Vultr Is Scaling AI Cloud for the Next Phase

[MUSIC] Hello and welcome to the Tech to Strupper's podcast, Host by Bloomberg Intelligence. In this podcast series, we speak with sea level companies, executives and management teams about their views and disruption and how striving their decision-making and strategy. Bloomberg Intelligence is Bloomberg's research arm. We have over 2,000 companies covered globally across multiple asset classes and it's backed by Bloomberg and third party data supported by nearly 500 research professionals. My name is Rujan Ho, analyst of Bloomberg Intelligence. We're actually really glad to have back JJ Cardwell CEO, a cloud provider of Olsher. Hey JJ, welcome back to the podcast. Well, thank you for having me. Couldn't be bad. Great. So JJ is back and when we had JJ and Volcher on around mid 2024, the AI cloud segment was still largely new. Quite frankly, I actually had a lot of excitement as well as uncertainty. Now, if we fast forward to today roughly 18 months or two years, the AI cloud segment is not only recognized as a viable segment. It actually has a name called NeoClouds, but one that's rapidly growing with lots of investment into the space to aggressively build out the infrastructure. Now, unlike other NeoClouds, Volcher didn't start off as in the cloud. It was a cloud service provider, but Volcher does come up with a little twist as an AI infrastructure provider. AMD is a key infrastructure partner and it back AMD has an equity stake in the company. Now, that contrasts the Nvidia infrastructure company, AI clouds that are out there and here to give me a deeper dive into the story, I'm going to pass it over to you, JJ. Look, it's been a year since we last, it's been a couple of years since we last spoke. And Volcher feels like it's operating a very different phase today than, than even a year ago. So what's the most important way the company has matured as AI workloads became a central option rather than just a, you know, just a, I guess a workload tool for some people. Because it seems it's very critical now that it was a year ago. Yes, yeah, absolutely. Well, you know, it's amazing. Two years of pass since we, when we had that conversation and so much has changed in that time, you're right. We've gone from a stage of the market where it was clear that there was an enormous amount of potential in, in applied AI. And there was a lot of capital chasing deployments. And there were a lot of new entrants. And in that sense, there was also, whenever there are new entrants and new capital formation, there's also typically some undisciplined things happen. And that certainly happened. What we've seen, though, since that period is obviously enormous market level growth. But we've also seen just what in most markets would probably be a decade worth of maturation at all layers. And when I say that, I mean, a few things. One, you know, at that point, we still had many, many foundation model companies all, all vying for potential market leadership. Since then, we've seen massive, essentially, consolidation of market share. We've seen in some cases, even aqua hires, where high potential heavily funded companies became part of the largest platforms in the world. And then we've seen enormous amount of growth in the unique layers across the stack. We've also seen just a cycle of enterprise adoption emerging more recently. And a lot of that is that the AI products and tools have become easier to use and incorporate your data. I've become more integrated through things like just pure performance and the relevant performance is something like Google Gen and I and then the integration of those capabilities into the applications, the Google applications that many businesses use every day. You start to see massive acceleration and uptake. So it's created this huge growth at the at the large end of the market for some of the most commercially successful applied AI products. And that's a really great foundational dynamic to observe because two years ago, a lot of the growth was very speculative. A lot of pretty revenue effectively R&D stage companies trying to train models and be competitive with compete with OpenAI and compete with Google and compete with better than others. And so that's a really important dynamic. We're going to much more mature and stable and therefore predictable place, even though the growth is still feels in some ways unbridled. God, so look, again, when we spoke to you a couple of years back, in this baseball season, right, pitches and catchers, right? And I almost felt when we spoke, it was pitches and catchers in spring training for baseball. I've heard the analogy that we're not even in the first setting, where we're heading into the baseball stadium in terms of the cycle of AI workloads. I mean, how do you characterize where we are in terms of the build-out cycle, as well as the adoption cycle for AI workloads from your standpoint? Well, I think the way you characterize it is pretty much on its. I'm certainly much closer to the truth than any fear that we're somehow lead in the cycle. And the simplest way to test that is just to ask yourself at an intuitive level. Do you believe the current AI applications that you're using, they were all using, represent the peak of AI innovation, that like text-based chat interactions, or five years from now are going to be state-of-the-art? If you do, then it's a question of how much more volume will there be on the current generation of products, and the reality is it'll be an enormous amount, right, as the adoption gets wider adoption and just the higher volume and frequency of usage. But in reality, that's not going to be the case, right? If you imagine the AI-driven applications that are even three or five years from now, they will be so much more resource-intensive than today's applications, that what is deployed and aggregate around the world will seem very small. And it's often hard people see the absolute volumes and the dollars and 10 figures, tens of billions being deployed by high-versal platforms, and they fear that, you know, must-that, not mean we're headed for an overbill, but it is truly a function of what resource will be needed and what will be the value