Inside the Data Center Boom: Understanding the Massive Infrastructure That Supports AI

[00:00:00] Steve Eisman: so the hyperscale is all of a sudden their need for more data centers exploded in 2022 their [00:00:05] Nick Del Deo: revenue was about 16 million dollars in 2025 5.3 billion frightening it's unbelievable there's nothing else like core we've out there a data center is basically a warehouse for computers land rich people poor the type of chip that you need to train and run these ai models is very [00:00:23] Steve Eisman: power hungry hi this is steve eisman and welcome to another episode of the real eisman playbook today we're going to explore data centers and all their tangents and with me to help us explore it is nick del deo who is the digital infrastructure analyst at moffett nathanson welcome nick yeah thanks for having me steve so i was thinking about how to start this and i was thinking you know if we had we were talking about banks and i had a bank analyst on you know the bank analyst and his team would have literally almost all the banks that are relevant under their actual coverage there'd be no bank out there in some outside world that wasn't under their coverage they would all the information they would actually be covering but in your case you cover data centers and a couple of other things which we're going to get to but so much of your space is actually not technically in your coverage that's right you say that's fair yeah absolutely you look at um [00:01:26] Nick Del Deo: the current currently publicly traded data center stocks there's digital realty and equinix just about it and there used to be more there used to be more that number that went private and i think what you've seen over the last several years is a lot of data center construction has been happening in the private markets rather than the public markets a couple reasons for that the sheer volume of capital required i think has spooked public investors um the the the accompanying cash flow profile is is not something that public uh investors tend to tend to like you know if you pitch a story that you know we're going to spend a bunch of money and be negative cash flow for years and years and maybe five or ten years down the road while the payoff public shareholders don't like that they don't like that at all [00:02:05] Steve Eisman: so let's take a step back for a sec let's lay out for everybody the entire landscape of data centers who what what are hyperscalers what are data centers what's core weave lay out for me all the players that that basically encompass your space whether you cover them technically or not [00:02:25] Nick Del Deo: sure so if we think about a data center a data center is basically a warehouse for computers it's got a lot of power available it's got power backup systems cooling systems security fire suppression systems basically a purpose-built space for it equipment players like digital realty and equinix are landlords who own the buildings just the building just the buildings then you have customers who use the buildings so that could be someone like a hyperscaler it could be a corporate enterprise fine hyperscaler so hyperscaler typically think of it as four players so amazon microsoft google and meta basically people who invest in it infrastructure at a scale that's just unmatched globally and when 10 years ago would have been unfathomable yes yes so you have a variety of different type of customers whether it's hyperscalers corporate entities network service providers and so on you have public data center operators like digital realty and equinix you've got a host of private operators cyrus one qts vantage down the list you have fiber operators uh who operate the fiber that connects the data centers to one another so that could be like lumen zeo cogent crown castle 18t verizon down the list you have players like dicom or maztech who will construct the fiber players like corning who supply the fiber both inside and outside of the data centers you have equipment suppliers nvidia intel amd cisco and so on that provide the dell um provide the equipment that goes inside gpus and the servers yeah the servers the gpus the cpus so there's a whole whole stack and then there's some um the lines aren't always clear cut in terms of who does what someone like a microsoft is a big uh customer of the data centers they lease a lot of space for them but they will also own their own [00:04:17] Steve Eisman: facilities in many cases so there's a mix of them so microsoft would might lease a building from digital realty and put all its stuff inside and then two miles down the road just simplistically they might own their own building and put all their stuff in there as well yes they do both they do both okay [00:04:34] Nick Del Deo: yes so sort of like how an amazon might own some and lease some warehouses the hyperscalers would [00:04:40] Steve Eisman: do the same thing can you give us an idea of how much capex is actually going on in this space right now because if when you add it all up who's spending what i mean the numbers that get thrown out are [00:04:48] Nick Del Deo: astronomical yeah i mean the simplest way that people look at it is to sum up the capex that the hyperscalers spend each year and that i think in 2025 that's supposed to be 350 billion dollars or so up substantially over the past several years and expected to grow further in the coming years now it's it's a little bit um i don't say misleading but it's a bit murky in terms of where that money goes so when you think about the cost of a data center it's generally in the low teens millions of dollars per megawatt i'm sorry say that again so so to build a data center you typically quote it in terms of megawatts the power capacity that the data center offers i see and it's usually they'll call it low teens or so millions of dollars per megawatt so if you build a 100 megawatt data center it's going to cost you what's that 1.2 billion dollars or so right that's more than obviously just the [00:05:44] Steve Eisman: building it's everything that's going to go into the building yes okay or actually no that that is [00:05:49] Nick Del Deo: the building just the building the stuff that goes inside is three to five times the cost of the [00:05:55] Steve Eisman: building so so so the the building that digital realty and equinex builds for this 100 megawatt entity you're saying is about 1.2 billion dollars to to build just the building yeah 100 100 megawatts [00:06:11] Nick Del Deo: yeah yeah and then and then you've got the um again the the stuff that goes inside is multiples of that right so when you look at the the capex of the hyperscalers um the bulk of that is actually on servers and it equipment networking gear that goes inside data centers rather than what they're spending on data centers themselves to the extent that they lease facilities that's generally not showing up in capex that's showing up as a lease right if you look at the capex for the data center operators right so again people who focus on the the buildings not the stuff inside that's obviously been up dramatically over the last several years but again a lot of it's happened in private hands as opposed to [00:06:49] Steve Eisman: public hands so just for viewers to understand there used to be a whole bunch more public companies that competed with digital realty and equinex and they got taken private by private equity and now they're private and they're operating that's right they're still competitors they're still competitors [00:07:03] Nick Del Deo: but like you know qts was public it got