Equinix, Inc. (EQIX) Earnings Call Transcript & Summary

All right. Well, I guess it's afternoon now. So good afternoon, everyone, and welcome to our 9th Annual Communications Infrastructure Summit. For this meeting, we have Equinix. And from Equinix, we have Jim Poole, their VP of Business Development; and Chip Newcom, who's their Director of IR. This session will be structured as a fireside chat, which will run for 40 minutes. I have a list of questions prepared, but to the extent that there are any questions from the audience, I'll do my best to take questions from you guys towards the end. So with that, Jim, Chip, thank you very much for being here. Really appreciate it.

Nice to be here. Thank you.

Awesome.

So let's start off. As we start things off, Jim, for those who may be a little less familiar with you, can you give us an overview of your role at Equinix? And as part of that, how your role at the firm has evolved over time.

Okay. Sure. So as stated, I'm a VP in the business development organization, which, at Equinix, is really around ecosystem extension since that is core to the business model. I joined the company 14 years ago, originally worked in the marketing organization where I ran service provider marketing. And today, I run a program inside of business development we call Evolving Edge, which is essentially what is the role that Equinix plays in things like 5G, subsea satellite and AI.

Okay. All right. And as part of that, I would ask you, where do you spend most of your time focused these days? What takes the lion's share of your time? Is it one thing? Many things?

It's probably two things. I hear this AI things become super interesting to people.

Yes.

I joke that I started working on that 4.5 years ago, and no one cared. And now, it's amazing. And then the other thing would be 5G. So I spent a lot of time on warehouse.

Okay. All right. So let's transition and talk a little bit about the demand environment that you guys are seeing. So 2022 was a record leasing year for the broader industry. But this year seems to be marked by some tougher macro conditions. We've been hearing more hesitation on the enterprise side in terms of their willingness to lease. We're going to put AI aside for the moment. Can you frame up the enterprise demand for digital transformation that Equinix has seen and perhaps how it compares to what you were seeing this time a year ago?

Sure, sure. So on a comparative basis, without a doubt, obviously, we've got inflation in a place that it wasn't 1.5 years ago. And that certainly, I think, has in the minds of many customers being super careful and cautious about where they spend their money. Now that being said, the Equinix value prop to our average customer is -- I'd like to say, is an ROI argument. It's really around how do I get the best yield for infrastructure that I have to deploy for some reason. And I know Charles used the statistic in the earnings call just recently, where we talked about the fact that with, say, the Fortune 500, we're looking at, what, 30% digital services revenue generation in terms of contribution to those companies in 2022, moving to 40% by 2026. And so the reality is digital transformation for our customer base is not so much a, it would be nice if I could. It's more of a reason for being and something that they have to invest in. So while I think we're seeing people try to be cautious, I don't see any slowdown in the demand for what we provide as part of that equation.

If I take that and dig a little bit deeper, what it sounds like is, even if there was a pullback, let's say, because of macro conditions, at the end of the day, it just leads to pent-up demand. And at the end of the day, they still need to move forward with digital transformation. Is that the right way to interpret that?

Well, yes. I mean, I don't see any customer going -- maybe I'll try digital transformation. And if it doesn't work, I'll go back to my standard brick-and-mortar. No, that does not happen, right? Either they're going to do it or they're not. The question might be the pace at which they do it. But when you take that against, say, us as the example of 10,000 customers, in any given moment, what portion of that customer base is in what portion of that journey? And if you think about like where our sales activity comes from, the vast majority of our sales is selling to customers we already have, right? So they're already in the system. They're already seeing the benefit, and so it's a land-and-expand sort of equation. So we have a lot of ways of sort of looking at these moments in time where the macroeconomic environment might look challenging from the outside, but it doesn't necessarily have a huge impact on us because either you're digital or you're not. You're not sort of halfway in between.

