Does Using AI Hurt the Planet? A Straight Answer

It is a fair question, and it is one we get a lot. You hear that AI is burning through electricity and draining water, and then you hear someone else say it is no big deal, and you are left not knowing who to believe. So here is our honest attempt to put it in perspective.

One query is about one second in the microwave

Let us start with the number that surprises us most. A typical text question to a modern AI model uses somewhere around a third of a watt hour of electricity. Picture that as a 1,000 watt microwave running for one second.

For a while the numbers floating around were ten times higher than that, and a lot of the worry comes from those older figures. But the models have gotten dramatically more efficient. One of the largest providers reported its typical query uses about 33 times less energy than it did a year earlier.

That does not mean it is free. Generating an image uses more, maybe ten times a text query. A few seconds of AI video uses a lot more, on the order of running the microwave for an hour. The point is not that AI uses nothing. It is that for the everyday business uses we help people with, answering questions, drafting copy, organizing information, the footprint of any single task is quite small.

The real lever: the right model for the right job

Not every job needs the biggest, most powerful AI.

Think of it like your vehicle. You do not fire up the work truck to go grab a gallon of milk a couple blocks away. You take the smaller car, or you walk. The truck is the right tool when you are hauling large items, and the wrong tool when you are not. AI is the same. A lot of the work a business needs is simple enough that a small, efficient model handles it perfectly, using a tiny fraction of the energy of the giant models that grab all the attention.

When we build solutions for a business, matching the size of the model to the actual job is one of the first things we think about. It is better for the environment, and it happens to be cheaper and faster too. The ecological choice and the practical choice point the same direction, which is the kind of thing we like.

You can already see the industry building toward this. There are platforms now, like FreeToken, designed so that an app runs AI right on your phone or computer when the device can handle it, and only reaches out to the big cloud models when it actually needs to. When the work runs locally, the energy cost and the dollar cost both drop close to nothing, and your data never leaves your device. Right model, right job, automatically.

The two concerns worth taking seriously

The per-query footprint is small. But zoom out to the infrastructure level and the picture changes, and we think you deserve a straight answer on both fronts.

Electricity. Data centers used about 415 terawatt hours of electricity globally in 2024, roughly equivalent to the entire annual electricity consumption of France, and the IEA projects that figure will roughly double by 2030, with AI identified as the most important driver. That growth is already showing up in utility bills: residential electricity prices jumped 7.1 percent in 2025, more than double the inflation rate. The mechanism is straightforward. Utilities build new infrastructure to serve data centers and ratepayers cover the cost, because that is how the utility business model has always worked. Your individual query is not what moves those numbers. The energy-intensive work is training the models in the first place, which happens once, not every time you use one. And between 2020 and 2025, the power density of AI servers increased eleven times, a dramatic increase in efficiency. But we will not tell you the concern is nothing. Of the two issues on this list, it is the one we watch most closely.

Water. Water is where you hear the most alarming headlines, but the per-query number is smaller than you might expect. Sam Altman published figures in June 2025 claiming a typical ChatGPT query uses roughly one-fifteenth of a teaspoon of water. Take that with some salt, the figure comes from OpenAI itself, what counts as an average query is unclear, and training runs are not included. But it is in the same neighborhood as other independent estimates, and it points in the same direction: for everyday text tasks, the per-query water footprint is genuinely small. The bigger and more legitimate concern is at the infrastructure level, and it is not evenly distributed. A data center drawing millions of gallons a day for cooling is a real burden in Phoenix or the Central Valley, and barely registers in a region with abundant water. The worry is not AI in the abstract, it is where these facilities get built and whether the communities nearby have water to spare.

The horizon: real intelligence, running local

The most encouraging part is where this is all headed. The frontier of AI research right now is not only about making models bigger. A growing share of the smartest work is about making them denser, getting more capability out of far less.

A company called PrismML, working out of research at Caltech, is a good example. They are building what are called 1-bit models, a way of shrinking AI down so dramatically that it can run on an ordinary phone or laptop instead of a massive data center. Their models take a fraction of the memory, run several times faster, and use a fraction of the energy, while holding their own against much larger systems. The whole idea is intelligence per bit rather than raw size.

If that direction holds, and we think it will, a lot of useful AI ends up running quietly on the device in your pocket, fast, private, and using barely any power or water at all. The future of this technology is not necessarily more sprawling server farms.

Where we land

So is AI bad for the planet? The honest answer is that it depends, and what it depends on is where these data centers get built and how they are run. The footprint of a single everyday task is small. That much is settled. The weight sits at the infrastructure level, and it does not land evenly. A data center on a clean grid in a place with water to spare is one thing. The same building drawing millions of gallons a day from a region already short on water, or dropped next to dense housing where the neighbors absorb the cost in their utility bills and their air, is another thing entirely. The shift toward smaller, local models pushes the right way, moving more of the work onto the device in your hand and taking load off the grid. The efficiency side keeps improving on its own. Where these buildings go does not. That part is a choice. And it is the reason for more optimism than dread: the heaviest costs are the ones we as a society can still choose to avoid through thoughtful state and local legislation.