Data Center Leaders on Building AI’s Infrastructure

[00:00:00] Speaker 1: well that was such a great setup by my colleague in Bloomberg intelligence we don't need to speak anymore no it was fantastic I actually do want to say before we kick off our panel the segue was perfect from the prior speaker because he was just saying that if we're polite to these LLM models it consumes more energy and becomes a lot more expensive so we've got a poll that should be coming up in a second do you say thank you to your AI model yes no I haven't used an AI model yet so scan the code and while you do that I'm gonna get started with the panel it's really good to be talking to you Navin I'm gonna start with you maybe just just set the stage for us because of course there's there's so much hype around artificial intelligence and the two things sort of go hand in hand you need the data centers to train a lot of these models how do you see AI impacting data [00:00:59] Navin: center infrastructure in the coming years yeah good question but I think if you take a step back you have to look at the whole demand for cloud storage to start with and and then how AI is having an impact on further increasing it just while we were waiting backstage I was just looking at the history of cloud in 2004 the word cloud storage did not exist smartphone maximum memory was five gigabytes now is 100 gigabytes and just the amount of data storage created by the smartphones that we all have now is 246 billion gigabytes add to that Netflix Amazon Spotify teams whatsapp zoom and you are and the fintech demand and you are at 3 trillion gigabytes add to that artificial intelligence now the the language learning models that you mentioned and that's supposed to be 50% of cloud storage demands in the next few years so from our perspective AI is yes enhancing the demand side but as QIA we've been investing in data centers for the last 10 years what we are focused is not so much on the demand side we are focused more on the capacity side the supply side that do we have the capability and do we have the right partners to deliver these data centers that are required both for hyperscale cloud storage and also for AI and there are [00:02:33] Speaker 1: various challenges around that I'm going to come back to that point Doug I'd like to move on to you you recently announced an expansion of your global data centers division land across pretty much everywhere from what I can tell North America Europe Asia so it sounds like you see very strong demand as well but not [00:02:53] Doug: specific to any region everywhere yeah so we look at everything from a global perspective and we believe that there's going to be growth all across the globe right now we're seeing strong demand signals for the last couple years in the US Europe's been constrained because of power availability some of those same issues exist within Asia and so we have a team of people that have literally scoured the earth to find 10 plus properties a little over two gigawatts of of building load capable in those properties to build out because what we're seeing as a company is our pipeline is swelled over 25 percent year over year we are seeing tremendous offtake from both enterprise and cloud as well as AI and if you look at the stats on AI and the projected growth rates Jensen Wong talks about his belief that cloud that AI excuse me will contribute to over a true it'll be necessary to contribute over a trillion dollars worth of data centers over the next four years just to meet the AI demand and and to put that in perspective it's taken 15 years in the cloud environment to hit that one trillion dollar data center number so it's about 4x what you're looking at for AI so we're extremely bullish on the global platform and we're seeing different areas of the globe that we're investing in for different reasons in the US India it's mostly hyperscale and that's we're seeing most of our absorption most of our offtake within Europe there's a fair amount of retail and enterprise and when then Asia we're seeing a tremendous amount of enterprise and we're starting to see some very large takedowns for AI and cloud and so the the properties we've built we purchased over the last year have been Milan, Frankfurt, Hillsborough, Oregon several in Tokyo we've we've built we purchased all across so we're starting production on those and and frankly on about half of those properties before we even put a shovel in the ground we've already leased the properties out either partially or completely so we're extremely bullish on the outlook of the industry if you just go back a year ago two years ago the CAGR was 13.5 percent the projected CAGR now is 23 percent so you know there's not a lot of industries growing at that rate [00:05:21] Speaker 1: and we're very bullish yeah the numbers are huge though one trillion dollars is a chart up there you know 230 billion dollars uh some of the the big spenders last year that those five that are listed uh Mark let me ask you uh there is a lot of money going into this space some there's skeptics out there they're nervous about their potentially being a capital glut and for all of this to come crashing down are we setting up for another bubble is this going to be another dot-com bubble in the space [00:05:50] Speaker 4: look to contextualize it i think that we're in a very different environment from a demand perspective i think you saw a lot of speculation in the dot-com era where people laid fiber cabling with no customers on the other end of those cables i think what's interesting having been in the sector for 30 years and having watched some of these cycles what is different about this cycle is we don't have a demand problem we have plenty of customers that require compute we