And the thing as well too that I'd add is, as you think about all the various different technology trends out there, all of them going back to Equinix when it comes down to it. And so, I mean, part of the durability that we've seen in our business for now, 25 years is the fact that because of this broader trend towards digitization, whether it's thinking about 5G or AI or hybrid multi-cloud, all of that requires doing more stuff with Equinix. Because we end up being the hubbing point for all of your networking traffic or where your private cloud deployment is going to sit because you want it to be proximate to the 40% market share that we've got for the on-ramps to the cloud service providers in the market where we operate. So I think that's part of what we benefit from is that if the thesis is that the world is not going to go more digital than admittedly. That's a challenging market environment for us. But even across macroeconomic cycles, enterprises right now are still trying to figure out how do they get more efficient, how do they get more digital, which means how are they going to work more with Equinix over time.

That makes sense to me. Sticking with the topic of digital transformation. This is a question that I asked earlier in the enterprise session. We've been talking about digital transformation and enterprise outsourcing for a while now. And I made the joke that, for a while, we've been talking about 70%. And it seems like when I go to these conferences, I hear the same thing over and over again. Jim, from your perspective, as you think about what inning we're in, to use a baseball analogy, where do you think we are from a digital transformation perspective? As part of that, maybe you could tell us what you think where we stand in the U.S. versus maybe other regions around the world?

Sure. Sure. So probably without being able to call the exact inning, I would say, we're -- we still have a lot of game left. Certainly, if you had to stack rank progression, obviously, North America is the most advanced, then followed by Europe and followed by APAC. As far as the -- how much more do we have to go, we use a really simple way to think about it, which is if you looked at our penetration into the Fortune 5, we're at 58%. If you looked at the Global 2000, we're at 41%, right? And that is the exact profile of customers that benefits the most from being an Equinix customer. So at the end of the day, there is a lot of market yet to tap.

It's a long runway of future demand for you, guys. Along those lines, when we've spoken with Keith and Charles in the past, they've highlighted the focus on signing the right deals at the right price points. How is Equinix is thinking about what the right deal is, and as part of that, also what the right price point is, over the past few years?

Sure. And so this goes back to my earlier point of the discussion that we typically have with a customer is a conversation around ROI for the deployment that they put with us, right? If you come to us and you say, I want to be in a single site, and the guy down the street is $10 per kVA cheaper, you should probably talk to him. You shouldn't be talking to us, right? And I use this with -- I gave an example in an earlier meeting today with network service provider customers that we have. If I said, if you stack ranked every deployment you have of infrastructure in your network by where you get the best return, who is the top 10? Equinix is, right? So the question we're having is, how am I helping you do your business better? So when we say get the right deal, that's exactly what we mean is because ecosystem effects are additive, right? It's not like a Venn diagram where there's this thing over here and this other thing over here, and they talk to each other. It's more like this one and then this one and then this one. And so they should keep adding, so it's incumbent upon us to maintain the value, right? If we keep customers who are super interested in the idea that says, hey, I've got a supply chain that involves 10 different companies, and the cheapest way I can interconnect to those 10 different companies is to do a deployment at Equinix, that is the perfect customer. They get the most value out of that because we make that part of what they're trying to do as simple as possible.

I want to stick on that because one of the things also I think about as you think of ecosystem effect is, one of the things I imagine that's important is propensity to consume interconnection, right?

Yes.

Connecting to other folks within the ecosystem. When you're having conversations with, let's say, a new logo, how do you frame up what the potential opportunity is on the interconnection side? And as part of that, how a customer could be actually additive to the ecosystem? Just curious how you think about that.