of the work being done on them. And the unlock that's happening is just enormous. So even in our own lives, as you think about, the resource usage that'll come from just long-running agents, just the difference between you and I entering something, having a session, the runtime is ends when we walk away versus 24/7 running, performing tasks and doing work on our behalf, let alone then a totally different level of work, which would include, if you said, five years from now, and probably three years from now, every single element of this business, every person's work, every interaction will be being, you know, pre-simulated, right? Effectively, think about digital twins as two years ago as like a very conceptual thing, but you said operationally what will that look like in the next couple of few years? It will just be running parallel to your actual operations, simulations, and that'll happen at company scale, right? At economy scale, at ecosystem scale, and those things, the resource intensity is just almost unimaginable. So we're underbuilt, we're like net short capacity, as they would say relative to where we'll need to be over the next couple of years, and the physical constraints of the market, both in terms of, you know, power, space and power, wrap-rate is, center capacity, and, you know, any, even just at the hardware layer, given some of the supply chain constraints that have emerged, there are massive blockers to us being able to keep pace as a market with deployments, as much as they're needed. Like we're not, the market in Adria is not only that overbuilt, it's not keeping pace with, you know, what the innovation should demand. Okay, I do want to circle back to that. I do want to talk about the company itself in terms of what kind of workload you're seeing. If I recall, you did start off as a, you know, a model building, workload type of company, more so, we were waiting for this inference curve. You're more inference driven than anything else. Is that still the case? Absolutely. Since the beginning, we've been inference focused, and part of it is, the history of this business, but we entered the AI infrastructure market in 2021, we started buying Nvidia GPUs at that point. We G8R Cloud GPU platform in early 2022, well ahead of the broad-based market awareness of ChatGPT and the broader AI boom. But before that, the business business was in the cloud computing market, the CPU centric market, and has been since the Vulture brand existed since 2014. And before that, the company had a history as a dedicated hosting company. So a longer history and in infrastructures of service than even the largest hyper scalers in the world. And that that that's remarkable. But that needs part of what we entered the AI infrastructure market with was hundreds of thousands of daily active customers, millions of instances deployed every month, customers in 185 countries around the world, infrastructure deployed in six continents and 16 countries around the world, like a full fledged cloud scale platform. And that that enabled us to have a huge advantage in, you know, in deploying AI infrastructure and meeting the market and the moment there. Yeah, but but let's let's talk a little bit of uptake, right? Because I also at the start most most of the workloads were, I guess, core AI, AI centric type of companies who seem to be doing a lot of the workloads. Has that demographic changed a little bit? Are we are we still more AI centric or have you started to democratize those workloads to, I guess, general type of corporations that are using more AI as as as part of the day to day process? Yeah, great question. Yeah, two, two years ago, the market was almost entirely essentially AI natives and then the largest hyper scale platforms, which were in many cases deploying to support AI natives, right? The market now is is very different. So they're, they're we've been working with inference focused platforms, including some of the best managed inference companies in the world, frankly, probably all of the best managed inference companies in the world work with us in some form or another. And they value the platform capabilities and the distributed edge focus for a plan. But to your point, what we've seen in the last year in particular is a level of uptake on the enterprise side. And the way to think about is two full. You know, there's enterprises starting to integrate AI into their standard workloads. Their infrastructure deployments are not in garnish, right? Because there are many cases implementing, you know, technology capabilities that you that existed, you know, two years ago, right? But they're bringing them into their operations or anywhere where humans representing knowledge on behalf of a company, a call center, a good market organization, they're integrating those technologies where a, you know, where a PiDi can help with fraud or risk detection, they're integrating those. But those are not enormous resource intensive GPU cluster intensive. So the scale of their deployments tend to be somewhat smaller. But we're absolutely just in the last year seeing, you know, real and fortune 500 companies that are not infrastructure companies, you know, buying and that and that's very, you know, that's a great sign to see. You know, we talked to, a touch about this in two years ago, which was, you know, in the markets, we're certainly much closer to it now. You'll, you'll, you'll know that we're underbuilt from an infrastructure standpoint when big public companies on a more frequent basis are announcing increases in their projected long-term earnings, margins increasing by 50 base points, 150 base points because of the realized benefits of AI implemented in their existing workforce. Not because they're selling some new AI based product, but just more efficiency. And that, that's when you just have a cascade across enterprise boardrooms and public boardrooms around the world, it's focusing not just like where my AI aren't the initiatives and are showing me an AI strategy, but where's my 50 basis points of net income margin, where where's my 200 basis points of permanent margin improvement? And that, that's when you'll see the really accelerate enterprise adapters when that's