By 2030, the data centres powering AI will consume water equivalent to the basic annual domestic needs of every person in Sub-Saharan Africa — all 1.3 billion of them. They’ll burn through 945 terawatt-hours of electricity, nearly triple what Pakistan, Bangladesh, and Nigeria use combined. Their carbon emissions will match the entire United Kingdom. And their land footprint will exceed 14,500 square kilometres — twice the Jakarta metropolitan area.
These are the numbers from a new report published June 3 by the United Nations University Institute for Water, Environment and Health (UNU-INWEH). And they’re the conservative estimates.
The Problem with Counting Only Carbon
Most assessments of AI’s environmental impact focus on one metric: carbon emissions. Training GPT-3 used an estimated 1.3 GWh. GPT-4 consumed between 50 and 70 GWh. These numbers get cited, debated, and optimised.
But the UN report’s most important finding is that carbon is the wrong yardstick — or at least, an incomplete one.
Every kilowatt-hour of electricity used to train or run an AI model carries three footprints: carbon from generation, water from cooling and electricity production, and land from energy infrastructure and supply chains. And they don’t move in the same direction.
Switching from coal to bioenergy cuts carbon emissions by 70%. Sounds great. But it increases water consumption thirty-fold and land use a hundred-fold. “Low-carbon” is not automatically “low-water” or “low-land.”
“If we keep judging AI sustainability by carbon alone, we might think that renewables make AI infrastructure clean,” said Dr. Miriam Aczel, the report’s lead author. “But that is solving one problem while creating other problems, often in places that didn’t ask for it.”
This is the uncomfortable arithmetic that the AI industry hasn’t confronted. When Google and Microsoft tout their carbon-neutral data centres, they’re telling you about one-third of the story. The water and land costs are being exported to regions that host the infrastructure but don’t benefit from the AI services running on it.
The Numbers That Matter
| Metric | 2025 (Current) | 2030 (Projected) | Comparison |
|---|---|---|---|
| Electricity | 448 TWh | 945 TWh | 3× Pakistan + Bangladesh + Nigeria |
| Water | — | 1.3 billion people’s domestic needs | Equivalent to all of Sub-Saharan Africa |
| CO₂ emissions | — | 400 million tonnes | Equivalent to the United Kingdom |
| Land footprint | — | 14,500 km² | Twice Jakarta metro area |
If data centres were a country today, they’d be the 11th largest electricity consumer on Earth — behind France, ahead of Saudi Arabia. By 2030, they’d crack the top five.
Inference Is the Real Problem
Here’s a finding that should reframe how we think about AI’s environmental cost: inference accounts for 80-90% of total AI energy consumption, not training.
All the hand-wringing about training GPT-5 misses the point. Training is a one-off cost. Every time you ask ChatGPT a question, generate an image, or run an AI agent, you’re paying the inference tax — and that tax compounds with every user, every query, every day.
A standard chatbot conversation uses 200 times more energy than a basic AI function like classifying spam email. Generating a synthetic image? Multiply that again. The Google-SpaceX $920M/month compute deal isn’t about training — it’s about keeping up with inference demand that’s growing faster than anyone predicted.
The Rebound Effect
The report also identifies what economists call the “rebound effect”: when efficiency improvements in AI hardware lead to more AI use, not less energy consumption overall.
Better chips don’t reduce total energy use. They make AI cheaper to run, which drives more adoption, which drives more inference, which drives more energy use. This isn’t a bug — it’s how technology scaling works. The same dynamic played out with air conditioning, cars, and computing generally. Efficiency gains get eaten by volume growth.
What This Means for New Zealand
NZ keeps getting mentioned as a potential data centre haven. We wrote about Datagrid’s Invercargill facility and the carbon arbitrage opportunity. The pitch is simple: cold climate, renewable hydro power, stable politics.
The UN report adds a critical dimension. Data centres need water — a lot of it. Southland has plenty now. But data centre clustering creates local water stress even in water-rich regions. The report specifically calls out that “the communities who provide the critical minerals for advancing AI and the ones that host its infrastructure and e-waste” should be “among those who benefit from it.”
If NZ is going to host data centres for the world’s AI companies, the question isn’t just whether we have cheap renewable electricity. It’s whether we’re getting enough back — in jobs, tax revenue, and local AI capability — to justify the water, land, and energy we’re giving up. Right now, the answer is unclear.
The 20 Largest Data Centre Hubs
The report analysed environmental footprints across the 20 largest data centre hubs. The variation is dramatic:
- Ireland’s data centres already consume more electricity than all Irish rural households combined
- Singapore has had to impose a moratorium on new data centre construction because of land and water constraints
- Virginia (US) data centre corridor draws enormous volumes of water from the Potomac watershed
- Chile and Uruguay are being pitched as green data centre locations, but their water stress profiles tell a different story
The geography matters. A data centre in Iceland has a fundamentally different environmental footprint than one in Arizona — even if both claim to be “carbon neutral.”
💰 Industry Impact
Who Benefits: Companies selling data centre cooling technology (liquid cooling, immersion cooling), water recycling systems, and energy efficiency solutions. Also: regions with abundant renewable energy and water (Nordics, parts of NZ, Canada) that can market themselves as genuinely sustainable AI infrastructure hubs.
What’s at Stake: The AI infrastructure market is projected to exceed $500 billion annually by 2030. But environmental constraints — water stress, land use conflicts, grid capacity — could become the binding constraint on growth. Companies that solve the water-energy-land tradeoff will have a structural advantage.
Key Risks: Regulatory backlash is building. Singapore’s moratorium is the canary in the mine. EU and US states are starting to require environmental impact assessments for new data centres. Water stress in data centre hubs could trigger local opposition that slows buildout regardless of corporate plans. This isn’t financial advice — it’s mapping the commercial landscape. Watch for water-rights legislation in data centre corridors as a leading indicator of constraint.
❓ Frequently Asked Questions
Q: How much water does an AI query actually use? A standard ChatGPT conversation uses roughly half a litre of water (0.5L) when you account for both cooling and electricity generation. That’s 200× more than a basic spam filter. Generating an image uses even more. Multiply by billions of queries per day, and you get the 1.3 billion person-equivalent.
Q: Does switching to renewable energy solve the problem? Not by itself. The UN report shows that switching from coal to bioenergy cuts carbon by 70% but increases water use 30× and land use 100×. Renewable energy still requires water for cooling and land for infrastructure. Carbon neutrality ≠ environmental neutrality.
Q: What should NZ do? NZ should require water and land impact assessments for any new data centre, not just carbon. The NZ AI Blueprint should explicitly address the water-energy-land tradeoff. If we’re going to be a data centre destination, we need to ensure the benefits flow locally — not just to foreign AI companies.
Q: Is inference really bigger than training? Yes. The report finds inference accounts for 80-90% of total AI energy consumption. Training is a one-time cost; inference is ongoing and compounding. Every new user, every new query, every new AI agent adds to the inference bill.
🔍 THE BOTTOM LINE
AI’s environmental footprint isn’t just about carbon — it’s about water, land, and where the costs land. By 2030, data centres will drink like a continent, burn electricity like a top-5 nation, and emit CO₂ like the United Kingdom. The UN report’s most important finding isn’t the headline number — it’s that measuring only carbon hides the real costs, and those costs are being borne by communities that didn’t ask for any of this. The AI industry needs to reckon with all three footprints, not just the one that looks best in a press release.
