The Environmental Inequity of AI

Shaolei Ren:

The preliminary study from our group shows that nine out of the top 10 counties being affected most by this AI expansion or data center expansion are low-income communities, predominantly black population. So I think that shows something.

Danny Caine:

There was nothing stopping me from going to ChatGPT and typing, “Write me an intro to a podcast about the environmental cost of AI-powered large language models.” Similarly easy would’ve been logging onto some AI image generator and ordering up a feature image for that same podcast. I could have even asked ChatGPT to draft a list of questions for a professor who studies the environmental impact of the technology that would make those questions appear on the screen almost instantly. Millions of people do this kind of thing every day. Many don’t even realize they’re doing it. I decided not to ask AI to help with this intro or this podcast in large part because of what I learned in the conversation you’re about to hear. But saying I decided not to ask AI to help isn’t quite as straightforward as it seems. Though I’m writing this episode’s intro, conclusion and questions using my old-fashioned human brain, I’m still using tech powered by what has come to be known as artificial intelligence, or AI.

I know for a fact that Riverside.fm, the platform I use to record and edit this podcast uses AI to instantly generate transcripts of every conversation I have in these recordings. Whatever aversion I might have to AI due to its climate impact, I have to admit that this feature of Riverside is immensely helpful. Is there room for responsible AI use despite the technology’s track record of climate harm? And that climate harm is pretty severe. Just one part of the AI process cooling the massive servers that power the technology has massive environmental impact using up a staggering amount of fresh water. And like so much else with big tech and monopoly power, these impacts are disproportionately felt by poor people and people of color. Here to discuss this with me today is Dr. Shaolei Ren, who has been doing great work studying exactly that, how AI’s environmental impact disproportionately harms certain folks and how to fix that. He’s an AI agnostic. He believes in the technology’s uses. He just wants it to be more sustainable and equitable. We’ll discuss that and much more on today’s episode.

Dr. Shaolei Ren is an associate professor of electrical and computer engineering at the University of California Riverside, where he is also a cooperating faculty member in the Computer Science and Engineering Department. His research focuses on responsible AI for a resilient, sustainable, and equitable future, and we’re thrilled to have him here to share some of his findings.

Shaolei Ren, welcome to Building Local Power, and thanks so much for being here. I always like to start with the question of our expert’s origin story. So can you talk to us a little bit about the path to your work and the environmental impact of AI? When and how did you become interested in this work?

Shaolei Ren:

Technically, I started working on sustainable AI back in 2012, the last year of my PhD study at UCLA, I did an internship in Microsoft Research where I worked on some research related to energy efficient computing. But I realized that energy does not only have an impact on the electricity bill, but also has a big impact on the environment, such as the carbon emission, which has the connections to global climate change, but also it has some regional impacts such as water consumption. That’s how I got into this environmental impacts of AI, or not AI computing, but more recently this as AI has been taking up a higher, larger fraction of the computing workloads in data centers. So that’s why I’m also starting to work on this AI computing as well. By the way, I want to share some of my personal experience of why I’m kind of interested in this local community impact.

I spent a couple of years of my childhood in a small town in China. It’s a coal mining town, so we only had water access for half an hour each day. And of course the city has changed. Now the town has water available 24/7, but when I was young, more than 30 years ago, more than 35 years ago, I was a really little boy, we had limited water access, so we really had to learn how to save water and try to minimize water usage very wisely. And also because it was a coal mining town, you could imagine the air pollution and, in general, I would say U.S. has a better air quality standard than many other countries, including China. So after I left that small town, I didn’t go back to that place until 2013. After I got my PhD degree, I went back to visit my parents and then I thought, okay, maybe just go back to where I grew up to have a visit. The air quality was still not really good, so you could see those black coals, black carbon everywhere on the street. So I think I just had that natural instinct, or I had the feeling from the local people’s perspective. So I know it’s really bad for people’s health if you live close to those places.

