Why Is AI Bad for the Environment: How Much Energy Does AI Use
7 sections
AI’s environmental cost comes mostly from electricity, not from any single chatbot reply: data centres used about 1.5% of global electricity in 2024, and the International Energy Agency projects that could roughly double by 2030. The technology to run AI on cleaner power and more efficient hardware largely exists already. What’s missing is regulation and the corporate will to use it, not the engineering itself.
If you run a SaaS, ecommerce, or agency business that now leans on AI tools every day, you’ve probably had the environmental question thrown at you by a client, a journalist, or your own conscience. This piece walks through how much energy AI actually uses, what its carbon and water footprint really looks like, and what’s a business choice versus what’s structural.

Why is AI bad for the environment#
AI isn’t bad for the environment by default. Unmanaged AI is, and that distinction carries the whole argument of this piece. I went through EIS by EFdeN, Romania’s first sustainability leadership program: a program built around five modules and a final project. That doesn’t make me an environmental scientist, and I won’t pretend otherwise. It did push me to read the underlying energy reports instead of repeating headlines. The picture in those reports is messier than “AI is killing the planet.”
AI’s footprint looks different from a normal web app because it’s never just one server answering one request. Training a model is a multi month compute binge. Running it afterward, what the industry calls inference, happens millions of times a day, every day, and that ongoing use is where most of the electricity goes over time. Then there’s the hardware: chips, racks, and cooling loops, all built and shipped before a single query runs. Put that on a global electricity grid that still leans heavily on fossil fuels, and the climate cost of AI scales with how popular it gets, not with how clever it is.
The technology isn’t what’s broken. What’s broken is that almost nobody building or regulating AI treats its energy bill as a design constraint. The tools to run it cleaner already exist. Very few companies are required to use them, and that gap between what’s possible and what’s mandatory is the real story behind AI and climate change.
How much energy does AI use#
Data centres, the infrastructure that runs AI training and inference alongside ordinary cloud computing, used about 1.5% of global electricity in 2024, roughly 415 terawatt hours, according to the IEA’s Energy and AI report. The same report projects that figure could reach around 945 TWh by 2030. That’s more than double in six years. It’s the single most important number in this whole debate, because it tells you AI’s energy problem is a growth problem, not a current crisis.

