft’s researchers estimate 0.16 to 0.60 Wh for a typical query to large models. Independent estimates for GPT-4o land between about 0.3 and 0.42 Wh. Long documents, reasoning modes, images and video use much more.
Google reports 0.03 g CO₂e for a median Gemini text prompt, a market-based figure that reflects its clean energy purchases. Independent analysts estimate a location-based figure of around 0.09 g. On Victoria’s grid, the same amount of electricity would produce roughly 0.19 g. Where and how the electricity is generated changes the answer as much as the prompt itself.
Google estimates 0.26 mL per median Gemini text prompt for data centre cooling. A UC Riverside study estimated roughly 10 to 50 mL per medium-length GPT-3 response when water used to generate electricity is included. Location, model and accounting method explain most of the difference. Microsoft’s researchers estimate up to 0.067 mL per typical query to large models.
Yes. Hugging Face’s AI Energy Score found reasoning models used on average 30 times more energy than models without reasoning, and the same model used 150 to 700 times more energy with reasoning switched on.
Per task, video is far more energy-intensive. MIT Technology Review’s measurements found a 1024 x 1024 image used about 2,282 joules (about 0.63 Wh), while a five-second video used about 3.4 million joules (about 944 Wh), roughly the energy of running a microwave for over an hour.
Far more than chat. Climate scientist Zeke Hausfather estimated his own coding-agent use averaged around 150 Wh per prompt, roughly 600 times a median chat prompt, because each request triggered many model calls that re-read large amounts of context. That’s a single case, though a KAIST lab study found agents used up to 136 times more energy per request than a chatbot.
Use the smallest model that does the job, switch on reasoning only when a task needs it, keep pasted context and requested outputs short, generate images and video deliberately, and scope agent tasks tightly.
| Date | Activity |
| 09 November 2026 | First published |