AI Speed Starts in the Data Center

AI isn’t slowing down. Your data center is.

As enterprise AI moves into real workloads, physical limits show up fast. Power. Cooling. Space. Timing. The conversation shifts quickly from what model to use to whether you can run it at all.

In this episode, Schneider Electric CTO Jim Simonelli and WWT’s Chris Campbell talk through what happens when AI infrastructure, data center design and workload placement collide in production.

Because “speed to first token” isn’t just about model performance. It’s about whether your environment can keep up.

More about this week's guests:

Chris Campbell is responsible for the strategy, development, and delivery of WWT's AIaaS/GPUaaS and Facilities & Infrastructure solutions, and their associated GTM to our clients. Campbell has also served as Sr. Director - Executive Engagement and Customer Advocacy at WWT, and was responsible for the strategy, development, and delivery of WWT's global customer executive engagement programs, and has also served as a Sr. Director of Engineering and Architecture - Mid America at WWT, where he was responsible for managing WWT's team of consulting systems engineers and architects in Mid America.

Jim Simonelli is a Schneider Electric leader focused on the infrastructure realities shaping enterprise AI. He brings a practical view of what it takes to make AI deployable at scale: reliable power, liquid cooling, flexible data center design, and the operational certainty leaders need before turning on compute. He speaks from both sides of the equation, as someone helping Schneider use AI while also helping enterprises build the environments required to support it. His perspective is relevant as organizations weigh cloud, hybrid and on-prem choices and try to move faster without locking into the wrong infrastructure decisions.

The AI Proving Ground Podcast leverages the deep AI technical and business expertise from within World Wide Technology's one-of-a-kind AI Proving Ground, which provides unrivaled access to the world's leading AI technologies. This unique lab environment accelerates your ability to learn about, test, train and implement AI solutions. 

Learn more about WWT's AI Proving Ground.

The AI Proving Ground is a composable lab environment that features the latest high-performance infrastructure and reference architectures from the world's leading AI companies, such as NVIDIA, Cisco, Dell, F5, AMD, Intel and others.

Developed within our Advanced Technology Center (ATC), this one-of-a-kind lab environment empowers IT teams to evaluate and test AI infrastructure, software and solutions for efficacy, scalability and flexibility — all under one roof. The AI Proving Ground provides visibility into data flows across the entire development pipeline, enabling more informed decision-making while safeguarding production environments. 

Transcript

AI Is Hitting Infrastructure Limits

SPEAKER_00 0:00

Enterprise AI is accelerating faster than most data centers can adapt, and that's why speed to first token has become an infrastructure problem. Because right now the gap between ambition and deployment is measured in megawatts, cooling capacity, and speed to outcome. From Worldwide Technology, this is the AI Proven Ground Podcast. This episode was recorded live in NVIDIA GTC, where one message seemed to be getting louder and louder. Enterprise AI is reshaping the data center faster than most organizations can plan for it. So on today's show, we're joined by Jim Simonelli, CTO of Schneider Electric's Secure Power Division, and WWT Senior Director of AI Solutions, Chris Campbell, to talk through the real constraints and real decisions leaders face now. Things like power delivery, liquid cooling, digital twins, infrastructure flexibility, and how to reduce deployment risk without slowing down the business. Our conversation starts with a very real question. Is the industry's infrastructure ready for what comes next? That's the work in front of leaders right now, and that's where we'll jump in. Okay, Jim, thanks for stopping by the booth here at the AI Broomy Ground Podcast. How are you today?

SPEAKER_02 1:13

Great. I'm doing great. It's a fantastic event, and good to get a chance to speak with you as well.

SPEAKER_00 1:17

Absolutely. Chris, as always, good to see you. Yeah, nice to see you as well. Yeah, I'm fantastic to be here. So much to talk about here today. I mean, the conference is just so focused on, you know, on infrastructure, agents, inference. I'm wondering, you know, Jim, we can start with you. What are you seeing that's kind of you know the evolving narrative as it relates to enterprise AI and how it you know affects what what you're doing with Schneider?

