How much water and power do AI data centers actually use and how will they handle the incoming spike in demand? Let's talk about it with Christophe Beck, the CEO of Ecolab, in a conversation brought to you by Ecolab. And Christophe is here with us in studio. Christophe, great to see you.
Yeah, good to see you as well, Alex.
So, I'm excited to speak with you because you know what's going on in the AI data center buildout, right? We hear all these rumors about how much water and how much power an AI data center actually takes to run, and you're actually working in the field and maybe you can help enlighten us in terms of when a company goes to stand up an AI data center, what is the lift of power, water, the consumption that they use, and are they this looming environmental disaster that some people claim they are?
It's the revolution of our times. AI started in big ways four years ago. It's like electricity suddenly got invented again, like 150 years ago when Thomas Edison invented the light bulb. Well, it's happening right now and it's going to happen not in 150 years but in 10 years probably. This is the really good side of it. At the same time, you need huge natural resources, to your point, in order to power AI.
We estimate that in the next four or five years by 2030, AI will need the incremental power of the whole of India and the incremental drinking water of the whole of the United States. That's what's at stake. And when we know that we already have a gap in terms of how much water nature can supply versus how much we need, we see that everywhere in New York City, in the US, anywhere around the world. Look at Europe right now. We know we have a problem. We had a problem, we have a bigger problem today. So when we think about the impact on natural resources, it's going to be huge. If we don't do anything, we're going to have a problem. We can dream about AI. Well, the physical infrastructure won't be able to follow.
And today 60% of the data center projects are delayed mostly because natural resources are not here or because communities are pushing back.
That's wild. So you said that it's going to take an incremental basically India in power and US in water. That's right.
That's per year. Yes. So, okay.
And it's growing. It's exponential. It's an AI story obviously. It's not linear. It's something that's moving every week. It's booming.
And now you work in water cooling of AI chips within data centers, and so you're going to explain to us a little bit about the solutions here. But I want to tell you I'm struck a little bit because when you hear folks on the pro AI side talk about the power and the water consumption, often times they make it out like it's not going to be a very big deal, but then I hear you talk about this incremental India in power and US in water per year and it seems like actually that is a very big environmental impact.
It is a big deal and the fact that so many of those projects are delayed — they're going to happen at some point but not in the time frame that we were all hoping — is an indication that there's an issue. It's a natural resources issue but at the same time communities are reacting. And when we talk about communities, well, it's you and I, it's our neighbors, it's our families, it's our employees. It's not people living on a different planet.
Obviously people are saying, well, you're going to take my water, my power bill is going to go up, I don't want to have that data center in my backyard as well. So when we think about the last few years, it's an innovation. AI is something that didn't really exist five years ago. So we had to try new ways. Well, they're imperfect at the beginning by nature. It's innovation. Well, it's to recognize that the way we've built them so far, the way we've operated them so far, was maybe not the ideal way. And once we recognize that, we can embrace new solutions.
And for us, Ecolab, we've built the world's water company. Over the last 103 years, actually, we serve three million locations around the world from data centers to fabs to hospitals to schools to breweries, you name it, in 40 different industries. The way we've approached the water challenge has always been, let's do it the same way as nature does. We always use the same water for millions of years.
We can't think too much where it's coming from because ultimately every water that you've been using today, you'll be reusing it at some point as well. So down the road our approach is to say, let's do the same in a data center, in a brewery, in a hotel. Let's reuse the water. But that's hard to do because it's effluent water. Obviously nobody wants to reuse effluent water. That's a little bit of a hard thing to fathom.
Obviously, but if you use the right technology, the right data technology as well, you can use water in a closed circuit, which is what nature does through the clouds, the lakes, the rain, and so on. And that's what we're doing with data centers. So, when we're looking at how data centers are being cooled today, well, it's big AC air conditioning in a room where you have a bunch of computers that keep getting hotter and you need to use water in cooling towers in order to cool the air. This is gone. This is not the way we're going to solve it tomorrow.
That's why we've developed solutions that we call direct to chip cooling technologies where you can bring a water-based product directly to each chip. You cool it again, you bring it back and you can get compute power and at the same time you need no incremental water. So you get in a place where you can get AI at the same time not using water. The way we share it with politicians in our areas for instance to say, well, the new data centers are not going to use more water than a car wash — do you have a problem with that? Maybe not, and the discussion changes.
Why do the data centers take so much water and power to operate?
It's different reasons. So first, if we look at the chips themselves, the way the technology works, well, you need power, you need electrons in order to get the system, the logic system to work in a chip. And the more of those chips you have, the faster they are, the more data they can ingest, the more power you need to make them compute ultimately. So the higher the compute performance of the chip, the more power you need at the same time. And that's the tricky component.
Well, the more heat you generate — you have almost a direct relationship between how much power you use and how much heat you create. So that heat, you need to remove it because if you don't remove it, the chip is going to melt. It's going to die. Not good things are going to happen. And when you cool the chip, well, so far we've been using water to cool those chips through a cooling tower like you have in an air conditioning in a normal building that uses a lot of water.
