HPC at the Edge: Where Data Turns Into Action

Artificial Intelligence is most valuable at the edge where data can empower productivity and efficiency. From healthcare and manufacturing to enterprise and municipal environments, edge locations are where AI applications become real-time decisions, automation, and human capability. Yet, these environments were never designed for the level of demand they’re currently seeing. Traditionally constrained by space, limited flexibility, and built to support lower kW loads, edge infrastructure now faces a new reality of higher density, increased criticality, and always-on expectations.

With the help of our guest Bryce Kleen, Field Engineering Manager (Vertiv), we’ll challenge conventional thinking and explore what it really takes to support AI/HPC beyond the data hall. What breaks first? What must evolve? And, how do you ensure your infrastructure doesn’t become the bottleneck to innovation?

Webinar Originally Aired Live on May 14, 2026

Transcript

SPEAKER_00 0:08

Thank you for downloading DVL's Beyond the Product Podcast. This podcast was originally released as a live webinar and has since been edited to fit this format. To view the original webinar and reference PowerPoint's live, please visit DVLnet.com slash resources slash webinars. Have a question or want to submit feedback, please email us at marketing at dvlnet.com. Thank you. We hope you enjoy the podcast.

SPEAKER_02 0:35

We welcome you to today's episode of the DVL Power Hour. Today we're going to be talking about high performance computing at the edge, where data turns into action. My name is Robert Leek. I'm the director of marketing here at DVL, and welcome. Appreciate you joining us. Really excited to talk about this particular piece in the world of data centers. This world of AI is redefining what's possible and driving massive increases to compute densities, power demands, cooling innovations. Naturally, there's a lot of attention that has gone into the data hall with this conversation. Hyperscale environments, they're considered the bleeding edge of infrastructure design. But it's really only part of the story. AI doesn't deliver value sitting in a centralized data warehouse alone. It delivers value when it interacts with the real world, when it informs decisions, enables automation, enhances human capabilities, and really drives outcomes in real time. And that's where the edge comes in, from healthcare to manufacturing, retail, municipality. I mean, there's really no vertical that I don't think hopes to be impacted by AI. But it's where machine learning turns into human capability, is where this will really be felt. Here's the challenge though. These edge environments were never designed for this type of density. They were built for lighter loads, traditional IT, really lower risk and sometimes less dependency on some of the network stuff. But now, you know, we're asking these facilities, we're going to ask them to support latency-sensitive applications, higher compute loads, and really always on expectations, which are really need to be immune to failure. But the question becomes, how do we evolve? How do we rebuild everything without rebuilding everything from scratch? You know, how do we bring these same principles that we've learned in the data halls with the hyperscalers about resilience and scalability and efficiency and put them into closets and IDFs? Because getting this right isn't just about performance, everyone. You know, it's about longevity, reliability, and ultimately, I think it comes down to whether AI initiatives succeed or really stall out at that point of execution. You know, that's what really what we're here to talk about today. Uh thank you for joining us, you know, because we're going to talk about where AI meets infrastructure and where Edge becomes the enabler of what's next. But before we dive too much into that, I want to give everyone just a little bit more information about who DVL is. Uh, if this is your first power hour, welcome. Uh, we were headquartered and started out in Philadelphia, that Delaware Valley, that's where the DV comes from. About almost 10 years ago at this point, uh, DVL started expanding westward, and we're now covering the entire Rockies from Mexico all the way up to Canada. Uh that being said, we are critical infrastructure experts, really focused on all those pieces to the infrastructure puzzle: power, cooling, controls, even the monitoring, but we're more than that. Uh, when you work with DVL, you really are working with engineers, people who understand what products need to go into the right facilities for the right types of solutions, um, which we can turn into turnkey and manage your entire project as needed. That said, after installation, the work doesn't stop and neither do we. We're happy to be a part of your maintenance and emergency services team for infrastructure and those same dots that you see on the screen where we have locations. So um, that's a little bit about who we are from a from a solutions perspective. We're also a great place to work now, nine times for my Ferris Bueller fans, or maybe I should just say my old colleagues out there, but nine times to for a great place to work. We're really proud of that, um, as well as being an employee-owned company. But with that, uh, I'd like to introduce our guest for the day, uh, Mr. Bryce Clean, field engineering manager with Vertiv. How are you today, sir?

SPEAKER_03 4:07

I'm doing well. Doing well. Thanks, uh, thanks for having me, Robert.

SPEAKER_02 4:10

Well, uh well, thanks for being here. Um, and you know, I know I've gotten to know you here quite a bit in the in the recent past, but for those uh on the line, would you share a little bit about maybe who you are, uh your experience in the world of data centers and and kind of this progression with Verti that you've been recently on?