Danny Caine:

Well, thank you for sharing that personal connection, and I can easily see how it’s informed your work on this issue. In 2012, it wasn’t specifically about AI, you were just looking at the environmental impact of computing in general. And then after that, as AI rose to prominence, you became interested in that in particular?

Shaolei Ren:

Exactly. So at that time in 2012, I think AI wasn’t as popular as it is today. So at that time, I think the main thing was the cloud computing. A lot of cloud operators are offering different type of virtual machines, cloud services to customers. So that drove up the energy consumption of data centers. And at that time also, the power usage effectiveness, which is a measure of data center energy efficiencies was pretty high around … The industry level was around 2.0. And so it really just means if you consume one unit of energy on the server, then you’re going to have one extra unit of energy for cooling for power distribution. So that’s why a lot of people started to work on the energy efficient computing.

Danny Caine:

So the 2.0 rating means that for every bit of energy consumed by the computing, you also have an equal amount of energy consumed by the cooling and maintaining of the servers?

Shaolei Ren:

Yeah. Yes, yes. And nowadays the ratio has, if you look at some state of the art facilities, we’ve seen that it was 1.1 and even 1.07, but that’s not the industry average. The industry average I think is around 1.5 or something, but it’s still much better than 2.0 10 years ago. Of course, the PUE itself is not a great metric. Somebody can game it, but still it’s a reasonable way to give you first order analysis or insight of the energy efficient of data centers.

Danny Caine:

Oh, that’s really interesting. I hadn’t heard that measure before, and thank you for explaining it. So let’s move on to AI then, which is your current work and what drew me to your work in researching this stuff. I have a general idea that AI has a large environmental impact and somehow water is involved. But can you explain to me a little bit more about what this particular technology, why it seems to have a higher than normal drain on the environment and how exactly it works? How do you get from typing in ChatGPT into elevated water usage? Just explain it for me and our listeners.

Shaolei Ren:

Sure. So AI, in general, it just means artificial intelligence. It means a lot of things like vision models, robotics, language models. But nowadays, when the general public talks about AI, they seem to by default refer to large language models, image generation models. So these are of course one type of AI, but they’re getting more and more popular now. So let’s maybe focus on generative AI like ChatGPT, language models like ChatGPT. Technically, ChatGPT is not the model name, it’s just a service, but let’s use ChatGPT for the sake of a presentation or talking. So underlying the ChatGPT is a large, huge model which has hundreds of billions of parameters. So that means whenever you try to generate even one single word, it needs to go through many billions of calculations. And this calculation consumes a lot of energy, and this energy produces heat. We have to get rid of the heat to prevent the system from overheating.

Just like your cars, your engine needs to dissipate the heat constantly to the environment to prevent overheating. So we need to do the same thing for servers. And to do this cooling, there are two stages. The first stage is to transfer the heat from the server to a heat exchanger at the facility. Sometimes, it’s just a cooling tower or chiller. And then the second stage is we’re going to move the heat from the heat exchanger to the outside environment. So the first stage, there’s no water loss, no water consumption, so either we can use air cooling or immersion cooling or direct-to-chip cooling. So in any case, there shouldn’t be any water consumption. If there’s a water consumption, that means there’s a leak. So the second stage is really moving the heat from the heat exchanger to the actual environment, to the sky. And most commonly data center use water evaporation either as the main or as the supplementary method to dissipate the heat.

If you use a cooling tower, basically that means you’re using water evaporation 24/7, and if you use water evaporation as the assistant, that means you only use water during the hot summer days. So this overall consumes a lot of water. And in general, the cooling tower has a higher water consumption than water evaporation assistance, but they both have pros and cons because water evaporation assistance has a low average, but it has a really high peak. So it has some other issues. But in general, this is what we call direct water consumption for data centers. And if you look at some of the tech companies, one company can consume more than 20 billion liters water directly, and this water is mostly potable water, drinking water. And more than 20 billion liters water is comparable and actually the same as one of the major beverage companies annual water consumption. I’m not sure if we can name it, but it is definitely a household brand.

Danny Caine:

You can definitely name it.