So, they're immersed in the water or?
They're not immersed in the water. So today, the way it is is that you have computer racks in a big room that's air conditioned and your air conditioning unit is connected to a cooling tower like you have in your buildings. Well, that cooling tower uses a lot of water and that's how the water is being used.
So far, and that's what's changing.
But at the same time, Alex, what's important to know is that when you power the chips, the second largest user of water in the world behind agriculture is power. So to power the chips, you need water. To cool the chips, you need water. And to make it even more extreme, to produce the chips, you need water because they're produced in ultra pure water.
And so when people talk about the water consumptions that AI data centers use, often times they'll be like, well, they're just using new water and it's coming in and they say they're wasting that water and it's just this constant supply of water coming into the data centers. But you said that the plan is to actually use all that and use the amount of a car wash. So how much of the AI data centers that we have today, how much of them are this more wasteful version? How much of them end up using the amount of a car wash and how do the data centers that are currently more wasteful end up moving to this more car wash volume version of a data center?
So that's the good news and the bad news. Since AI was kind of truly invented four or five years ago, most of the data centers that have been built are not using that direct to chip liquid technology. So 10,000 data centers around the world today, one is being built roughly every week as we speak. So it's a lot.
Mostly air conditioning that they're using to cool.
Most of the ones that have been built so far. So we're in a place where 95, 97% of the data centers that are existing out there are cooled by air, the air conditioning approach. The newer ones are liquid cooled and the very newest ones are liquid cooled in a closed loop circuit that doesn't use any water.
And that's what you're working on.
That's what we're working on. So in a way, you can look at the not so good news: 97% of what we've built, not that great. Lot of water, a lot of power, not always built at the right places. The very good news is that the latest ones that are being built with the direct to chip circular water cooling technology, ultimately are using less water than a car wash.
Okay. So, then talk a little bit about the technology progression from the air conditioning style version of this to what you're going to do. How does the data center change physically when you go from like let's just fill a cooling tower with water, blast it through ACs and hope the chips don't melt to what you're trying to do?
So the change is quite dramatic. And interestingly enough, Alex, the technology got invented 20 years ago for GPUs, for video games. Interestingly enough, and we know the whole Nvidia story is based on GPUs more than the CPUs, which were related to graphics, to games back then because we knew that those chips were super powerful but using a lot of power and creating a lot of heat and needed to have a different way of cooling them. Hence the direct to chip cooling.
You fast forward 20 years later, Nvidia and others coming with those super powerful chips, creating a lot of heat that needed to be removed. Air conditioning didn't work anymore. So the chips would have been melting or disappearing while they were processing the data. So the idea was to say, okay, if it worked for those GPUs, those graphics processing units in the past, could we use them for the latest generation of chips?
How does that work? Ultimately, with the difference of air conditioning, well, you blow cool air on computers. That's not enough anymore. What about bringing water directly on a chip? They're not immersed. That could be a solution as well. You just have some technology challenges. Just imagine, you need to fix the computer — you need to remove all the liquid, you need to dry it. Kind of a crazy thing to do. So why don't we do something which is more practical where you bring the water directly to the chip through a small radiator.
So where you have thin tubes that are coming on top of each of those chips — but it can be thousands, millions of times in a data center — you get that liquid close to the chip. You bring it back to a cooling distribution unit. It's like a big fridge, a bit more complicated obviously than that. You reprocess the water. You make sure it's at the right temperature. You make sure that it's clean to get back as well. So to the chip, and you keep doing that over and over again. And that's how you get higher performance of the chip and ultimately less water than a car wash.
So there could be just a closed circuit of water that's recirculating through this system that goes onto the chips, cools the chips down, when it heats up enough, goes back into the center distribution tank, and then comes back through.
That's right. That's the easy part. The only thing is that as we all know as humans, water doesn't like to be recycled. When you have water that's staying, especially if it's warm temperature — usually the chip operates around 30 Celsius, for chips it's roughly 75, 80 degrees, something like that. Just imagine water at that temperature staying in that state for a long time.
You can think about the pond. You can think about your coffee pot. Well, bad things happen. It turns green because you have biological contamination or you have scaling, the white stuff that you see in your coffee pot as well. You can have corrosion as well. You can have leaks. Water doesn't like to be reused. That's why it's such a big issue.
And we need to make sure when the water comes back from the chip, it needs to be standardized again where you remove all biocontamination or risk of scaling and making sure that it's as pure as it can be before it gets back to the chip to get the max performance and that you don't clog anything or create a bigger problem because the whole data center is going to stop.
I think something that people don't fully appreciate, at least I didn't know until I started looking into it further, was the water that's used to cool chips needs to be many times purer than even the purest drinking water.