SPEAKER_03 4:25

Sure. Yeah, so uh Bryce Clean, I'm I manage uh actually two teams here within the Vertive organization. Uh one of the teams is is focused on uh we'll we'll say data center white space products, um, the the field sales engineering team, they work with our LVO representatives, um, our local reps, uh, our RSMs and end users and customers in the field to really deliver any kind of a white space solution that you might uh need technical assistance with. So that's that's one part of what I do. Um the other part of what I do is kind of a newer role. Uh we're we formed a new team called the AI Solution Architect Team. Um and that team is really focused on a lot of the things that we're talking about today and and a little bit beyond that as well. But we help a lot of our uh we call them advanced compute or accelerated compute partners work with customers to develop AI solutions, whether it be a small edge environment, a corporate data center, um, uh presence in a colo facility, or you know, some combination of those. So a lot of uh a lot of exposure to that here recently. Um, a little bit about myself and my history. Um, I'm actually from Lincoln, Nebraska, home of the Geist manufacturing plant where they make the RAC PDUs. Uh, I've been with uh Geist Inverted for about 15 years, just kind of working in this in this same industry. So lots of experience with you know the the what's happening in the rack, um what's happening in the data center white space. And I'll tell you, it's it's it's been quite a journey here the last five years, seeing how things are are evolving and and just the the pace that things are are moving these days. So pretty exciting time to be here.

SPEAKER_02 6:11

Well, yeah, to say the least, and we're glad to have you here with us today. For everyone on the line, I've got some questions to kind of help guide our conversation today. Uh, I think Bryce has a few slides that we may show. Let's dive in. And Bryce, before we get too far into this idea of high performance computing at the edge, let's start with the edge itself, right? Because I mean, listen, I've been in the data center industry now for a while. And it seems like everybody has their own unique definition of what the edge comes down to. But maybe can you start us with just like kind of your perspective as far as what is the edge?

SPEAKER_03 6:43

Yeah, no, that's that that's a great observation. You're you're right. There are a number of different number of different definitions for what that can mean. Um, I I guess what what I and the and and my teams have kind of started to consider the edge is is just really anything that's not a corporate data center or a hyperscale location, um, a big colo or something like that. So, you know, it could most people would think of you know your network closets as the edge. Certainly uh that's a that that's right in the middle of it. Um but then you know, a lot of things that we'll talk about today, um, you know, there's there's a four to six rack deployment um in in a small room or you know, in a um in a makeshift, we'll call it data center room, um, or maybe a modular facility that's that's sort of you know uh on the factory floor or in a warehouse or something like that. So so those are some other definitions of what edge are. Um I don't really think that there's a perfect answer for that, but um that that's kind of what what we uh consider it on our team.

SPEAKER_02 7:50

No, that's that's fair. I don't I agree. I don't think there is a perfect answer while there's about you know there's an answer for every human on the planet. But it's interesting you bring up the the modular aspect because you know prefabrication and and the efficiencies that those bring to construction in general are one thing, but modularity can bring edge environments to pretty much brand new IT locations as related to the edge and kind of how in the history aspect a little bit. How were they traditionally built for you know? I was talking about kind of those traditional compute demands uh in that early part of this podcast, but what would you say is kind of how the the kind of strategies behind how edge environments were originally built?

SPEAKER_03 8:27

Yeah, so I I think when people first considered edge environments, it was um you know, network closets, Wi-Fi hotspots, things that supported primarily your network and and the the corporate data center or your your data center was not necessarily called an edge uh environment. You would just have a you know a small data center in each one of those locations. But you know, as as things became a lot more um centralized and internet became you know such a big thing uh with high speed everywhere, um you were able to kind of pull most of your compute into one central location, and then your remote locations were were uh you know just just collecting data and and sharing data. Um so now I I think it's kind of come full circle to that, but rather than just having those hotspots and and network access points and things like that at the edge, um, we're seeing situations where there's a need for local processing for latency reasons, like you said before, for reduction in bandwidth. I mean, video applications and and uh inspection systems and things like that are huge bandwidth consumers, it it just doesn't make sense to send that across the country or across large distances when you can process it locally. So I think edge has kind of grown. Um, and and it it's kind of up to you as far as how you define it and and and what your business needs.

SPEAKER_02 9:53

Well, let's kind of shift into the high performance compute aspect of this then. You know, as we know, as we you and I discuss, and as I'm sure the people on on who are listening, you know, this whole data hall getting all the attention, it's where high performance computing is has been born and things of that nature. But curious about what you're seeing, and I'm sure we're still in the early stages to some degree, but what are you seeing as far as HPC actually getting deployed?

SPEAKER_03 10:16

I I would say you know, the AI and and this whole revolution, we we've been talking about this for probably three to five years, um, pretty regularly. The idea of having localized AI servers and and uh compute has been there for you know since since the beginning. I had the you know the the privilege to be kind of uh involved with this from the very start. There was a lot of talk um for probably for the first three years, I would say the last year and a half, um, we're we're really starting to see that talk turn into deployments. So I think a lot of that what are the early signals that you're seeing?

SPEAKER_02 10:53

I mean, are you actually are you guys working actual projects or what's what's kind of turning the talk into reality?

