Perfect. Well, good morning, everyone, and welcome to TD Cowen's 10th Annual Communications Infrastructure Summit. My name's Michael Elias, and I'm the communications infrastructure analyst here at TD Cowen. For this meeting, we have Equinix, and from Equinix, we have Jon Lin, who's their EVP and GM of Data Center Services. This session will be structured as a fireside chat, and we'll run about 40 minutes. I've prepared a list of questions, but I'll do my best to pause, and I promise I will try and pause for questions at the end. With that, Jon, thank you very much for being here.
Oh, my pleasure.
Really appreciate it. Now, Jon, I think you have some things that you need to lead off with, so I'll turn it over to you.
Yeah, just so Chip doesn't yell at me or hit me, and we don't get hit by legal, some of the conversation today will involve some forward-looking statements. Please read our SEC disclosures about risk factors, et cetera
All right. Well, I think, I think we're covered. We're ready to rock.
Okay.
Okay, so
Like the one public guys, right?
you know. Among a bunch of private guys. Let's get started with this. You know, Jon, I think you're very well known in the industry, but just maybe for some of those who may be less familiar with you, can you just give us an overview of your role at Equinix, and as part of that, how your role has evolved over time at the firm?
Sure. No, appreciate that. So I've been with Equinix 15 years. I currently run the data center business, so kind of 93% of the revenues, look at the global estate, what we're doing in terms of development, planning, capital allocation, and et cetera. Prior to this, I was president of the Americas, but again, been there 15 years. I started as a director in marketing, actually. Went to corporate strategy, had a finance role, and then, ran corp dev, did the Verizon acquisition, Infomart, a couple deals there, and, been there since, so
All right. Now, as part of that, where are you spending most of your time these days? Like, yeah, where are you spending your time now? I'm curious. Curious.
It'll be a shock, I think, for everyone to hear, like, utility and AI are kind of top of mind.
Okay.
In some ways, those are kind of obviously tightly coupled, but I think, you know, even prior to the GenAI kind of rush, so to speak, utility and the forward thinking about land banking and power banking has been something that Equinix has been focused on for probably, like, three or four years now, especially as we saw kind of some of the utility dislocations with Western Europe, et cetera. So it's been a huge focus area for us, right? [The question] is how do we make sure we're securing the future of the Equinix platform for long-term shareholder value?
Okay. All right. I will come back to that.
Yeah.
I have some questions on that, on that topic. But maybe let's kick things off to talk about demand, right? In 2023, I think the state of the enterprise, broadly, it was characterized by cost optimizations. You know, we saw a pullback in gross demand for Equinix off of the record levels we had seen in 2022. As you look at the landscape currently, how would you describe the state of the enterprise customer currently?
Yeah, I think there's the needs of digital transformation and digital infrastructure for the enterprise to just be viable into the future feels like, I mean, that's almost table stakes, which provides us, like, a very good, solid, like, fundamental base for which we can kind of build value off of. And so we still see a healthy amount of conversations there, like, still great engagement with both new customers and kind of forward expansion of existing. New use cases and workloads are particularly interesting to us. I think there's I might observe, like, two things. One, coming into this year, I think there was a lot of concern, I think, about the macro picture, right? It's, you know, I think 70% of the world's population or something was having an election this year.
I think interest rate dynamics were kind of question mark, question mark, and I think as, like, each month has passed and you've kind of seen a, a little bit better certainty around what that global picture might play out as, I think that's given additional confidence to, to the enterprise on how they're gonna invest around that spend. I think early in the year, though, there, there's still a lot of question about, you know, kind of what's the right emphasis and focus for the ability to deploy AI inside of the enterprise?
What we're seeing is, in, like, particular verticals where it's very, very high yield to delivery for, for the value that you can get out of there, like, particularly around, like, financial services, insurance, and the, and kind of life sciences, both on the healthcare side and, and on pharma. Like, there's big fundamental transformation that's gonna be, like, really, really, really massive top-line and, and bottom-line drivers for those industries.
I think across the rest of the general enterprise, though, there's a bit of a, "We need to get our data house in order to understand how to use AI, like, most, most, powerfully, and safely." And so that idea of structuring and making sure that the right governance mechanisms are in place for all of their data sets, understanding where they currently live even, and how to, like, get into and out of those estates, and then how to, like, move those data, sets into AI for ingest, again, in a safe way, and, on, on what we're increasingly seeing, for, for the large multinationals as on a private AI basis, is, like, really interesting. I think there's, again, a tremendous amount of AI activity that's, like, consumer-oriented, public-a, like, GPU and service stuff, like, huge amounts of workloads there.