So I guess you could do it in two ways. One would be sort of on the consumer versus the supplier side, right? So on the consumer side, start with the enterprise side. You're looking at generally a customer where you're saying, okay, you're in some phase of your digital transformation, who are your biggest suppliers, right? And so that's where we always get back to, as much as it's an over abuse term of -- I always say it's not just hybrid multi-cloud, it's hybrid, multi-cloud, multi-network. Why? Because all of those customers have 3 characteristics in common. They all use multiple networks. They all use multiple cloud providers, and there is something that they do that they do not want in the public domain, but has a propensity to interconnect to the public domain, right? So if you look at our enterprise customer, the most common use case is private storage, public compute. They're fine using the hyperscalers for the compute cycles, but they don't ever want the data to say resident in that location. So right away, you can qualify out a whole lot of potentially bad deals by looking and finding that customer who's like, yes, that is exactly the profile I fit. On the supplier side, it's almost the inverse. And so I'll go back to my network analogy, is that if you're a network operator or you're a CSP, a cloud provider, deploying at Equinix, you're fishing in a barrel. The highest propensity consumers are the ones that sit in our buildings. And I've seen it on both sides. If you were to ask a network provider, they've got the most customers possible from doing a deployment. I could do one deployment in an Equinix campus, and I can go after hundreds of customers. My joke is a data center without a network is a refrigerator, right? So at the end of the day, for them, that's a big deal. Same thing on the cloud side. If the cloud providers told you how they would characterize an Equinix customer versus an average customer, an Equinix customer is a way more sophisticated, pushes way more traffic, right? And so in the cloud space, how do they make their money? They make their money on traffic, right? What goes in and out of the cloud. And so therefore, we are the perfect sort of combination of characteristics for whether you're an enterprise looking to talk to a service provider or vice versa.

Well, I want to take that point and build on it and tie it to AI. So when we think about what they're putting in your facilities, it's the stuff they don't want to relinquish control over. To your point, they can -- they'll use the public compute, but they don't want that sitting in the public cloud environment. When I think about AI, I'd imagine, there is sensitive data that they want to be able to do something with, but they don't want to relinquish control. So as you've spent time over the last few years thinking about AI and how this works. How do you see the enterprise deployment architecture or AI different from what -- if at all, from what we've seen with the cloud, essentially hybrid multi-cloud?

Yes. Well, the reality is, at least at this point in time in the market, it doesn't actually look any different. Why? Because the vast majority of GPU capacity sits in the hyperscalers, right? So if I just looked at the AI customers that we're talking to today, what are they doing? They're basically saying, I have private data that I do not want because that's the secret sauce of my company, sitting resident inside of one of these AI services. However, I'm fine interconnecting to it. And so they're asking us, they're saying, "Hey, where are all the Azure gateways, because I want access to open AI?" Right? It's a very standard thing. And then on the flip side of that, where we're starting to see some new comers come in, some of these upstarts who are starting to deploy GPUs as service, they may decide to deploy somewhere in the middle of nowhere where there's cheap power. And then the next day, they show up at us and they're like, I need to note in Chicago, Dallas, Silicon Valley and D.C. because that's where all the customers are. And the training function within AI is not a horribly latency sensitive thing, right? But then the next phase of my deployment, after I get my model trained and I do everything I want is, okay, where do I put things like the inferencing functions, which is a far more distributed function. And they look at Equinix and go, "Oh, you're within 10 milliseconds of 80% of the GDP of the planet." Right? So where else am I going to go?

So as we consider that, we're talking about training being less location latency sensitive, right, relative to the inference side of the house. Would you expect Equinix to play in that side of the market to go after the training opportunity? Or should we think of it as strictly inference as what you'd pursue?

Well, our default is inference, without a doubt, because that is purpose made for what we do. And just to give some insight to that, the inference function within sort of AI more generally, is not a GPU-specific function, right? Most inferencing functions run on CPUs just fine, right? And so a lot of that has to do with the fact that training is a highly parallel process, and so that's what GPUs are very good at. Inferencing is a serial process. You ask a question, and it gives you an answer, right? And so that lends itself more to distribution. However, per the last earnings call the other day, Charles did acknowledge the fact that we are now thinking about xScale, which is essentially our exposure to the hyperscale market, which traditionally for us has been a non-U.S. activity. And we are looking at that in the U.S. at this point. But we were looking at that in the U.S. already, [ Sons ], the whole AI demand. So does AI have a potential influence on that. Yes, potentially, and we're -- that's part of the calculus as far as us thinking about it. But I don't know if you have any other...

