Cooling the AI buildout

0:00Hey, everybody. It's Ronnie. It's July 8th, and I'm here in our studio in New York with Will Thompson from our thematic investing group. We've got a very interesting topic for you today. We want to get into a recent report that Will wrote on demystifying data center cooling and water use. Thanks for having me, Ronnie. Will, when I tend to think about a lot of the complexities around data centers, I naturally go to power constraints. I naturally go to permitting. But this water issue feels like a pretty serious issue. Can you walk us through the needs to cool data centers and the approaches that folks are taking?

0:33Yeah, maybe let me just set the stage. And the fact that cooling and water use is one aspect of some of the constraints we see around the digital infrastructure build out. We do think we'll get increasing focus. And that's why we want to address it in this report, because based on our conversation, there is a lot of confusion around this topic. Keep in mind that a data center is effectively a large warehouse with a bunch of electric boilers. That's a weird concept to think about when we think about digital infrastructure and compute. What actually happens is the vast majority of the power that's serving the servers that are in a data center turn heat as a waste byproduct.

1:01If you don't remove that heat, you have performance degradation in the compute. You can have downtime of the compute and you can have effectively complete failure of the equipment, including melting the chips. These things are made of silk. Thermal meltdowns. Is that what they're called? Well, I think that's a nuclear topic. I like the analogy. So why is this historically hasn't been a big issue? Because, you know, when we think about server power density, it was 10 to 20 kilowatts, which wasn't a lot when it comes about conventional cloud compute. With AI, we've already jumped to over 100 kilowatts per server rack.

1:31Next generation of GPUs, we're talking about hundreds of kilowatts. And, you know, I think even beyond that, people are talking about one megawatt per rack. That has a ton of power. That's a ton of energy and a ton of heat that needs to be removed from the servers. So you bring that in and then water use, there's a lot of conflation around water use. So obviously a lot of investor focus on the transition from air to liquid cooling, right? That doesn't use water. It's water used in it, but it's a closed loop system. Once you fill the system, you don't need a replenished system. That's one part.

1:59So there's two parts that need to happen. Cool a data center, move heat from the servers. That gets the most attention. The other part that doesn't get a lot of attention, you have to reject the heat into the atmosphere. That's where water is used. So why is water used for that process? It's called evaporative cooling. It's a conventional HVAC technique, you've maybe been in a city and you've seen steam come off the top of a commercial building during the middle of summer. What they're doing is they're introducing water to heat. The water turns from a liquid to a vapor. And what you do is you're rejecting heat much more efficiently with less energy than you would if you didn't use water.

2:29That is the cause of water use for data centers. That's what's being scrutinized. And how much water? Put it in context. Yeah. So Google just put out their sustainability report. They disclosed that their water use jumped over 30 % in 2025. They disclosed 50 billion liters of water used in that year. That sounds like a lot. Yeah. I mean, to put in context, it's about the irrigation needs of 90 golf courses in Southern part of US, but it's concentrated during the summer months, right? That water use is not used throughout the year. Water is concentrated during the hottest month and it's overlapped with obviously drought season.

3:02So that's what creates the complications about water use when it comes to data centers. Probably drives a lot of the backlash as well. Yeah, it's becoming, it's a local issue among another issues around higher utility bills and the implications of having a lot of construction jobs just show up in your communities without housing, without the public services. That's a not my backyard issue that is also layered into this. But water, the higher utility bills that come with power demand and that local permitting issue is becoming a big issue than when we talk about the execution risk that comes with the digital infrastructure build out.

3:30You also illustrated how there's choice here. A lot of the major data center operators are choosing different paths. Can you talk about those paths? Yeah, the hyperscalers have been really interesting. The hyperscalers have really diverged around their cooling and water use strategies. On one camp, you have Microsoft and Meta who said, we will stop using evaporative cooling for our future data centers. So they're trying to reduce their direct water use. On the other camp, you have Google and Amazon say, look, we should use water when it's available and we're in low risk watershed regions because if we don't use water, we need more power.

