No, your AI prompts aren’t causing an energy crisis

0:00 I'm not a big chatbot person.

0:03 I'll use ChatGPT every once in a while,

0:06 but recently I've been hesitating because I keep hearing

0:11 that using AI has a hidden cost of energy and water,

0:14 things that I try not to waste.

0:18 I started to wonder, should I treat using a chatbot,

0:20 like leaving the lights on or running the faucet?

0:23 So, I talked to a bunch of experts

0:25 to better understand how I should think about my AI

0:28 energy consumption and to figure out exactly how

0:30 much energy it takes every time I do this.

0:33 (enter key clacks) (mysterious music) (underscore beeps) The first

0:40 thing I wanted to understand was whether there

0:43 was a specific measurable number for just how much

0:46 energy it takes to ask a chatbot a question.

0:49 So I went looking for one.

0:51 (mysterious music) It was hard to find.

0:57 The companies behind AI chatbots generally

0:59 haven't been very transparent about those numbers,

1:02 but then I found this, it's a leaderboard ranking

1:05 large language models by how energy efficient they are.

1:09 That leaderboard is run by Jae-Won Chung,

1:11 a researcher at the University of Michigan.

1:13 I asked him if there's a useful average number that could

1:16 help me understand the energy cost of using a chatbot.

1:20 I am indeed asked that question a lot, and typically,

1:23 my answer is that I avoid the answer

1:26 because there's too much nuance to give one number.

1:29 (calm jazzy music) There's models with vastly different sizes.

1:34 You can run them on one hardware.

1:36 You can run them another hardware.

1:37 Some requests are super long.

1:39 Some people just shove in an entire book into ChatGPT or Gemini,

1:43 whereas someone just say, hi.

1:44 There could be easily a hundred X difference between

1:47 the lowest energy consumption and the highest energy consumption.

1:51 I really try to avoid giving one number,

1:54 although, many people would want that from me.

1:58 Okay, end of story, I guess.

1:59 Thanks for watching.

2:01 But then in the summer of 2025,

2:03 some companies actually did start to disclose some numbers.

2:08 In June, OpenAI CEO, Sam Altman, published a blog post claiming that the average

2:13 ChatGPT query uses 0.34 watt hours of energy.

2:17 And Google released a paper estimating that the median

2:20 text query on Gemini consume 0.24 watt hours of energy.

2:24 We can average those out for a simpler estimate, 0.3 watt hours per query.

2:33 Okay, but what does that actually mean?

2:35 Watt hours are one way to measure energy.

2:38 The classic 60-watt light bulb, if you leave it on for one hour,

2:42 consumes 60 watt hours of energy.

2:45 So, a typical chatbot query is roughly the same

2:48 as running that light bulb for 18 seconds,

2:50 brewing coffee for 10 seconds, playing a gaming console for six seconds,

2:54 or running a microwave for one second.

2:58 It's worth mentioning that the numbers

2:59 are higher for generating images or video.

3:02 One third-party estimate,

3:03 put the energy cost of generating an image at three watt hours.

3:06 That's like running a light bulb for three minutes

3:08 and generating a short video at 90 watt hours,

3:11 which is like running that light bulb for an hour and a half.

3:14 That's definitely not nothing.

3:16 Now, the number for a text prompt, 0.3 watt hours,

3:19 that only accounts for the cost of answering a prompt.

3:22 It also takes energy to train AI models,

3:25 and that's an important part of the total cost to consider.

3:28 Research from Google estimates that 60% of its AI

3:31 energy use goes towards answering prompts and 40% towards training.

3:35 So, the full cost per prompt is closer to 0.5 watt hours.

3:40 Those calculations come from Andy Masley.

3:42 I just meets so many people who are under the impression that if

3:44 you send a single chatbot prompt or if you just use chatbots regularly,

3:48 this is somehow especially harmful to the environment.

3:51 [Christophe] On his substack, he wrote a comprehensive guide on how to put

3:54 the environmental cost of a chatbot prompt in perspective.