certainly don't have a lack of land and we certainly don't have a lack of capital what we do have is a lack of power we can certainly drill down into that but on the capital side it's actually interesting i think most of the allocators globally and i bet you could you could vouch for this i think investors have put a lot of capital to work in data centers particularly in infrastructure funds and a lot of these platforms that they've developed have reached a level of 10 15 20 billion dollars and ultimately lps are saying the same thing which is okay this has been great we've enjoyed riding this ai wave but at some point we need capital back how do we create dpi and how do you recycle capital investors obviously want to invest in great ideas in ai they want to be exposed to investment grade leases but at the end of the day we do have an obligation as a gp to return capital and to create returns so one of the things that i hear most recently in the last sort of six months of fundraising is what are you doing for us in terms of returning capital by the way it was the same thing that happened in cell towers in the late 90s it happened in fiber in the early 2000 it happened dug with cloud computing in the last 10 to 15 years people want that capital back and i think ultimately it's about where you sit in the capital structure and if you think about the progression of how digital infrastructure has been built it started almost as venture capital it migrated to private equity and then ultimately it now sits with infrastructure funds and now there's the next leg that's going to play out which is 90 to 120 billion of stabilized data centers that ultimately has to be in real estate vehicles whether it's a publicly traded s reit or if it's a continuation fund or if it's really insurance capital that's coming into the sector for the first time there is going to be a recycling of capital that is going to happen and we're in that moment right [00:08:08] Speaker 1: now would you say there is a funding gap though given given the needs and given your projections [00:08:14] Speaker 4: i don't think there's a funding gap i mean we we put we put out about 26 billion of capex last year just at digital bridge alone we have another 19 billion that's that's funded this year and i think the way that we're funding those assets today are perhaps a little bit different from how we funded it three years ago i think certainly the financing market has matured we're talking about equity we haven't even begun to talk about debt i think the way these assets are now being deployed in the securitization market and the depth of that securitization market particularly through investment grade is about as deep as you've ever seen it just in the last 30 days alone we had six data center securitizations that got launched and closed across many different platforms across many different gps these if you aggregate all of that securitized capital which is predominantly insurance companies or opportunistic debt investors there was close to almost 12 billion dollars of securitized paper that's traded hands in digital infrastructure in the last 30 days so there's a coming of age of this asset class that i think is really interesting i've gotten to watch it because i started in the early 90s when i was backed by family offices and venture capital and now today we're an infrastructure fund but i think this progression of capital is really important and i think the movement of capital and the maturation of the debt markets for digital infrastructure is something that's really important so we don't have a funding gap as long as capital keeps moving capital has to keep moving [00:09:38] Speaker 1: in this market yeah goodwin one of the big events earlier this year was uh deep seek and for those in the space it wasn't a massive surprise but for the rest of us who are sort of tours tourists and in the worlds of llm it did come as a surprise that china were becoming competitive and foundational llms how does that change the picture for china and for the broader asia pacific region to have china really powering ahead with their own foundational models right i think historically we've always [00:10:11] Speaker 5: seen our china is actually very good at taking what's uh new technology and then adopting it adapting it to create a actually better user experience quite often in this particular case i think there's a lot of excitement in particular whether it's the west apac or china about inference and so when you go back to the supply demand question you also also have to ask yourself because data center you have the old architecture are done first is co-location secondly is cloud and then thirdly the new the new big use for ai is training which actually the black world chips are what 132 kilowatts right demand going to 600 kilowatts of ruden but cloud architecture is only 10 kilowatts right less than 10 kilowatts so you look at that and then you're going to ask you have to ask yourself where is the supply demand imbalance is it too much supply potentially on the hyperscale training site right in china we always the word is east side it's for data west side it's for compute because east side is where it is cooler much cheaper power for example china doesn't have a power constraint issues two percent of the power today is consumed by data center going to four percent by 2030 going to 2030 by 2035 is expected to be 10 percent of the total power and it has a much more modern power grid when you take all that in the equation then you realize in china you're seeing a little bit of how the world actually could play out because china is one big country you have the training being done on the east you have compute being done on the west where it's closer to population so