Well, the thing I'd add to is, as you think about AI, it's not going to be homogenous. So there's going to be different types of training that's required. There's going to be different types of inference that's required. So as you think about like what we've been doing in AI for years already, what you now call effectively machine learning sits within our facilities, and you can do training where you've got structured data coming in from a manufacturing for or something that is easier to train on relative to unstructured data like what you're going to have in a large language model and do that very easily and efficiently within an Equinix facility. So we're already supporting different forms of AI, not necessarily the high, high compute needs for something like a large language model. But for portions of AI, that's already sitting in our facilities. So it's not necessarily a new function. It's just because of open AI and because of what has happened sort of without exploding on the scene, it's put a spotlight on one very specific portion of AI models, but we're supporting a whole host of other different types of AI applications already.

Yes. It's important to note that -- and some people don't necessarily appreciate this, but solving the language problem from an AI perspective was the hardest problem they could think to go after, right? Because the variability in language is almost infinite, right? On a comparative basis, and I use this as an example, genomics has less moving parts, right? So you can do genomics jobs on a fraction of the GPU capacity that you would do something like a large language model, because they went after something that was a really hard problem to solve. And so it's drawn a lot of attention. But to Chip's point, it's not the only thing happening in the market.

That's a fair point. And then as I think you brought up xScale when I asked about the training side of it. Where I thought of xScale is I don't necessarily think of it as like commoditized hyperscale. The way I think of it is whatever workload is going in xScale needs to sit proximate to the -- essentially to the retail campus, because there is some interconnection that's going to be happening with the existing ecosystem. From what you're saying, it sounds like there is going to be a subset of training workloads that will need to leverage the existing ecosystem on your campus. And as such, it makes sense to do xScale in the U.S. as you look to service that type of demand. Am I interpreting that correctly?

If you looked at our xScale footprint internationally today, that is exactly what it conforms to. We do not build xScale facilities in the middle of nowhere in France or Germany or any place else. They always sit in markets where we're already deployed. So that is the model by which we are not chasing where is the absolute cheapest place we can deploy so that we can sell at the absolute lowest dollar. That is not the model for the xScale business from day 1.

Okay. And then just talking about the demand related to AI. You've made the point that you've been servicing these types of workloads. I think now it has gotten a lot more attention. As we think of kind of the evolution of demand related to this specific type of workload over the course of the last 6 months, when it really became popularized with ChatGPT, have you seen an inflection higher in terms of the demand from your customer base associated with this essentially doing leveraging AI? I think Charles said on the call, like they want -- enterprises want to integrate AI. Curious what you're seeing in that.

Yes. So that example I used earlier, where we are seeing customers who want to access foundation models that were built in something like a hyperscaler, like open AI is built inside of Azure, right? And so customers are specifically looking to say, "Hey, I've got data." And oddly enough, the customers that are asking, these are existing customers, right? So it's already that their data has been sitting in an Equinix facility. And now all of a sudden, they're like, "Yes, I think I might want to try this out. So how do I do that?" Nothing faster than a cross-connect or a fabric port to connect you to that hyperscale infrastructure as quickly as you want to. To the extent that we are seeing some new-ish things, like I said, those upstart companies coming out looking for comms nodes, we're starting to see that. That's activity that you didn't see, say, 6 months ago? And then also cases where you do have certain types of inference jobs that are heavy and need to be proximate. So the example I've been using for people is rendering. Rendering is a very video-intensive. So even the inferencing function tends to use a lot of GPUs. And it's not the case that the guy in Hollywood sends the workload to Salt Lake and back again. That doesn't make a lot of sense. Instead, they come to somebody like us and they say, "Hey, where can I put this in the L.A. market, because just the sheer size of the amount of data that they're trying to process requires timeliness. So it's not so much a latency thing for latency sake. It's more of a size and heaviness of the workload. Data gravity is a real thing.

That's completely fair. It sounds like from what you're saying, you have seen an uptick in demand associated with these kinds of workloads. So I'm curious, in terms of the power density on the inference side specifically, propensity to consume interconnection and the cooling needs associated with it, how does it compare to the traditional enterprise hybrid multi-cloud deployments that you've been seeing?