Like, those are, like, interesting, but not, like, super compelling to us, I think from a durable trend perspective. Those are, like, you know, kind of as, as we've seen with a lot of these, consumer trends and cycles, like, it could be interesting for us. I don't know that it'll be there for the next, like, 7-10 years, right? Whereas I do believe, like, a lot of the longer-term AI value that's gonna be unlocked, like, this will be with us forever and fundamentally transform a lot of the economics of many industries.
Okay, I wanna take what you said, kind of disaggregate it, and then we can take it maybe a layer deeper. So, I met with Keith in June, and one of the things he was noting was the challenged macro backdrop, yet in spite of the challenged macro backdrop, you guys put up record gross bookings in the second quarter. My question to you was going to be, can you help me reconcile that tough macro backdrop with kind of the record pipeline that you guys are seeing? But I think at the beginning of your answer, you said that there's just better, there's more certainty-
Yeah
... around the macro backdrop. Would you say that's kind of the change there, that's driving this better bookings and better pipeline, that just people feel a little bit more comfortable?
I think that's helped. I'd say, look, the market dynamic has also changed over the course of the last, like, three or four years, but now, in particular, in the last, like-
Mm
... probably two or three quarters, right, you know, the large public players that are, like, out there and were, like, going after retail enterprises and, like, doing a lot of speculative development to, like, take those workloads. There aren't many of us anymore.
Mm.
There's very few. And then the amount of, like, incremental demand that manifested based off of GenAI and some of those workloads, et cetera, that we talked about, they've ended up consuming a lot of the capacity that's, like, out there in market, right?
Mm.
So when you think about, okay, you look at the macro landscape on the data center side, I don't know that there's any data center capacity coming online in 2025 and 2026 that isn't already pre-leased.
Yeah.
Except for maybe ourselves and maybe one or two other players in various markets doing speculative development there. Like, that means that that workflow is also starting to come to us, right? And so in ways that, I think both the velocity, kind of like the pricing firmness that we can see around that, and our ability to address that, and kind of like willingness to take those workloads is increasing. So I think that that's also part of the reason why we're seeing a lot more of that activity this last quarter, and probably going into the second half.
Okay, so keep. We're, we're just gonna keep stacking it. We're gonna keep building it here. Okay, so we got the better macro. Now, as part of that, I think what I just interpreted, or I think is implied in what you're saying, is that, like, look, you have seen so much hyperscale demand, that there's almost been a forcing function away of the enterprises outside of their traditional strong, some of their stomping grounds.
Mm-hmm.
And now that's creating almost a, you know, tailwind benefit to Equinix. And it sounds like from what you're saying, you are actually seeing that manifest.
Yeah.
In addition, you talked about kind of financial services, healthcare, which I would say those have been staples of the colocation market for a while. So you put those things together, and that helps drive the, you know, an increase in demand. One of the things that you mentioned as part of the initial response, which I thought was really interesting, is, you're talking about some of the other enterprises outside of that group getting ready for AI. It's almost like you're getting your data in order because you know the data is going to be key, be key in the AI, in an AI world. What are you seeing out of that group that maybe is like: "Okay, we need to do some re-architecting and figure out how we're going to have an AI strategy"? What are you seeing out of that?
Yeah, I think, you know, we've been on the journey with the enterprise for a long time, talking about, like, hybrid multi-cloud and, like, hey, making sure they understand what workloads are in which locations. A part of that for a long time has been, like, this notion that they need to have storage that's cloud adjacent. Like, that was probably one of the first use cases for hybrid cloud, was, hey, put your NetApp filer directly adjacent to AWS, and then you can have control of that data.
Mm.
I think that data aspect, though, and like, storage aspect is increasingly interesting, and so you've seen a lot of our activity out in the market around partnerships with Dell, partnerships with HPE
Yeah.
and NetApp, and Pure around, hey, that data element as a fundamental layer for the underpinning of what's gonna. go on top of that, I think is a really powerful thing for, for again, like, that every enterprise has to get ready. Again, those industries that I mentioned earlier, like financial services and healthcare, as you guys can imagine, like, they've been pretty mature on that data landscape, and that's why they've been so ready to deploy these AI technologies faster than most.
Mm.
Like, they've... Either from a regulatory basis or from a, you know, kind of statutory basis, they've needed to have better control of that data estate, and so they're, they're ready to go with it.