acquired by blackstone who's funneled uh enormous sons of money to grow that business cyrus one who's taken private by kkr and gip so let's dig down a [00:07:15] Steve Eisman: little bit to um the two data center companies that you cover their reits which basically means they pay out on a very high percentage of their cash flow as dividends they don't really retain capital so if they want to grow do they have to constantly raise capital to basically keep growing is that is that is that the way it's structured or or can they yes and that's been one of the sore points among [00:07:39] Nick Del Deo: public investors is that uh data center operators have been public have had to consistently access the capital markets to grow given the magnitude of the investment opportunities they have um and having to pay out a substantial portion of their cash flow as dividends now i think at the end of the day it's a more nuance than that right so if you're a real estate operator and you can qualify as a reit uh you don't have to pay corporate level income taxes so you eliminate double taxation there's a real value to that so i think you know if you zoom out you're basically trading off not paying corporate income taxes for maybe having to issue equity over time i think net net that's that's a benefit for the industry but it certainly means that they have to lean more on the public markets to access capital than than most [00:08:23] Steve Eisman: other providers would i mean these companies years ago were not reits then they converted to reits [00:08:28] Nick Del Deo: yeah well you know if you go back to digital realty when they launched because it was probably 20 years or so ago as a public entity they've always been a reit equinex was not equinex converted that's right they took a similar path as as what the tower operators did qualify as a reit but so long as they were early in their growth cycle and had negative uh taxable income there was not necessarily a point to be in a reit because it simply imposed that dividend payout requirement on you without a commensurate benefit once they got to a point where their taxable income was going to come positive then it made [00:08:59] Steve Eisman: sense to flip over to reit and avoid paying it so are the um the private entities that used to be public growing a lot faster than digital realty and equinex yes yes how much faster hard to put exact numbers [00:09:12] Nick Del Deo: on it but qts for example i saw a stat that they've leased you know multiple gigawatts of capacity that that they plan to bring online over the coming years which would actually from the time that they were public they were substantially smaller than digital realty i think if you zoom out from a capacity perspective and look forward a few years they could be you know twice as large as digital realty so really pressing the gas in terms of the amount of capital that they're willing to put into the business the way that they finance it so willing to put a lot more debt on it and use some creative financing structures like abs and and you know potentially explore some different markets to really you know accelerate that growth if you're a public operator again you know you've got to worry about your affo or ebitda trajectory over time and it makes it harder to really um lean as as heavily into the uh investments as you as you might would as a private entity if you're willing to sustain higher leverage or willing to defer the benefits from a cash flow so the private companies they're more [00:10:12] Steve Eisman: levered than the public companies yes how much more it depends but you know it could be twice as high twice as high yeah so are there any shareholders out there who are advocating that these companies de-REIT [00:10:26] Nick Del Deo: no not why not well then you you incur the um but you could retain capital you could grow more yes but again the effective solution to that is going private going private and digital realty and equinix in practical terms are too big to go private how big are they now yes from an enterprise value perspective i'm going to say equinix is probably in the 90 or 100 billion dollar zone and dlr is probably [00:10:50] Steve Eisman: in the 75 billion zone and how big were qtx and cyrus one when they got taken private oh good question they were probably i'm going to say in the 20 billion range okay so much more manageable yes yes yes [00:11:04] Nick Del Deo: yeah so still still a a a decent take private valuation but not in the same league in terms of [00:11:10] Steve Eisman: size yes it was it was easier to fund interesting okay um let's talk about uh let's go to digital realty and equinex right now sure um so why don't you describe the difference between the two companies in terms of where they focus their business models because historically equinex was always used as a better company used to trade a higher multiple these days the multiples are roughly the same let's just walk us through the whole um story of these two companies what do they do how are they different what are the [00:11:38] Nick Del Deo: management's like etc yeah so i think you're right you know historically equinex was viewed as the clearly superior asset and the reason for that is because equinex's niche is quite a bit different than digital realties um people tend to use data centers as this monolithic industry but in reality there are a number of different business models or data center types that are out there and they have different profiles to them you can kind of break it down into three categories right call it hyperscale data centers or very large data centers uh where you would host acres of servers for folks like microsoft [00:12:13] Steve Eisman: or google or amazon and there might be only one tenant in the building yeah in some cases one some [00:12:17] Nick Del Deo: cases a handful okay uh but historically that's been a fairly you know maybe not commoditized but a very competitive niche in the market and why is that well because you know if you were to build a data center and i had a data center there's not a lot to differentiate it right so long as they're in similar region similar proximity to a key interconnect point which we'll get back to similar size similar age not a whole lot different about them you've got a another group of data centers that are focused more on um serving enterprises so think what do you mean by enterprises so think of jp morgan or walmart big users like normal companies yes exactly exactly you know so maybe as opposed to the amazons of the world exactly exactly so think you know jp morgan wants to run some of its internal systems but it doesn't want to put it in the cloud right that might go in a corporate type data center or an email system a disaster recovery system those sorts of things and then the third type is um you know an interconnection focused data center and these facilities are are ones that really cater to companies looking to be in close proximity to other uh to some of their counterparties with whom they want to exchange data like give me an example so you know if we go back to the dawn of the internet right so the dawn of the internet as we know it the late 90s um let's say that at&t needed to exchange traffic with verizon bt deutsche telecom comcast down the list the most efficient way to do that is for everyone to put their equipment in one data center and then run cross connects or cables from there in between their servers exactly right exactly so a business model kind of cropped up and equinix really led the charge here but to to build these purpose-built