And so on the inference side, in particular, it is almost completely identical to everything before. Like I said, there's that rendering case, which is a very unique sort of a case, and there's only a handful of customers that are going to need that kind of processing. For the vast majority of customers that are doing more structured data, machine learning type workloads, those actually run on CPUs. And so the fact that we -- that's what we do, right? If you need to put infrastructure proximate to end users for some sort of latency or performance reason, then you tend to come to Equinix because you need to deploy that in 20 metros around the world. AI doesn't look any difference from an inferencing perspective. In fact, actually, one of the benefits we believe we can bring to this, and this is a large part of -- if you listen to our Analyst Day when we had NVIDIA on stage...

Yes.

Was this idea that says Equinix facilities are all interconnected by a fabric, right, which is the network that connects everything together and connects into the hyperscalers, all on an SDN sort of API-driven basis. So this idea that says, "I'm going to develop this great training thing off somewhere in the middle of nowhere where it's super cheap to do." Sure, you can go do that. But the reality is I got to take that algorithm, and I have to deploy it proximate to all the places where somebody is going to consume that function. And I want to do that as easily as possible. Or am I going to go out and say, I need to go find 4 network providers and 3 cloud providers that are the same in each of these locations? Or I can just go to Equinix, and they can tell you exactly how that's going to happen?

What I find interesting about what you're saying is, I mean, you talked about inference essentially leveraging CPUs rather than GPUs. I mean, now you're saying that the workloads that you're seeing for AI inference are essentially very similar to what you saw in terms of hybrid multi-cloud deployments. From that perspective that suggests to me that we've been talking about higher power densities, but on the inference side, maybe we don't need to essentially reinvent the wheel in terms of how the interconnect-oriented data centers are architected in order to support it. What are your thoughts there?

Yes, I mean, we've been asked the question that says, "Hey, do you need to go back and retrofit a bunch of your facilities?" No. No, we don't. So the first thing to level set with is Equinix, as a retail colocation provider, exists in a demand market where demand is heterogeneous, meaning we don't know until we talk to the customer, what the customer wants to bring to us. So if you walked into a hyperscale facility, you would see rows upon rows upon rows of the exact same thing. If you walk into our facility, you see, you name it, right? Supercomputers on one side of the room. And so we have to be able to accommodate liquid cooling today. We have to be able to accommodate HPC today. The average may have sort of stayed in this kind of 4 to 6 kVA range for a very long time, and that's been the case. And certainly, I think on a go-forward basis, as we expect GPUs now to be starting to be used in more and more different types of applications and the -- like we talked to NVIDIA, Intel, AMD. And if you ask them about their HPC systems in the next, say, 3 to 5 years, they all go to liquid cooling period, right? They are right at the point now where air cooling, they're at the last generation that can be air cooled. So on average, on a go-forward basis, to the extent that HPC will be some portion of a number of workloads, it's not like all the regular IT workloads banish. That's still growing, and we're still benefiting from that. The portion that we require for the HPC, we have to be able to accommodate. And so we'll build that into -- as we go forward, new buildings will have better capacity to be able to handle things like liquid cooling. So for example, we have a co-innovation facility in Ashburn, Virginia, where I live. One of the things we've been working on is liquid to the rack. Why are we doing that? Well, we're doing liquid to the rack because liquid to the rack basically allows for us to handle customers who both do radiators on chips, right? So it's literally -- or radiators on the back of the rack, what they call a rear door heat exchanger. Either way, I have to be able to put a cooling distribution unit on the floor, and I'm going to be able to push out water to that particular rack. Those kinds of innovations will start showing up in new builds as we move forward.

That's interesting. So to take that a step further, it's like you will see power densities in your data centers increase. Like if you build right now the 4 to 5 kW per cab on average, maybe that goes to like 5 to 7 or something like that.

Right.