Shaolei Ren:

It’s PepsiCo. If you look at the PepsiCo’s annual report, they consume 24 billion liters of water, which is the same as one of the largest technology companies’ water consumption directly. So they are comparable. So this is direct water consumption, and in addition, there’s also an indirect water consumption associated with the electricity consumption. Just like when we talk about data center carbon footprint, by default, we’re referring to the carbon emission from the power plants, because data center do not have direct carbon emission unless they operate on diesel for a couple of hours each year. That’s just basically literally zero. So the carbon emission for data centers is coming from the power plants, of course, also coming with coming from the hardware. That’s the supply chain. Let’s ignore the supply chain for now. So data center also has indirect water consumption associated with the electricity generation.

That’s what we call indirect or scope two water consumption. And that’s a lot. So if you look at the US average to produce one kW hour energy from the power grid, there will be roughly 3.14 liters water being consumed. And when we say water consumption, it’s actually a technical term that means the difference between water withdrawal minus water discharge. In the data center contact, that just means the water being evaporated into the atmosphere. So when we say residents take some water for a shower, they use a lot of water, but that water is considered water withdrawal, and most of the water is going straight into the sewage and can be discharged for reuse shortly afterwards. So on average, only 10% of the water withdrawn by a residential setting is considered water consumption. But for data center, if you look at Google, Microsoft, their water withdrawal, 80% of the water withdrawal is consumption.

So these are just two different concepts. By default, we refer to water consumption, not water withdrawal. So sometimes we see this comparison, wrong comparison say, “Oh, if I take a shower, I’m going to be able to use ChatGPT for a whole year.” That’s not really the right way to do the comparison. And also here the number we are talking about in our research and in many of the papers, it’s actually only the operational water. We’re not talking about the water for making the AI chips, which is the supply chain water, or sometimes we call it embodied water. You might have heard that producing one kilogram of beef requires several thousands of liters of water. But most of that water, I think if I remember correctly, between 94%-97% of the water for beef is green water, the rainwater containing the soil used by the plants to feed the cattle.

Essentially, that’s the supply chain water for beef. So again, this is not a fair comparison. So because we are not talking about the water needed to manufacture the AI chips, that’s another huge part. There are just not much data. So we are considering that, but it doesn’t mean it doesn’t exist. It’s still a huge part we just don’t have a good idea. So overall, if we focus on water consumption and do not consider supply chain water, if you have 10 to 15 queries from GPT-III, which is the older version of the model being used by ChatGPT, there will be roughly 500 milliliters of water being evaporated, and that’s water consumption.

Danny Caine:

So 10 queries in the ChatGPT is equivalent to 500 milliliters of water?

Shaolei Ren:

10 to 15.

Danny Caine:

10 to 15. Wow. But these staggering numbers, these 20 million gallons, that’s only part of the picture because you’re not even considering the supply chain water that takes to get these data centers operational.

Shaolei Ren:

No, no, we’re not considering the embodied water or the supply chain. That’s a big part. If you look at the companies, technology companies, the ESG report, the carbon emission from the embodied part, that’s a huge fraction.

Danny Caine:

Wow, that’s incredible. Already that tells me that this technology has a really steep environmental cost, but I think there’s another layer to it, and I find it very interesting. I’m going to quote an article you co-wrote for the Harvard Business Review where you say existing approaches to deploying and managing AI computing often exacerbate environmental inequity, which is compounded by persistent socioeconomic disparities between regions. Can you explain how AI’s energy needs disproportionately impacts some communities, and do you think this could fall under the category of environmental racism?

Shaolei Ren:

First, I want to clarify that I’m not an anti-AI person. I’m focused on the negative side because I want to make it better. So that’s my first disclaimer. So let’s come back to the question. Well, if you look at the water, water is a regional resource, and we try to minimize the total environmental cost, typically carbon, and some people do consider water as well. But if you just minimize the overall total amount, it doesn’t really mean every region is decreasing, is seeing a decreased environmental cost of AI. So for example, last year, Uruguay had a mega-drought crisis, and the national water authority in the country even was using 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 use many million liters of water each day, and that could definitely further exhaust the already limited freshwater supplies in the country.

