Short Wave - What's The Environmental Cost Of AI?

Episode Date: May 7, 2025

By 2028, Lawrence Berkeley National Laboratory forecasts that U.S. data centers could use as much as 12% of the nation's electricity. The reason: generative AI. Since 2022, AI innovation by four leading tech companies — Google, Microsoft, Meta and Amazon — has led to annual increases in both energy and water consumption. So, in this episode, Short Wave co-host Emily Kwong probes huge water footprint of AI. We begin with the rise of data centers, then look at how computers came to need so much water and, finally, what tech companies are doing to try to turn the ship around. P.S. Part 2 talks about the leading solutions in the green AI movement. So don't miss our Friday episode! Curious about tech and the environment? Email us at [e-mail] — we'd love to hear from you! Listen to every episode of Short Wave sponsor-free and support our work at NPR by signing up for Short Wave+ at plus.npr.org/shortwaveSee pcm.adswizz.com for information about our collection and use of personal data for sponsorship and to manage your podcast sponsorship preferences.NPR Privacy Policy

A quick note before we start today's show. You may have heard that President Trump has issued an executive order seeking to block all federal funding to NPR. This is the latest in a series of threats to media organizations across the country. Whatever changes this action brings, NPR's commitment to reporting the news without fear or favor will never change. Even as paywalls arise elsewhere, we offer this vital resource to everyone, regardless of their ability to pay. This is a pivotal moment. It's more important than ever that every supporter who can contribute comes together to pitch in as much as they are able. Support the news and programming you and millions rely on by visiting donate.npr.org.

And if you already support us via NPR Plus or another means, thank you. Your support means so much to us, now more than ever. You help us make NPR shows freely available to everyone. We are proud to do this work for you and with you. You're listening to Shortwave from NPR. Hey, ShoreWavers, it's Regina Barber with my co-host, Emily Kwong. Hey, I'm. Hi, Gina. So today, our episode starts with water.

And someone who's been thinking about water for a long time. He says maybe that's because of where he grew up. The official name is Kang Ar-Cheng. And the town only had like 50,000 people at that time. This is Shalajan. He's from a coal mining town in northern China where growing up, water was really scarce, so he learned how to make every drop count.

We only had water access for like half an hour each day, so we just had to use water pretty wisely. So he grew up very water conscious, and now at UC Riverside, Chalai studies the water footprint of the tech industry, because as you know Gina, as the tech industry has grown, so too have data centers. Right, these data centers that are those huge buildings

filled with hundreds of thousands of computers that store cloud data and do a lot of computing for AI, those computers can get really hot. Right, which is why water, you know, chilled H2O, has become an ally in keeping those computers cool. And Shale wanted to know exactly how much water was being used. But as early research, some of the first ever studies on water efficiency and data centers was kind of met with crickets. Back in 2013, there was no attention at all. Zero.

But then, in 2022, OpenAI's ChatGPT took the internet by storm, and people started to look at Shale's work. The amount of water that AI uses is astonishing. AI needs water. People are saying that every time you use ShatGPT, you're using this much water for a hundred million. And where will that water come from? Just to train a large language AI model and keep a data center cool can consume hundreds. hundreds of thousands of liters of fresh water. And by consume, I mean that the water evaporates and doesn't necessarily return to the local watershed.

Like the water turns into vapor, goes up in the air, and does not come down to that location. Not necessarily. That's water consumption. Okay. Yeah. It's where the water is no longer available for reuse. In 2023, for example, Google's Data Center in Council Bluffs, Iowa, consumed nearly 1 billion gallons of potable water. Wow.

Okay, so I know data centers also use a lot of energy, primarily like fossil fuels, but I guess they're also using like a ton of water. Yes, and it's because of AI infrastructure. Now, unless you live near a data center or a power plant, AI infrastructure is mostly invisible. And my goal with this reporting was just to pull back the curtain and ask what toll this is all taking on the environment. Today on the show, the first in a two-part series on why the true environmental footprint of AI is so elusive. Starting with the rise of data centers and how computer architecture got to the point of needing gallons of water in the first place. Then we'll talk about how big tech is trying to turn that ship around. I'm Regina Barber.

And I'm Emily Kwong and you're listening to Shortwave from NPR. All right. So all of these headlines about how AI is using water, it's because it takes a lot of energy to compute and solve really big problems, right? Right. So data centers, they grew from these like single rooms. to whole buildings during the dot-com boom of the 90s and aughts. And now these big buildings contain hundreds of thousands of computers. If they get too hot, the servers can shut down or suffer damage.