You'll be able to support the higher power density workloads in pockets of the data centers. But to your point, GPU compute is not taking over the entire data center. So there's still going to be that 5 kW workload that's going to go into the facility. And as part of supporting those higher power densities, you bring in liquid cooling into the facility in order to support it. But it's -- this doesn't seem like we're going to go to now you need to build 30 kW per cab data centers. That's what I'm taking away from this composition.

Yes. That is -- it's not like a binary flip where all of a sudden, the whole world just decides, oh, everything needs to be done on a GPU. GPUs, like I said, they're very good at doing something very specific, which is parallelism. That is not the case for the vast majority of IT workloads. They don't need that. So it will be part of the equation. It will be a bigger part of the equation than it was in the past. But it's not the only part of the equation.

Okay. Now as we think about bringing in liquid cooling into these facilities, I'm curious to know, in terms of the costs associated with the cooling architecture, liquid cooling versus traditional air cooling, right? Is there a material difference in the cost associated with deploying liquid? And if so, is there any kind of framework you can provide for thinking about it either on a per megawatt basis or on a per cab basis?

I mean, it is not different enough than what we're doing already for customers today to worth noting it doesn't show up as a -- oh, it's going to cost 20% more or nothing like that. Because like I said, we've been doing this all along. You have to think -- data centers already consume a lot of water, right? That's how the cooling function happens. So the fact that all of the water infrastructure is in the building, that was the biggest part of the cost to begin with, right? So to the extent like when a customer comes to us and says, "Oh, I want to do 20 racks and I need this very specific thing." Okay, there might be some NRR associated with some very specific infrastructure that they need for that customer. That may show up in the mix, but it's not like it's going to cost us huge amounts more money to accommodate that for our customer.

So the chillers are on site already. So really the incremental -- any incremental CapEx cost would be just bringing literally the liquid to the cabinet?

Yes, correct.

That's where it is. And then on the OpEx side, is there any material change between the two? I'd be curious.

Well, I mean, most of the beating and care of the infrastructure that sits inside of an equity facility is done by the customer, done by their contractors. What I can tell you, just as a perspective on the market, is that one of the things we have noted is that the industry notices that there is a skills gap in the maintenance of very high-end GPU-based infrastructure, right? Not every IT guy in an enterprise knows when InfiniBand is or knows any of the sort of resource management tools that you use for GPU clusters as opposed to what you would do for CPUs. So there is, I think, a lot of companies, why you're seeing some of these start-up -- GPU as a service company, this is for exactly that reason. It's a very sort of a unique set of hardware and aren't as good, but we don't materially participate in that end.

Okay. Last question on this topic. And this is something that I've been reflecting on as I think about Equinix. So I think the moat around Equinix's business goes back to the essentially pioneering the carrier neutral data center, right, and getting all the networks into the facility. That then had a derivative benefit when it came time for the cloud. If you had to pick one -- the cloud have to pick one place to deploy, a network pop where you'll put it where all the networks are, right? So I think that led, that translated into the market share that you have of cloud [indiscernible]. To the extent that AI represents the third cycle, first was internet, then was cloud, and now AI is the next one, how do you position Equinix to make sure that you have the market share that you had in the first 2 legs, in the first 2 big cycles?

Yes. Well, I think some things are sort of nicely built in. So back to my comment about HPC systems and where they're going as far as power consumption and cooling. If I had to say that normally, historically, there were sort of 3 buckets of demand for places you could put IT, it was on-premise, it was in a multi-tenant or it was in a hyperscaler. And then AI came along and this just disappeared, right? There is no option now to stick it in my building because I can't get the power of the cooling, and I'm not willing to even attempt to try to figure that out, instead I'm more am because of things like the skills gap are more likely to go to a service provider who is set up in a multi-tenant data center or who happens to be a hyperscaler. So from our perspective, that's just fuel to the fire, right? So to your point, do we think it is going to be additive to the business overall? Absolutely, in part because you have one of the alternatives just came off the table.

It's fair point. Let's switch gears and talk a little bit about power. One of the things that I was most surprised about through my conversation today is when I ask folks, what's the greatest challenge or bottleneck in their business. They all say the same thing, utility power. So as we think about Equinix, how do you position the company to ensure that you have long-term access to power to support the continued growth of that retail business?