neutral facilities what do you mean neutral neutral as in they didn't have a stake in in the business one of the challenges before this time was that if um let's say at&t wanted to get access to a sprint facility well sprint would kind of try to use it as a competitive advantage really jamming them on the price make it inconvenient to to do that equinix said you know we're going to treat everyone equally we're switzerland exactly they're switzerland so what happened is uh for these interconnection oriented facilities in which equinix is the leader uh networks came in to exchange traffic with one another because none of no network is ubiquitous they all need to to hand off traffic to one another to get data where it needs to go it's a you know we learn about network effects in business school and this is a literal network effect right if you go to an equinix facility or a digital realty facility that's oriented along these lines you know there could be hundreds or thousands of cables between customers in the facility just a spider web of of [00:14:58] Steve Eisman: cables between their gear so you're talking about it basically a data center where almost all the servers from different companies are actually connected to one another yes primarily the networking gear right [00:15:08] Nick Del Deo: right that's right and and you know so what you've seen historically is that these interconnection oriented data centers kind of form the hubs in key data center markets because that's where data comes in and out and then these other facilities the enterprise facilities or the large footprint or hyperscale facilities will tie back to those interconnect points i'm not following what do you mean they tie back to those interconnect points so in other words um a network service provider's network is going to terminate in equinix and then there's going to be fiber from a hyperscale facility back to equinix i see right okay so so equinix is sort of the connectivity hub and and what you've seen happen over time is with that network foundation in place a lot of content and digital media companies decided to co-locate in equinix to take advantage of the network access pump out their content to eyeballs over time you saw cloud service providers put in um on ramps to their platforms so on ramps are basically the the physical equipment where customers can directly access the cloud you think of it as a as an on-ramp to the jersey turnpike okay right um and then you saw enterprises go into equinix so that they could access the cloud providers [00:16:20] Steve Eisman: directly so this service that i guess what equinix was equinix was was providing was proximity to everybody else yes that your data flows a lot more quickly has has real value yes it's think of it as [00:16:32] Nick Del Deo: as um akin to the chicago merc or ebay in its essence it's it's a marketplace right so if you want to if you want to exchange traffic directly with the counterparties in equinix you go to equinix right and everybody goes to equinix because everyone is at equinix so it feeds on itself exactly okay so equinix has historically been the leader in that space and and it's a legitimately differentiated model that's hard to replicate if you or i built a data center next to equinix and tried to offer [00:16:58] Steve Eisman: but we don't have that we don't have the network at that we can't do it you can't really had a real competitive advantage with this with the creation of this network that feeded upon itself everybody wants to be an equinix because of the interconnectivity exactly which breeded more interconnectivity [00:17:12] Nick Del Deo: precisely got it so what you've seen historically is that equinix or other interconnection focused operators there was core site which was public that was acquired by american tower there was interaction in europe that was acquired by digital realty and a handful of others um they tended to generate returns [00:17:28] Steve Eisman: that were much much better than the remainder of the industry now explain how much more profitable it is to have this type of networking data center versus a plain vanilla data as i measure it you know [00:17:39] Nick Del Deo: if you were to go back say three or four years before the ai boom which has flipped things around a little bit and we'll talk about that the return on invested capital for for guys focused on hyperscale facilities or more vanilla enterprise facilities wasn't terribly different than their cost of capital [00:17:54] Steve Eisman: and the cost of capital would be roughly what yeah i'll call it 10 percent at that time no you know probably seven percent eight percent um so your roa is equal to seven percent it's not great yeah it's [00:18:07] Nick Del Deo: not and that's why if you looked at the say the afo per share growth for those companies over time or their stock price performance it was a struggle for them whereas equinix um and and you could look at core site again similar business model was also public at the time their roic was probably in the you know low to mid-teens so basically the difference yes that's meaningful yeah okay uh and only just [00:18:29] Steve Eisman: because you're putting wires in between people in the back of the servers exactly to charge more exactly [00:18:34] Nick Del Deo: it's a classic network effect right uh now fast forward to three years ago you know where the ai boom really started and it's continued through today um the sheer volume of capacity that players like microsoft and amazon or open ai are demanding has kind of shifted the landscape a bit so the hyperscale is all of a [00:18:54] Steve Eisman: sudden their need for more data centers exploded yes because they just need they need to create more and [00:19:00] Nick Del Deo: more yeah and and more specifically the the gpus the type of chip that you need to train and run these ai models is very power hungry and the way you get to more powerful models is by putting more and more of these chips together in one location so it sort of upended the data center landscape a bit in that it introduced legitimate supply constraints so in other words the amount of capacity that the hyperscalers and others were demanding was greater than what the industry could supply and you saw pricing for um larger footprint space which historically declined year over year um you know over the last several years it's it's probably doubled okay you know so you've seen returns on these larger builds improve dramatically [00:19:47] Steve Eisman: how much more would you say from the seven percent return yeah it's probably yeah you call it from 78 [00:19:52] Nick Del Deo: percent you know everybody defines it a little differently but it's probably low teens today [00:19:56] Steve Eisman: okay so almost up to the equinix return on investment capital yeah and and but not only that is you [00:20:01] Nick Del Deo: know you've got a bigger spread between your cost of capital and your return but a much larger volume of capital that you can invest behind it so you get both the the return benefit and the amount of capital you can deploy has gone up so that's really benefited players who focus on that part of the market which is digital realty yeah as a public player digital realty has been a real beneficiary plus a lot of these private players equinix nothing about their value proposition has been eroded but at least to date ai has really benefited um or it's really been focused more on model training creating these models as opposed to inference or the usage of these models training ai models it's not latency sensitive