Note: This transcript is auto generated and lightly edited.
SHAOLEI: If you write two emails using GPT-4, the amount of energy could be enough to drive a Tesla Model 3 for one mile.
I think that's really a lot, a lot higher than what many people believe.
MONICA: Welcome to Policy for the Planet, a new bimonthly podcast exploring the global response to the climate crisis.
I'm Monica de Bolle, a best-selling author and senior fellow at the Peterson Institute for International Economics, based in Washington, D.C. My work bridges the fields of economics, science, public policy, and public health–all under an international lens.
In each episode, I speak in-depth with experts to understand how governments are responding to the monumental challenges of the climate emergency. We'll unravel the complex tradeoffs of different policy choices to steer us toward sustainable practices and public well-being.
Welcome to the conversation!
While the promise of artificial intelligence seems limitless, behind every AI query lies a hidden environmental cost that few of us consider.
Data centers powering AI systems consume massive amounts of electricity — in some regions, up to 30% of the state's power usage. Even more surprising is AI's water consumption: Just 50 queries to an AI system can use as much water as a two-liter bottle of water. As companies race to develop ever more powerful AI models, these environmental impacts are set to grow dramatically.
Today we'll explore the true environmental cost of our AI future with Shaolei Ren, an Associate Professor at the University of California, Riverside. Shaolei's groundbreaking research has revealed the startling water and energy demands of AI systems, and his work is helping us understand the broader environmental and public health implications of our increasing reliance on artificial intelligence.
His insights come at a crucial moment as we grapple with how to balance technological progress with environmental sustainability.
Shaolei, welcome to the show. It's a pleasure to have you. You've done a lot of research into artificial intelligences or AIs, energy and water use, a black box for many to be sure. What have you learned and can you please summarize some of your key findings?
SHAOLEI: I think the key finding that we have so far is AI consumes a lot more energy than we thought and also a lot of water than we previously estimated. just to give you a context, so last year our estimate shows if you just have some 10 to 50 queries with GPT-3, the water consumption was about 500 milliliters. But it turns out based on the new data that we had this year, the water consumption is actually should be four times as much as what we previously estimated. So essentially, 10 to 50 queries will probably consume two liters of water. And that's just for the water.
And if you look at the energy consumption, first, we have absolutely no idea about how much energy this AI model consumes exactly, especially for those large but closed AI models like GPT or like Google Gemini systems. But based on the open source studies, it turns out that these models could consume a lot more energy than many people believe. Just to give you a context, if you write two medium-length emails, let's say 100 to 200 words emails, just two emails, the amount of energy required to write two emails using Lama 3 7dB is enough to give you a full charge on your iPhone 16 Pro.
MONICA: Wow.
SHAOLEI: Pro Max actually. And what about bigger model like GPT-4? Because although nobody really know the exact size, but it's commonly believed to be 1.8 trillion parameters.based on some interpolation, because all we know is the energy consumption for some small open source models. So we have to do some interpolation for those bigger models. So based on this interpolation, the GPT-4, if you write two emails using GPT-4, the amount of energy could be enough to drive a Tesla Model 3 for one mile.
I think that's really a lot, a lot higher than what many people believe. Exactly.
MONICA: That's astounding. That is really astounding. I had no idea about that. I mean, I had, I knew it was a lot, but I think when you put into context, know, two emails is going to take, if you use ChatGPT to write two emails, the amount of energy you're going to be consuming would charge a Tesla. You know, that is something to make you stop right in your tracks and think, am I going to need ChatGPT to write two emails?
SHAOLEI: Not really give you a full charge on your Tesla, but just drive Tesla Model 3 for one mile, which I think is still quite a lot. And again, I have to highlight that this is based on interpolation. So because we don't really have the transparency from OpenAI or Microsoft, so all we did was based on the measurement of smaller models, open source models, and do some projection and interpolation to get an estimate of those larger models.
MONICA: So there is an AI supply chain, but it would be great if you could explain what the different portions of this supply chain are. And in what way these different AI production stages or processes impact the environment.
SHAOLEI: So supply chain is part of what we call lifecycle environmental impacts of AI computing. so for making the AI chips and also recycling the chips after the retirement, we need energy. We generate carbon emissions. We also have water consumption. And additionally, we have a lot of air pollution, for example, PM 2.5., sulfur dioxide and some hazard gases. So if you look at the energy consumption overall over the life cycle, it depends on the model, depends on the specific hardware. But generally, the supply chain energy could be roughly 30 to 70 % of the life cycle energy consumption. If you look at the carbon emission, again, this varies a lot depending on the the models and devices you're using. But generally, this scope three or the supply chain carbon emission takes about 30 to 70%, similar to the energy part of the whole life cycle carbon emission of these AI models.