This is where I push back on how most people frame it. “Is AI using too much energy right now?” is the wrong question. The efficiency of AI models keeps improving per task: newer chips and better architectures do more useful work per watt than older ones did. The problem is that total electricity use keeps climbing anyway, because demand grows faster than those efficiency gains can offset. That compounding drives the jump from 415 to 945 TWh. AI didn’t get less efficient; far more people are running it, more often, on heavier tasks. That growth curve doesn’t look like anything else in enterprise software.
Reasoning models make this worse before it gets better. A model that works through a problem step by step generates many extra tokens behind the scenes before it writes the answer you actually see. Each hidden token is another pass through the model, and each pass costs electricity. Swap a quick, standard model for a reasoning model on a task that never needed one, and you’re spending a lot more compute for the same output. Teams make that design decision. Physics doesn’t.
What is the carbon footprint of ChatGPT#
Nobody outside the companies running these models can give you an exact CO2 number for a single ChatGPT query. Anyone handing you a precise gram figure is estimating, not reporting. Training runs and inference workloads for a model like ChatGPT draw on the same data centre electricity described above, and that electricity’s carbon cost depends entirely on what’s powering the grid underneath it. Same model, same query, different country, different footprint. That’s the current state of public disclosure, not an excuse.
Zoom out to the IEA’s view and the picture gets calmer than the headlines suggest. The emissions increase tied to data centres is small in the context of the entire energy sector’s emissions. It isn’t nothing, but it isn’t the standalone climate emergency some coverage implies either.
The more interesting part of that argument is the offset side. If AI gets adopted widely across the economy, in things like optimizing power grids or cutting industrial waste, the emissions reductions it enables elsewhere could offset the increase from running it. That depends on adoption actually happening at scale, and on companies building those efficiency use cases instead of just shipping chatbots. I’d rather make a business decision on that honest condition than pretend ChatGPT is secretly carbon neutral, or that it’s single-handedly cooking the planet.
Lifecycle emissions of AI chips explained#
Every GPU in a data centre carries a carbon cost before it ever processes a token. Chip fabrication is one of the most resource and energy intensive manufacturing processes there is: wafer plants run cleanrooms around the clock and draw on global supply chains stretching from mines to foundries to assembly lines, each step adding its own emissions.
The mineral extraction underneath all of it gets less attention. The IEA’s Energy and AI report flags that data centre equipment, AI chips included, depends on critical minerals whose supply is highly concentrated among a small number of countries and producers. That concentration is a supply chain risk on its own, separate from the climate question. When a handful of suppliers control the minerals behind every AI chip, whoever holds that chokepoint sets pricing, availability, and extraction standards, not the companies buying the finished hardware.
None of this shows up on a monthly cloud bill. It’s baked into the hardware long before a company logs its first API call, which is exactly why most sustainability conversations about AI leave it out.
Cooling systems in data centers and water use#
A modern AI data centre needs serious cooling capacity because GPU racks pack far more heat into the same square footage than traditional servers did. Heat is the hard limit on how much compute a facility can run before hardware fails. Operators handle it with a mix of air cooling, liquid cooling, and evaporative systems that use water to pull heat out of the building. How much water a single site actually uses varies enormously with the cooling technology chosen and the climate the site sits in. A facility in a dry, hot region behaves nothing like one in a cool, wet one, even running identical hardware.
Now the honest counterpoint to everything I’ve argued so far: water and grid strain are real, local, immediate problems in a way global electricity percentages are not. A community living next to a data centre doesn’t experience “1.5% of global electricity.” It experiences its own water table, its own grid connection, its own construction traffic. Global averages can look manageable while a specific town genuinely struggles with supply. Nothing in the global numbers above resolves that tension. It’s a siting and permitting problem that has to be solved place by place, and right now it mostly isn’t.
Renewable energy for data centers#
Renewable power can cover a meaningful share of AI’s growing electricity appetite, but not all of it on its own. The IEA’s Energy and AI report finds renewables are set to meet about half of the growth in data centre electricity demand, with natural gas and nuclear also expanding to fill the rest of the gap. So the real split isn’t “AI runs on coal” or “AI runs on solar.” It’s a mixed grid where clean and fossil sources grow side by side.

That half-and-half split comes from choices, not from nature. A company deciding where to put a new data centre decides, by that choice, how much of its electricity load lands on a clean grid versus a gas-heavy one. A government deciding how fast to permit new transmission lines or nuclear capacity decides how fast the clean half of that split can grow. The technology makes neither decision. Both get made in boardrooms and permitting offices, which is exactly why I don’t buy the idea that AI’s energy mix is somehow fixed or inevitable.
Sustainable AI practices for businesses#
Most of the technology needed to run AI cleanly already exists:
Efficient chips: newer hardware does more compute per watt than previous generations.
Renewable power contracts: data centres can buy clean electricity directly instead of defaulting to the local grid mix.
Compute scheduling: heavy workloads can shift to hours or locations where the grid is cleaner.
What’s missing is regulation that forces companies to use these options, and enough interest from corporate leadership and governments to pay the premium for the cleaner default. Regulation is still catching up to how fast AI usage has grown inside companies. Clear, consistent rules on disclosing AI-related energy and emissions, or on how data centres get sited and cooled, exist in patches rather than as one coherent standard. Until that changes, the responsible choice stays voluntary, and most companies that aren’t required to disclose anything won’t.
I haven’t built a formal sustainability program around how difrnt. or GEOflux.ai use AI day to day, and I won’t pretend otherwise. What I do apply, from running a business that watches how brands show up inside AI tools, is the habit of measuring before claiming anything. Treat AI’s energy use the same way: measure it before deciding it’s fine, and before deciding it’s a disaster.
I co-founded GEOflux.ai to track exactly that kind of question for brands: how they actually show up inside ChatGPT, Gemini, and Perplexity. If measuring AI’s footprint matters to you, measuring your own footprint inside AI search results is probably worth a look too.