Anyone with an internet connection can now fire off a query to an AI agent – but what’s the real-world environmental impact of the ‘fourth industrial revolution’?

This story was first published on RNZ

Conversations about AI in New Zealand have centred more recently on its effect on the world of work.

The government plans to save $2.4 billion in baseline costs over the next four years by cutting nearly 9000 public sector jobs – and says in many cases AI will fill the gap.

But a new global report, published by a UN think-tank, has laid bare the other real-world cost of enmeshing AI into everything we do: its environmental and carbon footprint.

It comes just a few months after Datagrid was granted consent to build an “AI factory” in Southland, which is set to become the country’s second-largest electricity user.

Local experts say the report holds warnings for New Zealand if we choose to encourage more hyper data centres here.

How much energy is AI using?

It depends what you’re asking about.

The report, published Thursday by the UN University’s Institute for Water, Health and Environment, says global AI-related use now gobbles up 93 terawatt-hours (TWh) of electricity a year – more than double New Zealand’s annual electricity generation.

If it were a country, its electricity use would rank it 39th in the world – ahead of industrialised countries like Finland and Belgium.

AI now accounts for 20 percent of all data centre electricity use, and the report says that’s expected to double to 40 percent by 2030.

At the same time, data centre capacity is rising – meaning that by 2030, AI electricity consumption could hit 374TWh.

Depending on the mix of how that electricity is generated, that would take AI’s carbon footprint to 158 million tonnes of greenhouse gas emissions.

If you’re just asking about a one-off text prompt to ChatGPT, then the report estimates that individual request uses about 0.42Wh – about the same as turning on an LED lightbulb for a few minutes.

So what’s all that energy being used for?

One of the most energy-intensive aspects of the technology is the process of training AI models, which requires processing enormous datasets – not just once, but multiple times.

“AI training is not a single, instantaneous task, or a one-off event;” the report says. “Training can run for days to months, with repeated, compute-intensive cycles designed to fine-tune accuracy.”

All this advanced processing uses high-performance graphics processing units (GPUs), clustered in large data centres, which draw vast amounts of energy and also give off heat that requires cooling using water.

Only a few models make enough documentation public to accurately calculate their energy use during the training process. However, based on the requirements for previous iterations of ChatGPT, the report says a plausible estimate for training GPT-5 is 100 gigawatt hours.

So it’s the model training that’s using all that power, right? Not me asking ChatGPT to compose a polite but firm email to my colleague?

Wrong.

Although the training phase is energy-intensive, the report says that is now far surpassed by daily use of AI – known as the “inference” phase.

“In the long run, as … the environmental impacts of AI increasingly arise from the continuous running of models to generate responses for billions of user interactions each day,” the report concludes.

“Each individual query consumes far less energy than training, yet when multiplied by billions of users, inference becomes the dominant contributor, estimated to account for 80-90% of total energy use.”

The environmental costs from all those tiny interactions are “staggering”, it says. “ChatGPT, for example, processes around 2.5 billion prompts daily.”

And while 380 million people actively use dedicated AI tools, many more people are interacting with AI with limited awareness.

“AI is now deeply embedded in features such as summaries, recommendations, and contextual responses,” the report says. “As a result, Google alone engages over a billion people in AI-assisted interactions each day.”

While a conventional search uses about 0.3Wh, an AI-enhanced generative search uses up to 3Wh – a 10-fold increase.

AI video generation is a particular “emerging environmental crisis”, the report says.

“A single high-resolution AI video clip can require more than 415Wh, making it more energy-intensive than the creation of hundreds of AI images.”

The higher the resolution and frame count, the more intensive the process becomes.

What about other resources?

AI and data centres don’t just consume electricity. There’s a hell of a lot of water that goes into the cooling functions, while data centres and the electricity generation used to power them also take up space.

The report estimates that by 2030, the land footprint required by AI’s electricity use alone is nearly 6000 square kilometres – an area a bit larger than Auckland – and 3.7 trillion litres of water will be needed.

Even if data centres draw on renewable energy, there are still environmental trade-offs – solar and wind farms require space, while hydro power demands even more water.

“‘Low-carbon’ is not automatically ‘low-water’ or ‘low-land,'” the report says. “Evaluating sustainability through a single metric can hide trade-offs and shift burdens onto places already facing water stress or land pressure.”

At the moment, there’s limited large-scale data centre infrastructure in New Zealand – although Datagrid recently gained consent to build an “AI factory” in Southland that’s set to become the country’s second-biggest user of electricity.

University of Waikato computer science professor Te Taka Keegan says local resources must be taken into account if New Zealand allows further AI infrastructure to be developed here.

“If we were putting in a hyper data centre then are we putting in surrounding infrastructure to generate electricity or are we expecting it to just tap off the local electricity? And are we putting infrastructure in to increase the water intake for that location, or are we expecting just to suck all the resources from the local water environment?”

The government could consider additional consent requirements, he says.

“We shouldn’t be treating them as any other large-scale development.”

Can we limit this energy use?

The report sets out six principles for a responsible AI ecosystem.

One of those is “efficiency by design”, and another is “sustainable use”.

University of Waikato law lecturer Amanda Turnbull-McRae says the second of those places responsibility onto us as consumers of AI to consider what we use it for, and how.

“Our default setting is always to improve, to get the latest and the greatest,” she says.

But businesses looking to integrate AI should be looking carefully at whether something less energy-intensive could better suit their needs.

“Do we really need the biggest large language models? Can we take advantage of smaller language models? We don’t necessarily have to have the most energy-hungry product on the market,” she says.

University of Auckland engineering professor Nirmal Nair says New Zealand has very limited AI infrastructure, so local resource use is a “moot point” at the moment.

However, he sees local development of AI as a solution to two issues: first, a way to ensure New Zealand does not remain a mere consumer of infrastructure and networks developed by a handful of countries – dominated at the moment by the US and China – and secondly, as a way to develop efficient, tailored AI for local purposes.

“We could develop our own thing for our own services,” he says. “That’s the kind of direction in which we need to really start talking, [but] at the current point, I don’t see any of those discussions.”

This story was first published on rnz.co.nz