01 logo

Every Time You Ask AI Something, Someone Somewhere Is Paying the Environmental Price

A new United Nations report has put a number on what AI actually costs the planet. The figure is staggering and it is about to double.

By Mark Lim Published 6 months ago • 5 min read
Every Time You Ask AI Something, Someone Somewhere Is Paying the Environmental Price
Photo by İsmail Enes Ayhan on Unsplash

There is no smoke coming out of your phone.

When you type a question into ChatGPT, ask an AI to write something for you, or generate an image from a text prompt, the experience feels weightless. Clean. Instant. A few words in, an answer out, no visible trace of anything having happened. It feels nothing like driving a car or running a factory, the kinds of activities we associate with environmental cost.

That feeling is wrong. And a new report from the United Nations University has put precise, sobering numbers on exactly how wrong it is.

The Scale of What We Have Already Built

Last year, global data centres, the physical infrastructure that makes AI and cloud computing possible, consumed 448 trillion watt-hours of electricity. To put that in context: only ten countries in the entire world used more energy. The electricity required to run these facilities produced approximately 208 million tons of carbon dioxide, roughly equivalent to the annual emissions of Argentina. Generating that electricity consumed about 1.2 trillion gallons of water.

These are not projections or estimates of a possible future. These are last year's numbers. The infrastructure already exists, already consumes at this scale, and is already leaving this footprint largely invisible to the people whose queries and prompts are driving it.

By 2030, according to the report, data centres will account for nearly 3% of the world's total projected electricity use, consuming 935 trillion watt-hours annually. If data centres were a country, they would rank sixth in the world for power consumption. The associated carbon dioxide output would reach nearly 440 million tons per year.

Kaveh Madani, a water scientist and director of the United Nations University Institute for Water, Environment and Health, was direct about what these figures represent: "If you look at these numbers, we're seeing scales comparable to nations. The demand is enormous."

AI Is the Engine Driving This

Not all of that energy consumption is attributable to AI. But AI is the fastest-growing part of it, and the trajectory is steep.

Currently, about 20% of data centre energy use is driven by AI workloads. By 2030, that figure is projected to reach 40%. The reason is straightforward: AI systems are extraordinarily energy-intensive compared to conventional computing. A typical ChatGPT-style query uses about 200 times more energy than a basic text classification task, the kind used in email spam filtering, for example. AI-generated images and videos require substantially more still.

Training large AI models is itself enormously costly. GPT-3 required approximately 1.3 billion watt-hours to train. The following version required between 50 and 70 billion watt-hours, a roughly fiftyfold increase for a single training run. These are one-time costs, but they are enormous ones.

Here is the part that often surprises people: training is not actually where most of the ongoing energy goes. About 90% of AI's power consumption comes not from training models, but from running them for the billions of daily queries that people send to AI systems. GPT alone handles approximately 2.5 billion prompts per day. Every one of those prompts consumes energy. Every one of those prompts generates heat that requires cooling. Every one of those prompts draws on water and electricity in quantities that add up, at scale, to something that rivals the footprint of entire nations.

The Paradox of Efficiency

The standard response from technology companies to environmental concerns about AI is efficiency. Chips are becoming more efficient. Data centres are becoming better at managing power. Renewable energy is increasingly being used to run these facilities.

All of that is true. None of it resolves the problem.

There is a well-documented phenomenon in economics and environmental science sometimes called the rebound effect: when something becomes more efficient, it tends to get used more, and total consumption rises even as individual use becomes cheaper or cleaner. More efficient AI chips mean more AI applications, more queries, more data centres, and more total energy consumed.

The renewable energy argument has its own complication. When a data centre powers itself with solar or wind energy, that clean electricity is no longer available for other uses, such as homes, hospitals, and public transport. The dirty energy that would otherwise have been displaced ends up being used somewhere else instead. The net environmental benefit is smaller than it first appears.

Madani is clear-eyed about what this means: "AI is not just a virtual thing. We're talking about something that has physics, something that has real impacts. There is infrastructure there. There is energy that is being used. A lot of hardware is behind all these operations that to us seem very, very clean because we don't see smoke out of our devices. But somewhere else, someone is suffering."

The Transparency Problem

One of the most significant obstacles to addressing AI's environmental footprint is that much of the relevant data simply is not publicly available. Many companies and jurisdictions do not disclose how much energy their data centres consume, how much water they use for cooling, or even where their facilities are located and how large they are.

The UN report is significant partly because of who produced it, a United Nations institution carrying global credibility and authority, and partly because it attempts to bring carbon, water, land, life-cycle impacts, and environmental justice into a single analytical frame. Fengqi You, an energy engineering professor at Cornell University who directs the university's AI sustainability research, noted that this is the first time a UN institution has done this comprehensively for AI.

"We cannot manage what companies do not disclose," You said.

Jean Su, director of the Energy Justice Program at the Center for Biological Diversity, called the report the first global examination to meaningfully shine a light on AI's environmental harms.

The technology industry, for its part, emphasises benefits. The National Artificial Intelligence Association pointed to AI's potential to improve safety, extend lives, increase efficiency, enhance food production, and reduce poverty, arguing that the energy return on investment is transformative and worth the cost. The Data Center Coalition said the industry takes its environmental responsibilities seriously and is committed to growing responsibly and transparently.

These responses are not wrong. AI does offer genuine benefits. The industry is genuinely investing in efficiency and renewable energy. But benefit and harm can coexist, and acknowledging one does not require pretending the other does not exist.

The One Thing You Can Actually Do

Here is the part of the report that is both practical and slightly absurd: you can meaningfully reduce the energy cost of your AI queries by being less polite and more concise.

Reducing the word count of your prompts by 30% reduces the energy consumed by approximately 25%, according to the report. That saving, aggregated across all AI users globally, would be equivalent to the annual electricity consumption of around 700,000 people in Africa.

Madani put it with characteristic directness: "If you're too polite, then that extra 'please' you put there can make a huge difference. You've got to be very precise and be short."

It is a small thing. It will not offset 448 trillion watt-hours of annual consumption. But it is a concrete illustration of something the report is trying to communicate at a larger scale: AI use has a cost, that cost is physical and real, and the choices made by the people and institutions using AI, from individual query length to corporate disclosure practices to government regulation, will determine how large that cost ultimately becomes.

The smoke is not coming out of your phone. But it is coming from somewhere.

tech news

About the Creator

Mark Lim

Hi I am mark an automotive student and a car, tech and food enthusiast ! Im gonna try and post daily & hope you enjoy what I write and do share my page with people you know. I would gladly appreciate it! Cheers

Enjoyed the story? Support the Creator.

Subscribe for free to receive all their stories in your feed.

Subscribe For Free

Reader insights

Comments

There are no comments for this story

Be the first to respond and start the conversation.

Sign in to comment
    Written by Mark Lim