And for water, unfortunately, we don't have good numbers for water. But what we know is a lot, because this part, we have very little transparency. some manufacturer companies says they recycle most of their water. But it turns out some government study from Singapore, because Singapore is another major manufacturing site, shows the average recycling rate for semiconductor industry is just about 30 to 50%. So it's very deep, much lower than what many other semiconductor manufacturers claim.
And in terms of the air pollution, again, very little data, but I believe it's a substantial part.
MONICA: So, Shaolei, you've brought up two very important points. On the one hand, the carbon footprint of AI and the electricity use and the connection between that, because as we know, the sources of electricity matter here. If AI is using renewable energy, that would be better.
If it's not using renewable energy, there's a carbon footprint. And this is usually what people are referring to when they talk about the environmental impact of AI. They very rarely talk about water, which is something that you in your research have done a lot of pioneering work on. can you tell us about the impact on water specifically as a major environmental aspect of the AI effect overall on the environment?
SHAOLEI: Sure. So maybe we can start with how AI uses water. It uses water throughout the whole life cycle from the very beginning. Scope 1, scope 2, scope 3. Scope 3, similar to how we define the carbon emission scopes. For scope 1, that's essentially the direct water consumption. So because these AI models consume a lot of energy during the training and the inference, so this energy produces a lot of heat.
And a massive cooling system is needed to move the heat from the servers to the outside environment. And typically, this involves two stages. The first stage is to move the heat from the servers to a heat exchanger. And this process requires no water. At least there's no water consumption, no water loss. And the second step is to move the heat from the server, from the heat exchanger to the outside environment. And this process usually requires water evaporation. Because water evaporation, water evaporation is the most efficient way to move the heat and to dissipate the heat from the facility to the outside environment. And if you look at those large technology companies, they also use water evaporation either as the main or supplementary method for dissipating the heat. And this is what we to as scope one water consumption, direct water consumption. But in addition to directly cooling data centers using water, there's indirect water consumption associated with the AI due to the electricity generation. Because generating electricity is similar to carbon emission, it also uses a lot of water. And for example, if you use a thermal power plants or nuclear power plants or even hydropower, there's a big water consumption. lastly, there is a supply chain water. Because when we make the AI chips, we need water for cleaning, etching, and washing the wafers to prevent contamination. And typically, a semiconductor manufacturing facility can use many millions of water each day. So overall, this is what we call AI's Water For Print. But it's very important to note that this Water For Print has three different colors, blue, gray, and green.
Blue water is the freshwater that we take from the rivers, from the lakes, or groundwater sources. And this is often more scarce compared to gray water and green water. Gray water is the freshwater needed to dilute the pollutants. And green water is the rainwater stored in soil and used by plants. For AI, most of the water is blue water. And in many cases, it's just drinking water.
So it's really scarce. And there was a study published, I think, last week by OECD Global Commission on the Economics of Water. They show that for the first time in human history, the global water cycle is out of balance. And this raises water risks over many parts of the world, including large area of the US, for example, the Western parts and the Midwest. So there's a big water risk. And AI and in general data centers should lead by example to reduce their water consumption, improve their water efficiency, just in the same way as they improve their energy efficiency and carbon efficiency. So I think this is really an urgent issue.
Turns out yesterday, JP Morgan published a study to show that AI and data centers are draining the water supplies in the US, which needs more attention and more investment, more efforts to reduce this water consumption. And by the way, when we say water consumption, we're talking about the evaporated portion of water. It's not the water that we use for showering or for washing…washing our hands because when we use water to take a shower, we use a lot of water, but that water is called water withdrawal. So water consumption is the water withdrawal minus water discharge. When you take a shower, most of the water goes straight into the sewage system. So that's not really consumption. And water consumption for data center is just the evaporated portion. And for data center, typically 80 % of the water they take into the facility will be evaporated.
By contrast, If in a household, urban setting, urban household, typically we only evaporate 10 % of the water that we use. So it's not really fair to compare the data center water consumption with the water that we use for taking a shower. That's a very important concept. And it's a technical definition.
