AI Is Everywhere Now—and It’s Sucking Up a Lot of Water

From our collaborating partner “Living on Earth,” public radio’s environmental news magazine, an interview by Aynsley O’Neill with Shaolei Ren, associate professor of electrical and computer engineering at the University of California, Riverside. 

Artificial intelligence has become a part of everyday life, but there’s little regulation thus far of its deployment and use. Currently, there’s no law on the books in the U.S that requires AI companies to disclose their environmental impact in terms of energy and water use. Concerned researchers rely on voluntary data from companies like Apple, Meta and Microsoft.

But research is showing that AI generation may be even more resource-intensive than originally thought. Imagine that you want to ask an AI program to write up a 100-word email for you. You get an almost instant response, but what you don’t see are the intensive computing resources that went into creating that email. At the AI data center, generating just two of those emails could use as much energy as a full charge on the latest iPhone. And according to a Pew Research Center study, that 100-word email could use up a whole bottle of water for the cooling that’s needed at data centers.

This interview with Shaolei Ren, associate professor of electrical and computer engineering at the University of California, Riverside, has been edited for length and clarity.

AYNSLEY O’NEILL: For those of us who are unfamiliar with the technical aspects of how AI works, why does it take so much more energy and so much more water than anything else you do on your computer? 

SHAOLEI REN: Well, because a large language model is by definition just really large. Each model has several billions of parameters, or even hundreds of billions of parameters. Let’s say you have 10 billion parameters to generate one token or one word: You’re going to go through 20 billion calculations. That’s a very energy-intensive process. This energy is converted into heat, so we need to get rid of the heat; water evaporation is one of the most efficient ways to cool down the data center facilities. That’s why we also use a lot of water besides the energy. 

The water evaporates into the atmosphere, so sometimes that’s considered as lost water, although it technically is still within our global water cycle system, but it’s not available for reuse in the short term in the same source. Water consumption is the difference between the water withdrawal minus the water discharge, and that’s very different from the water that we use to take a shower. When you take a shower, you withdraw a lot of water, but there’s not much consumption. 

O’NEILL: From what I understand in the United States, at least, the water that’s used in these AI data centers to do this cooling comes from local or municipal sources. What impact does an AI data center have on the local community around it? 

REN: In the U.S., roughly 80 percent to 90 percent of the water consumption for data centers is coming from [public] water sources. We did some preliminary study and it shows that currently in the U.S. the data centers’ water consumption is already roughly 2 percent to 3 percent of the public water supplies. So here we’re talking about consumption, not water withdrawal. Based on the estimate by EPRI [Electric Power Research Institute], AI energy demand by [the] 2030s could go up to 8 percent.

O’NEILL: There’s an ongoing debate in Memphis, where tech billionaire Elon Musk is trying to build a massive server to accommodate AI. The local utility estimates that this system is going to need something like a million gallons of water per day to cool it. From your perspective, how should local communities weigh the benefits versus the cost of having these local AI data centers?

O’NEIL: Now, the companies that created and operate these AI systems themselves have an interest in making the technology more efficient. What kind of possible improvements in AI technology could make it more energy or water efficient over time?

REN: They definitely have the incentive to reduce the energy consumption, reduce the resource consumption for training and inference. We have seen a lot of research proposals and solutions that promise to reduce the energy consumption, but it turns out that in reality, the systems are not that optimized. 

I saw a paper from a top tech company’s research team, and it shows that the energy consumption is 10 times higher than what we thought before, even though they’re using state-of-the-art optimization techniques. So they do have the incentive to reduce the energy and resource usage for AI computing. 

However, the real world is a different story, partly because they have strict service-level objectives to meet, which means they need to return the responses to users in a short amount of time—and that limits how well they could optimize their system. If they are just doing batch processing, they could be very energy efficient, but it turns out, in reality, there are a lot of constraints which prohibit them from using those optimization techniques. 

Maybe we can compare the bus versus a passenger car. In general, per passenger, the bus should be more energy efficient than a passenger car, assuming the bus is fully loaded. But in reality, due to the user requests, random patterns and some other constraints, the bus is not fully loaded at all. If you have a 50-passenger bus, usually it’s just loaded for five passengers, and on average per passenger, fuel efficiency is much worse than the passenger car.

O’NEILL: AI has become a really immense part of many people’s day-to-day lives. It’s supposed to make our lives easier, but it comes at this sort of tremendous cost to the environment. What’s the solution here? If the technological advances aren’t working out the way we’re hoping they will, what’s the fix?

REN: One potential fix is, instead of using larger and larger models, we could be using smaller and smaller models, because usually those smaller models are good enough to complete many of the tasks that we actually care about. 

For example, if you just want to know the weather, or a summary of text, using a smaller model is usually good enough, and a smaller model means you’re going to save a lot of resources and energy consumption. Sometimes you could even render small models on your cell phone, and that can further save the energy by, say, 80 percent very easily, compared to running a larger model on the cloud.