Interesting. So maybe building on that point, that. What you're saying suggests to me that, okay, like, for let's say maybe some of the laggards who still have on-premise deployments, and their data is, let's say, more disaggregated, essentially bringing all of that data, and in turn, which creates the need for storage next to the compute, so that then you could leverage the next generation of compute, the AI inference, and so on, you actually are able to unlock the value of your data.
That's right.
That's what I'm interpreting from
I think that's right.
from what you're saying.
But it's less than maybe about, like, data storage sprawl on the enterprise estate. 'Cause again, I think most customers, if they have storage platforms, they've got some level of maturity around that. It's more as they've developed the applications over the course of the last five years, and many of them have been cloud-first in nature, but, like, the data underpinnings of those application and workloads and, like, where is that data stored? Well, it's in the cloud.
Mm.
It's adjacent to that workload.
Yeah.
And now thinking through like: Okay, how do I start converging those sets of data like, into the rest of the data estate? Whereas before, they might have thought it
That's right.
as, like, a standalone piece.
Got it, okay. That's, that's helpful. Maybe staying maybe on a similar topic, but just a different approach here, is one of the things I've noticed on the last, you know, call, two earnings calls, has been, you've talked more about a greater contribution from large footprint deals, which I believe Equinix defines over 500 kW. Is that, is that right?
250, yeah.
Say you... Over 250-
Yeah.
... would be the point of demarcation. So earlier today, we had the enterprise data center roundtable, and that was one of the things that came up, seeing a skew to larger deal sizes. You know, my question for you is: Okay, deal sizes are getting larger. What, in your view, is underpinning that trend?
Yeah, well, here's a couple of different, like, really interesting areas, again, over the last couple of years that we've seen. One is, in the past, when you've thought about some of those large footprint deployments, it was more around, like, hey, migration of existing data estates from their on-prem location into something else. I'd say the last, like, two or three years, like, you don't really see that very much anymore. It's really, these are net new core workloads, and just the size of compute and storage associated with that data set is larger than it has been. And it's like, that, one of that is the proliferation of data and data capture is greater.
I think the kind of complexity of models and the amount of economic value that they can drive by having higher amounts of compute applied against that, whether that's machine learning or like, you know, early-stage AI, so to speak, et cetera, I think that's greater. And so when you look at that, then you're like: Okay, that's very different than in the past when we might have been skeptical about some of the larger footprints. We were like: Well, you know, half of that workload should go to the cloud anyway. Like, if it should be in the cloud, like, the enterprise has already made that decision. I think if there were,
If the enterprise is willing to buy equipment, expend capital on their precious balance sheet, especially in the kind of like a raising interest rate environment, and actually deploy that, that server networking storage, et cetera, that is a core workload. And so that's a change in kind of the character on how we think about that and our relationship with the enterprise. And like, "Well, we want to help you support that." Like, this is no longer some fringe thing that may end up going away. This is durable revenue. It's an important workload for the customer. Like, we think that both kind of the longevity of that and the importanc of that for their enterprise architecture is more important, so we wanna make sure we're leaning in to help support that.
As part of what you're saying, it almost suggests that... Do you think we're at the end of that, let's call it, that on-premise to colo/cloud migration? Are we, are we towards the ending stages of that, and now it's new workload creation in a colocation environment that's driving the incremental growth?
I think so. It's, you know, I think the pundits have been kind of predicting the end of the enterprise, like data center for too long to count. But I'd say, like, we certainly started seeing that motion. Like, COVID was one piece of it, like, people needed workloads deployed. They definitely weren't gonna build out new stuff in their office prem. I think the second part that was really powerful around this has been everyone has sustainability targets. Building out a sub-scale data center is probably the worst thing that you can do
from a sustainability perspective, and, like, offloading that to an Equinix, where we have, like, science-based targets. We can go ahead and provide that. Attestable on green power utilization, like, you can carry that through. That's been a really powerful lever. And then the third element that we're seeing in particular is just the density requirements of, like, compute in general has gone up, but particularly around any of these AI use cases. I mean, our partnership with NVIDIA was originally founded. One, they're obviously an important technology, and we, we're really interested in understanding it. But they came to us because they were selling their boxes, and then they were getting to the point when they were getting ready, like: "Well, where are we shipping it to?" And then they're like: "Whoa, whoa, whoa. 40 kVA a cabinet, like, where...
I can't put that in my enterprise premise.
