(2025-05-20) We Did The Math On AI's Energy Footprint; Here's The Story You Haven't Heard

We did the math on AI’s energy footprint. Here’s the story you haven’t heard. This story is a part of MIT Technology Review’s series “Power Hungry: AI and our energy future,” on the energy demands and carbon costs of the artificial-intelligence revolution.

we found that the common understanding of AI’s energy consumption is full of holes.

From 2005 to 2017, the amount of electricity going to data centers remained quite flat thanks to increases in efficiency, despite the construction of armies of new data centers to serve the rise of cloud-based online services, from Facebook to Netflix

In 2017, AI began to change everything. Data centers started getting built with energy-intensive hardware designed for AI, which led them to double their electricity consumption by 2023.

The latest reports show that 4.4% of all the energy in the US now goes toward data centers.

According to new projections published by Lawrence Berkeley National Laboratory in December, by 2028 more than half of the electricity going to data centers will be used for AI.

Meanwhile, data centers are expected to continue trending toward using dirtier, more carbon-intensive forms of energy (like gas) to fill immediate needs, leaving clouds of emissions in their wake.

We’re taking a different approach with an energy accounting meant to inform the many decisions still ahead: where data centers go, what powers them, and how to make the growing toll of AI visible and accountable.

leading AI companies and data center operators disclose too little about their activities.

This leaves even those whose job it is to predict energy demands forced to assemble a puzzle with countless missing pieces, making it nearly impossible to plan for AI’s future impact on energy grids and emissions

Before you can ask an AI model to help you with travel plans or generate a video, the model is born in a data center.

it’s estimated that training OpenAI’s GPT-4 took over $100 million and consumed 50 gigawatt-hours of energy

It’s only after this training, when consumers or customers “inference” the AI models to get answers or generate outputs, that model makers hope to recoup their massive costs

inference, not training, represents an increasing majority of AI’s energy demands and will continue to do so in the near future. It’s now estimated that 80–90% of computing power for AI is used for inference.

All this happens in data centers. There are roughly 3,000 such buildings across the United States

A growing number—though it’s not clear exactly how many, since information on such facilities is guarded so tightly—are set up for AI inferencing.

A single AI model might be housed on a dozen or so GPUs, and a large data center might have well over 10,000 of these chips connected together.

many buildings use millions of gallons of water (often fresh, potable water) per day in their cooling operations. (Update: see (2026-03-05) AI Data Centers What To Know About Their Water And Energy Use)

Depending on anticipated usage, these AI models are loaded onto hundreds or thousands of clusters in various data centers around the globe, each of which have different mixes of energy powering them.

In reality, the type and size of the model, the type of output you’re generating, and countless variables beyond your control—like which energy grid is connected to the data center your request is sent to and what time of day it’s processed—can make one query thousands of times more energy-intensive and emissions-producing than another.

Factors like which data center in the world processes your request, how much energy it takes to do so, and how carbon-intensive the energy sources used are tend to be knowable only to the companies that run the models.

Without more disclosure from companies, it’s not just that we don’t have good estimates—we have little to go on at all.

So where can we turn for estimates? So-called open-source models can be downloaded and tweaked by researchers, who can access special tools to measure how much energy the H100 GPU requires for a given task

But even if researchers can measure the power drawn by the GPU, that leaves out the power used up by CPUs, fans, and other equipment. A 2024 paper by Microsoft analyzed energy efficiencies for inferencing large language models and found that doubling the amount of energy used by the GPU gives an approximate estimate of the entire operation’s energy demands.

Here’s what we found.

Text models

their ML.Energy leaderboard.

The smallest model in our Llama cohort, Llama 3.1 8B, has 8 billion parameters

When tested on a variety of different text-generating prompts, like making a travel itinerary for Istanbul or explaining quantum computing, the model required about 57 joules per response, or an estimated 114 joules when accounting for cooling, other computations, and other demands. This is tiny—about what it takes to ride six feet on an e-bike, or run a microwave for one-tenth of a second.