SPEAKER_02 1:38

Yeah, so a few things here. So one is you think about enterprise AI, it really is like I'm living and breathing it on both sides of the wire here. So Sneider is an enterprise. We are using AI for many different factors as a design enhancement for our engineers, uh for our call centers, to be more effective and efficient as authoring and translation. The classic use cases, but it's it's kind of happening on steroids right now. So as a consumer of AI, and it also is changing our ability to kind of how do we deploy AI, how much cloud, how much hybrid is on property. So as we see a bit of a pivot or an index towards specialation, specialization on inferencing AI, you're starting to see a bit of that inflection point becoming more real in the enterprise side. Now I see the other side of it, which is also now how do we deploy the Armframe or hybrid or even cloud data to deliver all those tokens as well? So I, you know, it's interesting. Uh for the probably the first time I'm seeing growth vectors on both sides. How do we consume AI as a company to be more competitive, more effective in the market? But also how do we use AI or how do we make AI available for enterprises as well? So you can kind of see a mix here because you get gentic, you've got inferencing, you got the new stuff that's dealing with hyperscalers, but also a lot in terms of how does it get used at uh by enterprises like Schneider and many others in the space.

Demand Is Outpacing Supply

SPEAKER_00 3:01

Yeah. I mean, and Chris, I'll build on what he's saying here. How what how does that change the equation or how is that reshaping the narrative around data center builds or just architecture in general?

SPEAKER_01 3:09

Yeah, you know, it's interesting. In Jesser talked about the five-layer cake. Yeah. And that what's the bottom? Power. Second is going to be infrastructure. Well, these are all things we work with Schneider on for sure, but it is a bottleneck right now. And so what I'm coming away with more and more is that demand is now exceeding supply for the first time in what we've considered to be the last year. Everybody was like, build, build, it's all available. You're getting it. And now it's becoming harder. And I don't know what your backlog is in terms of data center builds, but what I'm seeing is that there's a massive investment being made in data centers right now, and that'll continue to feed that demand over the next few years. It's just going to take a while to build them all.

SPEAKER_02 3:49

Yeah. Yeah.

unknown 3:50

Yeah.

SPEAKER_02 3:50

And I think, you know, it in and again, I think that there is certainly, you know, things have normalized a little bit, but there's a lot of demand just on the data center supply side as well. But I think, you know, what we're seeing with WTT and WWC in particular is that look, is no one's no one's building a data center because they love physical infrastructure and compute. It's about how do you deliver that? Yeah, yeah, yeah. As much as I would like to say that, it's about how do you deliver that compute on time, scalable at all dimensions from an enterprise AI to a cloud data center. That's where WT really specializes, kind of helping put a wrapper around that. You know, partnerships to kind of make sure that those uh chunks of compute can be delivered reliably. We're always going to have challenges checking up with the gigawatt data centers. But again, that is not everybody. We have that side, but I I want to you know be clear that is how we deliver compute and how those get integrated and wrapped and to be able to deliver it on site, that's kind of where I see a big shift in action in that space as well.

unknown 4:47

Yeah.

Data Centers Weren’t Built for This

SPEAKER_00 4:47

I mean, Jim, maybe stick with you here. Tell me more about how the and you know, just the design of the data center is evolving kind of in real time alongside as as organizations try to adopt and scale their ambitious AI plans.