Yeah.
Like, we need to find a place to land these things on a consistent, reliable basis." And so I think that, like, last piece, in particular, as you start seeing some of that, you start seeing the complexity about, like, liquid cooling delivery. I mean, an enterprise trying to deploy liquid cooling into their data center, like, that's never gonna make economic or sustainability sense. And so I think all of that will, will end up skewing us towards, like, maybe finally, at last, the, the death of the on-prem data center for, like, most of the use cases.
Okay, that is very helpful context. Thank you for that. Again, further building on it, Equinix has historically focused on signing the right deal with the right customer, which from my vantage point, has been focusing on customers that have a high propensity to consume interconnection.
Mm-hmm.
Right? So as we get these larger footprint deals, what are your observations related to the propensity to consume interconnection?
Yeah, I'd say on a per deployment basis, again, what we saw historically on some of those larger footprint deals was that they were relatively low on an interconnection basis compared to, like
Mm-hmm.
these core network nodes. What we're seeing now, increasingly, is, like, those are converging. Those, like, larger footprints that we're looking at are still... Again, when you look at it across, on a per deployment basis, the amount of interconnection, both on cross-connects and VCs, is right in line with the rest of the deployments. Now, on a per cabinet basis, like, the interconnection will be a little bit lower 'cause they're bigger. But, I think what it points to, though, is the strategic value and importance of kind of the data flows is still very strong. And kind of the importance, then, to the overall, like, ecosystem of interconnection and kind of, like, that supply-demand balance that we're trying to help between our suppliers on the cloud and network side and the enterprise is still very healthy, right?
And so, like, that, that all feels like: "Okay, well, that's, that's business that we want.
So almost different from the large footprint deals that you would have seen in the past, where there would have been, let's say, less interconnection associated with
That's right.
these new core workloads, they are consuming interconnection, let's say, at a rate similar on a per deployment basis with what you would have seen in the past from Equinix
From a standard perspective.
Yeah. Got it. Okay. You know, I had this question also earlier today, which is that to the extent that we are seeing dea sizes go up for the reasons that you mentioned, let's say, you know, just higher densities and and so on, you know, does it make sense for Equinix to just start building larger data centers? Like, how does it impact the way you think? And when I say size, I mean on a megawatt basis. Right? Just to be clear.
Yeah.
Like, how does it impact the way that you think about build size on a megawatt basis?
Yeah, I mean, the answer is yes. Like, and we have been already, over the course of the last, like, 2 or 3 years, been, you know, our overall construction size for our data centers has increased over time. And it's also starting to increase kind of the size that we're thinking about in terms of the phases, especially in kind of the highest velocity markets for us, where we're even thinking about, like: "Should we be phasing some of this construction at all?" Especially kind of given what we're seeing in terms of the demand characteristics and competitive supply in these markets. In... You know, and New York's a great example of this, and certainly our last development in Ashburn.
Like, historically, when we were doing data center development, we wouldn't do any amount of, like, pre-sale/pre-leasing into the facility. It's like: "Hey, we're spec building." When we, like, maybe three months before launch, when we had, like, a really, really firm ready-for-service date, we would go out and say like: "Okay, sales, like, go start selling into that now." I think there's been so much latent kind of demand, that we've actually started pre-selling into our facilities, like, six, nine months early even in some cases. And so that means, you know, again, that may not translate into monetization, like, day one, but certainly within the first quarter, like, that acceleration is greater than we've seen in the past.
And so the additional part is that gives us both kind of, like, comfort and security to say, like: "Okay, then we should build big- bigger," 'cause we know that we can actually take some of that earlier on the front side. And so it's... But it's hard, right? I think when you think about data center construction, as and I see many folks in the room, like, this is, like, it's not like you're doing a change order on your kitchen, and, like, you can go ahead and, like, get a new sink. Like, you know, you're already a year into a data center project. Your MC, your equipment ordering is, like, a 2-year lead. You can't just magically accelerate the development that you're already doing and, like, double or triple the size
Yeah.
no matter how many times I ask my construction team to do that. Like, in some cases, we actually have reallocated... And again, part of our flexibility of the platform and the scale that we're doing is we have standard equipment that we're doing across multiple sites. I can move gensets from one location to another on some of these construction projects, and, like, when I see higher velocity, I can do some of that, but, you know, not to the degree that we would love. And then you always end up with, you know, we got to do re-permitting, you got to do all that work with just the bricks and sticks construction elements that can slow you down.
