Every time we ask an AI a question, somewhere a data centre draws power, generates heat, and consumes water. We don’t see it. We don’t pay for it. That’s the point.
This invisible cost has a name in economics: an externality. It’s what happens when the price of something doesn’t include the damage it causes. We’ve encountered it before. In the plastic bottle we threw away, in the carbon from the flight we took, in the river downstream from the factory that makes the things we buy. The cost is real. It just lands somewhere else, on someone else, usually later.
AI is the latest industry to be built on this logic. We have been here enough times to know where it goes.
The numbers nobody is talking about
The AI industry is growing faster than almost any technology in history. Data centres are expected to more than double their electricity consumption by 2030. A single large data centre can use up to 1.5 million litres of water per day. Not to produce anything. Simply to cool the machines. That water evaporates into the atmosphere. It doesn’t vanish from the planet, but it leaves the local watershed, often in places where water scarcity is already acute, and it returns as rainfall somewhere else, on no timeline that helps the community next door.
To put that in context: global AI water consumption in 2025 could rival total annual global bottled water production. The bottled water ends up in someone’s hands. The data centre water ends up as waste heat dispersed into air.
The carbon picture is similarly confronting. AI systems alone could be responsible for between 33 and 80 million tonnes of CO2 emissions in 2025, comparable to a large global city running for a year. Unlike a city, which has been emitting for decades, AI’s footprint was essentially zero four years ago.
The big tech companies, Google, Microsoft, Meta, Amazon, all claim to run on 100% renewable energy. What they mean is that they purchase certificates equivalent to their consumption. The actual electrons powering their data centres at any given moment may come from coal or gas, depending on what the grid is doing. It is an accounting solution, not an engineering one.
We have seen this pattern before
Coal did it. The Industrial Revolution externalised the cost of burning fossil fuels into the lungs of city dwellers and the atmosphere of the planet for 150 years before regulation caught up. Plastic did it. The cost of single-use convenience was paid by the ocean, by marine life, by the microplastics now found in human blood and infant lungs. Industrial agriculture does it still, responsible for nearly a quarter of global greenhouse gas emissions while the price of a burger reflects none of that.
In every case, the pattern is the same. The industry scales fast. The cost is made invisible at the point of use. Transparency is resisted until catastrophe or legislation forces it. By the time the reckoning comes, the damage is embedded in systems too large and too profitable to easily change.
The people who saw it coming were not wrong. They were just early, and they were not the people with the power to act.
What would actually have to change
The AI energy problem is not without solutions. They exist. The question is whether the people with the power to build them are being asked the right questions.
The most promising engineering development is immersion cooling: submerging entire server racks in non-conductive fluid that absorbs heat directly from the chips. The chips produce the same amount of heat regardless of how they’re cooled. What immersion cooling does is capture that heat more efficiently, reducing total energy consumption by roughly 50% compared to conventional air cooling, eliminating the need for evaporative water, and critically recovering the heat in a form that can be reused rather than wasted. Microsoft committed to zero-water cooling across all new data centre designs from August 2024.
The energy sourcing problem is being worked too, though unevenly. Data centres need consistent, reliable power, what engineers call baseload, and most renewables are variable. Wind doesn’t always blow. The sun doesn’t always shine. The energy source that comes closest to solving this for a data centre is geothermal: consistent, clean, and available around the clock regardless of weather.
Here is where Aotearoa enters the conversation.
What we have that most countries don’t
Aotearoa generates approximately 85% of its electricity from renewable sources. It has around 200 megawatts of consented geothermal capacity sitting idle, waiting for investment. Its geothermal resource, unlike the expensive deep-drilling projects being developed in the United States, exists naturally, in hydrothermal systems that have been operating for millennia.
A $3.5 billion data centre has just been consented in Southland. Microsoft has already backed a new geothermal unit near Taupō to power its Auckland data centre region. The economics work. The resource is there. The cable infrastructure connecting the South Island to global networks is being built.