SPEAKER_02 5:00

Yeah, and again, here what I say is that there's kind of two sides of the equations like there's the hyperscalers and how they're doing that, it but it actually relates to everybody, to be clear. So as you think about you know being the most productive in terms of and I and I'll speak in token language just because we're AI, we're an AI kind of I won't speak about compute in general, but in token language, you know, the if you kind of look at the most effective use of tokens, whether it be for training or inferencing, is it is around packing the the the the best level compute in a scale up manner, putting as much as closely as acts as one one uh one GPU. Now that itself is not terribly disruptive, but what's interestruptive is as a concentration of power comes in, that's what's driving us to rethink data center designs, certainly at the hyperscaler level, because just the magnitude of the quantity gets so high, it's forced us to quickly index and put solutions out that are pure native liquid cooling, which has been around at medium scale or low scale for you know for high performance computing, but now it's at massive scale for for cloud for for hyperscaler and neo-cloud type spaces as well. The second dimension is that you know getting all that power into the rack. So used to be you could do it by rule of thumb, put up on bus away, 30 ep tappes, a couple of whips, you're done and go. You can't do that anymore. You really have to think about that. And that you know, we're at the realm now where you know, I ironically, in my view, getting heat out of the rack with lipid cool is a little bit easy to getting power into the rack. That was never the equation before.

SPEAKER_01 6:31

Yeah. Yeah. Well, what's also interesting is the ability now is NVIDIA innovates. And so we work with them to say, well, we've gone from 32 KW racks to 600. Yeah. And now we're going to a thousand kw rack. And so the density of the data center changes and the math changes for what you can actually put in these data centers. And so, you know, it's a challenge for any data center provider or builder today to say, well, do we build for the future of what NVIDIA is trying to do, or which takes three years, or do we try to build for what's NVL72 liquid cooled and have a little more flexibility? So we can do it all, yeah. But it's an interesting dilemma is for data center providers, yeah, because you're having to keep up with that scale of innovation. It's every 18 months instead of a software provider that's every three, four years. Yeah, right. Yeah, right.

SPEAKER_00 7:15

What's the right answer there? Are you designing for today?

SPEAKER_02 7:17

So the flexibility, the the the flexibility is the key, right? I and I and again, I think you know, it's very we while we know roadmaps, you know, everyone works a little bit different forever. So you know, there's data centers that are are perhaps you know have uh different frameworks for resiliency. You know, they're able to they're maybe pure training, there's others that are very high resiliency dimensions in that in that space, and all this stuff kind of in is there's there are different uh thoughts on all liquid cool, but what's the temperature? You heard us you know yesterday with 45 C was the new hot water. Well, the industry doesn't all jump there all at once. So, you know, the idea is that you have to build the name of the game is out building flexibility into your systems while not sub-optimizing, you know, for the future. And that's a that's a tough balance, you know, to to to work with, not to mention the least of which is that when you get up above, like we can deal with I say basic AC coupled designs, AC power delivery, up to maybe 300 kilowatts. You get above much above three, four kilowatts, it now flips to DC, 800 equals DC. It's just another dimension that the industry is ready for, but just not at scale. And this is kind of where we have to be careful as we collaborate more with WT is that look at we can have the perfect design at that space, but if we can't, you know, deploy it quickly and scale, that's where things break down. So it is not just about the best tech, it actually is about the best system that can be built at scale. And that's where I think WWT and Schneider really do a good job collaborating because we are trying to supply, we understand that there is the the the hyperspace, but then there's the masses that actually we have to solve for as well.

SPEAKER_00 8:54

Yeah. I mean, Chris, you're you're boots on the ground talking to clients and executives kind of you know every day, every week of the year. How is the reality that Jim's describing right now kind of jiving with with their plans?

SPEAKER_01 9:06

Yeah, you know, a lot of them are having the same challenge, and their existing data center does not have any capability to cover the new stuff. So traditional infrastructure is fine, but anything with high performance compute requires brand new build. And so they're trying to plan for it. And and this is an AI workload placement exercise, which is where are you gonna go put this stuff? And so it can't be on-prem, maybe a cloud, maybe a private hosted cloud, right? And all those instances they're gonna need to build data center underneath it. So there's all of them have the same concerns. Speed to first token is the number one thing we talk about, which is if you're gonna go build or you're gonna go put this somewhere, you better do it where you have the right timing. So that's that token you get profitable, you get the token operating. So what Jim said is right, which is they're looking for structures and framework that we can provide jointly that says how fast can I get there? Yeah. Because time is of the essence for everyone.