Mm-hmm.
But I would just say in general, like, you are seeing us build, like, you know, much bigger and much... and also designing for flexibility so we can support much higher densities. That doesn't mean for us, like, "Hey, we're going to build a data center that is designed with an average of 100 kW a cabinet, and it's, like, a tiny square footage, and it's all liquid-cooled," because, you know, we're building for assets that need to be here for the next 10, 15, 20, 30 years. We want to make sure that it's as flexible as possible, and so, like, we want to be able to support workloads that can go up to the, you know, the 100, 200 kW a cabinet.
Mm.
We don't think that that's going to be the majority of the workload
Mm
or even if it's the majority in the near term, that may not be for forever. And so we're seeing that both, obviously from the enterprise side, but even on our hyperscaler deals that we're looking at, when they're asking us to be able to support that, they're not asking for a smaller shell. They know that this tech refresh, you know, five years down the road on the next thing that comes, it may not be at 100 kW a cabinet, and they want that flexibility to say, "I can still put it," you know, general kind of compute workload into that and not end up being hamstrung by having too tiny a shell.
Got it. So with that, like, I had, like, five additional questions pop up in my head. So the first is related to, you know, you're talking about essentially pre-leasing, and that jives with what we've heard around the conference. It's like there, we are seeing a greater degree of enterprise of enterprise pre-leasing. You know, I guess my question then becomes, like, that would naturally elongate your book-to-bill, right? And I think we're what investors are used to is that short, you guys say, "Hey, we had strong bookings
Mm.
and then you kind of see it show up relatively quickly. When did that dynamic with the enterprise start to change? Was that in response to kind of the boom in hyperscale demand and them going out and leasing 2-3 years in advance, so it's a relatively recent shift?
Yeah.
Like, when did that change?
Yeah, no, I, I would say that's definitely been a relatively recent shift, like, in the last 2 or 3 quarters.
Okay.
Yeah, it's, and again, yeah, I guess that's the basis of it.
Okay.
I think, you know, I was just thinking about what... As we report and think about bookings, though, we're still kind of timing that based off of the ready-for-sale date.
Mm-hmm.
So it's even though we've taken it off the street, we're not kind of putting that embedded into our bookings numbers for that particular quarter, just like as we think about the treatment around that, so.
Okay.
But-
Got it. One of the things you and I, we've discussed in the past, we're just talking about general rack densities and then how
Mm
you guys think through building. My understanding is, like, the historical design for Equinix data centers, you could accommodate double-digit densities in, like, a portion of the data center, right?
Yeah.
Not uniformly throughout the facility. I think 4-5 was the right number on average across the facility. As you think through the next generation of data centers, are you essentially looking to build data centers where you could uniformly support, bear with me, 50 kW+
Yeah.
per cabinet across an entire facility? Or are you still taking the approach that, "Hey, we will probably have some customers that need a flex up from a density perspective, and we'll have a sleeve of this data center where they can go and run higher densities?
Yeah. What we're trying to do in the flexibility is, like, not create hard demarks in terms of like, "Hey, this data hall is going to be this density versus this." We're planning around, call it, like, somewhere around 8-10 kVA per cabinet as kind of the across the average of the estate. If customers were consuming at that, we would basically, like, perfectly match the amount of space that we have as white space, along with the amount of electrical, et cetera.
Mm.
Now, that doesn't mean that it needs to. At the end of the day, a lot of our customers are, you know, the fundamental unit of capacity that we're selling at is kind of the power limitation around that.
Yeah.
So if they ended up chewing up, let's call it 20 kW a cabinet across that entire state, we have full modernization of the assets, but we'll end up with stranded white space that, you know, gives us opportunity then to say, "Can we add another augment block? Are there future technologies or efficiencies that we can do? Can we upscale?" And again, that's a lot of the work that we're doing even in our existing sites of, like, "Hey, it used to be a 4 kW site, now can we add another power block in? How can we supplement the mechanical to be able to generate more, more revenue off of those, existing facilities and, and keep, like, keep their, their longevity?" But, that's how we're thinking about it again. And we were looking at some of our latest liquid cooling designs.
Like, I don't know, the piping is good enough that you could have 5 megawatts in a cabinet if you needed that. We don't think that that's likely, but
Yeah.
you know, you know, the math works. Like-
The math checks out.
