What's the Impact of AI? (Part 2): What is the energy consumption of AI?

This post first appeared in our weekly Make Waves Mondays email series on August 31, 2026.

---

Hello hello friend!

After taking an extra week to do some extra research and calculations, we are BACK with Part 2 in our Impact of AI series 💪

✨ If you haven’t read Part 1 of this series yet, you can do so on our blog here. I’d definitely recommend reading (or listening to!) that one first!

If you recall from Part 1, I mentioned that this would be at least a 3-part series, potentially 4-part depending how deep I go on a few things, and wouldn’t ya know it, I’ve already had to adjust my plans to make this a 4-part series, officially 😅

So instead of today’s blog post being about both the energy consumption and water consumption of Artificial Intelligence, today we’re just gonna talk about energy consumption, and next week we’ll get into water consumption (and other water impacts). 

Finding accurate representations of the energy use associated with AI proved significantly more challenging than I anticipated — for several reasons:

First, because while there are a plethora of articles that say things like “One ChatGPT query uses about as much energy as it takes to use a light bulb for about 20 minutes,” (6) or “One ChatGPT query uses 10x as much electricity as a single Google search query,” (6, 10, 23), basically none of these articles actually give the numbers to back up those claims.

Which makes you wonder where those comparisons originated from. Who was the first one to make those claims, and were they ~just~ plausible-sounding-yet-click-bait-y enough that they’ve been recirculated again and again all over the internet? I had to find out.

And second, AI is moving so damn fast. By the time I finish writing this blog post, there’s gonna be something new. It’s almost impossible to keep up, and a big part of why it took me two years to actually start this series.

So given all of that, I’d like to start off today’s post with a ‘lil Editor’s Note (that’s me 😁):

All of the information included in this post, including calculations and estimations, is based on my research and data analysis across dozens of articles and studies, using the best available data we could find. These calculations and estimations should be viewed as contributory insights rather than definitive facts. There are many assumptions that need to be made when conducting an analysis like this, and I’ve done my best to make clear those assumptions throughout this series. Given how much this analysis relies on various assumptions and how quickly Artificial Intelligence and energy efficiencies are developing, the numbers included today may be out of date tomorrow, but the goal of this post is to provide context and overarching trends surrounding the environmental impact of AI, based on the most recent available data.

And so with that out of the way, let’s dive in 👇

What is an AI query?

Let’s start with the basics: what exactly is a query? Because everything else we talk about today is going to build off of one query.

💡 A query is what you’d type into an AI platform or LLM to get a response from the AI. And each time you respond to the AI, that’s another query.

So a single query could be something as simple as asking Gemini or ChatGPT what the weather is supposed to be tomorrow, or something as complicated as asking Claude to create for you an image of a porcupine doing ballet on a balloon. Then let’s say Claude makes your image with a yellow balloon, but you think a purple balloon might look better, so you ask Claude to change the balloon to purple — that’s another query.

Every single day, ChatGPT alone processes at least 2.5 billion queries. (81) That’s ONLY ChatGPT — it doesn’t include any of the other popular LLMs. And that number is from a year ago.

We’re still sprinkling in pretty nature photos throughout this series because we deserve pretty nature. Photo credit Fabian Quintero via Unsplash.

What’s the energy consumption of a single AI query?

Okay so here’s the thing about calculating AI energy consumption…

There is very little information available about how much energy is used by LLMs, because the companies behind these platforms have little incentive to share that information. (19) Basically everything these AI companies do, they do behind closed doors in the name of “proprietary information.” Which, again, is STEEPED in irony, considering their whole schtick is that they need to be able to steal copyrighted media in order to have a “viable” business model 🙄

But here’s what we do know:

In the fall of 2025, Google released some limited data, which included that the “median prompt” that’s given to Gemini consumes about 0.24 watt-hours (Wh) of electricity. (63) This does not take images or videos into account, and we don’t know what the high- or low-end numbers are. In this report, Google also claimed that the energy used for a Gemini query was 33x higher in May 2024 than in May 2025. (63) 

We also know that in a June 2025 personal blog post, Sam Altman claimed that one “average query” in ChatGPT uses about 0.34 Wh of energy. (65)

But I have a very hard time believing these numbers.

Last year, MIT did a heck of a deep dive into the energy consumption of AI — and if this kinda stuff interests you, it’s definitely worth a read!

The researchers tested an AI that contained 8 billion parameters by asking it to do various text-based prompts, such as creating a travel itinerary or explaining quantum computing. (19)

💡 A parameter is “essentially the adjustable ‘knobs’ in an AI model that allow it to make predictions.” (19) It’s the stuff that the AI uses to map the inputs to the outputs (which we talked about with how AI is trained in Part 1). The more parameters an AI has, the more accurate its answers tend to be — and the more energy it consumes to produce a response.

