Why Meta Builds Its Own AI Data Centres

[00:00:00] Rachel Peterson: What we had constructed over the first 15 years of our data center program probably constructed over the last one to two years alone. My name is Rachel Peterson. I lead the data center team here [00:00:11] Speaker 2: at Meta. Why does Meta have their own data infrastructure? It deserves us to go back in time a little bit. What would your response be to people who are concerned around strain on their [00:00:21] Rachel Peterson: local infrastructure? So we pay our full share in this process. Is there anything about this [00:00:26] Speaker 2: industry that perhaps keeps you up at night? Any hope that everything's going to work as [00:00:29] Rachel Peterson: expected? Those are the kinds of things that I think keep me more up at night. [00:00:36] Speaker 2: If you follow me on social media you will know that over the past few months I have made multiple visits behind the scenes at some of Meta's most impressive facilities. All the way from their silicon labs where they are designing bespoke AI hardware up to their data centers powering the services that we use on a daily basis. And since I've been finding out so much about these facilities I thought that it would be a really interesting idea to sit down with some of the people that are behind the planning construction and deployment of this infrastructure. So without further ado this is my conversation with Rachel Peterson from Meta. So Rachel can you first introduce yourself and tell me about [00:01:14] Rachel Peterson: your role at Meta? So my name is Rachel Peterson and I lead the data center team here at Meta. And so what that means is that my team is responsible for basically the life cycle of the data center. So from the very start where we choose the locations that we want to go into to the actual data center design to the energy we procure to power the data center through construction and operation. So it's really the entire life cycle. [00:01:43] Speaker 2: Okay. So I want to start a conversation off by asking why does Meta have their own data infrastructure? Why aren't you using one of the cloud providers that are already out there in the marketplace? [00:01:55] Rachel Peterson: Well it's a great question because it actually it deserves us to go back in time a little bit. And so I joined back in 2009 and at that time Meta Facebook at that point was just kicking off their sort of data center program which is why I joined. And at that time we were very small compared to where we are today and we were in a bunch of rented facilities of data center providers. But even though we are a lot smaller we saw a huge opportunity in that we were looking at the landscape of the data centers we were in and we saw that we could do it probably less expensive. There was a lot of redundancy built in the data center and we didn't think was actually necessary. We thought we could do it faster. We basically saw an opportunity to sort of redefine how data centers are built and constructed and designed from day one. And also we wanted to control our own destiny. That meant not just building the data center but really controlling the full stack. That goes to everything from the hardware that's going into the data center to the software that we're designing to building a really custom built infrastructure that supports the apps and services that we needed to support in our own unique way. And so we had an opportunity to do that in a meaningful way even at that small scale. Going back to 2009 I never would have predicted we were at the massive scale that we are today. But I'm really happy that we made the decision to do that at that point in time because it absolutely set the standard for us to be able to create the amount of compute in the way that we thought back then to scale in a way that's really important for us in the AI world that we're living in. [00:03:42] Speaker 2: Yeah, I think it was a really efficient decision, you've got a really efficient outcome. [00:03:47] Rachel Peterson: Yeah. [00:03:48] Speaker 2: So during this trip, I've heard the term personal super intelligence quite a few times. And I'd like to gain an understanding of what your vision of personal super intelligence is. [00:03:59] Rachel Peterson: The way I think about it is having the utmost expert who understands your unique preferences, your unique ways of, you know, wanting to, you know, to operate that is personalized to you that makes you the most effective and efficient person that you can be. So for me, I look at that as super intelligence. And I think that has different meanings for different people and how that plays out will be very different for, I think, different companies and applications. But I think it is having that unique expert who really knows you and can really work on behalf of you. [00:04:35] Speaker 2: What does the infrastructure behind personal super intelligence look like in your mind? [00:04:41] Rachel Peterson: Yeah. Well, we really look at, first of all, personal super intelligence, I imagine, will require a massive amount of infrastructure and compute. And so to make that happen, you really have to look at it from the ground up. And for us, that means finding data center locations that we can provide ample amounts of compute in a timely efficient way that has a custom power infrastructure that we need. Mm-hmm. That has the data center designed uniquely for the needs of that location. We have all the state of the art technology within the data center, state of the art cooling technology that we use to cool these massive supercomputers that require a lot of power to be able to heat the GPUs or to cool down the GPUs. And there we look at it from sort of like the full stacks. So not only do we have a custom built data center that can deal with all the compute needs that we have there, but that we