far more common. The other big bucket of enterprise demand is, you know, it's through the existing products, whether it's Google Gemini or Anthropic and you know, and opening eyes enterprise offerings where those are getting pulled into and used by enterprises, really just in the last six months, right, with Gemini's latest model in the apps of performance there and you've just seen a huge uptake. I mean, business leaders saying, this is the performance is at a level that's compelling relative to, you know, other alternatives. And this is integrated from a security standpoint inside of the workspace and the app, set of applications where it uses a business and that concern around security and trust and not having to integrate a new thing, a new integration into stack. That moment has actually been so important and powerful in the market because you have folks saying that I can start operationalizing this in my business without deploying a new thing and that, that is, that's driving an enormous amount of on. So those, those B2B usage, that B2B usage, which is less visible, it's, you know, it doesn't show up in the form of like millions of downloads of an app or installs and developers, you know, reducing a new tool, but quietly the B2B AI volumes are growing enormously and, you know, more peer-played B2B coming as like anthropic or some of the best, best places to see that growth, showing you. God, it's so, so, so let's, let's delve into that a little bit. For the, for the B2B, right, for, for your B2B customers, there needs to be a build versus buy. I'm assuming there's a build versus buy discussion that they do internally before they wind up coming to you. I mean, clearly, the economics have to be in your favor to work with you or is it a supply constraint on their end because they can't get the hardware to build the AI infrastructure in-house. Yeah, it's a great question and, you know, it's a, it's a, it's a, it's a very good question because a lot of people struggle with this idea of like, why, why don't these folks just do it all themselves? You know, why, why does the independent AI infrastructure market even exist at all? Like, why, why don't they, and really, what folks are asking is it's the fundamental question you ask. It's like, why don't they engage in a 100% build strategy? And that, you know, the funny part to me, how many spent a lot of years around infrastructure is, that question's already been answered at other layers of the stack for decades. Like, just go down to the data center layer and say, for the largest deploys of infrastructure, pick all the hyper scale, you know, clouds for Microsoft, for, you know, for Google, for others, for Neta and the largest scale deploys of infrastructure. They've always had a mix of essentially first party and third party infrastructure deployments. There's a set of data centers that they build and own themselves. And then there's some mix of third party. It's why, you know, it's well understood that, you know, Microsoft is, you know, it's public knowledge is when the largest customers of most of the largest data center operators in the world. And that's no surprise. And there, there's value, there's flexibility. There's, to your point, there's speed just time to market of pulling for, of working with third parties. There's, and, you know, there are markets that they just don't want to be and they don't want to build a their own path. They don't want a 2015, 10, 15, 20 plus year lease on, on that specific facility. Maybe it's somewhat subscale. It's still relevant and a needle mover for them, capacity wise, but it's not big enough to be a full scale pop for them. So there's a very long list of reasons why they'll eat it always has been an infrastructure and why there always will be a meaningful mix of third party. So the fun and out question, when people are asking that existential question of, like, should independent infrastructure companies exist at all? I'm pretty laying down infrastructure. But the question should really be, what do you think is the steady state mix? Because is it a 50, 50 mix over time? Like, let's just say, let's just say five to 10 years from now when the technology matures. Right? Because in my mind, I'm with you. I think it's 100% on the infrastructure side just because of the shortages of GPUs and just harder to deploy. But I mean, where do you think the steady state mix may be five, six, seven years from now? I think it's certainly a meaningful portion. Is it 50 or is it a third? I don't know, but intuitively, I won't suggest that I have that exact number of thought, but it's a meaningful portion. And the reason is just there are so many markets where they don't want to be even in five to 10 years is a good way to test it because it's hard to imagine the dynamics of not stabilized by that point versus two, three, four, five, it's harder to, it's easy to see a world where we still have a big supply to the end in balance. So the other piece is even in a steady state environment, there's value for the largest companies and still leveraging third parties and having that be part of their mix because of the ability to modulate, right? Just to say, if you do everything in house, you are staffed for max throughput. And that means it's you know, and yet you can't surge up without adding people who you know you will have to then terminate later as you reduce back down to normalize levels. So just that ability to surge up ends up being very, very valuable and having some mix. And then frankly, even for the biggest technology companies in the world, most would self-identify as being engineering resource constraint, right? Like, that's truly, if they would say that, that's a strain and this is about augmentation. What is important is by two years from now and probably you did somewhat sooner, the landscape will have largely crystallized. So what you see is two years ago, the biggest technology companies in the world were working with a very, very short list of their kind of traditional third party providers. And for the most part, that didn't include independent AI infrastructure companies. Right now, there's never been more openness for large companies, including very conservative and