SPEAKER_03 11:00

The the the main thing that's turning the talk into reality is just the availability of of a lot of the chips. Um for the first for the first several years, you know, the the uh all of the GPUs uh were pretty much uh allocated to six six to ten companies um and with not much left over for the rest of the world. So now that uh you know that that that manufacturing process, that that uh supply chain has sort of matured, now there's a lot more availability. The price has come down quite a bit, and and I think people are now really starting to look at okay, those projects that we identified before, maybe we can now kick those things off. So I think you know, a combination of that and and plus just people's general comfort and the ease of deploying AI-assisted projects, and everything from coding to building your own solution to deploying it to you know building a chat agent, all those types of things have just become so easy and so quick to deploy that it's making financial sense nowadays for people to really start moving that stuff into place where they were maybe just theorizing it in the past.

SPEAKER_02 12:08

So, from an infrastructure perspective, and we talk about kind of how edge environments were traditionally built. What have been some of the breakpoints you've seen as it relates to readying or preparing an edge location for high performance compute? What's kind of some of the big infrastructure gotchas that are popping up?

SPEAKER_03 12:26

Yeah, so and and this kind of varies by site. I mean, if we're talking about just a simple closet or something like that, that's that's maybe one or two servers, it may not apply, but for more of those you know, medium-sized deployments, there you know, I'll I can probably just separate it into the different categories from a cooling perspective. Uh on the cooling side, we we tend to see the traditional perimeter cooling is not going to get us there. Um, so NRO uh has to become an option. I'm sure a lot of people here you know remember back in the day where rear door heat exchangers were a big thing and then they kind of went away for a long time. Well, now we we can't keep those on the shelves anymore. So rear door heat exchangers uh have have become you know a big filler for that for that need uh when the traditional cooling doesn't have what it takes. Liquid cooling in in the white space for for all of my career, we have done everything we can to keep water um out of the data center. Uh, you know, we're actually installing it and and and having it within of you know a few fractions of an inch from from hundreds of thousands of dollars worth of equipment. So that that's been a challenge um uh for a lot of people, uh, just being able to get that that liquid into the data center white space. Um so there's there's a number of ways that we can address that. Getting a building chiller loop um into the white spaces has been a real challenge, I think, for a lot of people, especially in those smaller data center environments. From a power perspective, you know, I would say fitting all of the equipment in a rack is not the easiest thing in the world, okay, especially if we talk about liquid-cooled environments. So where we used to have two PDUs, two rack PDUs on one side of the rack, and then cable management on the other side, now we also have to find maybe room for maybe more than two rack PDUs and liquid cooling manifolds in a rack. So that presents a big challenge to people, just being able to fit all that equipment in the back of a high density rack. The other challenge that is associated with that is okay, now we we've added all this additional equipment to the rack, all the additional piping, all the additional large cables and cords and things like that. Now we create an airflow problem. Uh so that that's a that's a you know something that they uh have to think about all the time when when we're deploying this type of equipment. That's definitely uh something you have to keep in mind, uh, something you have to be very aware of, and and and just uh neatness and uh you know proper lacing and things like that are just that much more important these days. Um from a rack perspective, uh so one last thing here, uh just just on the racks, with with the the challenges of if you look at a traditional server um and and you open that server up, you're gonna see you know half of that server is airspace, uh an air-cooled server. Now that we have liquid cooled equipment, there's no more room for air. So that for the per cubic inch, that server is much heavier. Um and and so our traditional racks, they're they're kind of at the limits as far as what they can support from a weight perspective. You know, that that's that's all also become uh something that a lot of people don't think about that we really really have to account for.

SPEAKER_02 15:46

To your point, you know, all of those challenges, right? Whether it's the cooling, the airflow, the power management, to your point, even the weight considerations, they do vary site by site, right? Especially that last point, I think on the racks, um, you know, high performance computing racks are doubling, tripling in weight, if I'm not mistaken. So I'm guessing like those raised floors for some of those facilities were a little bit challenged. But would we kind of reel it back into the edge piece of this? I'm curious, and I'm sure, you know, I again the edge can be a million things, but just call it what's either what maybe you've seen and been a part of, or maybe heard kind of in strategy strategic conversations. The cooling aspect specifically, and I know you mentioned the rear door heat exchanger, right? I know that really helps amplify some of the density capabilities of of airflow management in the rack. But what are some of these call it edge locations seeing from a density perspective? Is it are we already jumping up to the 40 and 50 kWs in the racks? Are we jumping higher than that at the edge? Um, because obviously depend, I think that there's a break point in there where liquid cooling absolutely has to be a part of it, if you know, depending and are those types of facilities at the edge able to manage, you know, or I guess easily evolve, you know, anything can be built from a greenfield perspective here, right? But thinking about edge and and retrofitting and things of that nature, are are those challenges being amplified at the edge? I guess is maybe the best way I can ask.

SPEAKER_03 17:10

Yeah, no, no, that that definitely we're we're seeing a lot of, I would say that the the the average deployment in an edge environment for AI uh compute, I think people it's it's very common to see something in that 30 KW or 30, 40 KW. Um, but as soon as I say that, you know, I can I can give you a a dozen examples where you know they're they're looking at 50, 60, 65 or more KW uh worth of compute per rack um in an edge deployment. So um that's kind of at the limit to where um it you know it makes sense to consider both liquid cooling, uh direct-to-chip liquid cooling, um, rear doors, or um maybe closed couple cooling or something like that. So that's kind of that area where where there's uh three different directions you can go. And if you take a step in one direction or another, you you kind of you know you're you're almost committed to that liquid cooling um at the at the chip.