i know what that means but i'm not sure all our viewers do sure explain what that means so latency sensitive means close to the people who use it if you were to look at where say equinix has its data centers they're in tier one markets so think new york silicon valley chicago london singapore tokyo uh you know they're not out in the middle of wyoming right because you know where the networks terminate is generally close to where they uh again you want to reach the eyeballs [00:21:08] Steve Eisman: if your eyeballs are in new york you don't want it you don't want your data center to be in alaska [00:21:12] Nick Del Deo: exactly because it because it takes time to get there right right so at least a poor a poor experience for for the customer ai training is not latency sensitive because it's a process that could take weeks or months it's power hungry so the cost of power and availability of power matters a great deal and you need a lot of land so what you've seen is that these so you could be in alaska yes in theory yeah in theory in theory in practice what you've seen is a lot of these facilities crop up in places like texas north dakota wyoming land land rich people poor yes yeah a lot of land and and power [00:21:49] Steve Eisman: availability right that's been that's okay we're going to talk about power we'll get to that so you [00:21:53] Nick Del Deo: know so we've seen um so the the folks who are more willing to focus on um or who focus more on these larger deployments have really benefited either because they they host these deployments directly or because uh these deployments have just sucked up all the available power such that there's scarcity for what remains what's happened is there's been a bit of a flip in investor sentiment where people view the players focused on the hyperscale opportunity or the ai opportunity is much more interesting than equinix which has not seen the same sort of benefit at least not yet and that's because [00:22:28] Steve Eisman: all the ai actions really in the light the language models that are being created as opposed to there's no greater apps yet yes there's some you know chat gpt is perplexity but but but there's no you know i'm not on my phone every single second of the day doing something in ai yet and when that happens [00:22:46] Nick Del Deo: that'll benefit equinix yeah a little more nuanced but yes i think some ai applications will remain latency insensitive like so think like deep research open ai deep research where you ask it a question five minutes later it comes back to you with a research it doesn't matter whether it's five minutes or four minutes yeah or or more precise you know five minutes or four minutes and 59 seconds exactly uh but if you're you know if you've got an application say um related to gaming or creating custom ads in real time things like that that's going to have to be more proximate to the user and it's got to be fast yeah so another the amount of time between when you submit the the prompt or the request and when it comes back to you is much more important okay uh and we're not there yet no mostly mostly yeah so i think that's that as that happens that will benefit equinix um it also benefit players like digital realty you know digital realty has not participated in these middle of nowhere data center deployments they focus on tier one markets as well so i think i think who's doing [00:23:51] Steve Eisman: the stuff in the middle of nowhere is it private companies yeah to a large degree and also the [00:23:55] Nick Del Deo: hyperscalers and the hyperscalers you know you've seen again a lot of hyperscale builds um there's less of a willingness on the part of third party developers to go into these oddball markets because they worry that 10 years or 15 years from now when the leases come up if the market changes they won't be able to release it got it uh so you've seen some private players that are maybe more aggressive the hyperscalers um and you've seen guys like this is another interesting angle but um bitcoin miners repurposing their capacity for high performance computing that was a sentence that was definitely in english i'm not quite sure i understood it all right explain what that meant all right so um cryptocurrency mining consumes vast amounts of power that i know and it's not latency sensitive so you saw firms again flock to places like west texas or north dakota because there was a lot of power land was cheap and they could set up these uh crypto mining facilities and what's happened is uh those sorts of locations have become more interesting to uh the folks developing ai models because they've got a lot of power in one location right so what a number of them have done is is say you know what we're better off taking this capacity this power capacity that we're using for crypto mining and again and becoming data center operators and leasing this space to folks who want to use it for that purpose i see core weave is an example of uh you know a gpu cloud provider who's core weave core weave yes they've leaned heavily on some of these reformed crypto miners to get access to so they throw the crypto miners out yeah well it's a you know they've been evicted that was a voluntary process okay yeah the the bitcoin miners did it by themselves because they saw there was a better return to be had leasing it out to players like core weave okay um but corey actually went so far they're um they're trying to acquire a company called core scientific there's a bit i saw that was about a week ago yeah yeah a couple weeks ago um there's also there's applied digital galaxy digital again sort of crypto miners are looking to repurpose this capacity the company [00:25:58] Steve Eisman: they're trying to come to what does that company do that they're trying to acquire core weave yeah so [00:26:03] Nick Del Deo: again historically it was a crypto miner okay and they procured lots of power in places like west texas oklahoma um georgia i believe in a handful of other locations and said you know we've got this is interesting to players like core weave so they struck a deal where they would basically vacate some of the crypto facilities that they had built upgrade them to a more traditional data center spec and then lease out that space to core weave so they could put their gpus in okay i will come [00:26:32] Steve Eisman: to core even in a second i mean if you look at digital realty and equinix this year the the thing that's kind of surprising to me and i know enough to be dangerous because i have tracked these companies for years but obviously you know a hell of a lot more is they're actually down this year and you know if you look at the whole ai digital infrastructure space like everything is up yep some of it's up insanely you know we could argue whether that's correct or not but you know the the sentiment for the ai space is huge and yet here like if you knew nothing you would and i said to you what do you think digital realty and equinix have done this year you would say oh they have they have to be up because everything is so good so why are these two stocks [00:27:12] Nick Del Deo: down it i think it relates in part to the dynamics that we talked about at the beginning of the discussion the public market dynamics so and you know we can focus on equinix so equinix has an analyst day every two years they had one in june right uh not far i did not attend it was at the time center a few blocks away uh but they laid out a growth plan over the next several years and it involved a substantial amount of capex and when you spend a lot on capex particularly