security and compliance-focused large companies to work with the dependents because, you know, obviously because of this applied demand in balance. Because that re-normalizes, you would expect the aperture starts to close a bit again and it will become very difficult for new entrants to come into the infrastructure market. It's very difficult for newer, less proven and sub-scale AI infrastructure companies to enter the approved vendor lists for those largest buyers in the world. And when I say crystallize, I mean, you'll have this handful, you know, if it's five or eight or who knows of large scale publicly traded independent AI infrastructure companies, some of which will be investment grade credits themselves. Right? And that matters a lot for cost-capling, for driving cost-capling, cost-capling is king and any layer of infrastructure. But that dynamic will be, yes, the mixed matters, but it'll also be who benefits from it. The universal who can benefit will be essentially froze at lock, well inside it, whether it's 18 moths or 24-hour ice. God, it's so, I think it's really important to note. And the big takeaway for me on our first conversation was that look, the compliance and auditing standpoint of your infrastructure and how you handle things in your own data center, as well as your global presence, that was a key differentiator. Does that remain a key differentiator? And why you think you'll be one of the few ones standing two through a year from now? Absolutely. You know, we have an unrival commitment to enterprise grade compliance, security trust privacy and that comes apart from operating at cloud scale. Many of the newer entrants in this market only have one or two or three or five or 10 customers at matter. Hundreds of thousands of users, you know, hammering on the platform every single day around the world. You don't really know if your platform is secure, yet you have five customers and you're not really an internet-facing business. So, as an infrastructure company, so what will happen is we'll see places where platforms are tested and failed. We're also seeing what we're seeing is. Right now, there's a tolerance for infrastructure providers that don't have technology platforms to, but do have access to power to just deliver single tenant, unmanaged, bare metal clusters essentially, deploy the, uh, uh, reference chip company reference architecture at a cluster level and then just kind of hand over the keys and, and there's a small set of buyers who are so sophisticated in their consumption infrastructure that they're absolutely, it's they're frankly their preference from a security standpoint to not have anyone in the middle. But over time, that becomes a massive, massive exposure for, you know, competitive deficiency for those businesses because so many large-scale buyers, they want to consume these as essentially as virtual machines, even though they're full system virtual machines, they want to consume through a cloud platform they want and need all of the ancillary services that it's like, you have a production AI service like it all doesn't need just GPU, uh, centric servers, it needs traditional cloud compute, uh, instances, it needs storage, all the other things that a full-fledged cloud infrastructure coming up to deliver. So what we will see is there are companies that are benefiting today, um, because, you know, they have access to power and that is a powered land and, and, and data center capacity and that's a scarce resource. And that means they can participate in the market. But the scary part will be as you get out a few years, um, in the first contract roll-off is what do they do with the capacity? Because if they don't have a technology platform, if they don't have a customer base, they don't have, uh, a mature go-to-market capability, how do you, how do you ensure you can monetize that gear in the out years versus us? I mean, that's the essence what this platform is is, you know, it's a software platform that maximizes, uh, the realization of yield from infrastructure assets and to do that over its full, full physical life. And that's, we've done that for over two decades. Yeah. So, so let's switch to careers because this is actually a very good segue in terms of monolithic, um, reference to architectures. You chose not to go that route, right? Uh, I know that you have a video, um, in video infrastructure, but you also like, and, in the lead in, I, I, I noted that you guys also have, uh, AMD, uh, AI infrastructure. Um, I used to cover Juniper before they were acquired by HPE, I mean, they have you as a key customer, right? So, so I'm just curious, uh, why, why AMD and, uh, why, and when do customers choose AMD infrastructure over in the, uh, Nvidia infrastructure? Great question. So, um, first off, we support both platforms. Um, we, we work very closely with both, uh, Nvidia and AMD. Um, you write that for most of this small kind of independent AI infrastructure companies, they, they, they essentially only have one product, but a lot of that's really around platform maturity because at the other end of the spectrum, right, the high per scale platforms, might have soft Google, uh, AWS, you know, Oracle, every single one of them sells both, uh, Nvidia and AMD based products and they end also many sell their own implementations of arm based offerings in the, the cloud compute work. So the, you know, the point is, yeah, in some ways, the anomaly is not selling both the anomaly is only selling one. It's an indication of just a lack of platform maturity because, you know, the, the, uh, you know, it did, it has much higher market share. The best, you know, the significant majority of demand, you know, continues to be, uh, the gear, there's a set of customers, though for whom, you know, the specific abilities and the, and the price of performance of the AMD gear is very compelling. So like, why do we offer, uh, both, well, first of all, as a company that's been around for 24 years, we are customer centric, first and foremost, right? Everything else is, is like, we want to meet the needs of our customers. Uh, and there, there's a subset of customers who, you know, who value different things. So we, you know, we, on the AMD side, we