SPEAKER_02 18:10

Are you and the team kind of getting any big lessons learned uh from some of the things that you've been going on and and the work you've been doing?

SPEAKER_03 18:18

Yeah, there's the there there is definitely some lessons learned over the last couple of years. I participate in several of our AI roadshows. So if you're not familiar with those, you know, they're they're uh sort of the the traveling circuit to talk about AI um from inverted perspective. And one of the things that has come up time and time again in our roadshow conversations, um, and I can absolutely back that up from our experience is communication is the key. Yeah, it it we used to build data centers, um, and we'd say, you know, here's your uh the facility guy would say, Mr. IT guy, here's your data center, you know, I can I can basically handle anything you throw at me within reason. And and everything, you know, even though we said we're gonna have high density load, um, you know, there was one or two racks that might be 20, 30 kw or more. Um and everything else ended up in that three to five kw range. Well now, you know, when when we're stretching that limit and we're putting in a dozen cabinets that are 60 kilowatts or higher, we're stretching the limits. We're really, you know, taking it to a point where you can't just build a room and then absorb whatever's the in there. So the IT guy and the facility guy, gal, really have to be you know talking uh from the very get-go when this when this project goes into into scope. You also have to make sure that you're on board with with your MEP contractor, with your your uh you know, your suppliers and so on and so forth. So communication is is really ultra, ultra important. And the days of I'll do my job, you do your job, and we'll stay in our corners, that doesn't really work when we talk about AI deployments.

SPEAKER_02 20:02

So to reiterate, yeah, knocking down those silos um internally seems I've heard that message before, so it's interesting that it's the same thing no matter no matter what part of the environment you're looking at. But I'm sorry, you were going to say more.

SPEAKER_03 20:14

The other thing uh that that we're seeing these days, uh that I I I can't I I can't really stress this enough is that you must plan ahead. So if you're if you're working on a project, there's so much volume being demanded from hyperscale, from NeoCloud, from our big partners that lead times, uh and I'm not saying this just about vertive, I'm not saying it to try to scare people, but it's the absolute truth that lead times are bouncing around, um, you know, doubling overnight in many cases. So if you have a project and you're ready to go, you're in much better shape getting that getting that plan in place, ordering the equipment as soon as possible, because it may not be what you're buying from us that has a long lead time overnight. It may be the switches, it may be the servers, it may be you know any any part of that infrastructure because everything is getting stressed to the max these days. On that same note, build flexibility into your designs. So if you're planning, you know, here's my ideal plan A, and it uses these racks, these PDUs, these servers, um, and then you know, something comes in overnight. Um, we got to have a backup plan. Um, and and that may not be a hundred percent perfect, but maybe it's 95% perfect. You have to be willing to turn on a dime um and and be able to maybe modify your design quickly so that you can get the job.

SPEAKER_02 21:44

All excellent points, Bryce. Thank you so much for sharing the you know, the communication and and knocking down the silos for sure. But I chuckled a little bit when when you mentioned plan ahead, right? And I was like, well, where's this going? Has he what kind of project had they been working on where they could but I get what you're saying. Right, it's not the fact that they have a plan, but there is a high demand for this type of infrastructure across the globe, mind you. Right, we're you know, we're a US centric kind of conversation right here, but Vertive is a global organization. Data centers are increasing not just in the United States, but around the globe as well, and and and all because of high performance compute. So when you talk about those lead times, right, that's not just a Vertive conversation. That is a no matter who your manufacturing is, they they are getting swamped from all types of requests. So it's it's a fair point. And you know, sometimes those lead times are not as sexy as maybe the customer would want them to be, which is that much more of a reason of why, you know, the the the further in advance I think that the planning stages can go, the better. And obviously, you mentioned the flexibility. So thank you for those public service announcements. I think hopefully everyone takes those to heart. And I'll also mention just a quick aside on the on the um vertive AI Roadshow. Uh we don't have one hitting and stopping in the DVL markets this year, but we have participated in the past. And and I would just say that if you're listening and go check out, go do vertive AI Roadshow, Google search will take you right to the page. I think there's 10 stops already planned for 2026. So maybe you could get to it because it's definitely a a day worth spending um and learning, even you know, diving into a lot of different topics. Back to our conversation today, and maybe it's what we were just talking about in regards to the lead times. Where do you see organizations underestimating these challenges? Is it just the power distribution? Is it the cooling capabilities? Is it those lead times, the physical space? I mean, where's where are the end users just kind of underestimating the project?