MONICA: And that is a critical point. I was thinking while you were speaking of the implications of this, because of course, if data centers are using water and water consumption is here, the water that is evaporating, it's a lot of water that they use and 80 % of it evaporates. That's it. That's water that's gone from the system as opposed to what we consume in our households which goes into the sewage, the water gets cleaned, recycled, all of that, and it comes back to us. And evaporation only constitutes 10 % of that consumption. So it's really a completely different picture when we're looking at how.
SHAOLEI: Exactly. There's actually a lot of misunderstanding among the public about water consumption and water withdrawal because they will say water, even though it gets evaporated, is still within the global water cycle. But this is out of balance. especially in places like California and Arizona, we only have a raining season for a few weeks a year. And we never know where the water is going to be going. It could go somewhere else.
And it could come back in a time when we need it, it doesn't come back. When we don't need it, it comes back. So this type of issue is creating a lot of uneven distribution of the water globally.
MONICA: That's fascinating and very worrying. I wanted to ask you because you have done, or you are doing a lot of research on this issue with regard to the US. So water usage, where things are happening, how they're affecting different places in the US, all of that linked to water consumption by data centers and all of that. Can you talk to us a little bit about what your findings are in that research? What areas in the US are most affected and so on?
SHAOLEI: So in general, the western parts of the US are more water stressed like Arizona. But it turns out Arizona also has a lot of data centers. Partially because Arizona has a lower energy price compared to many other states. And also Arizona has a lot of semiconductor manufacturing plants. That is also water intensive.
So overall, this is draining the water supply in Arizona. also because they all take water from the Colorado River basin. And along that river, there are a few power plants, coal power plants that are constantly evaporating water. But this water consumption is not only issue in Arizona, but also, let's say, in Pennsylvania, in parts of Pennsylvania, not everywhere.
For example, there is a big tech that buys nuclear power plants to power their data centers. And that nuclear power plants, although it generates no carbon, but it's constantly evaporating water. And it has some, by the way, water vapor is a greenhouse gas. Although it doesn't last forever in the sky, but...it's constantly evaporating. So overall, it's very stable amount of greenhouse gas.
MONICA: Yeah, that's a point that's well worth reminding people of because we often forget that water vapor is a greenhouse gas. Of course, it's not like carbon, as you just said. All of this that we're talking about, especially the water question, brings up the public health angle and public health concerns. And you have also done research into that. So can you talk about that? Because that I think is absolutely fascinating because then we're connecting AI, climate, water usage in public health.
SHAOLEI: Sure. So this is something that we have been working on in very recent month. So I'm really excited to share that findings. So when we talk about this AIS environmental impacts, carbon, of course, that's very important. Water, also important. But there's actually a public health impact due to, again, three different scopes. When we have data centers, we typically have diesel generators, even though they are alternative fields for backup generation, but that's very little in the US. So these diesel generators, we don't operate it constantly, but we do some maintenance and some testing. That testing can generate a lot of pollutants like PM2.5, sulfur dioxide and nitrogen oxides. This is scope one, direct emission.
Another one is due to electricity generation. Although in the last few years, the US power plants emission level has decreased quite a bit, but still is a major source of pollution. And the third scope is the semiconductor manufacturing. And in some parts of the US, like Arizona, some semiconductor manufacturing facilities are asking for raising the limit on their emission because they want to expand the production, which is needed for the industry to grow.
However, this raises a lot of healthy issues. And our studies shows that by 2030, each year, the US data centers could contribute to about $12 billion US public health cost. And this sounds like a big number, or maybe it's a small number. But let me put it into some context. So this number, by the way, this number is only for scope two, not for scope one or scope three yet. So basically this is only due to the electricity generation pollution issues. So if you look at the transportation on-road emissions in the US, on-road emission means any EVs, gas cars, diesel cars, any vehicles running on the road.
This public health cost of, so this data center public health cost is about 13% of the US overall on-road emission cost. Or the same amount of California's on-road emissions. So it's like 30 million cars on the road.
So we know that these vehicles have a huge public health impact. And think about you have additional 30 million cars running on the road. And that's the impact of AI computing by 2030 based on the EPRI projection.
MONICA: These figures are mind boggling. Very mind boggling. And you've brought up several of these and all of it is extremely worrying. So I want to ask you something because these companies are not, the AI companies are not providing us with the data on what kind of electricity usage and what's the volume of electricity usage that they actually have. What the sources of that electricity are.
