AI Has a Resource Problem Nobody Wants to Talk About

Everyone is talking about AI productivity.

Almost nobody is talking about AI consumption.

That blind spot is about to cost companies real money.

The conversation is dominated by smarter models, faster agents, autonomous workflows, and trillion-dollar opportunities. Every keynote, every board deck, every earnings call points in the same direction. More capability. More speed. More scale.

But there is another side to the story, and it rarely makes the slide.

AI is becoming one of the largest consumers of physical resources on the planet.

And most boardrooms have no idea.

When executives evaluate AI, they ask three questions. Can it grow revenue? Can it cut costs? Can it lift productivity?

All valid.

There is now a fourth question they cannot dodge.

What is the resource cost of intelligence at scale?

AI is not digital. It is physical.

Most people picture AI as something that lives in software. A clever answer appears on a screen, and the cost feels like nothing.

It is not nothing.

Every prompt travels through servers. Every image burns compute. Every agent running inside your business draws electricity. Every model sits in a data center built from concrete, steel, water, rare minerals, and an enormous amount of power.

The more AI grows, the more physical infrastructure it demands. Chips. Cooling systems. Grids. Pipes. Land. Supply chains that reach all the way back to a mine.

The numbers are no longer abstract. They are starting to look like the consumption of nations.

The electricity story

A new report from the United Nations University puts global AI data center electricity use at 945 terawatt-hours a year by 2030.

That is nearly triple the combined annual electricity use of Pakistan, Bangladesh, and Nigeria, three countries that together are home to more than 650 million people.

A technology serving the few may soon consume more power than nations holding hundreds of millions.

And this is not a distant projection. In 2025, data centers already used an estimated 448 terawatt-hours. If they were a single country, they would rank as the eleventh largest electricity consumer in the world, sitting behind France and ahead of Saudi Arabia.

By 2030, AI and its infrastructure could account for close to 3 percent of all electricity used on Earth.

Read that again. A single category of computing, climbing toward 3 percent of global power.

Not all AI use costs the same

Here is the part most leaders get wrong.

The public debate fixates on training large models. Training is expensive, no argument. But the research is detailed that everyday operational use, the prompts, and the agents’ accounts for close to 90 percent of the energy demand.

The environmental challenge is not building AI. It is using AI.

And usage is exploding. ChatGPT alone is estimated to handle around 2.5 billion prompts every single day. For that one product, that translates to roughly 383 gigawatt-hours of electricity a year.

One product. One slice of the market.

It also matters enormously what you ask AI to do. The energy cost of different tasks is not close to equal.

A typical conversational query is around 200 times more energy-intensive than a basic text classification. Generating a single AI image can require nearly 1,450 times more energy than that same classification. A short AI-generated video can burn as much electricity as 200,000 spam-filter checks.

So when an organization shifts from simple text tasks to image and video generation at scale, it is not nudging its footprint upward. It is multiplying it.

Most companies deploying generative AI have never run that math.

And the shift to autonomous agents raises the stakes again. A single agent does not make one call. It plans, retries, checks its own work, and calls other tools, often firing dozens of inferences to finish one task a human asked for once. Put a workforce of agents to work, and the consumption curve does not rise. It steepens.

The hidden cost is water

Electricity is only the start.

Data centers generate heat. Heat has to be removed. That cooling process consumes water, both directly inside the facility and indirectly through the power plants feeding it.

A single large data center can draw up to 5 million gallons of water a day to keep its servers cool. That is the daily water use of a small town, spent on machines.

The pattern repeats at the level of a single request. A high-complexity AI video can consume up to 4.1 litres of water per generation. Roughly what a person needs to drink across two days, spent on one clip.

And video is the fastest-growing category of all. If only a fifth of AI video requests run at that level of complexity, the annual water cost climbs past 13 billion litres. Most of it is spent on clips that viewers forget within seconds.

Scale that across millions of users, and the totals become hard to look at. The UN study estimates AI’s water footprint could eventually match the basic annual domestic water needs of 1.3 billion people in Sub-Saharan Africa.

We talk about AI as if it floats above the world. It drinks from the same rivers we do.

The footprint you can stand on

Land is the third cost, and the most invisible of all.

The same study estimates AI’s land footprint could exceed 14,500 square kilometers by the end of the decade. That is roughly twice the size of metropolitan Jakarta.

Every server hall, power plant, transmission line, cooling facility, logistics route, and mineral extraction site occupies real physical space. Space that was something else before, and that someone lives near now.

The cloud is not a cloud.

It is somebody else’s land, somebody else’s water, and somebody else’s electricity bill.

The part nobody recycles

The footprint does not begin when a server switches on, and it does not end when it switches off.