SPEAKER_03 23:38

Sure. Yeah. So, like you said, I mean, lead times are probably the number one thing. Um, but I think I've I've I've talked that one to death already. There's a couple other things that I could say. One of the things, and and this really kind of varies on on how much AI load you you plan to have. Most people don't know this, but an AI job when it when it processes, it has a very unique load profile. The the load profile uh is we we call it a dynamic load profile, but there there's uh an immediate spike that happens for like a half a second or less than that, where the load goes from 80%, where it sort of just chugs along, to over a hundred percent, uh like 120 or more percent. It's a little bit deceiving because when when you look at the um the rated load on this equipment, uh that they go for the average. So they say 100% is sort of the average of the 80 and 120. And and so when when you're scoping out these projects, it's it's very important that you you make sure that your UPS has the capability to absorb those little spikes that happen, uh those load spikes that happen every once in a while. Not every once in a while, but very periodically throughout the processing cycle. So that's one thing that I think a lot of people will say, okay, you know, I need 34 KW of of power. My UPS has let's say 35 kw. I'm more than, you know, I've got more than enough. Well, that may not be the case. And your UPS may not be able to handle that without having to hit its batteries every half a second, uh, which obviously is not good for many of the UPS battery technologies. So that's one thing. The other thing, um, sort of uh also on the power side of things, is if we get into a liquid-cooled environment uh where we have CDUs or coolant distribution units, there's a lot of talk that we could you know go into detail here. But the CDU is is the pump that circulates all the liquid coolant through the circuit. When that loses power, you lose cooling that instant. And so you really can't afford to do that um in a liquid cool environment, like uh as opposed to an air-cooled environment where you have the thermal mass of the room that can kind of absorb things when you have a blink in power. So the fact that a CDU, which is a mechanical pump, uh, has to be plugged into a UPS, that's something that a lot of people don't think about. Can in some cases be plugged into your normal UPS uh power circuit, in other cases, it has to require a dedicated unit. So uh that that's another thing that kind of gets overlooked, um, sort of not really part of a traditional data center approach.

SPEAKER_02 26:22

I think that's that's a that's an intriguing part too, right? Because we've all been working in this world of infrastructure for a while, but now in like the landscape is changing, especially on the cooling side, there's just more pieces to the puzzle that has been traditionally been required. So no, I thank you very much for pointing that out. Given your background with Geist and the rack level power distribution, how is that specific layer kind of evolving to support higher density applications? Because listen, I'm not a technical guy, I'm a marketing guy, but I know there's RPDUs back in the day, they were they probably weren't hitting 50 to 100 plus types of KW applications. But what's how how are things evolving today?

SPEAKER_03 27:00

Uh on the PDU, there there's some significant changes. Obviously, the the the power that goes through or the amount of power that goes through a rack PDU that that's changed significantly. So, yeah, in in the past we used to uh we used to go through 30 amp, 208 volt units. Uh, that was that was our bread and butter. Now we're looking at 60, 80, 100 plus amp rack PDUs, and and they're kind of bumping up from that 208-volt delta to 277 in some cases, two 240, 415 um in other cases. So the amount of of sheer power that we're trying to pull through a rack PDU, you know, is has gone up significantly. I I I have a slide if you want to share that. I uh I've got actually kind of an old versus a new uh rack PDU that we can we can look at just to kind of contrast how things have changed. If if we look at an older PDU, we we kind of talked about several of these things, but um the voltage is going up, um the the the uh the amperage has gone up dramatically, and in many cases, you know, we have to have more than more than your traditional two rack PDUs in a rack. So to make room or to make that all fit together, sometimes you have a combination of horizontal, maybe you have your two horizontal rack PDUs, and then you have a couple vertical units that are slit into and open our use space. So we're just squeezing stuff in as best we can, uh obviously in a in a very organized way, try trying to make room for more power. So this is what an old uh uh you know traditional PDU used to look like. It would be a five kilowatt unit, 42 outlets, two breakers. If you look at the bottom of the page now, we have you know an AI focused or you know, a PDU that's that's suitable for an AI environment. Um, and and you see in instead of two breakers, we have 18 breakers, 18 receptacles. So it's a one-to-one relationship. And and then you know that's a that's a 34 and a half kilowatt unit. So it's just uh you know seven seven X the amount of power uh that we're we're we're trying to push through that unit. I'll kind of go through a couple of other points that are that are really stressing that world out. So the breakers obviously I just talked about that. The input cables on a rack PDU, uh a 30 amp cord versus a 100 amp cord, a hundred amp cord is about as big as my uh as big as my arm. So trying to route a cord like that, I got small arms, but still um, but uh trying to route four of those cords through uh a rack that's already full of equipment is is a big challenge. Another thing that we're starting to see a lot of um, and and a lot of this is is it has started in the government world, they have started to push out 277, 480 volt power at the rack level. So we've built a whole new line of rack PDUs uh that can uh support that voltage, that increased voltage. That's great because we're we're adding you know to the amount of power that we can uh push through those units. Um the downside to that though is legacy designs don't support 277. So now you have one PDU that you can, or one set of PDUs that you can use for your your power hungry equipment, but now we still have to have something in that rack that can support like an older technology switch or um maybe a rear door heat exchanger or something like that that is is still you know living in the world of 250 volt maximum. So that's a that's a challenge. The last thing that I guess I'll point out here is is that the the final bullet point, and that's redundancy. If I go back to the old days three years ago, we had ever everyone would would call and say, you know, I've I've got 30 pizza box servers that I've gonna install in this rack, um, and there's an A and a B power supply in that server. I need two PDUs. If one side goes down, everything moves over, we have full redundancy inside the rack. Well, a lot of this new AI equipment, we're looking at six power supplies as an example in many of those units. And in order for that unit to run at 100%, we have to have five of the six power supplies that are energized. So you can't just have that world of, okay, I have a hundred percent redundancy on my power power sources. I have to look at, okay, what happens if I do lose half of my power? Is it gonna go into a throttled mode? Is it gonna just black out, or what's gonna happen? So that's something that you have to really, from a design perspective, make sure that you understand exactly how those servers perform when they're losing one side or the other.