So yeah, the math checks out. And so we're like, that's kind of how we're thinking about these design characteristics. Again, like, ultimately, we think it'll be, on an air-cooled basis, 8-10 as an average. On the liquid-cooled, I mean, who knows? But, you know, it once you start going, like, the directly with the chip, like, you know, you can support anywhere, like, north of 100. Like, the numbers just get silly pretty fast.
Mm-hmm.
And you're like, I think the real constraint's going to be more around the enterprise customer or the buyer of that equipment. Like, the failure domain for if something goes wrong in that rack, like, and just for their own gear, like
Right.
what's the blast radius
Yeah.
and the expense associated with that, right? And it's like, if there's a thermal event with a runaway and you're running gear that hot, and, like, you start melting down silicon or something like that, that's obviously going to be pretty problematic, so.
That's that, no, completely fair point. Okay, at the beginning, you talked about utility and AI. Okay, so, on the last quarter, I believe you announced that, you bought 200 acres, I want to say, in Atlanta. Am I remembering that correctly?
Yeah. Multi, multi-hundred megawatt campus.
Oh, okay. Multi-hundred megawatt campus, okay. Which to me, is a step function increase in the size of, like, the parcels that you've been looking at for xScale. I'm getting the sense that you're more actively leaning into this. I'm gonna put my tinfoil hat on for a second, all right?
Mm-hmm.
So just bear with me. Is it possible, or is maybe one way to think about that is the reason for doing more xScale in the U.S. is, yes, you can go after and capture part of the market for hyperscale, but also, you do more of this in your major markets, it actually, on the back end, provides Equinix further runway for power, and you could potentially become a tenant of xScale for your retail, for your retail footprint. Essentially, xScale gets the power, and then over time, as you need it, Equinix itself can draw down that power. Is that something that is... Am I crazy?
No, yeah, I mean, you're not crazy. I mean, and, like, you know, Equinix Inc., on the retail side, we're a tenant in a couple of existing, like, of our xScale properties internationally already, so... And we've designed the structure of the JVs and awareness of that so that we have that option. If, again, it's market rates, essentially, so it's not like, you know, our capital partners obviously don't want us to, like, you know, get any incremental upside benefit of that, but our credit's pretty good. We're gonna be here for a long time. Like, our JV partners are happy to take us as a lease inside of that. The negotiations internally of the team between our real estate team and our real estate team can be tricky, though.
It actually does make for some weird contentious discussions where you're like: "What are we doing right now?
It is possible.
But it is possible. Look, I think that there's a lot of different option benefit there on the xScale side. Look, we're leaning in. I think, one, we feel like we've built a really, really powerful mousetrap, just from a financial structuring and, like, ability to solve for what our customers are asking us to build perspective, where we can create outsized returns for us. Like, the 10-to-1 gearing on capital for us to be able to create this capacity is really powerful. I think there's supply chain benefit of, like, having the scale of doing xScale inside of any of our markets-
Mm.
gives us incremental buying capacity on our, our pre-inventorying, our pre-purchasing of equipment, et cetera, that, otherwise, just on the retail basis, we wouldn't have as much access to, because the velocity of development is much higher. And, and like you said, on the utility side, I, I think it does give us some, some option value there, and I think, you know, we're, we're all gonna... Like, we'll be put to the test when we end up saying, like, "Hey, there's X amount of 100 megawatts of capacity." Like, how much is Equinix gonna take if this hyperscaler is willing to take all of it? And I think it'll just come down to, you know, who's gonna give- who's gonna help get to that power co- consumption faster, right?
Yeah.
I think at the end of the day, capital partners just wanna make sure we're monetizing that as quickly as possible, and again, like, what we're seeing in many of these, like, you know, 100-megawatt-plus deals with the hyperscalers is they're still putting a ramp in, right? And so, you know, our ability to take multiples of those kind of customers into that campus, inclusive of ourselves, creates a better monetization event for, like, that total development.
Got it. That's very helpful. Okay, let's talk a little bit about AI. Seeing very strong demand on the hyperscale side, on the enterprise side, particularly related to training. You know, what are you seeing there? Is it a meaningful driver of demand where enterprise is trying to essentially build their own training, train their own model? Are you seeing that within your facilities, or is that still not a meaningful driver?
I think we're seeing an uptick on that. Is that, like, a massive boost to the overall demand? I think the answer to that, right again, would be no. I think, you know, it's in line, it's durable, it's important. I think, though, like, you know, again, if you ask any CIO: Do they have a mandate to do AI stuff? Yes. Were they given any incremental funding to do that? Like, no, right? And so there's a little bit of balloon squeezing, of like, "Hey, this is the nature of the demand that the enterprise has now." They've got dollars to spend on general purpose compute. They got dollars to spend across their entire estate. If they need to do AI, it's coming from over here.