The largest of our text-generation cohort, Llama 3.1 405B, has 50 times more parameters

On average, this model needed 3,353 joules, or an estimated 6,706 joules total, for each response. That’s enough to carry a person about 400 feet on an e-bike or run the microwave for eight seconds.

The parameter counts for closed-source models are not publicly disclosed and can only be estimated. GPT-4 is estimated to have over 1 trillion parameters.

Generating an image

The energy requirement instead depends on the size of the model, the image resolution, and the number of “steps” the diffusion process takes (more steps lead to higher quality but need more energy).

Generating a standard-quality image (1024 x 1024 pixels) with Stable Diffusion 3 Medium, the leading open-source image generator, with 2 billion parameters, requires about 1,141 joules of GPU energy.

That means an estimated 2,282 joules total

Improving the image quality by doubling the number diffusion steps to 50 just about doubles the energy required, to about 4,402 joules. That’s equivalent to about 250 feet on an e-bike, or around five and a half seconds running a microwave. That’s still less than the largest text model.

Making a video

The new model uses more than 30 times more energy on each 5-second video: about 3.4 million joules, more than 700 times the energy required to generate a high-quality image. This is equivalent to riding 38 miles on an e-bike, or running a microwave for over an hour.

It’s fair to say that the leading AI video generators, creating dazzling and hyperrealistic videos up to 30 seconds long, will use significantly more energy

All in a day’s prompt

“We should stop trying to reverse-engineer numbers based on hearsay,” Luccioni says, “and put more pressure on these companies to actually share the real ones.” Luccioni has created the AI Energy Score, a way to rate models on their energy efficiency.

Now that we have an estimate of the total energy required to run an AI model to produce text, images, and videos, we can work out what that means in terms of emissions that cause climate change.

*Most electrical grids around the world are still heavily reliant on fossil fuels. So electricity use comes with a climate toll attached.

“AI data centers need constant power, 24-7, 365 days a year*

carbon intensity of electricity used by data centers was 48% higher than the US average. Part of the reason is that data centers currently happen to be clustered in places that have dirtier grids on average, like the coal-heavy grid in the mid-Atlantic region

today, nuclear energy only accounts for 20% of electricity supply in the US, and powers a fraction of AI data centers’ operations—natural gas accounts for more than half of electricity generated in Virginia, which has more data centers than any other US state

In 2024, fossil fuels including natural gas and coal made up just under 60% of electricity supply in the US. Nuclear accounted for about 20%, and a mix of renewables accounted for most of the remaining 20%.

In April, Elon Musk’s X supercomputing center near Memphis (Colossus) was found, via satellite imagery, to be using dozens of methane gas generators that the Southern Environmental Law Center alleges are not approved by energy regulators to supplement grid power and are violating the Clean Air Act.

The key metric used to quantify the emissions from these data centers is called the carbon intensity: how many grams of carbon dioxide emissions are produced for each kilowatt-hour of electricity consumed

This intensity varies widely across regions. The US grid is fragmented, and the mixes of coal, gas, renewables, or nuclear vary widely. California’s grid is far cleaner than West Virginia’s, for example.

what does this all add up to?

In December, OpenAI said that ChatGPT receives 1 billion messages every day, and after the company launched a new image generator in March, it said that people were using it to generate 78 million images per day

In February the AI research firm Epoch AI published an estimate of how much energy is used for a single ChatGPT query—an estimate that, as discussed, makes lots of assumptions that can’t be verified. Still, they calculated about 0.3 watt-hours, or 1,080 joules, per message

One billion of these every day for a year would mean over 109 gigawatt-hours of electricity

If we add images and imagine that generating each one requires as much energy as it does with our high-quality image models, it’d mean an additional 35 gigawatt-hours

These estimates don’t capture the near future of how we’ll use AI.

leading labs are racing us toward a world where AI “agents” perform tasks for us without our supervising their every move. We will speak to models in voice mode, chat with companions for 2 hours a day, and point our phone cameras at our surroundings in video mode. We will give complex tasks to so-called “reasoning models” that work through tasks logically but have been found to require 43 times more energy for simple problems