3:58 My best guess right now is that each one

3:59 of those is probably gonna be around 150,000 of your daily emissions.

4:05 You said one 150,000.

4:08 Yeah, uh-huh, yeah, yeah, yeah.

4:10 You could save so much more CO2 emissions by just

4:12 like switching out a single burger for a plant-based

4:14 burger or just like driving a few feet less

4:17 or just choosing to take a walk somewhere, instead of driving.

4:19 Like that can buy you like tens of thousands

4:21 of chatbot prompts that you can then like feel fine about.

4:27 I just got a new passport.

4:29 This one is expired, but it is now a beautiful little diary of all

4:35 the places that I've been over the last 10 years,

4:39 I loved getting to fill this thing up.

4:41 I think travel can be incredibly rewarding.

4:44 It can also be very stressful.

4:45 But when I'm going abroad, there's one app that I use to make sure

4:48 that staying connected is checked off of my to-do list.

4:50 This is Saily.

4:51 It's an eSIM service from the creators of NordVPN,

4:54 and they have a special offer running right now at saily.com/tunnelvision.

4:59 Saily lets you instantly download an eSIM straight to your device.

5:03 You just open the app, pick from over 200 countries, and you're online.

5:06 You can even keep multiple carriers or numbers active at the same time,

5:10 and it works on both iOS and Android.

5:12 But if for any reason, your phone doesn't support eSIM,

5:15 they will give you a full refund.

5:16 Right now, you can save 15% on your plan

5:18 at saily.com/tunnelvision using the code tunnelvision at checkout.

5:23 It's a great product and a great way to support the show.

5:25 So, the next time you want to get

5:27 out and fill up your passport, check out Saily.

5:30 (mysterious music continues) Talking to Andy,

5:32 I felt most of my concerns about my individual footprint disappear.

5:37 Generating the response to a text query takes a negligible amount of energy.

5:42 Generating an image or a video takes a lot more,

5:45 but it is still almost nothing compared to plenty

5:48 of things that we do in our day-to-day lives.

5:51 But I also started to question how useful

5:53 of a measurement that an individual footprint even was,

5:57 because even if the individual footprint of using this technology is small,

6:01 the technology's total footprint isn't.

6:04 (mysterious music continues) Data centres,

6:09 the massive computing facilities that power AI tools made up

6:12 more than 4% of the US' total electricity consumption in 2024.

6:17 Some projections estimate that that number could grow to 12% by 2028.

6:22 Data centres aren't new, but the quantity and scale of their construction is.

6:27 If you go to datacentermap.com, you can see the quantity.

6:30 The site maps, the locations of more than 4,000 data centres

6:34 in the US zoom into any one of these locations on Google Earth,

6:37 and you can see the scale.

6:39 In Richland Parish, Louisiana, Meta is building a data centre big enough

6:42 to cover this much of the island of Manhattan.

6:45 It'll consume as much electricity as 2 million households.

6:49 Most of the electricity for data centres currently comes from fossil fuels.

6:53 Their power sources are 48% more carbon-intensive than the national average,

6:58 which is part of why big tech companies are investing

7:00 heavily in nuclear and renewable energy to power data centres.

7:04 Eventually, the carbon emissions of any

7:06 individual chatbot prompt should get smaller

7:08 as we start to power data centres with cleaner sources of energy,

7:11 but for the people living close to where those data centres are being built,

7:15 the impact is already here.

7:18 [Protesters] These data centres have got to go.

7:19 [Reporter] Residents are worried about the economy, water, and, electricity.

7:22 This is a bad deal.

7:24 We know the grid cannot take this,- And they

7:26 don't care about the negative impacts on the environment.

7:29 Electricity, the noise levels, the whole environmental issue.

7:33 Don't do this.

7:35 Don't sell out our town.

7:38 [Christophe] One of the major concerns for the communities living

7:40 near data centres is that their electricity bills will increase,

7:43 and it is true that in the US, the average household's electricity costs have

7:47 slowly climbed in the past decade.