And so we wanna make sure, though, that, you know, we're able to capture that and, and also provide a compelling, like-
Mm
... reason for that to need to live at Equinix, which we think we have.
You know, get it, not a meaningful driver, but it seems like you are seeing just some of
Yeah.
some of it.
Absolutely.
What is, like, what is the form factor that that takes on within the data center? Like, you know, any, any color on rack densities that you're seeing,
Yeah, I mean, it's like some of that estate's gonna look... We haven't seen crazy rack densities. I'd say, like, call it in the, like, the high end of that is the 15 kW-30 kW range, where we're again still generally delivering that via air cooling. Sometimes we're doing liquid augments there. And then across the rest of the estate, it's generally in the, like, the 10 kW-15 kW. Like, you know, that provides kind of enough, both from a data storage plus some amount of training capacity.
Yeah. Helpful. On the inference side, anything of note there, or are we still too early? We're still in the middle of the training glut, and we're not really seeing a meaningful, meaningful inference demand currently.
I think the folks that are, again, like, the verticals that I mentioned earlier, like, it's a combination of training and inference, right? Inference is the usage of that training to drive the business result, and for those, like, you know, financial services, healthcare
Mm.
like, life sciences, they are absolutely driving business, like, economic benefit, and that's why they're spending the hard dollars. So that inference is happening. I would say, in general, what we're seeing is most of that inference is converged with the training platform, and in some cases, using that same technology stack. What we're seeing, though, is, like, a lot of conversations in market and with the technology suppliers in chain around, hey, like, the nature of inference from, like, just the compute and algorithm perspective is very different than training. And, like, we've said this for a while, is like, hey, inference can run on CPUs. Inference runs on your mobile phone. Inference runs on
Mm.
a lot of different technology stacks quite efficiently. And so we don't think, like, inference in the long run necessarily will manifest as, hey, 120 kW liquid-cooled GPU deployments.
Mm.
It probably looks like the same kind of edge compute deployments that we've been seeing across our estate today, and in similar form factors.
Could it look similar to cloud compute nodes? Like, you know, where essentially that hosts the cloud. Could it take on a similar form factors to that, do you think?
I think that's right, like or like private cloud deployments for the enterprise, and like, hey, some of the, again, some of those edge, kind of like hub sites. I think that are what it would manifest into. The latency characteristics are also really interesting, where, again, training is basically latency insensitive. You can kind of have that in not anywhere, but pretty close to. And I think it's more about, like, fiber capacity into those locations. And so, for fiber folks in the room, like, fiber's hot again. It's awesome. I love it.
Yeah.
I love to see that. Congrats, Lumen. It's, we've seen the path. It's. And we're seeing similar trends in Western Europe, right? Where it's like, hey, you know, fiber development now is being done on an infrastructure basis, not like a spec-build fiber to the house. It's like, no, this hyperscaler is gonna pay for, like, this fiber build-out, and you can underwrite that. I think that that's really, really interesting. On the inference side, again, we're seeing it looks like web-type behavior, right?
Mm.
And I think, you know, maybe a year ago, people were thinking about inference and, like, AI as, like, hey, it's gonna be the autonomous car, and you need, like, five, sub-five millisecond access for inference to make any sense. We're, we're not really seeing that as, as kind of the meaningful trend, and so I think that, that notion, I think, is gonna be similar to, a lot of maybe the, the hype around, like, 5G, far edge, yada, yada.
Okay.
It's like maybe it manifests, but it'll be a very niche application for that.
... Okay, I wanna put this to you. Bear with me on this.
Yeah.
As I think of cloud architecture, right, I think of you have your compute nodes and availability zone, right? Which are all connected back to your network node, which would, essentially, your cloud on-ramp, right?
Mm-hmm.
The way that Equinix benefited in the cloud was not through hosting the compute nodes within the facility, but rather through the networking nodes, and the ecosystem that that created, right? It was essentially the lighthouse that attracted the enterprise workloads, and that's how you benefited from, through the growth of the cloud. As I think of AI and AI inference, I think of, kind of, the way I see it is, it's the same thing except with training as an additional leg.
Mm-hmm.