particularly see a lot of, uh, uptake for inference focused users who, you know, who value, you know, the large memory footprint and, and, you know, and the price to perform it to the AMD year. And on the inside, we continue to see the most training centric users tend to, uh, absolutely, you know, be more focused there. Uh, and they're, they're marriage to both and both, we see, you know, a lot of growth on both, both segments, but, you know, it's really again, it's our, our, we're, we're, because of how long we've been in business in the scale, the platform were more like the, um, the more mature companies in, in the market, more or less like a company that sells one product and just sells it is unmanaged, their metal. I think, yeah, of course, they only do one thing because they don't, they don't really have a technology platform and they don't really have a large customer base. They have one or two or three contracts that are constructively the entire business. So, yeah. So, so I suspect if we have the same conversation two years from now, the infrastructure conversation will matter because, uh, it's going to be, uh, predominantly inference. And if, if it's going to be the case, and videos, I guess differentiation somewhat diminishes, I think time, time will tell. I mean, they, you know, they, they're both very focused on that part of the market and, you know, they're, they're very compelling economics for, um, you know, certain types of inference workloads on, on maybe a year and, you know, even with the RAC scale systems, which many people equate with, you know, training focus, you know, those deliver some of the, the most impressive, um, you have price performance on, uh, and, and just absolute performance on your friends. So, you know, well, I think we'll, we'll see the overall market is growing so much that, like the relative mix of market share between the two is, isn't, that's not like a big conflict zone, right? You'd see more, uh, friction at the margin in markets that are more mature and whether, there's less organic, or if there's so much organic growth that, you know, there's, there's plenty of room for both of them. And obviously there are altered platforms, non, non, uh, GB based architectures that were, you know, we're seeing, you know, interesting traction with, we saw the, uh, GROC acquisition and we've seen the public news around, uh, Sree Bros and, and others. So, you know, that there, there's room for, for other, other solutions to the market. And then the last piece is, um, you know, we're, there, there are many workloads that people think of as a part of kind of the AI factory set of use cases that are, you know, best run even on CPU. And, you know, and you've seen in the recent news, you know, some of the videos increased focus on, uh, delivering their, uh, CPU, uh, offering the, the VR CPU on a standalone base. So, you know, they're certainly focused on that and obviously MD has an extremely strong calling card on the CPU side. So, you know, the market is getting, we have to be more and more and more efficient about where we stick these workloads. What's the most efficient way to execute these? And then increasingly as we move into a world of distributed inference and disaggregated inference, you know, in some ways platforms will want and need to abstract away from the specific GPU, even after all in on one company, even if it's just all in video, they'll want for their end user sake to be able to abstract away from the generations of products so that they can add to the same work going across an A100 or an H100 or a B200 or a GB300. And, you know, the economics are somewhat different, but, does end users shouldn't have to worry about the differences between those, right? You know, the best luck inference focus, the infrastructure platforms will increasingly make the differences between the generations are hard where less visible when less of a headache, frankly, for end users, that'll be the true value that it has to happen, given the nicks of generations of product that are getting built up over time in everyone is any of these, you know, heterogeneous fleets. So, let's see, about 15, 20 minutes ago, you said you were underbilled, quote unquote, a net short capacity, right? And, clearly, there's multiple bottlenecks in the infrastructure side, whether it's construction, land, power. I don't think GPUs as much as a problem as it was two years ago, but things have started to ease, but it's still tight. I mean, what are some of the bottlenecks that you're starting to face today that you didn't worry about two years ago, and you think that the audience should be wary of as part of this conversation? Yeah, it's an important question. There's those things, those dynamic shift in your ride. It's the GPUs themselves are less of a gating item. And it makes sense. You think about how much growth and potential there is in the market. You know, there's plenty of economic incentive for that layer of the market to have figured out, you know, the GPU-specific supply chain constraint. Those are things that are largely controllable by the chip companies. Now, as you get outside of that to whether they're not control where third parties make decisions about capital allocation and risk dating, that's where you end up with the constraints because those things are disconnected and not centrally controlled. So number one is back ready data center capacity. So, and if you said, "Would the things that go into that while is land the shortage?" No, land's not it. If you said power, yes, absolutely, in particular grid power. So, and then built AI-optimized data centers. My data centers, I can, you know, not a data center that has a powered land that has an equine mining facility. You can't deploy AI, you're in that. You could build an AI data center, X to it, and then switch the power over, and that's certainly playing your half doing that. But the really important things are, if you said, "Where's the constraints there?" Grid power is a huge one. And we are seeing right now in the markets, not that focused on it, but we're seeing a shift in acceptance of non-grid power sources. That's the acceptance of it, and that takes happened remarkably fast