Artificial intelligence is growing in popularity, with users asking for its assistance with all kinds of tasks. But AI uses a lot of water, which is in increasingly short supply in places like Arizona.
Shaolei Ren, an associate professor of Electrical and Computer Engineering at the University of California, Riverside, works on machine learning optimization and sustainability. Ren joined The Show to talk through some of this, starting with a broad question: what exactly is the relationship between AI and water usage?
Full conversation
SHAOLEI REN: Sure, so AI mainly uses water in two different ways. The first is cooling the data center facility. The second is for generating electricity. So the first part is for cooling the facility. Since data centers use a lot of energy for AI workloads, energy consumption turns into heat, which we just have to get rid of from the data center facility to prevent overheating, and this involves two stages. The first stage is just called server level cooling, where we were going to move the heat from the server to the facility or to a heat exchanger. And this stage involves air cooling or liquid cooling, like immersion or direct to chip coding, but this process doesn't really lose or consume water. The second stage is facilitated level cooling. So essentially we have to move the heat from the facility to the outside environment. And typically, many data centers, including those operated by big tech companies, use water evaporation. So this is the direct water consumption, and some of the tech companies can consume over 20 billion liters of water each year. And this amount of water is equivalent to some major beverage companies' annual water consumption. So it's very different from the water withdrawal. When we take a shower, we're going to be withdrawing a lot of water, but 90% of the water that we withdraw will be discharged and go back to the sewage immediately. So those are not considered as water consumption. This is the first part of the water usage for AI, like essentially just cooling down the data center facility. The second part is the electricity generation. So if you use thermal electric thermal power plants like coal or natural gas or nuclear, there will be a huge amount of water consumption as well.
MARK BRODIE: So when you talk about potentially needing more water to cool things down on particularly hot days, I'm thinking about how Phoenix and Arizona are popular places for data centers, because we tend not to have natural disasters. What we do have, though, are very hot days for a good part of the year. It seems like kind of a conundrum there.
REN: Yes. If you build data centers in Arizona, then probably you're going to be needing water for most of the year if you use water evaporation. But there is a technique called a dry cooler. Essentially, you don't need any water at all. But the downside of this type of technique is there will be more energy consumption, and it turns out that it will be probably consuming even more water off site. So overall, the impact on the water resources in Arizona, if you use a dry cooler could be even higher than using the water evaporation method.
BRODIE: Is there a way, or is it something that you know the folks who are developing AI like, is there an effort, or is there a way to use less water? Because it doesn't seem like people are looking to use a whole lot less AI.
REN: I would say these tech companies are taking water issues more and more seriously. So, for example, they are trying to use recycled water instead of drinking water for on-site cooling. And they're also purifying the water a lot more than before. So essentially, this can increase the utilization of the water and reduce the water consumption from the outside part and for the offset part. I think the most commonly used technique is reducing the energy consumption, reducing the model size and speed up the process. I think generally, these are the techniques companies are using to reduce their environmental impact.
BRODIE: Okay, how good of a sense do we have of exactly how much water these systems are using?
REN: This pretty much depends on the cooling technologies. If you're using cooling towers, I saw a research study for Arizona data centers. They demonstrate that during the summer days, it could use nine liters of water for one kilowatt hour of energy. So what does that number mean? If you have a medium sized data center with about 20 to 50 megawatt then you'd be needing, like, a millions of gallons of water each day. So that's, I think that's quite a lot, and most of the water is actually drinking water.
BRODIE: Yeah, I mean that would seem to be a really big problem, especially in a place like Arizona, where water scarcity is a big issue?
REN: Yeah.
BRODIE: Is it safe to say, or is it too much of a stretch, that as artificial intelligence continues to be used in more and more ways, and maybe in more and more complex ways, not just writing a paper, but, you know, making movies, for example. Is it safe to say that that will only continue to use more and more water down the road?
REN: I would say most researchers agree with this statement. So if you look at some tech companies sustainability reports, they publicly acknowledge the connection of AI computing and water consumption. So they say this expansion of AI workloads is driving up their water consumption. And also, last year, there was a report from the department of energy. And that report shows in 2028 just a few years down the road, the water consumption by the U.S. data center directly, will be doubling, or even quadrupling, the 2023 level. So I think that's a really fast increase. And mainly, and I will say, actually, primarily due to AI computing.
BRODIE: Does that lead you to think that, going forward, new data centers might be located or built in places, not in a desert, not like Arizona, not like Nevada, maybe not even in California, but in places where water scarcity isn't such a big concern?
REN: Water is an important problem, but it's not, at this point, on the top of their list when they choose Data Center locations. So I think we're still seeing data centers, more and more data centers being built in desert areas like Nevada and Arizona,
BRODIE: The lack of natural disaster is more important to them than the possibility of using too much water?
REN: I would say so. On top of that, this cheaper land and cheaper electricity prices are also attractive for data center constructions.
BRODIE: Sure is there other computing technology coming down the pike that you think is going to add to this even further, like something beyond AI.
REN: There was some studies about Bitcoin mining, and that could potentially be increasing the water consumption a lot as well. Beyond that, I would say AI is a primary driver for data center water consumption.