So what is the method of cooling down these computers? Well, every data center is different, but I'll describe the basic principles of a mechanical cooling system. Okay, picture a room with rows and rows of computers on racks. Yeah, I've seen them before. It makes me think of like a library. Yes, yes, it's like a computer library, except the floor is raised, so there's this void below that allows cool air to flow up through a bunch of grills and chill the computers. Benjamin Lee is a professor who studies computer architecture at UPenn, and he explained to me how air cooling basically works. You push the cool air through the front of the machines, and all the warm air gets pushed out the back.

And then what happens is a refrigerant takes the heat outside the building where, It gets dissipated into the air. But the thing about an air cooling system like this is it requires a lot of electricity. So some systems also use water to help pull heat away from the data center. Yeah, which is smart because water is so much better at transferring heat than air. Yeah, your physics degree really pays off at a time like this. Just in these moments. But like, where does this warm water go?

Well, a lot of it gets sent to a cooling tower and is evaporated. You can think of it like sweat. The data center is the brain. It needs to be cooled down because it's getting hotter and hotter in this era of AI. I think the difficulty has been that the air conditioning infrastructure is having trouble keeping up with the latest in GPUs and how closely packed GPUs are. Benjamin is talking about microprocessors. And a certain type of microprocessor known as a GPU is widely favored for running AI. They are delivering more performance, but they also may be drawing more power, which is why we are now taking under-examined.

and press out-handed steps to cool them. Now, the thing about data centers, Gina, is that some are more energy-efficient than others. There's even free air cooling systems which pull in air from the outside and use no water. But the point I really want you to remember is that in order to reduce the electricity demands of data centers, some have turned to water.

And that has meant the overall water consumption, like the number of gallons getting evaporated away, has gone up. Because of AI? Because of AI. getting integrated into products from the four biggest data center operators, Google, Microsoft, Meta, and Amazon. Which quick sidebar, we should note that, like, they're all financial supporters of NPR. Like, Amazon also pays to distribute some of NPR's content. Yes.

And Amazon does not disclose how many gallons of water they consume. They only report their water usage effectiveness or WUE. So we don't know how much water they consume. We do not. Oh, wow. Okay. We have a better sense from Google. Google, Microsoft, and Meta. Since 2021, all three have reported a bigger water footprint,

meaning they are consuming more and more water lost to evaporation every year. So who's consuming the most? Google. Okay. So in 2023, and this is according to their own report, consumption across all their data centers totaled 6.4 billion gallons. That's enough to irrigate 43 golf courses in the southwestern U.S. Wow. Although keep in mind, that is nothing. compared to how much water is used by agriculture. I mean, 43 golf courses sound like still a lot of water to me. It's a lot of water.

Yeah. And the concern, of course, is that once the water is evaporated, it's not available for reuse. Right. So just to give you an example of how this can play out badly, the Dow's, that's a city 80 miles east of Portland, Oregon, is where Google built its first data center. And residents noticed a change to the local water supply. The water level in our wells drop 15 feet. This is Dahl's resident Don Rasmussen, talking to the AP in 2021.

When you have dry conditions, you know, it's stressful on the plants, the animals, and the people, and the community. So the Oregonian, the local paper, asked Google, hey, what are your water numbers? And Google said, no way, we're not going to tell you, it's a trade secret. And after a year-long legal battle, it came to light that Google was using a quarter of all the water available in town. That is so much. Now, this surge of water use, I was like, why? Why so much water? It can be directly traced to the AI renaissance. And that's because tech companies are searching for what Benjamin Lee at UPenn calls the next killer app. The search engine was a killer app. Another example of that would be a recommendation system that social media feeds use to recommend ads and content. That was a killer app. But we don't have that for generative AI. Ben says that's why you're seeing things like AI overviews in Google Search or AI chatbots on Instagram or AI product summary reviews on Amazon.