So this shows that, I mean, they could show in their balance sheet that, “Oh, we’re reducing this water globally by this amount,” but you’re disproportionately hurting certain regions that are coincidentally experiencing drought, and this is for the water. And also, let’s look at the power generation or the location of the data centers. When power is generated from fossil fuels, they not only have carbon emissions, which definitely contributes to the global climate change, but there are also a lot of other things called criteria air pollutants, including fine particular matters like PM 2.5 and sulfur dioxide, nitrogen oxides, DOCs. So these particular matters, or basically these pollutants can travel many hundreds of miles. They can affect a large region, but it’s not global. It’s a regional thing, and they will disproportionately affect the regions, especially those communities living closer to the power plants. And we are doing a study in collaboration with Caltech.

Our preliminary analysis shows that based on the project of the AI demand in 2030, the U.S. Data Center, excluding Bitcoin mining, only for AI and other general computing, this U.S. Data Center industry could be attributed to about $15 billion of public health cost. And what does that really mean? It’s equivalent or comparable to the on-road emissions of the largest U.S. states like California, which has around 30 to 35 million vehicles on the road. So if you can pull up 30 to 35 million vehicles from the road, the emission impact you get, it’s the same as the U.S. data centers in 2030. The dollar value, just a single measure to give you a simple idea to show the equivalence. But if you look at the actual health impacts, that means we’re going to have a hindrance to roughly 1,000 premature mortality, and many asthma symptoms, cardiovascular diseases, heart attack and lung cancer, and even cognitive decline. This is based on decades of research from atmospheric research and public health research. So this has very strong scientific evidence to show not only correlation, but the causality.

We show that out of this national number, the total number, the regions that are particularly affected include many low income communities in Virginia, West Virginia, and Pennsylvania. The preliminary study from our group shows that nine out of the top 10 counties being affected most by this AI expansion or data center expansion are low income communities, predominantly black population. So I think that shows something.

Danny Caine:

Thank you for sharing that new data. I’m particularly struck by this nine out of the 10 top counties affected by data center expansion are poor and, perhaps, black communities. Why do you think these big tech companies are … Is it on purpose? Do you think they’re picking these counties on purpose?

Shaolei Ren:

Well, I don’t think they pick those places on purpose. Siting the data centers has a lot of factors to consider. They will consider the price, the energy bill, where the power is available, whether it’s close to the user base, because especially for inference data centers, they will need to consider the latency, and natural disasters, even tax credits. So a lot of factors will play into the decision process. So I’m not saying they are really on purpose or knowingly hurting those communities, but just happens to be disproportionately affecting them. But they do have a choice. If you look at the different location, I mean in our study we also show if you compare two different companies, we can look at the U.S. map. They are having different impacts, basically different distribution of the impacts across different parts of the country due to their different locations. So clearly, the location matters a lot.

Danny Caine:

So yeah, certain companies have different approaches. I hear you with not being anti-AI in particular and wanting to make the technology better. So I’d love to end our conversation there. Can you tell us about efforts that are currently being made to help with this problem and whether you think they’re sufficient. And if not, what would you rather these companies be doing? What do you see as a positive solution as the next step in rectifying the environmental impact of AI?

Shaolei Ren:

My personal take is that they’re not aware of this issue, probably. So they’re still in the arms race for AI to make more powerful, larger models, and that are more capable, and approaching what they call superintelligence. So I think they’re not really taking these environmental impacts into serious consideration as what they should be doing. So if they are aware of this issue and they put it into a major factor in their decision, I think there are a lot of different approaches to address this issue and these mitigate this issue. For example, they can prefer locations with the less environmental impact, with the less health impacts on local communities, and they can also balance their workloads. Right now they’re using what they call carbon intelligent computing, so they can use the health aware computing, health informed computing to reduce the health impacts on local people, and that’s something they can do, which I think is not rocket science for them. These things are just not coming into their radar yet.