for for facilities uh retail oriented facilities like what equinix offers so lots of small customers means it takes time to fill them up uh takes time to get them online so if you start spending capital to build a new facility today uh there is a drag that that imposes on your results interest expense when you buy the land and start construction then you bring the facility online until you reach a certain utilization rate you're going to be um either cash flow negative or ebitda negative or at least below where you were where you were um and they're putting enough money to work that it's actually depressing their afo per share growth trajectory why do you define for people what [00:28:20] Steve Eisman: affo is because that is the metric that that matters in this space yes since everybody's on the same page [00:28:25] Nick Del Deo: yeah so so for real estate more broadly and for data centers specifically afo stands for adjusted funds from operation and what it's supposed to present is a measure of how much cash flow the companies um produce setting aside growth related capex so in other words in a in a mature state or in a state where they stop reinvesting in their businesses uh how much cash flow does the company what is it what is it spin off okay so for for equinix which has historically grown its average share it call it a 10 percent rate or so at least for the last several years you know they expect that to tick down to like a mid single digit rate next year before kind of that's significant that's significant and and then so before going back to you know sort of 10 or eventually double digits if the plan works so [00:29:11] Steve Eisman: that that freaked people out now yeah but by the way if if this was a private company equinix yes and this was the plan this would be approved by by in a heartbeat of course the opportunity here is massive we have to spend some money in the near term and our cash flow is depressed who cares exactly [00:29:31] Nick Del Deo: but it's a public company that's the problem that's right and that's why so many of the other operators have gone private right yeah so i so i think you know that what happened with equinix certainly dragged down their stock and i think with digital realty they were sort of affected and you know in sympathy um and you know to some degree dLRs had a very good run over the last few years i think it got down to as low as 80 a share or so and it probably traded up to 180 you know so yeah it's down [00:29:58] Steve Eisman: are the valuations now roughly comparable in terms of how people value these stocks because in the past equinix used to sell at a pretty significant premium that's right if you look today on a headline [00:30:09] Nick Del Deo: basis dLR trades at a premium of a few turns to equinix off of a ffo yeah ffo or or ebitda okay um people look at both i'd say you know it's a little more nuanced than that so with um equinix has higher stock based comp which you want to take into account dLR hyperscale leases tend to be longer than retail leases so again retail is mostly what equinix does and what a portion of dLR does they have much more exposure to hyperscale leases the hyperscale rents or the large footprint rents have really been what have skyrocketed so they've got digital realty has a greater opportunity to release some of uh leases that are coming up for renewal in its base over time that helps and they have a proportionally larger development pipeline than equinix so equinix again earns better returns than the dollar it puts to work but because it's a niche its investment opportunity [00:31:02] Steve Eisman: is more constrained than what you see at other players is equinix in the hyperscaler space at all or or oh yeah so small way if you want to roughly 40 percent a little under 40 percent of [00:31:14] Nick Del Deo: cloud on-ramps globally are hosted in equinix facilities so what that means is that if you want to access aws your traffic is being routed through equinix okay so in other words you know if if this is an aws data center where they have all their servers and this is an equinix facility if you want to get if you want to get here you've got to run through equinix from it from a hyperscale perspective or from a cloud service provider perspective it's very efficient to be able to put your gear in in a facility where you've got access to many many networks many potential customers um even if even if your competitors have their on-ramps there because that makes it more attractive to the customers that want to use your cloud and i think they've come to the realization that um multi-cloud is how a lot of customers want to approach uh approach the architectures a big enterprise isn't going to lean solely on aws they're going to want to use aws for some things azure for some things google cloud for some things core weave for some things um so being able to to put equipment in equinix and directly access all their cloud providers is very important to them so yeah so equinix's largest customers are the hyperscalers but it's a much smaller share of their total revenue than what you would see at others have much a much more diversified base of business but hyperscalers are a key magnet to get customers [00:32:36] Steve Eisman: into equinix so would you would you say it's fair to say that when not if but when there are tremendous number of ai apps that's going to be the time when equinix shines yeah i think that'll absolutely help [00:32:50] Nick Del Deo: their business um again i think it'll be the more latency sensitive apps or maybe private ai deployments meaning uh customer doesn't want to use the public cloud to do it for whatever reason security privacy you've seen some um equinix gave an example of bristol myers for example putting a private gpu deployment in one of their facilities for that reason um yeah i think that'll benefit equinix's business it's never going to lead to the sort of like rocket ship results that you've seen at some others because equinix has a very diversified business you know cloud cloud adoption multi-cloud adoption that's still a big growth vector for equinix networking is a big vector financials right equinix will host uh like the bats exchange in secaucus you know so there's so everybody goes to equinix uh secaucus because they want to access the trading exchanges so they've got multiple pillars of their business of which ai is one as opposed to say you know osiris one or qts which is disproportionately benefiting from ai as a growth driver because they're so hyperscaler oriented yes [00:33:51] Steve Eisman: okay let's talk about core weave for a minute sure and then i want to talk about power or lack of power okay um so when did core weave go public it was the end of march okay so it's a very new company and yet when i go on to watch cnbc every single day i don't think a day goes by that that company is not mentioned i think that's right um so why don't you tell people like what exactly does core weave do and [00:34:15] Nick Del Deo: why do people care so much sure so um core weave operates a gpu cloud so a gpu graphics processing unit so those are the chips that are really um necessary for training and inferencing ai models okay so they've been in high demand from the hyperscalers or from open ai anthropic and so on um aws azure gc google cloud they operate more general purpose clouds so more you know think of it as traditional storage or compute with cpus now that they obviously play in the gpu space in a big way but core weave is purpose built for for ai purpose built for gpus so that's some definitions [00:34:53] Steve Eisman: sure so digital realtor and equinex they own the building yes but they don't actually operate what's in the building that's correct but core weave has the owns the building or leases the building to date [00:35:06] Nick Del Deo: core weave has exclusively leased its facilities and it owns the equipment and it owns all the equipment [00:35:12] Steve Eisman: inside which is exclusively gpu oriented that's right so it's it's the the newest version of a data [00:35:18] Nick Del Deo: center that's what they operate generally generally you know they'll have some smaller clusters and you know older facilities but but they're working with a lot of partners to build facilities that are um call it you know for with ai specs so much higher power densities liquid cooling right uh stronger floor totally state of the art yes yes yeah so they operate gpu cloud uh in terms of their customers you know microsoft is currently their biggest customer what percent of their customers is microsoft recently it's uh 70 percent wow now that's going to shift in the coming years they've signed some big deals with open ai but it's it's interesting because you know go back a few years and microsoft was struggling to bring on as much gpu capacity as its business needed in large part because it worked with open ai it leaned on core weave to help core we've had a good relationship with nvidia right nvidia doesn't want to only sell gpus to the big cloud providers they want to seed some new providers and you know core we've really seized the opportunity built a business from next to nothing to something quite substantial how young is the company it's several years old you know i'll go back find several i i think they started in 2017. so it's pretty new as a crypto miner as a crypto mine and that's how the management team learned about ai because crypto mining uses gpus yes okay uh they really you start out over they started out over here and they ended up over here and and but you know what's really fascinating is core [00:36:44] Steve Eisman: weave in 2022 their revenue was about 16 million dollars well i said 2022 the entire revenue of the [00:36:53] Nick Del Deo: company was 16 16 million dollars about this about as much revenue as two chick-fil-a locations okay and that is that that is accurate that is accurate uh in 2025 5.3 billion oh my god and if you looked out to 2028 so three years hence it's probably in the mid-20s billion revenue yes okay they don't make [00:37:16] Steve Eisman: many money yet yeah well it depends on how you define it okay good old good old-fashioned earnings [00:37:21] Nick Del Deo: per share yes no no no but you know what what's you've had other very fast-growing companies in history but they tend to those tend to be more like intellectual property what's really interesting about core weave is that they've grown this quickly with um and it's an infrastructure business so their capex has been astronomical like this year i think they're spending 23 21 to 23 billion in capex well [00:37:45] Steve Eisman: let's put that a new gpu core gpu facility that that core we've created let's it's from the de novo you know they here's the land we're going to build it up here what does that cost just one facility well [00:38:04] Nick Del Deo: depends on the size of the cluster right a different way to put it might be if they sign a they signed a contract with open ai for i think it was 14 billion in total it was like two pieces i think it was 11 and another four maybe 14 or 15 billion total contract value over five years um you know it might cost them if i remember correctly 40 billion or so in equipment to support that contract 40 billion [00:38:30] Steve Eisman: to basically put all the stuff that needs to go into the building yeah yeah i mean it shot you [00:38:34] Nick Del Deo: know in in just a handful of years their capital budget is in the same league as at&t walmart exxon [00:38:43] Steve Eisman: mobile right it's crazy crazy from so i gotta repeat that again 2022 16 million in revenue yes and today 5 billion in revenue 5.3 this year yeah it's frightening it's unbelievable yes this is where i i i personally have always had a problem with high tech growth companies in terms of like how do we even begin to fathom the valuation so when you look at core weave how do you even think about value i mean you've got a company whose revenue is exploding yes doesn't make any money yet yes but the growth opportunity is enormous yes walk us through like just conceptually how do you how do you even begin to think of okay i i think this thing could be worth x because of y like what what walk me through that process that that that's a terrific question um because because in bank world we just say what's my tangible book value and you know depending upon the return on equity of the company that that'll tell me what multiple of tangible value end of discussion walk me through the your your thinking yeah so [00:39:46] Nick Del Deo: what's fascinating here is there's nothing else like core we've out there right either in terms of the business model or um or it's you know the size what a lot of my peers have done which which i would not endorse is they've they've looked for other names sort of that are tangentially related like data center operators like equinex or microsoft google and and sort of slapped a multiple on core reasons [00:40:08] Steve Eisman: results uh what what result are you slapping a multiple on yeah either operating income or ebitda [00:40:13] Nick Del Deo: they do have ebitda yes yes again because their cost is all the it's it's the gpus it's the capex right uh so you know their their ebitda margins are probably 65 or so but they're obviously a lot of depreciation expense that weighs on operating income um i think that's totally the wrong way to [00:40:32] Steve Eisman: look at this business let's just backtrack yes so if core weave buys a billion dollars worth of gpus et cetera et cetera that's not part of ebitda what is part that all the depreciation from all that stuff [00:40:46] Nick Del Deo: is added back to the ebitda that's right that's right so in a way but it hits it hits operating income [00:40:51] Steve Eisman: and hits operating but if you're valuing the company on the ebitda on ebitda in some respect you're not taking into account like you know what it's costing you to create this thing and what return you might you might you're expecting to get on that that's the weakness of it yeah it's a silly [00:41:05] Nick Del Deo: it's a silly way to look at it right you know i think what you need to do is look at the unit economics of a deal you know so a gpu it remains to be seen exactly what the economic life of a gpu is core we've depreciates them over six years like i said the chips the chips um they signed that deal with open ai which is arguably the most sophisticated user of chips out there they signed a five-year deal with two one-year renewal options so they've got some life to them but at the same time you know you're still in a supply constrained market nvidia is trying to increase the pace at which it releases new chips so it's a bit of a question mark as to how long these things are going to be economic but uh what core we've does to reduce that risk to some degree is they will sign multi-year deals with their customers let's call it four or five years of locked in revenue and then at the end of the term they either try to contract it out again or or rent it on a spot basis but the approach that i took is is to say okay what are the unit economics of these deals you collect some cash