SPEAKER_02 31:46

We we know that the power is such a I mean, power and cooling are really what's driving this world of infrastructure with high performance computing. We know usually it's about the power availability and can you get it into the facility and things of that nature. But my goodness, I I kid you not, I've had light bulbs going off my head over my head the whole time you're talking just because you mentioned the CDU and it's important, integral piece as far as that cooling aspect. I always kind of compare that to the heartbeat of that cooling system because that heartbeat has to go and keep that fluid moving. What it seems like is these RPDUs are like the nervous system of of the AI and the HPC infrastructure, right? Because it's one thing to get the power in, but you've got to distribute that power at all the way down to the server and have really specific types of equipment, redundant equipment and backed up equipment that that really is uh is integral just as OCDUs are. So I thank you. I I I really uh you know I feel like sometimes within our worlds there's a lot of pieces that get overlooked and they can be just as critical uh and integral as as the other piece. But thank you. Moving into the applications themselves, because listen, AI and high performance compute goes well beyond Gronk and Chat GPT. It's people wanting to make money or save money and figure out new medicines. But at the edge, what do you see kind of moving in actual applications at the edge that that that you guys have been working on?

SPEAKER_03 33:04

I would say as a general statement, um, and and I I hear this from a lot of customers, a lot of end users, data gravity. And so it was a new term for me. Data gravity is a big thing. And really the the point is if you have all of your, if you have your database, if you have all of your your information stored locally and you want to perform some sort of AI action on that data, the cost to move it off-prem or to a colour to the cloud or something like that is very expensive. So if you have your operations on site, you're gonna really look hard at having your AI deployment on site just so that you don't have to move that equipment. Likewise, you know, if it's in the cloud, um, that's probably where you're gonna stay. Now that's not an absolute, but that's that's one big factor that has really caused people to you know really start to look at having on-prem AI operations. So some examples uh of where we're hearing about this smart cities, traffic patterns, kind of uh, you know, a lot of little in little bits of information that need uh, you know, if you look at how much data you're gathering from the video that's coming from traffic cameras and then detecting traffic patterns and things like that, that's a lot to send up to the cloud, have it processed, and bring it back down. So it makes sense to have something with low to no latency within the city. They they talk a lot about you know the next generation of AI is gonna be robotics. Um, again, robotics have so much local information that they're that they're processing that it makes a lot of sense for for that to be uh for a portion of that, not not necessarily all of it, but for a portion of that information to be computed locally. A few other applications in the medical field, you you brought something up, but medical imaging. I mean, if you look at the amount of information that comes from a CAT scan, you know, from X-rays or from whatever kind of imaging, there's a lot of that that uh has to be processed. Um, and there are a lot of applications that allow you to do that locally where it makes more financial sense. Another industry uh that is incredibly vulnerable to latency um is high frequency trading. So the financial world. Uh latency is is key, and they they will they will build local AI mega server systems right across the street from whether it be the the Board of Trade or the New York Stock Exchange or the NASDAQ, so that they can get that you know small fraction of a second of an advantage over the rest of the world. So that's another good example of of where we're seeing a lot of local AI deployments. And then one last thing, one last that I guess I'll mention is financial. And this is not necessarily because of latency, but security. So there's so much regulation in the financial industry that they won't allow financial organizations, they have to have a lot of their operation internal. And if they again, they sort of that whole going back to the data gravity thing, if their their information, if their data has to be stored on site, they want to process it on site. Those are just a few examples, but that list continues to grow and evolve every week.

SPEAKER_02 36:22

I've got just a dumb guy question here, though. So listen, I'm not haven't really jumped into the driverless cars as of yet. My kids have the generations behind, are you just going to continue to embrace this type of stuff? Are those types of applications, the Waymo's, the other self or whatever the other brands are out there, are they processing high performance computing at the edge as well? Or is that going back to sort of a more centralized data center? Because I feel like of all the latency applications, that's probably one of the biggest ones.

SPEAKER_03 36:48

Yeah, absolutely. No, um, if if you look at like a Waymo car, um, there are a number of GPUs that are that exist on board in a in a Waymo car. Yeah. So they are doing a ton of local processing. So yeah, that's that's kind of like a little data center um on wheels, so to speak. You know, if if we talk about the edge, that's the ultimate edge, right? That's a portable device that's moving around um and and making real-time decisions and then checking in with the cloud periodically. So, yes, that's another great example of where that has to make split-second decisions. Um, and it's gonna, you know, it there's it's not an option for that to go offline and and lose connection when it goes through a tunnel. I mentioned I live in Nebraska. If we go to western Nebraska, the the the number of times that I have one or less bar on my phone is is is pretty pretty regular. So, you know, uh you you you wouldn't want your car to just shut down um and and not be able to do anything. So, yes, that's another great example of where we're seeing that.