One query in this 8-billion-parameter model consumed about 114 joules of energy, or about 0.032 Wh.

Then they tested an AI that contained 405 billion parameters.

In this 405-billion-parameter model, one query consumed about 6,706 joules of energy, or about 1.86 Wh.

But according to this same MIT study, ChatGPT-4 has about 1 trillion parameters.

While MIT didn’t do the math on an estimated 1-trillion-parameter model, I decided to do some back-of-the-napkin math:

• The 8-billion-parameter model used about 0.0000000143 joules of energy per parameter for one query.
  
  

• The 450-billion-parameter model used about 0.0000000166 joules of energy per parameter per query.
  
  

• If we average those two numbers out, we can estimate that one AI query uses about 0.0000000154 joules of energy per parameter.
  
  

• That means that a 1-trillion-parameter model uses about 15,404 joules of energy per query, or about 4.28 Wh.

And let’s also not ignore the fact that MIT consulted with experts on these numbers, and all of those experts believe that the actual energy consumption numbers are higher than the estimates.

💡 I also just learned today (after reviewing more than 80 articles at this point!) that ChatGPT is actually now on model 5.5, and based on some quick Googling, estimates are currently putting GPT-5.5 parameters around 9.7 trillion. For the purposes of this blog post, we’re gonna move forward with the 1 trillion parameter estimate for GPT-4, but this really underscores my editor’s note at the top: things are moving FAST, and a lot of this information is not transparent or available to us.

The MIT study also looked at image and video creation with LLMs.

They found that one 1024x1024 pixel image produced with Stable Diffusion 3 Medium (an open-source image generator with 2 billion parameters) uses about 2,282 joules of energy, or about 0.63 Wh. With improved image quality, the researchers said, those numbers jump up to 4,402 joules, or about 1.22 Wh.

And again, these numbers are from early 2025. The tech has improved a lot since then — think about how much more realistic AI images have gotten in the last 18 months — so these are definitely low estimates.

Doing some more back-of-the-napkin math for AI image creation:

• The 2-billion-parameter model consumes about 0.0000022 joules of energy per parameter for one improved-quality image.
  
  

• If ChatGPT-4 has 1 trillion parameters, that would be 2,201,000 joules of energy per image, or about 611.4 Wh per image.

For video creation, the MIT researchers found that a single 5-second video created in OpenAI’s Sora consumes about 3.4 million joules of energy, or about 944 Wh.

But again...estimates for these kinds of numbers vary WIDELY, and are changing constantly.

• In 2023, a peer-reviewed study from a PhD candidate in Amsterdam calculated that one ChatGPT query uses about 2.9 Wh of electricity. (70)
  
  

• In 2024, the International Energy Agency agreed that one ChatGPT query uses about 2.9 Wh of electricity. (66)
  
  

• Also in 2023, someone much more technical than I am did their own deep dive into the energy impacts of AI, compiling numbers from various sources, and found that estimates ranged from 1 Wh per query to 10 Wh, but the average was about 2.67 Wh. HOWEVER, the author updated their article in January 2025 that their original numbers were based on GPT-3.5, and based on several other sources, GPT-4 “likely consumes up to 3x more energy than GPT-3.5,” which would bring that average ChatGPT-4 per-query energy use up to 8.01 Wh. (69) And remember…we’re now up to GPT-5.5.

So when Google and OpenAI are putting out claims that their LLMs use somewhere between 0.24 and 0.34 watt-hours of energy per query, but all of these other reputable sources are putting the numbers somewhere between 1.86 and 8.01 watt-hours (with data, reasoning, and references clearly laid out), I have a real hard time believing Big Tech.

Photo credit Leonard Cotte via Unsplash.

How much energy does one AI query use compared to one Google search?

Lemme tell you, friend, despite the claim that one AI query uses 10 times as much energy as a Google search being EVERYWHERE, finding the actual numbers behind that claim was not easy 😪

The only time Google has reported how much energy a single Google search uses was way back in 2009, at 0.3 watt-hours per search. (68)

But the internet and energy efficiencies have changed TREMENDOUSLY in the last almost-two-decades. And especially if Google is self-reporting that one Gemini search consumes 0.24 Wh of electricity, this 0.3 Wh number for a Google search feels way off.

Thank goodness I’m not the only one desperately trying to find an updated number for this, because that same 2023 deep dive I just mentioned above also did a whole boatload of calculations to determine that these days, a single Google search uses about 0.04 Wh of electricity. (69)

So when we compare that updated number to our per-query estimates above, we can estimate that one AI query uses anywhere from 6 to 200 times as much energy as one non-AI Google search — even including the super-low AI estimates from Google and OpenAI!