latency matters inference it's actually where it's driving the demand right now so the supply demand imbalance could very well go back to which part is the supply demand imbalance very well we could actually be completely under supply on the cloud architecture side right where [00:11:59] Speaker 1: it drives application address inference where do you see growth uh where do you see the fastest [00:12:04] Speaker 5: growth in the asia pacific region right now um certainly china is playing catch up where china this year they're talking about 50 billion us dollars plus of capex spending so china actually went through a very high cap expanding during covet because of um stay at home in e-commerce adoption and so on and so forth and then went through allow because of the chip war with us so lack of ability to get chips and also the big players couldn't quite figure out is it going to be a winner takes all on llm so not all the biggest company were focused on creating or chasing llm because the business model is less clear on the llm right so you have a bit of a low and deep seek actually seem to have all of a sudden create urgency for everyone to play catch up so the 50 odd billion u.s dollars in capex in 2025 it's a 59 year-on-year increase from last year and then three major players alibaba tencent and byte dance alone uh account each spending over 10 billion u.s dollars actually 100 billion renminb so 13 14 billion us dollars per company and then there are another three that are baidu uh meituan and kwaishow are actually over 10 billion renminb in spending alone in 2025. wow the numbers so it's a big catch-up [00:13:22] Speaker 1: okay let's bring up the results of the poll that i asked in the beginning if anyone is a polite polite user of ai most most of you are polite okay so about two-thirds of the room say thank you to the ai model so this question is relevant because uh ai is is notorious for consuming a lot of power navita i want to come back to you you touched on this uh in the beginning and you're saying you know one of the considerations is the sheer power consumption that goes into powering these data centers um is that [00:13:53] Navin: the only consideration uh that that's certainly one of the main ones that uh mark mentioned um the the the other one is just uh in terms of the supply chain dynamics uh with regards to fitting out a data center has changed dramatically um we are looking at a data center right now which is trying to do an lng powered uh behind the meter uh supply um foreign hyperscale cloud center and they went to buy a gas turbine only three companies in the world that make the gas turbine for data centers siemens ge innova and the waiting list is four years so so even if they do get the power supply and the lamb bank suggested they have to wait for four years before they can kick it off um in another transaction we are involved in the power generators coming from caterpillar there's a nine month waiting list for that right there is also another constraint which is human capital required to deliver these data centers in the right specifications you know there are a lot of new players that have entered the space were traditionally real estate players and are trying to do data centers now i mean we're lucky to be partners with people like digital bridge and go capital who have experience in this uh but finding the right [00:15:13] Speaker 1: capital a human capital to deliver it is a challenge also yeah i'd like to ask you about the heat that is generated by powering these data centers as well because um from what i understand you also need to have effective cooling systems in place is that acting as a constraint over how quickly you can build up [00:15:33] Doug: capacity it's a huge constraint because if if you think about the conversations in data centers they often lead to power which is our number one constraint as an industry to be able to find powered land is extraordinarily difficult and when you look at ai data centers versus standard data centers from my perspective there's really three main areas where they differ first is on scale because a ai data center generally is much larger put that in perspective the data centers that we've built for the last 10 years have been around 36 megawatt buildings the new ones we're building are 100 megawatt which sits sit on gigawatt campuses put that in perspective a hundred megawatt building powers a hundred thousand homes so you're using an extraordinary amount of power so the first issue scale around is power scale and getting that power available the second issue is uh being able to have the capital to scale which mark talked about earlier and then the third issue is around innovation and that innovation is all around liquid to the chip or what's called dlc direct liquid to the chip cooling which basically instead of using the technology we've used for 25 years for data centers which is air cooled to remove the heat because power equates to heat and you have to remove that heat we use direct liquid to the chip which is glycol device attached to the gpu and it removes that heat that's the technology that allows ai data centers if this were a data center hall and a standard data center you'd have 10 racks of ai sitting in the middle of it in all blank space in a today's modern data centers that we're building the entire floor would be filled with racks and so that is really the gating issue that we've looked at for the last three years the claim to flame claim to fame for gdc is that we're the third largest global data center provider um not including the role of excuse me mark but standalone data centers we're the third largest we've got about three billion dollars and excuse me three billion dollars in revenue 3700 employees 