really design it with flexibility in mind. Mm-hmm. So that means being able to deploy all the servers, the custom servers that we've designed for this facility, to be able to take the different types of silicon that would be required. Mm-hmm. Mm-hmm. So that means that we're able to build our own custom chips that we're designing or going with our providers such as AMD or NVIDIA in terms of the actual GPUs that we're putting into these facilities. To the network stack, which is really our custom design as well. And then, of course, to the software stack that is enabling us to be able to serve to our users or develop the models that we need at these locations. [00:06:15] Speaker 2: So you touched on the bespoke aspect of a data center being largely dependent on the location of that data center site. Yeah. So how do you pick a site? And I'm asking that question from a broad point of view of maybe like state or region. Yeah. But also then down to, you know, deciding do we pick this town versus this town that might only be 10 miles apart, for example. [00:06:38] Rachel Peterson: So we start off by looking at a lot of parameters. Mm-hmm. And so we'll look at everything from a lot of it's time to market. Okay. So, you know, where can we find the amounts of infrastructure that we'll need for this site? So we'll look at power infrastructure, like I said, water. We'll look at how quickly we can build the data center. So oftentimes look at labor, look at how we can attract construction labor, the employees who may live there at the end of the day, as these are 20 year investments. So we have to really be thoughtful about the time horizon. We will look at where ease of doing business or if there's any sort of regulatory or legal issues that we may need to look at. And really, we look a lot at community fit. I mean, the community has to be sort of a place where, like I said, we are going to put our roots down for 20 years. And that means we want to be part of the community. It means that we not only work there, we live there as well. So we try to employ from the community as much as possible. [00:07:36] Speaker 2: On the topic of community then, what responsibilities would you guys say that you have to the local communities that you might be building a data center next to? [00:07:43] Rachel Peterson: For example, one thing that we've done for many years now in support of our water goals is to restore more water than we're actually consuming in these communities. And so we do that, for example, by, you know, giving back through restoring the aquifer or could be something in the local habitat that we're contributing to that helps from a water standpoint. Right. We also have felt very strong that we're, like I said, we're not only working there, we're part of the community. So that means that we are actively hiring from the local community for our construction and operations personnel. Or we're hiring from any of the jobs that we have across the data center to support this massive campus that we have in the region. Or it means that we are contributing through our local grant giving back process, which we've done for many, many years in each of these communities. So we look at it really as us being a vital part of giving back in the communities as well as hiring into the data center. And these are really clean operations that we have set up. So we feel like it's been a really good program. And we've had a lot of communities where we've had a lot of success, where they've really welcomed Meta in, and we've really done our best to show up and be great community partners. [00:08:57] Speaker 2: Now I want to talk about the infrastructure that supports all of all of your data center sites. And I think there is quite a valid perspective around the infrastructure strain that a data center might cause. And I think that simply comes from the fact that people see that gray box on the landscape. So what would your response be to people who are maybe concerned around strain on their local infrastructure? [00:09:22] Rachel Peterson: We start early on working with the utility grid providers in these locations, first around feasibility, and then how we figure out how do we get supply to the data center. That entails a long process and working in quite close partnership with an ultimate goal, not only of providing power to the data center, but how do we ensure that we're not passing on any cost to any consumers around the data center? Okay. So we pay our full share in this process. And then we want to work with the utility provider to make sure that we're actually, like I said earlier, giving back and be part of the community. For example, in our Louisiana data center, we work closely with Entergy, the utility, ensuring that we're paying for the infrastructure that we're putting in place. But we actually are going to be saving their customers more than $2 billion of what they would have paid otherwise had Meta not gone to that location. [00:10:22] Speaker 2: On that $2 billion stat, you said that that's money that customers would have otherwise paid had you guys not been deploying into that area. How would they have been paying that? Is that through increases in energy bills or what's the reason for it? [00:10:39] Rachel Peterson: Yeah, increases in their energy bills. And so when we're coming into these communities and our infrastructure investment in Louisiana was so substantial, we were actually making the grid more efficient. Okay. [00:10:51] Speaker 2: So you kind of needed this upgrade already. [00:10:53] Rachel Peterson: Exactly. And more efficient upgrades needs that would have otherwise happened and been passed on to the consumers. We're actually picking up the bill for a lot of that through our development. So I think, again, this is a misconception that people may have of us and other data centers coming in and taking