AI's Water and Minerals Problem: Why Global Averages Hide the Real Cost-Part 3

A litre of water evaporated in a drought-stressed valley in Arizona is not the same thing as a litre evaporated beside a river in Finland, even though every global AI water statistic treats them as identical. That single confusion is responsible for most of the public argument over how much water AI actually uses.

The same is true underground. AI's mineral demand looks small as a share of global supply — a few percent here, low double digits there — right up until you ask which three countries actually refine that mineral, and what happens to the towns sitting on top of it.

Water and minerals are the two resources AI cannot simply generate more of by building a bigger power plant. Source credit to be added on upload.

Why One Water Number Can't Tell the Whole Story

Headlines about AI's water use tend to reach for a single dramatic figure — millions of litres for one model's training run, a specific number of millilitres per prompt — and leave it there. Those numbers are usually accurate as far as they go. The problem is what they leave out.

Global estimates put AI-related water withdrawals at roughly 4.2 to 6.6 billion cubic metres a year by 2027, a range equivalent to four to six times Denmark's annual water consumption, or roughly half of the United Kingdom's. That's a genuinely large number. But treating it as a single undifferentiated pool of water hides the question that actually determines whether it's a problem: where, specifically, is it being taken from, and what else depends on that same source?

A litre withdrawn from a water-abundant river basin barely registers. A litre withdrawn from an aquifer already being drawn down faster than it recharges is a different story entirely, even though both litres look identical in a global tally. That distinction is the organising idea for this whole chapter.

↑ Back to Table of Contents

The Three Water Footprints Behind Every AI Query

Researchers generally split AI's water use into three scopes, and most public arguments about "how much water does AI use" are really disagreements about which scope is being counted. Scope 1 is direct, on-site water — mainly evaporated in cooling towers to keep servers from overheating, and it can run anywhere from roughly 1 to 9 litres per kilowatt-hour of server energy depending on climate and cooling technology.

Scope 2 is indirect water: the water consumed by the power plants generating the electricity a data centre draws, which in the United States has been estimated at around 1.2 gallons per kilowatt-hour on a national average basis. Because this water is consumed somewhere else entirely — at a coal or gas plant, possibly states away — it's the piece most often left out of a company's headline sustainability numbers, and it's frequently larger than the water used on-site.

Scope 3, embedded water, is the least visible of all: water used years earlier and often a continent away, to manufacture the chips themselves. A single advanced microchip can require on the order of 2,200 gallons of ultrapure water during fabrication, and at least one major operator has disclosed that supply-chain water accounts for more than 99% of its total corporate water footprint. Reported industry figures almost never separate these three cleanly, which is exactly why single-number water claims deserve scepticism.

↑ Back to Table of Contents

Where Scarcity Actually Bites

US data centres directly consumed roughly 17 billion gallons of water in the most recent reporting year, a figure some projections put on track to double or even quadruple by 2028 as AI-driven demand accelerates. A single large hyperscale facility can draw up to 5 million gallons a day — comparable to the water needs of a town of 10,000 to 50,000 people.

The geography of where these facilities are being sited makes the scarcity question sharper still. Many new campuses are landing in drought-prone regions — Arizona, Texas, and similar climates elsewhere in the world — precisely because land and power are cheaper there, not because water is abundant. That mismatch between where AI infrastructure is economically attractive and where water is actually available is becoming one of the more visible flashpoints in local opposition to new projects: water-use concerns have been cited as a factor in the cancellation of roughly 20 proposed US data-centre projects worth a combined $41 billion in just the first quarter of this year.

↑ Back to Table of Contents

The Minerals Nobody Sees

Water is only the visible half of AI's material footprint. Every data centre is also a claim on copper, aluminium, silicon, gallium and rare earth elements — and by 2030, the IEA estimates data-centre buildout could lift global demand by roughly 2% for copper, 3% for rare earths, and as much as 11% for gallium specifically.

Those percentages sound modest until you look at the absolute tonnage and the supply concentration behind it. Data centres could be consuming more than 512,000 tonnes of copper and 75,000 tonnes of silicon annually by 2030, at a time when global copper supply is already projected to face a shortfall by 2035. And for gallium in particular, one country accounts for roughly 98% of primary low-purity supply capacity — which means an 11% demand increase isn't landing on a diversified global market; it's landing almost entirely on one.

Copper illustrates the squeeze best because it sits at the intersection of every major growth story at once — AI, electric vehicles, grid buildout, and renewable energy all compete for the same mined and refined tonnage. The IEA projects that by the 2040s, clean-energy technologies alone could account for over 40% of global copper and rare-earth demand, 60–70% of cobalt and nickel demand, and close to 90% of lithium demand — meaning AI isn't drawing down some separate mineral pool. It's one more claimant on a supply chain that's already stretched by the energy transition it's supposed to help.

Recycling offers a genuine partial answer. The IEA estimates that scaling up recovery and reuse — recycled chip materials, recovered gallium and germanium from discarded electronics, reclaimed copper — could reduce the growth in new mining supply needed for critical minerals by 25% to 40% by 2050. That's meaningful, but it's a multi-decade curve against a demand increase that's already underway now.

↑ Back to Table of Contents

Two Resources You Can't Manufacture More Of

Electricity and heat, the subjects of Part 2, are ultimately solvable with enough capital and enough turbines — you can build your way out of a power shortage, however expensive that turns out to be. Water and minerals are harder, because neither one can simply be manufactured wherever it's convenient. A watershed either has spare capacity or it doesn't. A gallium refinery either exists in a given country or it doesn't, and building a new one takes years even before geopolitics gets involved.

That's why the framing this series keeps insisting on — water stress times location times technology choice, mineral demand times refining concentration times recycling capacity — matters more here than almost anywhere else in AI's supply chain. A global average genuinely obscures more than it reveals when the resource in question is this unevenly distributed.

Part 4 follows this same logic onto land itself: who gets displaced when a data centre and its power and water infrastructure arrive, and who actually captures the economic benefit versus who absorbs the physical cost. It's the environmental-justice question this whole series has been building toward, alongside a fair look at where AI's own tools are genuinely reducing environmental impact elsewhere.

↑ Back to Table of Contents

K M Manohar’s Insight

AI is often described as a story of computing power. I think the deeper story is becoming a story of resource geography.

We can build more turbines.
We can add more solar panels.
We can expand the grid.

But we cannot simply manufacture a new watershed when a region runs short of water. And we cannot create a new supply of gallium, copper or rare earths overnight when refining capacity is concentrated in a few places.

That changes how we should think about AI infrastructure.

The question should not be only “How much water or how many minerals does AI consume?” It should be “Where does that consumption happen, and who else depends on those resources?”

Recent research increasingly supports this location-first view: AI's environmental footprint varies substantially with where data centres are built, how they are cooled, and what electricity system supplies them.

My biggest takeaway is simple:

The next AI bottleneck may not be silicon or compute. It may be geography.

The winners of the AI infrastructure race will increasingly be the regions that can offer reliable power, sustainable water, mineral security, strong recycling systems and social acceptance — all at the same time.

That is why AI sustainability cannot remain a data-centre engineering problem alone. It is becoming a national resource-strategy problem.

And for countries like India, this is an opportunity as much as a warning: build AI infrastructure where the resource equation works before the infrastructure locks the geography in.