All of this energy consumption comes from data centers.

Artificial Intelligence (and the entirety of the internet and the cloud) doesn’t just appear from nowhere — it all lives on servers that are housed in data centers.

💡 Data centers are “facilities that house IT infrastructure for building, running, and delivering applications and services. It also stores and manages the data associated with those applications and services.” Today, most of the data centers being built for AI are hyperscale data centers, which are MASSIVE facilities that can contain thousands of servers. (82) For example, one of the largest data centers in the world, Stargate, is currently under construction in Abilene, Texas. Once it’s completed, Stargate will include 4 million square feet of buildings across 1,100 acres — that’s larger than Central Park in New York City. (27)

Data centers have been around since the 1940s, but in the early 2000s when “the cloud” came onto the scene, the need for data centers grew rapidly, as did their size. The first hyperscale data center was built by Google in 2006 in The Dalles, Oregon — which we’ll likely hear much more about in Part 3 of this series. (82)

Because data centers keep all of our devices and systems running all the time, the data centers themselves need to be running all the time.

Every minute of every hour of every day, data centers need power. (19)

And because of that, data centers currently can’t rely on potentially-intermittent clean energy sources, like wind and solar. (19)

However… Despite more and more cloud-based services being used (things like social media and Netflix), all of the energy consumption by data centers from 2005 to 2017 remained pretty flat, thanks to increases in energy efficiencies.

But in 2017, AI changed the game. By 2023, data center energy consumption had doubled. (19)

A few quick stats:

• In 2015, there were about 3,600 data centers worldwide. In 2024, there were more than 7,000. (6) Today, there are more than 12,250. (56)
  
  

• In 2022, data centers were responsible for about 3% of all energy use in the United States. (6) By 2024, that number had grown to about 4.4%. (19) And by 2030, it’s currently predicted that data centers will consume 8–12% of all energy in the US. (6, 12, 19, 21).
  
  

• Energy consumption from data centers is growing faster than any other segment of our society. (67)

So how much energy are these data centers actually using?

According to that MIT study from before, data centers in the US used somewhere around 200 terawatt-hours (or 200 trillion watt-hours) of electricity in 2024. That’s roughly what it takes to power the entirety of Thailand for a whole year. Zooming in on AI-specific servers in these data centers, those alone are estimated to have consumed anywhere from 53 to 76 terawatt-hours of electricity in 2024. On the high end of that spectrum, that’s enough energy to power more than 7.2 million homes in the US for a whole year. (19)

And that was 2024. From Part 1, we know that ChatGPT users grew from 300 million a week to 900 million a week between December 2024 and February 2026.

By 2030, the data centers in northern Virginia alone are anticipated to consume the same amount of energy as we currently need to power 6 million homes. (6) That’s just one region of one state in one country…

On the whole, AI energy use is growing at an estimated rate of 26–36% annually. (21)

Our demand for AI (or rather, Big Tech’s incessant need to shove AI down our throats at every turn) is growing our energy demands at a rate that’s simply unsustainable.

Where is all of that energy sourced from?

Making all of this worse is the fact that data centers aren’t being built in areas with predominantly renewable grids.

A Harvard study found that the electricity used by data centers actually has a carbon footprint that’s about 48% higher than the average in the US (19), because combined with the fact that they need energy 24/7/365 and therefore can’t rely on potentially-intermittent energy sources like wind and solar, they’re also just located in areas that have dirtier grids — like coal and natural gas. (6, 19, 22, 67)

Now proponents of AI love to argue that the rise of AI and the need for data centers is going to fuel and accelerate the transition to clean energy, but we’re actually seeing the opposite: coal-fired power plants are staying online longer than planned specifically due to increased energy demand from data centers (7, 20), and many of these hyperscale data centers are building their own power plants with natural gas, diesel, and methane (19, 27, 67).

And because our current administration is just ~the bestest~ (HEAVY sarcasm here if you didn’t catch it), the EPA has even drafted a plan to eliminate all limits on greenhouse gas emissions from power plants, and now the demand on power plants is growing at a rate faster than grid capabilities, and faster than we’re seeing renewable energy growth. All-in-all, there’s very little regulation around data center energy use. (23)

As Mike Weinstein, the director of sustainability at Southern New Hampshire University has said: 

“If all of our electricity supply was clean and renewable, like that provided by solar or wind power, we could look at the other impacts of AI in a different context. But unfortunately, we as a planet are tremendously behind in transitioning to a clean, non-carbon energy grid.” (22)

Photo credit Mark Basarab via Unsplash.