more than their fair share, passing on costs. When in reality, our investments are oftentimes helping the electric grid, bringing reliability and consuming costs that otherwise would have been passed on to everyday residential consumers. [00:11:23] Speaker 2: I want to talk about water now because I think it's an unavoidable topic in the current day when you're talking about data centers. So are there any misconceptions that you perhaps want to debunk or talk about for a few minutes? [00:11:36] Rachel Peterson: I think there are some misconceptions that would be helpful to clear up. One is that we design our data centers to be a completely closed loop system. So what does that mean? That means that we have a closed loop that cools the GPUs and cools the facility and it circulates within the facility and it doesn't go anywhere. It's basically used within the facility and it's not discharged anywhere. And then we have coolers. We have to cool sort of the general facility with additional coolers. Those specific coolers that we use have minimal water usage throughout the year, basically zero. There could be hotter days in the year across some of our locations where we may draw on some of the water, but it's pretty minimal. Let me give you a stat around that. One of our recent facilities that will be operational soon is in Beaver Dam. And that facility, for example, the amount of water we will be using there per year is going to be less than two full service restaurants. Okay. So if you look at it from that standpoint, and I think that maybe puts it in perspective for people, the water usage is very minimal. And then beyond that, for the water that we could be using, for example, on our Richland Parish site in Louisiana, we're actually drawing from sources that are non-potable water. That means that the water we're using for that facility, the amount that it is, which is pretty minimal, is coming from sources that people are not going to be drinking from anyways. Right. Okay. So we're going to be very efficient. And then for any disposal of the, any excess water that may have to get disposed of will go into this water sewer treatment system. It does not go into some sort of community, you know, system that gets shipped to somebody's household. Yeah. So I think that's a big misconception on how the actual water consumed by these data centers and what actually happens to the water if we need to dispose of it. [00:13:34] Speaker 2: As we head into the end of this conversation, I want to talk about the future, specifically five years into the future. That sounds, does sound like a long time, but on the topic of data center construction, it really isn't because, you know, these are multi-year construction projects alone, right? Correct. So what do you believe Meta's data center fleet is going to look like in 2030, for example? [00:13:56] Rachel Peterson: You know, it's such a great question because again, going back to where we started the conversation in 2009, 2010, I would have never projected that we would be at this point now. Yeah. And in fact, it's interesting because what we had constructed over the first 15 years of our data center program, we've probably constructed over the last one to two years alone. Okay. And so that's a significant scale up and I never would have expected us to be able to do that. Even the size of our campuses have grown multifold only over the last couple of years. And so what I find fascinating about that is just the rate of change in the industry and the rate of development of our own program itself. And so we've really shifted looking at maybe ourselves as just a software company in some ways to a full on infrastructure company, which is a very big shift in how we think about not only environment we're in, but a necessity to be able to scale and support AI at scale and super intelligence. And so if I look forward to 2030, I'm not exactly sure what to expect because the rate of change is so significant right now. But I do think the amount of infrastructure that we are going to be deploying and of course the industry in general is at this very steep trajectory and I think we'll get more efficient. We'll do more interesting things and then we're going to be doing our most to be able to develop what's needed to support AI growth and to support meta, of course. And so I think it's very exciting to think about what might be around the corner. [00:15:28] Speaker 2: I have one final quite personal question then. Is there anything about this industry that perhaps keeps you up at night? [00:15:36] Rachel Peterson: I think the pace and complexity at which we're evolving is something that is both very invigorating, but also something that I think about a lot because we're, like I said, we're a new territory. I think everybody is a new territory. It's an unknown in terms of the scale that we're delivering at. Some of the things that we're doing have never been done. They're first of kind from a design standpoint to the power plants that we're putting in on site to, like I said, the technology that we're using. It's all first of kind. And so when you're doing it at the rate of pace that we're doing it, plus you're putting a lot of new technology together, you know, you wonder and you hope that everything's going to work as expected. Those are the kinds of things that I think keep me more up at night thinking about what are we putting and what's to come. [00:16:25] Speaker 2: I want to say a huge thank you to the team at Meta for inviting me in to learn more about their infrastructure. I know I really enjoyed this conversation. And if you did, too, you can let me know by leaving a like, subscribing if you're new and tapping the bell icon to get notified every time I post a new video. If the feedback is really good for this video, we will definitely schedule in more conversations like this.