up front you have a lot of capex up front you've got locked in revenue for a period of time you've got data center rental expense power costs and so on and then at the end of the deal term you have some ability to potentially squeeze some more juice out of these things towards the tail end of their lives my sense is that for every dollar of capex they spend they're probably generating something like 15 to 20 cents in value in in net present value terms so i kind of look at it you think of that like as a return on invested [00:42:36] Steve Eisman: capital exactly yeah it's a very dollar they buy simplistically for every chip that costs them a buck [00:42:42] Nick Del Deo: they make 15 to 20 cents off the chip yeah over the life of the life of the contract present value their net benefit is 15 to 20 cents okay so so the way that i've tried to approach it is you know how much capex do we think they can put to work over time um what's sort of a steady state that they eventually reach and think of that that value creation as a mechanism to value the company um it's a bit of a nuanced approach but i think for something that's growing this rapidly in a new space that's it struck me as the logical way to do it even if it's not a traditional so what kind [00:43:15] Steve Eisman: of valuation do you come up with or range of valuations you come up with for core weave yeah so [00:43:19] Nick Del Deo: our target price uh we just updated to 65 and stocks trading it moves a lot yesterday it was around 90. okay so so we're a bit below i think um look i mean the stock is growing like like a weed um at the same time and i think they've done a great job of of seizing the opportunity to position themselves as an alternative to the big cloud providers at the same time i think that um you know the big csps are big battleship businesses what's csps a cloud service provider so i think aws azure gcp so i think in time they've been a bit slow out of the gate a little bureaucratic but they have essentially limitless resources and i think i think they'll get better in this space i think chip constraints are easing i think net net that will probably put some pressure [00:44:06] Steve Eisman: on corey's returns over time so what you're saying is that right now because of the shortage of chips and core we have a good relationship with nvidia they have this kind of competitive advantage to seize the gpu data center space but the other guys are going to come yes yes i mean and i think you [00:44:23] Nick Del Deo: know core is a legitimately good operator operating these clusters is not easy and i think a lot of people will think of corey as simply you know buying chips from nvidia and releasing them but there's really an operational component too that differentiates them and they've been good on that front what's the operational component well then you know if you've got tens or hundreds of thousands of gpus in a data center linked by hundreds of miles of fiber and they tend to have the chips tend to have relatively high failure rates you have to have the right networking equipment i haven't heard that before what are the [00:44:53] Steve Eisman: high fail what's the failure rates of these gpus yeah you get different answers from different folks [00:44:57] Nick Del Deo: i'm sure you do but you know whether it's because of the heat they produce or simply the complexity of the system they tend to go offline at a faster rate than cpus now in some cases it's a matter of tweaking it to get them back online in some cases they literally burn out and you gotta replace them but it's much more complicated to operate a gpu cluster than a traditional cloud okay so they've done a good job there but yeah i think that in time the incumbents will be more i think more effective here i think chip supplies loosen up and that probably weighs on pricing and returns to some degree and you know from a buyer perspective you know like open ai is probably going to be corey's biggest customer in the coming years if you so you got to believe in open ai now look they're clearly a leader from a tech perspective they created software that could change the world they're also very cash flow negative you could see models like deep seek or other open weights models come out that maybe aren't as good as open ai but they're fast followers you know which could you know impact open as ability to monetize it's frontier models um or if you believe that open ai becomes you know the clear leader it kind of gives them monopsony power in terms of buying their who they buy chips from right so you know whether they've done that already they buy from microsoft they buy from oracle they buy from core weave you know so there are ways to manage that to ensure that your suppliers don't generate excess returns at [00:46:24] Steve Eisman: your expense so let's change gears for a second switch to um a topic that's not technically under your coverage but but uh if it's not solid your coverage is going to go down the toilet ball yes which is power sure walk us through the landscape now of every day you hear stuff about you know nuclear getting involved people bloom energy i mean the list is endless walk me through the landscape of of how much is this industry growing so that the power capacity of the united states that's not going to meet it what are people doing to try and get power um where are we yeah i think if you look over [00:47:01] Nick Del Deo: the last several years i think the data center in the u.s data center capacity has probably grown at like a 15 kager or so and projections obviously wide ranges but maybe 20 is a reasonable benchmark okay now within that you've got the ai component growing a lot faster you still have growth from cloud and then you have some older applications that aren't really growing but that's kind of how you get to that sort of forecast the data center industry in the u.s i'm going to say is in the 25 gigawatt range in terms of installed capacity today um i think the total generating capacity in the u.s is probably one one point i think it's one point two five terawatts so 1250 gigawatts uh obviously generation is lower because power plants don't run 100 of the time you've got peaker plants and solar and whatnot that are that are more sporadic but i think data centers are probably a low low single digit percentage of power consumed in the u.s and that and that's going to grow again a lot of challenges in terms of getting the generation online the transmission online the equipment for substations to support all that my you know my hunch is that we live in a market economy and and it responds to signals and you know you see things like capacity auctions at pjm which is the uh the organization that runs the power grid and you know a lot of the mid-atlantic including virginia which is the biggest data center market you know i have capacity auction the prices that just seem to be like screaming for people to come in and build generation so i suspect that in time we'll see more balance there you're seeing data center operators go uh uh try to circumvent the utilities in a way and how so so for example build your own generation you know buy gas turbines put a data center near a gas pipeline and generate power [00:48:57] Steve Eisman: yourself until you can get a grid connection so buy the gas turbine and operate yourself yes yes your own [00:49:02] Nick Del Deo: mini utility yeah i mean it's not what they want to do but it's right they need to but you know i think you know it creates an interesting question for the data center operators because if we go back to what we're