Who pays for all of this increased energy use?

I can’t move on from energy use without touching on something that really flippin’ irked me from all of this research: these AI companies aren’t really the ones paying for all of the energy they’re using.

Let’s look at Santa Clara, California, for example.

Silicon Valley Power, Santa Clara’s utility company, gives data centers discounts on their electric rates.

For individual households, the per-kilowatt-hour electricity rate increases above a certain threshold. (So for example, let’s say that threshold is 1,000 kWh per month. If you use less than 1,000 kWh in a month, Silicon Valley Power will charge you less per kWh than if you use more than 1,000 kWh.)

But for data centers, it’s reversed — data centers are charged LESS per kWh the more energy they consume. And all the way back in 2023, these data centers were already using 60% of the entire city’s electricity, in a region already prone to blackouts. California ranks 49th out of 50 states in their energy resilience, or their “ability to avoid blackouts by having more electricity available than homes and businesses need at peak hours.” (12)

Historically, electric rates in Santa Clara rose steadily by about 2–3% each year. 

But in January 2023, they rose by 8%.
In July 2023, they rose by another 5%.
And in January 2024, they rose by another 10%. (12)

And this isn’t unique to California. This is happening everywhere.

Big Tech is forcing AI on us everywhere we turn, and then sticking us with the bill.

How does all of this translate to carbon emissions?

Google, Meta, Amazon, and Microsoft had all at one time pledged to use 100% renewable energy, but with the rise of AI, those commitments have fallen off. (20)

Google’s own 2024 sustainability report stated, “As we further integrate AI into our products, reducing emissions may be challenging,” and Microsoft’s stated that, “The infrastructure and electricity needed for these technologies create new challenges for meeting sustainability commitments across the tech sector.” (6) 

YET WE CAN’T OPT OUT OF IT. As of earlier this year, Google phones are even completely eliminating Google Assistant (aka the voice-activated tool that lets you do simple things like “call mom” or “add bread to my grocery list” while you’re driving so you don’t have to touch your phone) and replacing it with Gemini. WHO ASKED FOR THIS?!

Anyhoops…

Between 2020 and 2023, Amazon’s emissions grew 182%, Microsoft’s grew 155%, Meta’s grew 145%, and Google’s grew 138%. (20)

Revisiting our previous numbers, in 2024, AI-specific data centers in the United States consumed about 76 terawatt-hours of electricity, and that energy has a footprint that’s about 48% higher than the national average (19). 

That means that in 2024, AI emitted about 44,264,000 metric tons of carbon dioxide equivalent. 

➡️ That’s the same emissions as more than 10.3 million cars emit in a year, or more than 9.2 million homes’ electricity use per year, and would require nearly 732 million trees to be planted and grown for 10 years to offset.

At what point do we ask ourselves if AI is worth all of that?

“Transforming the electric grid into the clean energy machine we need and deserve won’t happen on private grid islands owned by tech companies. It will emerge from coordinated efforts by electricity grid operators, regulators, utilities, and policymakers. Data center developers will say they’re building the future, but we should first and foremost be asking: what kind of future do we really need?” – Project Drawdown (67)

Meanwhile, other industries that rely on data servers, such as large telecom companies, are successfully lowering their emissions! Green America’s Hang Up On Fossil Fuels campaign urged telecom companies to switch to 100% renewable energy and as a result these companies have reduced their emissions by 94% since 2020. (20) So it is possible!

How does AI’s energy use stack up against other everyday activities?

You see it all the time on social media: some AI bro thinking he’s really got you by saying something like, “If you’re so against AI you better not be watching Netflix!”

So to close out today’s blog post, I went on the hunt for the energy use data from a bunch of everyday activities to see how they really stack up compared to using an LLM (and remember — these estimates are low-end estimates based on GPT-4, when we’re now on GPT-5.5):

So yeah, sure, one singular ChatGPT query may be relatively low on the emissions scale…but BOY does it add up FAST.

The average ChatGPT session is just 6 minutes long, yet uses 43% more energy than watching Netflix on a 50” TV for two hours.

An entire 8-hour work day on a laptop uses 92% less energy than creating one 5-second video with AI.

I could blow-dry my hair for almost 9 hours before I used the same amount of energy as creating one AI image.

So before I hand you back to your day, I’ll leave you with a ‘lil reminder from the great Jane Goodall:

“You cannot get through a single day without having an impact on the world around you. What you do makes a difference, and you have to decide what kind of difference you want to make.”

Is AI the kind of difference you want to make today?

 

💡 View all of our sources used for this blog post series here.