150 countries excuse me 150 data centers in 21 countries and our claim to fame is that we have 250 megawatts of production liquid to the chip cooling running in our data centers which outside of the hyperscalers is the greatest amount and i can tell you it's been an incredibly steep learning curve because for 25 years we did it air cooled and now all of a sudden we're using this new technology it's not proven it's not standardized and so that to me is probably the biggest challenge in our industry is creating a standard so that we can replicate these data centers in an efficient standardized manner and [00:18:17] Navin: drive the cost out yeah interesting can i just add something to that which refers to your question earlier about whether there is enough capital available for data centers and i think for the real estate people who might be in the in the audience one of the biggest challenge that real estate investors face when looking at data centers is what happens at the end of the lease so these data centers when we invest in it you know they'll be rented by google or aws for 15 years and and like doug mentioned the technology is continuously changing um the efficiency is changing so so if i spend one billion dollars on building an 85 megawatt data center in the what will happen to it at the end of its life if i buy a hotel in park lane mayfair as real investor i know that 15 years later would be worth more than what i bought it for but the data centers it it's a slightly different calculation slightly different returns so that's where the conflict comes in in terms of your core data centers and it's interesting you say this because we [00:19:21] Speaker 4: we had this debate in the early 2000s around cell towers people said gee uh you know we're really concerned about the residual value of a cell tower and what ended up happening was we found out that the location was really important and the interconnectivity of a mobile network is a part of an ecosystem that's daisy chained to many other cell sites and what we're finding with data centers as you think about the architecture of data gravity and where data goes these data centers are highly interconnected not only through their fiber optic networks but the amount of investment that happens at the core of the network that sits in these very difficult to pick up 50 million dollars of infrastructure and move it across the street and so what we found is again we've been in the sector for over 10 years doug you've been in it longer than i have but we've already gone through our first round of renewals of these leases and what we discovered is churn is less than 100 basis points across our global data center portfolio we have 86 basis points of churn and the other key thing that we found out is the stickiness of these locations rental rates have moved up on average about 22 percent in the renewal now most of what we modeled is we thought rents would go down we modeled three percent churn and we modeled 15 discount on rents and what we found out is that rents were higher and the [00:20:36] Navin: stickiness of these locations was really important that's absolutely location is critical right and i think a lot of the new investors and data centers might not realize the importance of ashman virginia where 70 percent of internet traffic passes by so absolutely right if you want something there 15 years later it would probably be more in demand but but there are a lot of data centers popping up in locations related to ai language learning models where one just has to be cognizant of the fact that the location might not be as critical in 15 years as uh as virginia might be yeah as the mark aptly [00:21:11] Doug: pointed out the guy with the most gray hair up here 25 years into this um i can tell you that we have recycled our capital through our older assets and to mark's point um never had any issues at all it's been very smooth we've we've lost uh you know less than 100 basis points on every asset that we've ever recycled um it's been very productive and it's about how you maintain the infrastructure over time so that you maintain the value of that infrastructure you know navid is very interesting right as a risk [00:21:46] Speaker 5: investor i started investing in at that time it wasn't even called data center when you were doing uh cell tower we were doing converting these office buildings in downtown la and downtown san francisco into telecom exchanges one wilshire when wilshire right those were the telecom exchanges the predecessor of data center back then so we always thought what are the alternative uses of these asset when the lease runs out all these data center get telecom exchanges or data center get obsolete right but they don't get obsolete because it's almost like uh 3g 4g 5g every time the pipe expand there's more content to fill it right in in telecom terms so in these data center even some of these so-called obsolete or legacy data center there seem to be more and more actually users because of the lack of latency so there are more and more applications that are coming up that actually need the lack of [00:22:38] Speaker 4: latency yeah and that interconnection at a data center that he's talking about those legacy data centers are where those fiber optic cables come together in interconnection and so as we watch the building of cloud public cloud the last 13 years as doug said what was interesting is as the cloud developed and as the cloud moved and proliferated to the edge these locations became more important because of that connectivity the same thing is going to happen in inference inference will follow the same ball flight as cloud and cloud computing why 90 percent of ai will sit on your mobile device inference by nature is going to be