7:49 I wanted to understand if that had anything to do with the rise of data centres.

7:53 Ari Peskoe leads the Electricity Law Initiative

7:56 at Harvard Law School and Michael Thomas runs Cleanview,

7:59 a clean Energy and data centre tracker.

8:01 And they told me the answer is yes,

8:02 data centres are driving up electricity bills, but not for the reason I assumed.

8:08 Most of the rate increases that we're seeing across

8:10 the country are due to utility spending on poles and wires.

8:15 The utility business model is to socialise the costs of new infrastructure,

8:21 that is to spread the cost to everybody who takes service from that utility.

8:28 [Christophe] That means the costs of any new physical infrastructure are spread

8:32 across the electricity bills of everyone in a utility company's service area.

8:37 I think that's what's happened in the places that prices are rising,

8:40 is the data centre comes and they say,

8:43 hey, we wanna build a gigawatt data centre.

8:45 The utility says, great.

8:47 I make all my money when I build power lines and then charge it to customers.

8:51 They suddenly have to build this new

8:53 power line that wouldn't needed to be built, if not for the data centre.

8:58 When that happens, then someone's bill goes up in a way

9:01 that it probably didn't need to, if not for the data centre.

9:04 [Christophe] So in the places where data

9:05 centres are making electricity bills go up,

9:08 it isn't just because those data centres require so much energy.

9:11 It's largely because of just how much new infrastructure utility companies

9:14 have to build to get those data centres up and running.

9:18 I really had no idea how much data centres

9:20 had actually changed the way power utilities were operating.

9:24 The utilities focus is now attracting data centres

9:27 to their service territory because utilities profit by building infrastructure.

9:33 That's their basic business model.

9:35 They make money by building stuff,

9:37 and there's no better excuse now for the utilities

9:40 to build something than to deliver energy to a giant, energy-sucking warehouse.

9:46 I set out to figure out whether the high-energy costs

9:48 of AI mean that I should try to use it more consciously.

9:52 Should I treat using a chatbot like leaving a light on for too long?

9:56 And I learned that if you're just submitting text queries,

9:58 the amount of energy per prompt is negligible.

10:01 It's not really worth worrying about.

10:02 Creating images on the other hand, can take 10 times more energy,

10:06 and creating videos can take hundreds of times more.

10:09 And given just how much platforms are pushing

10:11 image and video generation just about everywhere you look,

10:13 that part does feel like a cause for concern.

10:16 But still, it's important to put those numbers in context because

10:19 compared to tonnes of things that we do without thinking twice,

10:23 it's ultimately not that much energy.

10:25 But the real story isn't about individual footprints.

10:27 It's that training and running AI systems is

10:30 becoming a sizable part of our total energy consumption,

10:33 and the consequences of that growth aren't evenly distributed.

10:36 They land hardest on the communities where data centres are being built,

10:39 not necessarily because the data centres demand so much electricity,

10:43 but because all the new infrastructure that will

10:45 be needed to power them is expensive.

10:47 And utility companies cover those costs by spreading

10:49 them across the bills of regular people.

10:52 Those costs may start small, and they might be invisible to most of us,

10:57 but this technology is radically reshaping the world

11:01 around us and someone is already footing the bill.

11:06 Thank you so much for watching.

11:08 This was a fun piece to produce.

11:11 It's a story with a lot of complexities that requires a lot of nuance.

11:15 I think that's really where the power

11:17 of journalism comes from, is following your curiosities all

11:20 the way through to a conclusion and coming

11:22 out on the other side with better understanding.

11:25 Doing this kind of journalism is why I'm so excited about New Press.

11:27 It's a new platform that I'm building along with Johnny Harris,

11:30 Sam Ellis from Search Party, and hopefully, all of you.

11:33 New Press brings your community input and your perspectives

11:36 into our reporting process from the very beginning.

11:39 You can sign up for our wait list today

11:41 and be one of the first to join when we launch.

11:44 All right, see you next time.