Before: the chips that power AI depend on critical minerals dug out of the ground, often in regions with weak environmental oversight and little share of the eventual profit. The supply chain reaches all the way back to a mine, and that mine has a footprint of its own.

After: hardware ages fast in an arms race for performance. Today’s cutting-edge accelerator becomes tomorrow’s electronic waste. That waste has to go somewhere, and it rarely returns to the companies that profited from the compute.

So the true cost of a model spans its entire life. Extraction, construction, operation, disposal. Most corporate accounting captures only the middle of that line and ignores both ends.

Why efficiency will not save you

There is a comforting assumption that better technology fixes this on its own. Chips get more efficient, models get leaner, so consumption must fall.

History says the opposite.

Economists call it the rebound effect. When something becomes cheaper and easier to use, people use far more of it, and total consumption rises rather than falls.

Engines became more fuel efficient. People drove more.

Computing became cheaper. We built far more software.

Now AI is following the same curve. Cheaper inference. Faster models. Lower cost per call. Higher adoption. Greater total consumption.

In AI procurement, your success quietly becomes your biggest resource liability. The better the technology gets, the more of it the organization deploys, and the larger the footprint grows.

Efficiency is not a brake. Without intent, it is an accelerator.

The benefits are global. The bill is local.

The next problem is who actually pays.

The benefits of AI spread across the world. The burden lands in specific places.

A data center draws water from one specific region. A power plant runs in one specific community. Minerals are pulled from specific lands. Electronic waste ends up somewhere real, on someone’s doorstep.

The people gaining from AI and the people carrying its physical cost are rarely the same. That is not a footnote. It is a question of fairness that will land on regulators’ desks faster than most companies expect.

The imbalance runs deeper still. Only 32 countries host meaningful AI infrastructure today, and roughly 90 percent of that computing capacity sits in the United States and China. Most of the world consumes AI without owning any of the systems behind it.

The result is two divides at once. A digital divide between those who build AI and those who only use it. And an environmental divide between those who profit and those who absorb the cost.

This is already happening

If this still sounds theoretical, look at Ireland.

In 2023, data centers accounted for 21 percent of all metered electricity in the country, more than every urban household combined. The national grid operator has since paused new data center approvals around Dublin until 2028.

That is not a forecast. That is a developed economy hitting a physical ceiling and pulling the brakes.

It will not be the last.

What I see inside the enterprise

In enterprise delivery, I keep seeing the same blind spot. Teams track model accuracy and cost per call with real discipline. Almost no one tracks the resource bill sitting behind those calls.

The footprint is real. It is rising. And in most organizations, nobody owns it.

That is a governance gap, not just an environmental one.

None of this means AI is bad

Let me be clear, because this is easy to misread.

AI may be one of the most important technologies humanity has ever built. It can accelerate scientific discovery, improve healthcare, optimize supply chains, cut waste, and unlock enormous value. The goal is not to slow innovation.

The goal is to see its full cost.

For years, sustainability meant carbon and almost nothing else. That frame is now too narrow. The real bill spans electricity, water, land, minerals, and electronic waste, and these footprints can move in opposite directions. Cut one, and you can quietly enlarge another. Low-carbon does not automatically mean low-water or low-land.

The UN researchers call for a responsible approach built on a few clear principles. Transparency about real consumption. Efficiency designed in by default, not bolted on later. Environmental justice for the communities that carry the burden. Responsibility across the full life cycle. International cooperation. And mandatory disclosure of carbon, water, and land impacts for AI systems and the data centers behind them.

None of that slows innovation. It makes it honest.

The ownership test

So here is the uncomfortable part for any leader deploying AI.

If you cannot see your AI’s resource bill, you do not control your AI. You are renting it, footprint and all.

The fix is not complicated, and it starts this quarter.

Measure inference, not just training, because that is where almost all of the consumption hides. Make your vendors disclose water and land, not only price and latency, and treat a refusal to disclose as a warning sign. Match the task to its cost, so you are not generating a video where a line of text would do. Put the resource footprint on the same line of the business case as ROI, because in time it will move the same numbers.

Then name an owner. Right now, in most organizations, the AI resource bill belongs to no one. That is the single easiest thing to fix and the single most important. Control follows accountability, and accountability follows a name.

The winners of this era will not simply be the companies running the most powerful systems.

They will be the ones running them with their eyes open.

Because the future of AI will not be decided only by how intelligent the technology becomes.

It will be decided by how intelligently we choose to use it.

https://news.un.org/en/story/2026/06/1167658

https://collections.unu.edu/eserv/UNU:10647/UNU-INWEH-Report-The_Env_Cost_of_AI-2026.pdf