It does, for the reason we just talked about — making sure that nothing bad happens when water is being reused, especially at lukewarm temperature. The very interesting thing, Alex, is that we've learned how to treat the most extreme water, which is in health environment or pharma environment, when you think that many of the vaccines, unfortunately we went through that with COVID five years ago, well, you have water with the vaccine coming into your bloodstream.
We've learned that the most extreme pure water that's being used on the planet is to produce chips. Okay? Because those chips need to be produced in ultra pure water because any impurity that is in the water that comes on the chip kills the chip. And that's how we've learned at Ecolab on how to manage that process, which is between a thousand or a million times more pure than what you find in a drug.
So when you think about that type of purity, one of our experts were talking about it. You would take the great lake, let's take Lake Michigan, you would think it's ultra pure water. You drop a cube of sugar in Chicago in the lake. That lake is not ultra pure anymore. Just to put it in perspective what it means ultra pure.
So that technology, we're going to use it in data centers as well to make sure that it's always the purest water that's getting back to the chip where you get not only the best cooling performance but at the same time you can drive the performance of the chip as well. And we can talk more about that too.
Hey Christophe, what happens to the heat? I mean the heat is coming out no matter what. So what do data centers do with that heat coming off of the chips?
There's a lot of heat that's coming out. When to put it in perspective as well, when a company is saying we're going to install one gigawatt of compute power out there. One gigawatt generates roughly one gigawatt of heat. One gigawatt is roughly one nuclear plant. So when a company like one close to us said, we're going to build eight gigawatt of power out there of compute power — it's eight nuclear plants.
Well, you need to cool all that to your point. So what do you do with that heat? Well, you have a bunch of options. Some are good, some are less good. Obviously, one of them, which is the advantage of not using water, you bring it back to the air. So you warm up the atmosphere to a certain extent but that doesn't change much from a climate perspective but you don't use any water.
A second one is to warm water that could be used when you have data centers that are being placed in cities, which is a new trend that's happening out there. You can use it for district heating or district cooling depending on the physical process that you're using as well. Well, that heat that's being generated by the data center can heat homes or heat any process in a plant.
And a third one, which is something that we're working on, it's to reuse that heat to produce electricity. And if you produce electricity, you reduce the electricity that's needed to power the data center as well, which I think is going to be the future because ultimately you're not only trying to get zero water being used, but you're trying to minimize the power that's being used as well in the data centers. So, you have many options here, some better than others.
That's pretty cool. So, the heat coming off the water could eventually be transitioned to power that's used to power the data centers. How would that work?
Because when you think about a data center, the power consumption — it's roughly 60% of the power is being used to compute and 40% is being used to cool.
Well, cooling is not helping any AI application because that's a byproduct obviously here. If you can move this 40% to 10%, well, you get 90% of the power being used for the compute, which is what we're trying to accomplish. So it's to really drive the AI compute performance.
To your question on how do you transfer the heat into a system that generates electricity — well, that's the core on how electricity is generated in a thermal system today, which can be a coal plant, it can be a gas power plant, it can be a nuclear plant. Ultimately you heat water, you generate steam, you make a turbine turn which generates electricity and you come back. So, it's to use that heat like you would do in a geothermal facility as well where you bring the heat from the ground. That's what's happening in Yellowstone.
We don't have any geothermal plants in Yellowstone. But from a principle perspective, you get the heat out of the earth. Well, you can get the heat out of a data center that ultimately warms a fluid that generates electricity that you can reuse in the data center or elsewhere.
You know, I want to talk a little bit about what you mentioned earlier, that this is going to be on an exponential, right? That the demand for data centers is growing exponentially, which tracks what we're seeing about the demand for AI technology. And I want to get from your perspective sort of how much demand we're seeing or you're seeing because you're in it, in terms of new data center construction. Like, can you help quantify that for us or give us a little bit about what you're hearing about? Like, there used to be maybe one a year, now there's 50 a?
And it's been changing every year. When we started on that journey, right before COVID, it was roughly one data center a year that was being built, and today one a week. And 50% of them are in the United States, 30% are in China and 20% everywhere else, which is kind of an indication as well where the trend is going. So from an AI perspective, that's obviously generating a lot of new power needs, a lot of new water needs, and a lot of new space needs, which is obviously creating all the community challenges that we're facing.
Well, in New York it's pretty clear what's happening right now as well, and we can talk about that. At the same time, you need to deal with the older data centers. We know from all the chip manufacturers or designers, all the famous names that we're familiar with, they're coming with new generations every 10 months roughly now. And every 10 months the power is multiplied by 10 and that's becoming six months and it's going to be even shorter going forward.
Well, you see that those new chips are going to be used in the older data centers to replace the old chips with the new chips. So we're going to have a combination of new data centers with the latest technology and the latest technology being replaced in the older data centers as well. Which means that the older data centers will have to be retrofitted with new technology otherwise the power and the water needs are going to explode once again. So it's a combination of new data centers and new chip generation in older data centers.