near the consumer and is going to be near the device and that makes these locations as you start to build these language models and they proliferate into generative ai real estate becomes more important and real estate that's close to the connectivity the customer the [00:23:31] Speaker 1: enterprise and iot devices i just want to ask a question about the power generation the assumption is that i guess most people just assume that it will be generated from renewables somehow is it going to be sufficient can renewables generate enough power based on your projections of how much demand is [00:23:50] Speaker 4: needed how much supply i'll take a swing at this because we've been spending the last two years working on this um i've been saying this um i think you and i talked uh down in brazil about this we don't have a generation problem obviously china's done a very good job with their generation and their transmission the challenge in europe as we saw on display in portugal and spain a couple of weeks ago is that these grids are aging these transmission grids were built after world war ii they never contemplated building transmission infrastructure for ai compute i can guarantee you that was not in the design 60 years ago and and today on a global basis data centers are consuming about 60 gigawatts of power so imagine new york city consumes about five and a half gigawatts per day so think about that that's almost 11x new york cities happening every day in compute going to 300 gigawatts so we know that there's 240 gigawatts of power that has to be created to make ai work that's a big task and we we put that up against a current market construct where essentially the data center industry is turning on about five to six gigawatts per year and over the next five years we're going to have to light up 22 gigawatts per year so you have this massive trade and supply imbalance that's only growing which is the demand is going from six gigawatts to seven to nine to 22 per year and we can only turn on about six gigawatts per year so to your point the traditional path of how we build power is completely insufficient we have to throw everything at the problem it's solar it's wind it's hydro it's as the speaker before said gas building micro grids building grid independent infrastructure is critical to the future of ai and there's not a one single bullet solution we've talked about nuclear power that's six to eight years away we have a real problem which is how do we get the generation capabilities adjacent or near or into the data center and this is the challenge that our industry has to deal with [00:25:55] Speaker 1: over the next two to three years do you want to weigh in on that as well just on the uh sources of power [00:26:01] Doug: sure so um what's interesting is we had a little debate out in the green room about what those sources of power b i'm a big believer in smr's um and i also am a realist so i understand that we're going to have to use gas and coal and other alternative forms until we get there but smr's have become so proven they've been on nuclear ships and and submarines for years they're very very safe and i think that we'd be naive to believe that we're we're going to be able to build the grids out without creating microgrids that feed data centers so i am a big believer and we are working down the path for integrating smr's into our portfolio over time i think they'll be controlled at the microgrid level by the power providers probably not the data center providers but we're getting ready because if you look at the other sources they're diurnal you don't have sun at night the wind doesn't blow 24 hours a day you look at what happened in spain and portugal and the issues that occurred there just recently we've got to be very careful how we approach this issue because today data centers are responsible for two percent of the global power usage it'll be 10 12 over time we don't know what those percentages would be but it's going to increase dramatically and so again to me this is the biggest issue facing our industry we're spending a lot of time on this issue yeah navid i'm just going to end [00:27:20] Speaker 1: with you i'm sure you get approached a lot as an allocator about these types of opportunities how do you distinguish when you're looking at investments uh what who the winners are versus um the other pitches [00:27:33] Navin: that come your way i i think we were lucky that we involved started investing in data centers 10 years ago so we developed quite a few relationships early on but when we when we look at a new transaction today uh the key factor is location and location and data center is all about latency and inference uh so there are certain locations whether it's ashpen virginia london frankfurt amsterdam uh texas where it's just having the powered land bank available to build a data center makes it a transaction attractive enough to underwrite secondly what the other change that has happened is uh it the we always did data centers that uh with good credit quality tenants like amazon google microsoft uh but now they have realized how powerful they are that when they assign a 15-year lease to our data center they're creating value so they it's quite important how the lease is structured how much risk are they passing to us and how how much returns are we generating based on the risk we are taking so it's a combination of factors uh but it's essentially about credit quality of the tenant lease and location okay thank you for the crisp answer [00:28:48] Speaker 1: i'm sure many people will be writing down notes in this room panelists thank you so much that was really thank you so much that was really insightful and thank you for your participation in the poll [00:29:02] Speaker ?: you