over the last, just with really been the last 12 months. And what I mean, what I'm saying is, behind the meter power, power generation, on-site power generation, using natural gas, it's not, it's removing this dependency on grid power. So we're, now, the key, one of the key things there will be, anytime you have a shift like that, people know what it means from an SLA, and just a reliability standpoint, to consume grid power, or redundant grid power. The market doesn't know as well what it means to be consuming on-site generation, run by, perhaps, a data center operator who's a newer operator, and maybe has never run power generation and is operating. In some cases, that's being managed by third parties, but this is a layer that will be really important for people to understand, from an investor perspective, there'll be opportunity there, there'll be companies that focus on delivering that layer attorney, and obviously, there are some that already exist, and that'll be high value, because the buyers, the leasers of data center capacity will want to know, okay, I'm not getting redundant grid power, but I still need to know that this power is just as reliable as it would have been, because the SLA's, that they're signing up for it, whether they're end users, or no different. So anytime something's changing below the surface, but then the frontline SLA's are not shifting, there's a potential for disconnect, so that's something we'll see play out here over the next year, and it's under-appreciated. The other huge one is on this massive increase in pricing in the memory market, and NVMe, and the drive prices are being impacted as well, but the memory in particular, and the dynamic there is, that's not that well understood, people see that stocks like micron have run up a lot, so they understand, oh, memory companies are benefiting, but if you said what's truly happening here is, the same inputs that go into standard RAM, which we see in every non-GPU server, and in our laptops and everything else, that is the same input that goes into creating the HBM memory that goes into GPUs. So in a world where it takes at least a year and a half to bring up new memory manufacturing capability, which means we are at least a year and a half and more realistically, probably into 2028 before we will see relief in the memory, what's effectively in memory shortage right now, and it's happening from two things. There's a demand side impact, where obviously there's a lot of growth in infrastructure deployments, and every GPU has both the HBM memory, and that's the part of the GPU, and then standard CPU memory, typically a standard RAM that you see in a CPU centric node, and often it's three terabytes of RAM in just a reference architecture, single node of a product like an HGXB200, or an AMD, and my 355X. So if you think about what's happening, those memory manufacturers are actually removing production capacity. In a world where demand is way up, and that alone would push prices up, the manufacturers are removing capacity for building standard RAM and they're reallocating it. So if you think about you have this twin impact of high ultra-high growth, a long lead time to bring up your capacity, and then supply actually being taken out of the market, that is resulting in a squeeze on memory prices that the market does not fully understand yet, and that is going to be one of something that gets a lot of attention at a market level over the next couple quarters, because eventually that has to, that's by getting in stock price in memory companies, you will see, what is that? Follow that through the rest of the chain and say, where does that show up? So that's obviously the memory gets bought by the OEMs who build into servers, and they're passing through that price hike with while preserving margin, which means system costs are going up dramatically. The market is still digesting, racked capacity in many cases in the inventory. Enterprises are stretching the refresh cycles to avoid having to buy as much, right? They'll spend their budget to fix dollars of budget. They won't buy more or less, but they'll consume more or fewer dollars, but they'll end up buying less infrastructure, because system costs are so much higher. And that issue has not yet flowed through to frontline prices for buyers of infrastructure service, and buyers of GPUs and service heads. So there will be a reckoning here in that it will squeeze many business models, for business models, others that are kind of hard-based, where they're reselling GPU hours, and they're not adding a lot of value. They're this memory issue combined with the memory pricing issue that's driven by the shortage of combined with very scarce data and capacity through 26 and 27. We're going to see a really-- You're going to see how underbuilt we are, because we won't be able to, as an industry, deliver the amount of capacity. Okay, so you're describing this perfect store, right? Under capacity of AI infrastructure, right? That's one. Higher cost to you, which you're not going to eat the margin, you're going to try to pass through. If I look at the GPU rental prices, after the first few weeks of initial GPU launch, pricing comes down roughly about 40%, 50%. It almost-- The December data that I've seen pricing has gone up, sequentially in December to January, could we see the age rental pricing, even the A rental pricing, or even the LS40 rental pricing go up because the man is so high, and there's just no workload, or that's available. Yes, we can see all those things. We can see prices go up on a older year, and I will tell you we've seen utilization go to the highest level on older gear that it's been at in years. And the-- the CEO AWS publicly said they're sold out essentially of a 100. So if you said, what happens after, what happens immediately after being sold out on a unit basis? Well, price hikes, right? So now will they increase price or not? I'm not suggesting that I'm not hypothesizing on that, but I'm just saying from the market pressure standpoint, I just like demand if the unit cap the units are sold out on older gear, there is price pressure. So that is one of the first places buyers will go is, well, I can kind of, I don't want