There's a lot of generative AI being invoked on your behalf as these companies try to figure out what it's good for. Which is, you know, they're prerogative. But in the meantime, there doesn't seem to be a standard for these companies to report the details of their water use. So that golf course number that you mentioned earlier, we only know that because Google freely reported it in like a progress report on their own. own climate pledges. Can you tell me more about those pledges? Like, what has each company promised to do for the climate? Well, all four have pledged to be water positive by 2030, which means they'd put more water back into the environment than they use. And they're trying to do this through partnerships with local watersheds. In the DALs, that city in Oregon, I mentioned earlier,

Google is now building a system to pump excess surface water into an existing aquifer for later use in dryer months. It sounds like they're trying to be water positive. Yeah, water positive, and clean energy. Google, Microsoft, and meta have all pledged to reach at least net zero carbon emissions by 2030. Amazon

has set their deadline for 2040. But again, Gina, because all of their energy and water data is shared voluntarily, the public has no way to wrap its arms around the scope of AI's environmental footprint. And computer scientists Sasha Lucione, climate lead

at Hugging Face thinks that is a problem. We don't have any mandatory reporting mechanisms for companies, for compute providers, so they tend to give kind of very high-level numbers on a company level, sometimes, not even all the time. So after realizing just the scope of AI, I had to ask these four tech companies, are your climate and water goals even realistic? So what did they say? Well, Meta said they, quote, remain committed. Google said they are fully committed. Microsoft said they remain resolute and, quote, are proactively.

actively working to address resource challenges associated with the energy needs of AI. And Amazon, Amazon actually sat down with me. Can Amazon meet its climate and energy goals, as stated? Yes. We are continuing on our path to meet our climate goals by 2040. Kevin Miller is the vice president of global data centers at Amazon Web Services. And he told me all the ways Amazon is investing in green energy infrastructure. And all the tech companies are. Right.

And speaking of like, you know, green energy and being. more carbon neutral. I read that Amazon meta and Alphabet, which runs Google, just signed an agreement along with other companies that supports tripling the global nuclear supply by 2050. Yes, it's very ambitious. Wow. And along with Microsoft, these four companies have signed agreements to purchase nuclear energy, but that industry has been stagnant for years. It takes a long time to get nuclear up and running. So computer scientists who study climate are doubtful. Here's Benjamin at UPenn. I think before generative AI came along in the late 2022, there was hope among these data center operators that they could go to net zero. But he's lost faith now.

As companies increase their energy use faster than they switch to renewables. I don't see how you can undercurrent infrastructure investment plans. You could possibly achieve those net zero goals. Sasha at Hugging Face agrees. I mean, for what it's worth, Microsoft and Google already failed to meet their own goal last year. So I think that the tendency is going towards no. I also asked Jesse Dodge, a senior research scientist at the Allen Institute for AI at MIT. And over email, he said to me, quote,

these companies are making non-binding pledges to get positive attention. And I expect that if or when they don't meet those pledges, they will simply move the goalpost. In the meantime, more data centers are being constructed. Yeah, where? All over the country. Jeffersonville, Indiana, Rosemount, Minnesota, and Abilene, Texas. On January 21st, the day after his second inauguration, President Trump announced a private joint venture to build 20 large data centers across the country,

as heard here on NBC. A new American company that will invest $500 billion, at least in AI infrastructure in the United States, and very quickly moving very rapidly. This new project, known as Stargate, would, together, consume 50,000, 15 gigawatts of power. That would be like 15 new Philadelphia-sized cities consuming energy. There aren't any state or federal regulations for AI or data centers. Some legislators at the state level have introduced bills to regulate AI and data centers in California, in Connecticut. And at the federal level, Senator Edward Markey of Massachusetts introduced a bipartisan bill that would set federal standards and voluntary reporting guidelines to measure AI environmental footprint. But there really isn't a legal framework in place.

But like until laws are in place, are tech companies like doing anything on their end to fix the problem? Like to train or to create more sustainable AI models. That is why there is a part two of this series. Next time on Shortwave, the Green AI Movement. I can't wait. This episode was produced by Hannah Chin, edited by our showrunner Rebecca Ramirez and fact-checked by Tyler Jones. Jimmy Keeley was the audio engineer. Special thanks to Brent Bachman, Johannes Durge, and our incredible standards team.

The chat GPT commentary you heard at the beginning of this episode came from TikTokers, Dylan Page, Carter Smith, and Nikita Redcar. You also heard tape from Morning Brew and now this. Beth Donovan is our senior director and Colin Campbell is our senior vice president of podcasting strategy. Emily Kwong, and I'm Regina Barber. Thank you for listening to Shortwave, the science podcast from NPR.