talking about before it's it's been this supply crunch that has allowed players focused on the hyperscale facilities or their corporate facilities to generate better returns if we see supply power supply come more into balance with demand i think we probably see a reversion to what we had before which is returns much closer to the cost of capital for that piece of the industry so in a way lifting supply constraints really helps from a volume perspective but it hurts returns so from a value creation perspective for the industry it may not be ideal i think if i were a data center operator focused on that part of the market i'd want supply to remain tight because that's my stick to get [00:49:56] Steve Eisman: better pricing and returns i see okay so you're not that all that worried about that that the energy [00:50:01] Nick Del Deo: needs will be met look i think in time i mean it's clearly tight and people are working around it you know you see some projections for like hundreds of gigawatts there's actually a lot of double counting in those numbers or a lot of uncertainty there i'll give you two examples let's say a customer is looking for a big deployment in a given geography they might go to four or five different data center companies and say hey what can you do for me well then they all turn around and go to the utility and ask for what the utility can do right so the utility gets five requests that really really just one customer or you know let's say that uh you know you and i wanted to get into business and we went to wyoming and we bought a ranch that was next to uh there's some power lines and then we went to wyoming power and we said hey we'd like two gigawatts please well that's not a real request you we don't you know we're speculative buyers um so there's a lot of speculation or double counting in some of the bigger capacity numbers that you see in terms of the requests the utilities are getting so they're trying to take steps um you know to filter out what's real and what's not or to put conditions on players like well if you want to do this you got to put skin in the game and pay for the study to support the um uh you know to support a potential expansion to your site or you got to pay for the substation you know which is not an insignificant sum so let's finish up um the other companies that [00:51:20] Steve Eisman: you cover any of them that are relevant to this conversation sure so um so we talked about the [00:51:27] Nick Del Deo: data centers that i cover in core weave i i think as it relates to ai some of the fiber operators are you know in the same league because like we said before you got to move data in and out of data centers right if you didn't have fiber going to a data center it's useless useless useless so um at least within my coverage you've got lumen technologies and cogent which are big suppliers of fiber that um you know stand to benefit from these trends if we drill down into each one a little bit so lumen is that like a five dollar stock well it's probably foreign change now a year a year ago i was close a year ago was a year ago was a buck okay so so what why is it a four dollar stock because that's not a good sign yes yes so lumen um they provide fiber optics that's a big part of their business yes so lumen you might remember level three i do yes so lumen uh used to be century link which merged with level three to become lumen okay um but really you know the most interesting part of their business is the old level three network level three was built during the telecom bubble right to build you know uh they built a national long-haul network with a lot of conduits in the ground so condo is basically a pipe through which you can pull fiber and on most of the routes and most of the city pairs they built between 10 and 12 conduits but only used one the idea being that the real cost of a network is putting this stuff in the ground so if they ever needed to expand the derad fiber in the future they could just pull fiber through the empty conduit at a very low cost those extra conduits basically sat unused for two and a half decades two and a half [00:53:03] Steve Eisman: decades this is this is a 90s story i remember yeah level three was late 90s yeah yeah um now now during that [00:53:09] Nick Del Deo: time you know the business wireline space which is the bulk of lumen's business is very competitive and you've seen declining revenues probably you know for the industry at three to four percent lumens declined more a lot of legacy revenue there um highly levered it was a tough situation for lumen with the ai boom suddenly the hyperscalers want a lot of fiber to tie together these facilities because the traffic needs are a lot greater lumen was the best option to deploy uh long-haul fiber quickly because of these conduits and they they very quickly struck deals with you know i think it's a total contract value of eight or nine billion dollars with all four of the hyperscalers to pull fiber through these empty conduits and suddenly lumens financial profile looked a lot more interesting okay and they sell other products that um hyperscalers or other you know ai driven entities are using like wavelengths not sure we're going to get into weeds on that but other products that are interesting so they're arguing that they can take a business that's been declining for a decade and more yeah and and turn it to pivot to growth in the coming years okay and what do you think i'm a little skeptical i think that the competitive dynamics uh are tough and there is a big base of legacy revenue there that will continue to weigh on results look i think growth will improve but not as much as the company articulates cogent is another one coaching is an interesting story we could it's the smallest company i cover but we could spend market cap billion and a half tiny yeah it's tiny um not a four dollar stock is it no no but they've had a very rough year but but to focus on the ai piece a couple years ago they acquired the old sprint network t-mobile okay right so remember the pin drop network right um so one of the first alternative fiber networks built in the u.s built and employed mostly in the in the 80s and the idea there was to take that network which was built for voice and essentially repurpose it to offer wavelengths so again wavelength a point-to-point high capacity transport product it's again popular with hyperscalers and other network service providers and big enterprises but that's again wavelengths as a product has really benefited from ai um so i think cogent is been a little rocky uh a little rocky start but they've effectively converted the sprint network to offer waves and they've come into this market that has historically been dominated by lumen and zeo so they're coming in as a third provider and i think they have a number of interesting advantages that should enable them to grow grow their wavelengths business into a meaningful meaningful ebit done cash flow country do you like that stock i do i do i've been wrong the last couple years but i think it's i think it's pretty interesting okay well this has been great [00:55:47] Steve Eisman: thank you nick that was really really informative and uh hopefully we'll have you back yeah thanks for having me steve thank you this podcast is for informational purposes only and does not constitute investment advice the hosts and guests may hold positions and stocks discussed opinions expressed on their own and not recommendations please do your own due diligence and consult a licensed financial advisor before making any investment decisions