to pay a brand new, you know, GB300 or a brand new, you know, AMD, am I $455,000 when it comes out? You, the prices for those will be higher, they will. And you use the prices GPU per dollar per GPU per hour will be higher for new deployments. So then if that gets things that get brought up of the next couple quarters, then they were over the last couple quarters, it is just undeniable. Data center prices are up significantly, base rents are up and then system costs are up significantly, and that has to flow through to end users. OEMs 100.0%, will not and do not need to consume that. So like as a pass it through while also adding their standard margin, right? And then infrastructure service companies, you know, will make decisions. In the short run, they might decide to eat more or less as they compete with others who are still selling down capacity, they bought with prior basis. But at some point all that's consumed as you get out, whether it's one quarter or two quarters, sometime inside of that window, you will see front line prices. And it'll show up as just saying what did a pick current generation year, a GB 300, what's the market standard price per GPU per hour, there has to be upward pressure on that given the other place will then show up as if you said, under pressure to maintain a price per hour, you'll see before that goes up, you will see the non-price terms get enhanced. Contract durations will go longer, right? Companies used to only sign one year contracts or two year with the infrastructure service provider who's taking a lot of risk, that gets extended significantly to ensure more reasonable payback. And then prepinements will go up. And those two things are already happening. That's the first way before you see hourly price go up, you see other dimensions that drive TCDE and RLI, you can pull a couple of other levels as an empire to fend off the hourly rate. So, but that's good. This is a healthier place for the infrastructure market to be is more versus where we were a year ago, where you had more undisciplined, you know, in frankly, equity financed companies in the infrastructure market where they're like entirely venture equity finance to make very undisciplined decisions, because in infrastructure, ultimately 80, 90% of everything needs to be financed or more with that capital. And that creates a constructive floor on the price in terms. And those who were burning venture equity only early are able to operate outside of the rules of gravity for a short period of time and they tend to do on discipline things. So that's kind of burned off quite a, that really burned off in 24 and 25. And now we're into a world where even those companies as they mature the incremental dollars are being deployed much more responsibly. And that that's ultimately a good thing for the overall market is more realistic and sustainable pricing and deal terms. - Yep, you know, this is really good segue. In terms of getting capital to build up the infrastructure 'cause it's not cheap, right? I mean, the numbers can be staggering anywhere from $10 billion for a one gigawatt to $50 billion to a one gigawatt data center. I have to imagine given the improving economics on your end, you've already done a couple of rounds, if I recall, gaining more capital has been a little bit, has been an easier conversation than it was let's just say when we first spoke two years ago. - I think so, the market, the capital markets, understanding of the economics is meaningfully advanced. Right? First two years ago, we were still at a point where many investors were still afraid to take any risk. And those who were taking risk really only wanted to take senior lander essentially risk, not take equity risk. And that's an important dynamic to just contextualize about where the market was. If you said there's still never been a majority, you know, financial sponsor and led majority deal in this sector. Nobody's been willing to take like control risk, right? The equity deals have gotten done have all been preferred stock, high up in the stock. You're to the debt, but above everything else. And that's been a, essentially a way to avoid having to underwrite what do the economics look like in your four or five or six or seven or eight after initial contracts? It's just been a constructively here's a loan that'll be paid back before I know if five is right or wrong about the future. That's part of why public markets have, have had so much pull is, you know, the public markets are willing to underwrite that. And they favor growth enough that their folks or investors are willing to take that risk in a way that, you know, private investors have it yet again. And that's important maturation. When we see the first control transaction happen a majority equity deal, you'll know that the equity capital markets are maturing and the private equity and the private capital markets are getting to a point of confidence and conviction around this market. But it'll be around this small, again, this whole favor consolidation and scale and operators with very high credibility because the dollars, like you said, are so large. And so that's in poor part. We'll see a couple of ideas likely here in the next year and the more public companies are in the sector, the better market will understand economics and become more comfortable than they even are right now with the overall sector. - Well, last couple of questions, one, are you gonna be one of those public companies? And two, has the capital markets finally wrapped their head or wrapped their arm around a five to six year depreciation cycle for GPUs? - So, you know, first on the depreciation cycle, like the short answer is yes, you know, like the market standard at this point is, you know, folks are depreciating these over six year cycle. You know, the interesting thing is, you know, what we've seen set aside AI infrastructure wherever possible when we can look to the cloud computing market, we can see a lot of what the future looks like 'cause that's a market that's fully matured before AI infrastructure even started boom. If you rewind over the last five, five, six years, you've seen progressive extensions in the estimated useful lives. Like from three