Yes, Chip. He's got it good...

Well, I think the start as we think about it is, first, scale really matters now and scale and having the on-the-ground up market knowledge and expertise. So the fact that as you look at our business, we're doing 53 major builds right now across 40 metros around the world. And we're building -- that's more than half of our metros around the world. So as we're thinking about it from a corporate development function, one is it is increasingly important, not just to think about land banking, but also thinking about power banking to make sure that you have that long runway. So as we're thinking about our largest metros, we're proactively going out there and making sure that we have the 5-plus year lead time for where we think demand is going to be. I think the other big difference too, as you look at our business relative to others, is the velocity with which we sell power is completely different. So taking the example of Northern Virginia, where we've got 2 facilities that were just finished wrapping up on, and we've got the power to be able to continue to operate there through the current challenges that domain means having in terms of transmission. Well, the reason that we can do that is because we're not selling sort of the 5- to 10-megawatt clips. We're selling at the 5- to 10-cabinet clip. And so when we're going to have a conversation with the utility operator, one, we're not taking down the same level of capacity where, yes, if you walk up to a utility provider and say, okay, I need 150 megawatts, and I need it in 6 months. Yes, that market is gone. And so the difference is when we're saying, okay, we need 20 or 30, and we're going to use that up over the next 5 years, that's a very, very different conversation. I think the other thing, too, is as we think about certain capacity constrained markets, so a Singapore, a great example where that has been a market with a moratorium. We're very, very pleased to have gotten an allocation recently. But the reason that we're able to continue to drive that is because the value that we're creating for the ecosystem there, for the economy there is huge. So to Jim's point, talking about the ROIC that a customer can create within our facilities, I think it's having that conversation with the local municipalities, with the local Boards, with the local utility operators. Because in many cases, it's not just simply power now. It's also then can you get the approval from the local counsel? And in Europe, for example, if you're not thinking about how are you sustainably building your data center into the broader ecosystem of an environment, you're never going to have any shot. So if you're a sort of smaller subscale operator building 2 or 3 builds a year, yes, it's going to be a lot harder relative to a scale operator who is building in multiple metros, building at size the way we are.

And sticking on the topic of power. One of the things that's risen in popularity, at least from the conversations I have had, is cogeneration or self-generation of power in capacity-constrained markets. I believe Equinix leverage is Bloom Energy fuel cells in certain markets to power part of its data centers. How do you think of the role that self-generation will play, if at all, in data centers over the long term?

It's going to get bigger. And so as we think about self-generation right now, Bloom, for the most part, is actually more for backup power in a lot of the markets where we're operating. In Silicon Valley, one of our facilities is actually operated primarily on Bloom, and we found that's working well. But take, for example, in the Irish market, where EirGrid has said no new data centers connected to the grid until 2028. We're still building. We just opened up DB5 last year. We're building DB6 right now. But the reason we're able to do that is because we're operating those off of a gas turbine power plant that we built. And the beauty of it is, from a grid perspective for EirGrid is over time, will we connect to the grid? Sure, we probably will. But we then also can become a swing consumer for the grid for them. So as they're making that transition towards renewable energy, which is part of the challenge that they're having, shifting towards wind and solar and other sources, if it's not shining, if the wind is not blowing, we can turn on our gas turbines, come off the grid. And then for a grid payer's perspective, it actually keeps their rates low because then EirGrid is not the one building a power plant.

Got it. Let's talk a little bit about being supply chain dynamics, and this is a point that I raised recently -- rather in the last session. So as you think about it, the supply chain kind of gapped out in 2022 for a lot of the key pieces of equipment. That was important, I think, in response to a demand shock that we saw at the end of 2021, a little bit of COVID supply chain issues sprinkled in on top of that. Now as I'm looking about at the magnitude of demand that's hit the market, has the way that you think about winding up your supply chain and maybe the capital that you're committing to walk in key pieces of equipment, has that changed, if at all, as a way to hedge yourself against an expansion in lead times for key pieces of equipment?