Danny Caine:

I think my last question is, say I am a fairly regular user of AI and similar technology, and I am also very concerned about the impact on the health of communities, especially disadvantaged or poor communities. Is there anything I, as a computer user, can do to help make progress on this?

Shaolei Ren:

I believe so. That’s what I’m working on, which we call health-informed computing. So essentially, if you look at the public health research or atmospheric research, they study the air dispersion, air quality dispersion model. They studied how these different air pollutants could affect the local people’s health team. But I think the missing part in the current public health and atmospheric research is we’re not connecting the action to this impact. So what is the impacting this air pollutant? What is the impacting the health? I think what energy consumption is a big part because fossil fuel is the predominant energy source in the U.S. Starting from next year, we’re going to see maybe even more fossil fuels. This fossil fuel is something that we can change from the demand side because, in general, the energy production is really driven by the demand. We know from the basic physics law that electricity production has to be equal to demand at all times.

So it’s driven by the demand. If users are aware of this health impact, they are informed of the health impact. I think this user side or demand side management can really play a significant role to drive the efficiency from the power grid, to minimize the potential health impact on the local communities. But we need to have this awareness. So that’s why what we are working on is health informed computing. We try to bring up this knowledge, awareness, and also tell people exactly quantify the health impacts of their energy decision. And we’re starting from the energy as the initial point, but we hopefully in the future we can also expand to other actionable information related to the health so that people can make decisions in response to the health information we provide.

Danny Caine:

Well, it’s amazing work. I thank you for doing it, and I thank you for appearing on Building Local Power. Shaolei Ren, great to talk to you. Thanks so much and best of luck with your work.

Shaolei Ren:

Great to be here. Thanks for having me.

Danny Caine:

A quick epilogue, as I reflected on my conversation with Shaolei Ren, one thing really stuck with me. He asserted that big tech companies aren’t purposely trying to harm these poor and black neighborhoods by locating their water-hogging data centers there. I think he’s right in a way. I don’t imagine that there’s anyone at Google who’s explicitly like, “Let’s really mess up this black neighborhood.” They’re probably thinking in terms of economics. “This county in West Virginia has cheap land in proximity to the power grid and the natural resources we need to run this data center.” To them, it’s a numbers decision, but here’s what hit me in reflecting on this conversation. Those numbers themselves can be a result of racist policies. The very things that make these sites attractive for big tech could be the results of decades of racist and anti-poor legislating by those in power. Maybe the property values are low because of redlining.

Maybe there’s a lot of that cheap land available because these areas have been bypassed by transportation projects and developer interest in favor of areas with more money. Because they’ve been passed over, these areas may be willing to throw more tax breaks at those companies, at any companies who want to build there, regardless of environmental cost. On top of that, we can’t forget that poor communities have less time, power, and money to use in mobilizing and resistance to these climate damaging projects. So even though nobody at Google is saying, “Let’s be racist in how we build this data center,” race and class still play a part in how they grow and attain their monopoly power. That’s exactly what ILSR’s Power Play report is about, and that’s the through line to the season of building local power.

You can learn more about Dr. Shaolei Ren at his website, which we’ve linked to in the show notes. Also in the show notes is the Harvard Business Review article that inspired me to call him up and ask him to appear on the show. If you enjoy the work we do at Building Local Power, please consider making a donation to Institute for Local Self-Reliance to help us create more content like this. Donating is easy. Just go to ILSR.org/donate. This episode of Building Local Power was produced by me, Danny Kane, with help from Reggie Rucker. It was edited by me and Taya Noel, who also composed the music. Thank you so much for listening.

• Dr. Shaolei Ren’s’ most recent study (9 December 2024), “The Unpaid Toll: Quantifying the Public Health Impact of AI“
• Dr. Shaolei Ren’s website
• “The Uneven Distribution of AI’s Environmental Impacts” by Shaolei Ren and Adam Wierman in Harvard Business Review