The water usage and what they're doing with the water, the water usage has public health implications as you've just outlined in a very detailed way for us. So at the end of the day, we're looking at a situation that is begging and as an economist, I have to say it, it's begging for regulation. This is an environment that's begging for that. But AI is not regulated just like other...tech companies or not. And here's my question to you. What are your views about regulating the AI sector?
SHAOLEI: I think we are seeing regulations from the environmental sector. So we do have strong regulations for emission, for power generation. But when it comes to AI, which is the demand side? We're not seeing many regulations yet, especially from the environmental and health point of view. I think having the regulation from the supply side, that's important. But from the user side, demand side, that's also important, especially if you have a really fast growing sector. So we need to have some, at least to have some more transparency and to know where this emission goes, where, who is a counter attributed, who should these emissions be attributed to. So this part, if we don't have the measurement, you can't really manage it.
And when we talk about this food item like sandwiches or beef, we know how much water it needs, how much energy, how much carbon. But when it comes to AI, which is data essential, right? So we don't have a clear knowledge. I think we need that knowledge to make better informed decisions.
MONICA: Yeah, absolutely. mean, at the very, very least, we need to know with precision how much energy these AI companies are using and also how much water, because you've raised this issue of water and pollution. And we know that we have a dramatic water situation in different parts of the country, and you're highlighting those vulnerabilities. So it's really rather critical. which makes me think about the arguments that, you know, AI companies usually put forth to say, no, no, we don't need to be regulated, which is, know, regulation, if you if if you put regulation on us, that's going to stifle innovation. not going to be able to do the things that we're planning to do. We're not going to be able to, you know, improve the products that we're offering in the AI spectrum and all sorts of things that, you know, they say in order to stave off the regulatory pressure.
Do you believe this stifling of innovation argument? Is that a valid argument against regulation from the AI companies?
SHAOLEI: I think regulation can span different aspects like safety, risk. I guess we're not talking about those type of regulation. We're focusing on the environment and health aspects. think for this one, this AI is already having an impact in the same way as all the vehicles in California on the road. And we do have regulation for vehicles.
I think it makes sense to have regulations for AI as well, given that they have the similar scale of impacts. So that's my personal opinion.
MONICA: Yeah, mean, so just for our listeners, Shaolei is an engineer, I'm an economist, and we're both coming out on the same side on the regulation question, because it is about, I mean, as economists would put it, it's about externalities. So at the end of the day, if you have AI companies doing what they're doing and having negative impact, having positive impacts, but also negative impacts which are not being properly measured and those negative impacts are public in scope. In other words, they affect the public overall. They're not limited in who they affect. It's a very general effect. We're talking about public health. We're talking about the environment. So it's an obvious case for regulation.
SHAOLEI: I actually also want to add that this health impact is not limited to where they are located. For example, West Virginia has literally zero data centers. It may have a very few handful of small data centers, but large-scale data centers, almost zero. But West Virginia has one of the highest public health costs due to AI computing in the country.
Partially because it's providing a lot of energy to data centers in Virginia, which is a neighbor. yeah, we've seen this disproportionate distribution of the health impact across the country. some places, in some counties, the residents are actually paying a hidden public health cost, which is about half a year of their electricity bill.
MONICA: So Shaolei, do you know of any examples around the world of countries that may have done some regulations on AI? And do you have a sense of how this might be emulated in other parts of the world? So for example, if there's something that some country did which can be done by the US as well, or could serve as a...as an example for the US to follow or other countries to follow. Do you know of anything like that? And if so, can you talk about it a bit?
SHAOLEI: Sure. Right now in Europe, there's a EU-wide directive that mandates reporting of energy, water, and carbon for data centers, including AI data centers. And in the US, we're not seeing any mandating rules for reporting, for measurement, for disclosure yet. But there was a standard bill proposed in February this year and asking for some voluntary reporting framework from these AI companies to disclose their energy and water and carbon. But this is just a voluntary reporting and this is just a bill. We never know whether it will pass or not. So I think Europe is sort of taking the first step in the world to regulate, sort of regulating the environmental impacts of AI computing.
In the US, we're not seeing that yet. And maybe we could see something similar in the future, but not there yet.
MONICA: So do you have any idea how much AI consumption is likely to grow going forward?