SPEAKER_02 37:50

Yeah, you know, it's one thing about the the remote areas of Nebraska, but what about just the dead spots around metros? I mean, I live in a populated suburban part of Denver, and there it's wild how bad cell phone reception is, just in this one little recoloted uh our black hole or dead zone. Um you mentioned robotics. Just a quick look aside, I just saw a story a friend said to me most of the time when we talk AI and high performance compute and robotics kind of leads to the nice Terminator joke. But robotics continue to be more and more real. I've seen some companies talking about like putting these things on a production line, like a cars. I mean, I don't know if that's really armies that will be created, but they are happening and they've now been kicked off of Southwest flights. Uh, if if you go out and I I don't know when the story happened, it was a it was a newscast. I saw it just this morning before our recording. Um apparently a fellow who's in the robotics industry out in Dallas decided to take his little robot, it placed a little higher than knee high, took it onto a Southwest flight, and it freaks people out. And Southwest has now put out a new policy that says your robots cannot ride on airplanes, uh, they cannot be stowed. I don't know what what it is, but it's wild. Go check it out. Back to the edge. We were kind of we're coming into the last quarter hour here of the power hour. So thanks everyone for hanging out. When we talk about the balancing of upgrades, not disrupting operations and and really kind of hitting on these critical environments. Do you see really kind of challenges in deploying or just getting these things activated? Or is it pretty cut and dry once you get the infrastructure there and the solution built?

SPEAKER_03 39:22

No, I I I think there's every data center and every environment is is going to be a bit different. Um, and we do see challenges, especially if if you get into sort of those upper level or elevated AI loads where direct-to-chip cooling is required. Um the idea of getting that liquid to the floor is is not always that easy, um both from a physical challenge and also a financial perspective. I go back to a point I made earlier, be flexible, um, you know, be willing to think outside the box. Um a lot of people are looking at, okay, I've got this, you know, this this new server and it's liquid cooled. What am I going to do? Um, there's a lot of different options that you can uh can explore to maybe shoehorn a liquid-cooled server into an air-cooled environment. So we have liquid-to-air heat exchangers, we have um liquid to refrigerant options. So may not be you know an ideal, you know, uh one megawatt solution, um, but as you're kind of getting used to it, figuring out how do I work with this liquid-cooled environment, um, you know, maybe you have to take one step after another, maybe take a baby step to get started.

SPEAKER_02 40:38

So I think we've talked a lot about these environments and kind of legacy, uh, certainly, and kind of we moved into the high performance computing world. What about future proofing? Where does modularity, we talked, you talked about modular solutions at the top, uh, you know, scalability is always a concern at the data center. Uh, where what's going on there as it relates to edge environments to maybe either make these deployments easier or really more scalable as customers see more demand uh at the at those areas?

SPEAKER_03 41:06

Future proofing, that is a very common question. There's not a perfect answer to that. If you uh if you speak with NVIDIA, um one of their recommendations is everything kind of goes in steps, um, and and it's kind of the 2x factor. Every year a new model comes out and it's twice as uh consumes twice as much power and probably you know produces eight times more results. But it probably the best way to future proof something like this is to just plan on you know your next generation potentially having a rack every other space because you know you're you you have that power load, you have that uh heat load, you could over provision. Um, that's definitely an option. But I think most people would look at that and say, gosh, I'm I'm spending a lot of money uh to build this infrastructure for something that I think is going to happen in two years, whereas opposed to you know kind of pulling out every other cabinet, that still allows you to continue to use that space. Space is not nearly as expensive now in perspective of the whole you know cost of doing the business, the the cost of the space versus the cost of the AI equipment is is much less. So it's a small price to pay, it's not perfect. Um, but I think that's the best way to future proof these these type of uh uh of deployments.

SPEAKER_02 42:29

Well, thank you. And as we talk about the future, you're talking about the old days of three years ago. I'm just curious what you see in that very far out distance of call it three to five years or even beyond. What what do you what's what's the pulse of high performance computing, much less high performance computing um at the edge?