4:00We need more power. There's higher emissions than that power generation because let's be honest, these guys aren't using 100 % renewables. And thirdly, we talk about thermal power. So think coal, natural gas plants and conventional nukes. That's an indirect water use. What you have is one camp saying, let's stop using water, direct water use. You have another camp saying, look, we have to think this more holistically in terms of the direct and indirect implications across water and emissions. I don't know what the right answer is, but it's really interesting to see how the hyperscalers can agree with the right approaches.

4:28This feels so multivariable, so complex to me. Let's take it to markets and investing. What do you think this means for investing in markets in the foreseeable future? Look, we've been very bullish on compute and power demand when it comes to AI. But I've also been to the view there's a lot of emerging execution risks. I can definitely feel a lot of caution in the way you're describing this. So clearly AI has had a big impact in the market in the last couple of years. But what we're seeing from the convergence of all these execution risks, the complexity of getting these all right and aligned at the same time to meet a surge in compute demand makes me more balanced of just understanding what are the risk factors that come at the price in the market.

5:08So we see a convergence of execution risks. One, power constraints. In my view, we've already hit the grid power wall, which has led to a lot of these data centers being powered behind the meter, which there's questions of whether that's sustainable. Two, you got labor shortages, right? We need a lot of skilled labor in terms of mechanics, pipe fitters, HVAC specialists, and plumbers. And thirdly, local resistance, which I alluded to earlier. We often heard about not in my backyard NIMBYism. The new acronym I've learned is BANANA, which is build absolutely nothing anywhere near anybody. We're starting to hear a lot more about our data center moratoriums at the local level and potentially at the state level.

5:39I think the hyperscalers have grossly underestimated how much data centers and power are local issues. And they have to get this right to be able to get the permitting and the right approvals to actually build these things. I can tell you that this is niche right now. It's really not in the market narrative. When do you expect this to hit the market narrative? When do you think we'll see signs of this potential slowdown coming? We are seeing signs of the challenges, right? It wasn't too long ago that they had ribbon cutting ceremonies for data centers. That is reversed, right? People are asking about the impact of higher utility bills that come with data centers, environmental implications, including water.

6:11So I think you're going to see issues show up in early 2027. This is all very fascinating. What do you think it means for us in our quest in the AI race as a country? Yeah, look, there's clearly national interest of AI, right? Where there's an AI arms race against China that really resonates in three places in the United States. You got Silicon Valley, Washington, D.C., and Wall Street. Real Americans just want to pay their bills, right? They don't want disruption to their way of living in their communities. Data centers are shaking that up. I don't blame data centers for higher utility bills.

6:43I think it's two decades of underinvestment in grid infrastructure. That's not helping. So I think there's going to be a national debate of the benefit of AI and obviously how we're going to be particularly autocratic versus democratic AI in the arms race of China. These assets are of national interest, but we need to get the local support to do it. We need to make sure it's not impacting people in terms of their utility bills, their affordability, and that we're doing this from an environmental standpoint the correct way so that it does get the local support. This was so fascinating. I really appreciate you spending time in the studio with us.

7:11It's clear to me that we're pretty far away from an equilibrium here on this topic. So let's keep watching together and probably have to bring you back soon. Keep the listeners informed. Thank you, Will. Appreciate the time, Ronnie. Thank you. In summary, this conversation with Will was a serious reminder to me that even with the power of AI that feels undeniable, there is still a lot of complexity in the phase shift and the build out needed to support it. In addition to the power constraints and regulatory constraints, we're learning more about the cooling constraints and a lot of the backlash from local communities.

7:45All of these dynamics will matter tremendously for markets. So let's all keep our eyes on them. Thanks for joining us. And please remember to hit subscribe wherever you listen to your podcasts to be notified of future episodes.