No. And this again goes back to my previous comment of having scale really matters. Because as you're building or as we're building right now 53 major projects, and know, by the way, there is a long tail of smaller projects that we're doing that aren't on our expansion sheet, we can then think about interoperability as we're looking at our supply chain and putting our capital to its highest and best uses. So if -- we talked about on the Q2 call, if to get us 3-megawatt generator right now is 120 weeks. So if you're not in line already, you're having a lot of challenges. For us, if we've got all these equipment already preorder, we've got our position in line, depending on what we need for highest and best uses, as we're seeing demand flow into our markets and doing our builds, we can go back to our supplier and say, "Okay, well, actually, that piece of equipment that we said we need you to ship to D.C., now we're actually going to need that in Chicago." Or instead of having that go to London, let's send that over to Paris. And so that puts us in a position to be able to continue to deliver on the types of time lines that we want to deliver. But also given just the scope and breadth of what we're doing, there are -- certainly, the hyperscalers are going to be bigger consumers of the plant and equipment associated with what we're building. But they are not going to be many other larger consumers of these key data center components the way we are.

All right. Let's switch gears and talk about something which I think we haven't talked about for a while with Equinix, which is what I would call tech-oriented M&A. We saw -- when you guys did the Packet acquisition, you're building out your digital infrastructure services suite. And somewhere along the ways we kind of stopped talking about it, I'm curious, have you backed off of your interest in kind of growing the technology or digital infrastructure services suite via M&A, and now you're just building organically? I'm curious where that falls in terms of the priority list for the business.

Well, certainly, [indiscernible] point of a view on this one.

I can start and then you can go ahead. So absolutely. I mean if you -- and I think we mentioned this in the last -- if you looked at the CAGR against digital services for us in the last few years, it's around 44%, right? So Fabric continues to be the gift that keeps on giving, and it's the fastest-growing product we had. And part of the reason we did Network Edge and acquired Packet was for the same thing. And why? Well, the thing that we realized was that the average customer, especially the average enterprise customer, what they're taking advantage of with Equinix is that sort of positional value of being in the right market, connected to the right counterparties, catering to the right set of customers, right? So they're trying to solve for those 3 things. What's the biggest barrier to entry for that customer? It's the CapEx associated with expansion. It's not us. right? So normally, they would come to us, and they'd say, "Oh, it's a few hundred dollars per rack." That's no big deal, except I got to go procure tens of millions of dollars kit to go do it. And so by offering things like metal-as-a-service and offering things like Fabric and Network Edge, what we're capable of doing is allowing a customer to essentially consume colo as a completely virtualized construct. And so what that does is it just simply speeds up the whole land-and-expand behavior that is basically most of our enterprise customers. It just speeds it up. So we are very committed to digital services for that particular reason. Packet was a newer acquisition. Fabric, it started off light as is what we call Internet Exchange, then it became Cloud Exchange, and then it finally became Fabric. And what we're seeing there is the fact that almost every single network provider, cloud provider, everyone in the market today is all around as-a-service consumption. I need to make everything go as fast as possible. And one of the things that a lot of our customers have realized is, "Hey, I'm going to say I'm a network provider, and I'm connecting to 10 clouds. Well, how am I doing that?" I'm probably doing it through Equinix. What if I want to connect to all the other customers sitting in Equinix facility? Well, Fabric will allow you to do that on an automated basis. So my choice is either I go off and build a Snowflake and then I individually try to figure out with every counterparty I deal with, how I'm going to actually automate that interconnection process? Or by default, I can just use what Equinix has already built. So part of the reason we're seeing such good demand for our digital services is that every network as a service, every compute as a service offer requires some amount of automation going on in the background that links their systems to the counterparty they're trying to talk to that's what Fabric does.

All right. Well, look, with that, we're just over time actually. So gentlemen, thank you so much for being here with us today. Really appreciate it. And again, thank you.

Thanks.

Thanks.