SPEAKER_03 42:47

I'm uh I'm actually in South Carolina where where our modular uh facilities are built. Um and and yesterday that question came up, you know, what what does Iraq look like in the next generation, two years from now, three years from now, five years from now. You know, the it's it's it's hard to say exactly, but what what you're showing here on the screen, this is this is from our friends, from our partners at NVIDIA, this is kind of what they're looking at for the next couple years. This is not what an edge necessarily rack, uh, you know, you you're not seeing necessarily 600 kilowatts of power being consumed in every edge rack. Probably pretty rarely will you ever get to that point today. But if you kind of look at this trend and and say, you know, whatever percentage I'm looking at today, you know, in in two years, we're gonna go from 600 to 1500 kilowatts or 1.5 megawatt potential per rack. If I look at when I first started my career 15 years ago, I remember I worked with a cooling product. Um, it was an airflow management solution. Um, but we were talking to um a data center in Texas who uh they they had a one megawatt feed, and that was a big data. Now we're talking about one and a half times that in a single rack, uh, where where things are are angling right now. It's it's just amazing to think that you know the size of a refrigerator is consuming the way more power than than most cities, um, which is crazy. You know, what what what that showed, what that last slide really talked about is kind of what's on the left-hand side of the page. Training operations, like you said, the GROCs, the the Chad GPTs, Meta, those folks. That's that's kind of the area that they live in. And there's obviously those aren't aren't the only players that are doing that. There's a lot of big organizations that are training their own systems that do require that kind of power. Where we continue to live uh right now, especially in the in the work that most of most of my team is on the left hand, or I'm sorry, the right hand side of the page where we see the you know on the the far right where we have robotics and smart cities and traffic cameras and things like that, where maybe a model uh has already been built, and now we're just doing the inferencing, the actual processing of data in those locations. So there's a lot of new platforms that the server uh manufacturers have come out with. Um, one to keep a couple to keep in mind, I guess, you know, the the Cisco Unified Edge solution. They they say that that would be an ideal solution for every franchise location, whether it be you know a coffee house, a cookie shop, um, maybe a grocery store or whatever, something that that coordinates all your point of sales, your your shopping trends, all that kind of stuff. Because AI is is is getting into every part of our business, not just answering questions that uh for your kids for their final exams and things like that. It's it's being used for just about everything. Um it's pretty crazy. I I've done a lot of research here in the last few months, and this is kind of an overlay to that last slide. So again, on the left-hand side of the page, um, we see the training. I mean, that's again, that's your hyperscalers. That's about 10% of the compute units that we expect, we'll we'll say now and for the next several couple, three years. Everything to the to the right of that uh makes up 90% of what's what's out there. And in particular, if you go, you know, the last two boxes, the small AI clusters, the edge AI clusters, what the enterprises are going to be purchasing, what the the small franchises, the you know, places like that, there's a a lot of volume. Um, maybe not the the you know, the the the the number of flops is way less for that, but the the advantage of having that locally, you know, you're looking at a lot of volume, a lot of work to be done at the itch. So that's kind of um exciting for us. Uh that's kind of the world that we live in. And uh all the attention right now is on the left hand side of the page, but there's a lot of work to be done on the right hand side of the page.

SPEAKER_02 46:54

Absolutely, because even if you just think about this in scale, right? Going back to the first the the two slides ago when you had those super jumps, 200 kw to 1.5 megawatt in a rack. Even if the I mean the edge will experience some of that jump and in and required demand and cap capacity and capabilities, right? Because these edge environments, as we've talked about, is only going to get it's kind of like the internet of things, right? The internet, we were already all subvert submersed in the internet, but somehow we got even more submersed in the internet when it started hooking up to our refrigerators and our TVs and everything else. And I think that's going to be high performance compute at the edge, too, is that you know, these mysterious applications that haven't really kind of surfaced other than within enterprise conversations and whatever the case is, but they are going to be changing the landscape all across the world of IT. And you know, Bryce, I really appreciate you coming on today and and talking about this and giving us a new perspective because again, I do think the edge has been gravely ignored in in this conversation up to this point. And uh I you know I don't know if we'll start trending um on TikTok or anything, but uh I do appreciate you bringing some some new insights to this world. Last thing, and I and and then we'll let everybody go. Just kind of but do you think the edge is the weakest link in AI infrastructure or the biggest opportunity?

SPEAKER_03 48:14

I'm a positive thinker. I think it's uh I think it's the biggest opportunity. Um I'm I'm really excited about it. I I I think there's just so much potential out there. And uh, you know, kind of going back to one of the things I said before, it's amazing the leap that we've made in making it easy to deploy an AI assisted application. Um, so the the more that gets simplified, the more that the common my mom or my grandma can uh install AI, um I think the more we're gonna see opportunities and and and the quicker it's gonna get deployed at the edge. So I think it's a huge opportunity, and I can't wait to see what you know what things are like a year from now when we uh discuss this again.

SPEAKER_02 48:56

Hey, Abe, hey, I'm gonna take you up on that for absolutely, Bryce, uh, because this thing, these things change so quickly. We'll have another conversation about the edge, and it'll be completely different because that's how fast this world is moving. I did want to do one quick thing that if you mentioned flops a moment ago, it's a term, the way that some of the processing is captured. For those out there, if you're unfamiliar with flops, we actually did an episode on the DDL Power Hour. You can find it on Spotify, et cetera, where we had a PhD from down at Sandia National Labs uh come on and kind of talk about a new metric that he and that team developed to really kind of look at the efficiencies of high performance computing. So not just PUE. Uh he and some of those smart guys down there developed the FTUE, uh, which is all built on flops and how many flops you're getting with the energy year. You thought I would take that to give a little promotion for to our past episode. But uh appreciate the mention, Bryce, appreciate the time. Thank you to everyone. I think we are right on the cusp of the hour. So um until next time, I hope uh everyone has a great day. Thank you.

SPEAKER_03 49:53

Thanks. Thanks, everyone.