SHAOLEI: So we have a different level of estimates, mid-level, low-level, high-level. So if you look at the low level, then probably it's going to be... So right now, data centers in the US consume about 4 % of the national electricity consumption. If you look at the low estimate, then probably in 2030, we still keep 4 % or 5%. The mid-level estimate is about 6 to 7%, but more… on the higher end, then it could go up to 9 % of the national electricity consumption. whether that's really going to happen or not, we don't know. But there's definitely a possibility given the arms race in the AI space.
MONICA: Yeah, and in any case, as you highlight at 4%, there's already a very big impact, difficult to measure, but nonetheless big on the environment through carbon emissions, on the environment through water usage. So if it's going to grow beyond the 4%, which it probably will, even if we don't know exactly where it's going to end up, then those pressures are to be even greater.
SHAOLEI: And yes, exactly. And also this 4 % hides a lot of details because it just gives you a national number. If you look at a particular region, say Virginia, it could go up to 30 % of the state's electricity usage. And in some other states like West Virginia, they don't have any data centers. And they're not projecting to have any data centers by 2030. But they're paying the environmental cost, health cost and pollution.
MONICA: Yeah, so you have places like Virginia where you have a massive amount of data centers. And of course, then the impact on everything is way greater because you have a lot of data centers located there. And then West Virginia, which is your comparison case, with no data centers, but still suffering the consequences because it's right next door, right?
SHAOLEI: Yes, yes.
MONICA: Yeah, so that is extremely concerning. And how do you think that?
SHAOLEI: Because these air pollutants actually travel by hundreds of miles. They're not staying within like a three-mile radius, which EPA uses to classify the affected range. But that's not really the case. It actually goes quite a long distance.
MONICA: Yeah, so in other words, it's not like a cartoonish cloud that kind of sits above whomever is doing what they're doing with the data centers, but it's rather a dissipating thing that just spreads everywhere. how do you think that You know, all of this will further impact the environment and health now around the globe, not just the United States.
SHAOLEI: Unfortunately, it's affecting certain regions disproportionately. For example, Uruguay had a mega drought crisis last year, forcing the National Water Authority to even use salt water for much of the municipal water supply. And at the same time, a large technology company from the US announced a massive data center project in the country, which could further exhaust the already limited freshwater supplies.
And many people took to the streets to protest against this project. Of course, today, I think starting from this year, the project changed the cooling system designs and reduced the water consumption a lot. But this shows that usually when companies plan for their data center, they don't really take this environmental impact seriously. They just try to meet whatever the regulation needs.
So unless there's additional pressure, they will not change their designs. They just use whatever is most profitable and most efficient way, most cost-effective, efficient way.
MONICA: And then on top of that, having a data company that's coming in to use water, most of which is going to evaporate, not come back into the system in any way, as you were saying in the beginning, is really problematic. So I think that case illustrates really well that, you know, this is a global problem that we're dealing with and we need to focus on it and be serious about it because it's urgent. It's not something that we can wait to resolve, right?
SHAOLEI: Yeah. Sure. Absolutely.
MONICA: Finally, is there anything you think we've missed in our conversation? What are you most concerned about regarding the future of AI and these environmental and health consequences that we've talked about?
SHAOLEI: I think this AI's environmental impact is very broad. It's just a single metric of carbon emission. Carbon emission is definitely important, but we have this water issue, we have this health issue, we also have the electronic waste issue. And there was a study coming out yesterday showing that by 2030, the electronic waste coming from AI could reach I think, 5 to 50 million metric tons of electronic waste. So that's really a lot. recycling this electronic waste has energy consumption, has a carbon issue, has water consumption, has health problems. So it's like cycling. This process just never stops, sounds like. So we need to have a more holistic way to assess the environmental impacts and health impacts of AI computing. And truly think about this circular economy.
MONICA: Well, Shaolei, thank you very much for coming on the show. This was fascinating. I hope you enjoyed it. For me, I learned a lot. So thank you so much.
SHAOLEI: Great to be here. Thanks for having me.
MONICA: Thank you for joining me on Policy for the Planet. Have a question or a topic to suggest? Email me at [email protected]. I'd love to hear from you.
Don't forget to rate, review and subscribe to Policy for the Planet today so we can keep this podcast going! And even better, please share our show with your friends and colleagues.
Special thanks to Jennifer Owens and Alex Martin, our producers, and Melina Kolb, our supervising producer. This podcast is brought to you by the Peterson Institute for International Economics. Learn more at piie.com.
Until next time, here's to creating meaningful impact. Stay motivated, stay curious.