You do the training. Training gives you some foundational model. That foundational model is exported via inference, and it lives locally within a major metro, right? And then from there, that is connected to a network node, likely to be in an interconnected facility like Equinix's. And as I think of the growth of AI and the way that Equinix benefits, and I actually see it similar to the cloud, where it's owning the network node within your facility that creates additional ecosystem and additional new core workloads on the part of the enterprises. So the way that you benefit in AI is actually the same way that you benefit in the cloud. Where would you push back on, along what I just said?
I like that theory.
Okay.
Look, I think many of these trends, one, as we're all saying, right, it's early days around this
Mm-hmm.
and like, the early signs, though, are, hey, if you're an AI service provider or a GPU as a service provider, you still have the characteristics of any other cloud service provider, which is, you really care about, like, high performance, like, low latency, and relatively low cost and efficient, like, network transport.
Mm-hmm.
And so we're, again, the deployments that we've won from, like, CoreWeave and Lambda, et cetera. It's like we're getting their network nodes, we're getting those cloud on-ramps.
Mm-hmm.
They're also thinking about enterprise traffic and those inflows and outflows in the same way. I think, you know, the early first waves of cloud were really driven around, like, internet as the primary transport, and I think what we're seeing is, on AI in particular with the enterprise, like, crown jewel data that is going to be fed into the like any kind of ingest engine, et cetera, is more sensitive than maybe some of that discrete set of workloads. And the interconnectedness of those data sets is way higher than, like, hey, like, just general cloud applications, you can have, like, some marketing app, and it can be sitting in AWS, and it's almost just like a standalone thing there, and it's spitting back some amount of data.
I think when you start talking about AI and how to unlock the real value of that, the web of interconnectedness on those data flows is substantially higher. And I think, you know, that sensitivity is something that we're watching. So I do think, like, yeah, that network and interconnection is gonna continue to be incredibly important. Whether or not that's gonna be, like, incremental volume, not sure, because, again, a lot of those are gonna be the same players that we know and love, the hyperscalers at large
Mm-hmm, mm-hmm.
the SaaS players that are of massive scale, like Salesforce, ServiceNow, like, they're gonna be big AI monetizers around that. I do think, though, that, you know, the... Our ability to think more about some of these core, like, data deployments and kind of like, hey, what that infrastructure and architecture means for us, again, you're continuing to see more and more talk from us about like, "Hey, we want..." Like, all of that estate is incredibly important to make sure we're supporting.
Okay. One of the things we've been seeing is, and I mean, this has come up on the earnings call, is the differential in density between what's getting churned and what's coming in new. My question for you is gonna be: What is driving that delta? Is it that you have a small number of, let's say, GPU-based deployments that are coming in, that are skewing the average higher?
Mm-hmm.
Or is it that, as you go through the refresh cycle on CPUs, the thermal design power of the new
Yeah.
chips is higher, so as a result, you're getting, on the refresh, a increase in rack
Yeah.
in rack density?
Yeah. As with everything, the answer is both.
Mm-hmm.
But its
Uh
I would say what we have seen, though, is, like, the general tech refresh rate explains the, hey, you know, call it a 5%-10% increase on an
Mm-hmm.
on an annual basis there. We are seeing, like, more outliers that are, like, very high density or, like, double what our embedded base density is than in the past. And I was just telling the other team, like, in the first half of this year, we have sold more megawatts in, like, call it the 8+ kVA per cabinet range than we did in all of last year. And so, like, that
Mm.
as a percentage, is definitely going up, which is, again, that number had been in the, call it, high single digits and steady around that for a very long period, and now it's, like, in the low double digits. And that's like, okay, that is a trend that, again, across the estate, like, it's, it's a very, very big embedded base and pool for us to, like, watch and manage and look at. But it is, that is an interesting trend that we're gonna keep an eye on. And again, that's why it's probably further exacerating some of the density dynamics that we're talking about.
Yeah.
'Cause the natural tech refresh probably would've driven us to, like, call it, like, 4.8, 5.5, or 5.25, or something like that. And you're like, "Okay," when you blend in the rest of that, it starts pulling it up. Because it, it's not linear anymore, right? It's not like, hey, that AI or those dense workloads are like, you know
Mm.
11 instead of, like, 10 KVA. It's like, no, it's 30. Like, it throws your math off a lot when you're dealing with, like, 4, so.
That makes sense to me.
Yeah.
All right, well, Jon, it's always a pleasure spending time with you.
Well
Thank you very much for coming in.
My pleasure.
Really appreciate it.
All right. Thanks, bud. Thank you.