year depreciation schedules by the hyper scaler, to four to five to six. And it's not, you know, it's not like accounting games. It's actually anything the short cycles or sandbagging and then as you approach reality, and at some point we get to it. As you approach reality, yes, it has the effect of releasing more dollars into earnings without producing any cash impact and that's, you know, viewed as a net good. But it's a sandbagging from a kind of a finance standpoint. But even six years will probably frankly still somewhat conservative and we've seen in our own business. - Really? - Yes. I mean, the reality is we've seen in our own platform over the decades useful life on servers, pre-AI infrastructure that's been longer, meaningfully longer than even that. And if you think about there's, you know, solid state drives and you know, standard server running in a data center environment, it's like the ideal environment, no moving parts, right? It's not like you know, spinning disk. So useful life will be longer. Now will the GPU element of those systems still monetize at a rate where like it's the power of consumes relative to what it can output? Will that still be compelling? That's a different business question than like is there still physical life in the system? And those companies, the few companies like us that are all have a long history and cloud computing, it will be able to monetize hardware in many many ways. As VMs, it's more instances fractalized plans, you know, and monetize the other aspects of the hardware. So that's, I mean, say yes, the market's comfortable with six. We'll know it'll take years before people get more comfortable and we'll frankly have to get to the end of the first way, the big contracts that the markets will wear out and see that you're still monetizing, probably before and for folks who looks is starting to accept something longer, but I wouldn't be, I would not be at all surprised if you see that go out to seven, you know, and once we get out there, when we get to your seven of those first common contracts, right, we're seven from the beginning of the year deploy, I guarantee you, the useful life will get extended if and when that year's still useful. And just keep in mind, because cloud computing was born out of the balance sheets of big diversified tech companies, right? There was not a single standalone pure play, publicly traded cloud computing company, like a scale until the last, you know, handful of years. The markets never really been, you know, had the equity research scrutinizing the economics in the way that you would if you had pure play companies, right? Google and Microsoft, Amazon. the cloud business are all embedded inside a giant, giant, diversified company. So you haven't had the scrutiny you would if these were like data centers and economics were all on display with many public companies. So do your other question-- does that make sense on that side? That makes complete sense. And just curious about your IPO intentions. Because you're running on a good clip. Yeah. Whatever it is, all the standard disclosures apply. We're evaluating all alternatives, no decisions, no final decisions have been made. For our company, the scale, the IPO, could be a very natural next step. The business has been public scale for, frankly, for years. And it's a pride to leave a company. And that'd be because the business was bootstrapped for until a year and a half ago. It just never really needed outside equity capital. The reasons for a company like us to go public despite that scale and that independence is-- it goes to financing, and it's a branding event. And market awareness has value. And this is a market where cost of capital matters a lot, just like in data centers and fiber and towers. And every layer of the infrastructure stack costs a capital is very important. So being public, having access to capital at scale, form capital quickly, do it in a efficient way, and ultimately be an investment-grade company yourself will be very important. And so there's all be very, very strong reasons to do it. And but as I said, final decisions around that sort of thing are still being evaluated. Well, if and when it happens, I promise be I will be there. You're having that conversation, hopefully. And that was my last question. I'm just curious to see if any parting words for the audience. I think I recognize for investors and for business leaders across the market, and people are routinely struggling, we're trying to understand what, how are they being too conservative or too aggressive. And in general, almost to a company, I haven't seen a business yet that's fully utilizing a plighting eye to the full extent of what they should be or could be, we're so early in the adoption curve. So given the massive demand, I don't have any reason to try to hype up demand. Their ability to sell is probably 10 to 100 X, the capacity that we have available to sell right now. So my point is mostly wherever people think we are in this cycle, it's where earlier to your point or early on. And a lot of that is just the folks who think perhaps were further along should the fundamental question should be, do I really believe that today's technology is like, that's going to be state of the art five years from now. If not, what do I, and if they start on the get their head around that, it's like, go and ask we'll chat to you here, Gemini, say, well, you know, will be the five most resource intensive AI work loads that are likely to exist with lives better not to find years from now and how well their infrastructure needs compared to today's and like, tell me why they'll be useful for me and for my business and why those will drive why widespread adoption. And then it's it's it's it's quite quite easy to understand. So as always, really good talking to you. Appreciate the very type of questions and look forward to talking again, hopefully before two years from now. Perfect. Thanks again for coming on, JJ. And thank you, everyone for joining us. We have a great lineup of future instructors in order to culture and JJ Cardwell. So hit the subscribe button to keep up to date with tech instructors, podcasts, and not miss an episode. With that, we'll wrap.