Running a large data centre in New Zealand on renewable energy rather than in China on coal could save the equivalent annual emissions of nearly half a million cars. We are, right now, in a position to export clean compute to the world, to let the most energy-intensive AI workloads run on our hydro and geothermal network rather than on coal. Energy sourcing is not the only variable. Data centres bring land use, water rights, noise, heat, and community disruption with them. The places that have had these facilities arrive without adequate consent processes know that clean energy doesn’t automatically mean clean impact.
That is a genuinely significant opportunity. Opportunity is not the same as outcome. The question of who builds this, how, and in whose interest is not yet settled.
Who decides
The Datagrid data centre consented in Southland is majority-owned by a French entrepreneur. The power agreement is with Mercury. The community benefit is economic: construction jobs during the build, fifty ongoing staff. That is a conventional commercial model applied to an unconventional resource.
It is not the only model available.
Aotearoa has something else the American and European players working this problem don’t have: a living indigenous framework for thinking about resource stewardship that is fundamentally different from the extractive logic that created the problem in the first place.
Kaitiakitanga, the Māori concept of guardianship, does not ask “how much can we extract safely.” It asks whether the resource is in a state of abundance that allows giving, what reciprocity looks like, and what obligation is owed to the land and to future generations. It is not environmentalism with a Māori aesthetic applied to it. It is a different starting question entirely.
The Waitangi Tribunal has released an interim report on Māori rights and interests in geothermal resources, finding that the Crown’s draft geothermal strategy would breach the Treaty by excluding it entirely. Ngāi Tahu, whose territory covers most of the South Island, has already been to the High Court over conservation law changes that threaten their settlement. These are not abstract legal processes. They are live questions about who has the authority to make decisions about resources that have been in relationship with tangata whenua for centuries.
There are already Māori-led energy companies working in this space. Nau Mai Rā, Aotearoa’s only kaupapa Māori electricity retailer, is ending power poverty in communities by distributing surplus energy from marae-based solar systems to whānau households. They are doing something the gentailers haven’t: building a model where community benefit is structural, not an afterthought. They are stuck, not because the model doesn’t work, but because they lack access to generation at a scale the wholesale market rewards.
A kaitiakitanga-governed, geothermal-powered data centre, with community energy distribution built into its architecture from the ground up, would change that equation. It would give Nau Mai Rā the generation source the big players currently deny them. It would give the AI infrastructure industry something it doesn’t have anywhere in the world: a model built inside an indigenous stewardship framework rather than retrofitting ethics onto extraction after the fact.
The question worth asking
We are not at the end of this story. We are at the beginning of it, the moment before the pattern locks in, when the decisions being made will determine whether AI’s environmental cost follows the coal and plastic trajectory, or whether something genuinely different is possible.
Aotearoa is small enough to move differently, and positioned well enough to matter. We have the resource, the engineering talent, the legal and governance frameworks, and the indigenous knowledge systems to build something the world hasn’t seen.
The question is whether the people making decisions about how this gets built are asking the right questions. Not just “is it commercially viable” and “does it meet the consent requirements” but “who has the right to make this decision, what does the resource itself require, and what do we owe to the people whose relationship with this land predates the data centre industry by centuries.”
Those questions are being asked. Not loudly enough, and not yet by the people with the most power to act on the answers. Here is where it usually starts.
This article was written in the context of an ongoing conversation about AI, energy, and environmental ethics. The author uses AI tools in her work and considers the environmental cost of that use to be a live ethical question, not a resolved one.
Statistics on AI water and energy consumption draw on research by Shaolei Ren at the University of California Riverside (2023), the International Energy Agency’s Electricity 2024 report, and the Digiconomist AI Sustainability Index. New Zealand energy data sourced from MBIE Energy Quarterly. The Datagrid project details are drawn from reporting by Joel MacManus at The Spinoff, May 2026.
Image Alt text: A workstation monitor and keyboard on a rolling desk inside a data centre, with rows of dark server racks receding into the background under blue lighting.