SPEAKER_00 9:59

Is the idea of getting the speed to first token, is that aligned to building for flexibility within the data center?

SPEAKER_02 10:05

It can be. So it starts with what I consider the collaboration on, we would call them reference designs, is you start with a design that you know is flexible from a start, and then you work out the details so I can I can deploy that, you know, again, from from chip from we always say grit the chip in chip the chiller type dimensions, but even on the on the the AI pod dimension space is that sure you can design for flexibility, but you have to put it as part of a what does it mean? And you might be able to turn knobs, like, okay, in this location, I might, I might want to, it might be feasible for me to run with 45 C supply because it's a very hot area, but maybe in other areas where I already have a cold climate, I could run more efficiently at a lower supply temperature in those spaces. Those types of knobs we have to bring available, but you have to have the infrastructure that supports that is you know as well. And that's where you think of, you know, one of the things that we've we've done, what we have done in the past, or we'll continue to do, is think about helping our clients and customers deploy their compute, not just time to token for sure, but also the with the with the consistency and the determinism that you know that when you turn it on, it's gonna be there. Right. So it's not just around it's a time-diverse token inclusive of making sure that you can turn it on and it works reliably. And that's our job to do is to make sure that make sure that it's not just supply chain speed and bringing the power and that when you turn it on, you get your outcomes right away. Because that can be a long cycle if you don't get that part right.

SPEAKER_01 11:28

Yeah, I I agree. I actually think this compute is you is becoming a utility and should be a utility as reliable as power. And so you're actually operating at that level. I've met with a number of power providers that have to operate at that, and they say, do we have to provide that same level to the data center? And the old answer used to be no. The new answer is yes. You know, it's okay to have a small 99%, you know, 99% uptime. Yeah, now they don't they want 99%, 99, 999. Okay, so it's does matter. Yeah.

Build Flexible or Rebuild Later

SPEAKER_00 11:55

So if the idea is to design for flexibility, are there where are the risks at in terms of being locked in anywhere? I mean, are there is there practical advice you can give to on areas to avoid, or you know, how should we be thinking about ensuring that flexibility moving forward?

SPEAKER_02 12:11

So I think there's a few things there. One is so one is thinking about, and I kind of take them in different planes, right? There's the plane of say the heat rejection and the cooling plane. So making sure that you are leaving uh either gallery or other place for like CDUs to make sure you can expand and contract, as maybe you might need different flow rates to build temperature. You don't lock yourself in spatially, you kind of leave space to build. You may not deploy all the space, but make sure you do that so you don't lock yourself in. Same thing on the heat rejection side is that look, it might be better off thinking about you know an economizing your chiller where it can do, it could do chilling sometime, but also many times so you can have flexible space so I can change the water supply. Those are the practical things that we see. Don't lock yourself in there on the power side. It also is about finding, I call it the line of demarcation of kind of make sure that if it is and it gets so big that you actually can't have you have the right PDU and RPP structure or busway structure, that you don't have to do a big rip and replace when the next generation comes out in that space. And that it, you know, then that is you know also about kind of forward sizing in terms of uh doing a reference design that actually can manage a wide dimension of stuff. Tend not to worry about you know stranding some degree of of floor space. So maybe it's probably the power and cooling in that space, that the the two dimensions that was that make most sense.

Digital Twins Change the Game

SPEAKER_01 13:32

I I a follow-up for you. I know digital twin is something you're using for that. Do you want to just talk a little bit about how Schneider is using digital twin?

SPEAKER_02 13:39

Yeah, so that gets back into the determinism piece, which is the fact that look, we can do all the reference signs on paper, the real engineering work space, but you never know until you know. So the idea here is but see with Omniverse, you know, what we're doing right now is we're able to kind of take a design that might start an Omniverse with sim ready assets, be able to kind of now use it and then simulate that in an environment where we can where we can know ahead of time what happens if if the conditions change. So I can know if I have enough flexibility put into my system before we actually turn it on with no surprises. So, you know, we've we've we've been sort of particularly at scale. We think about this as not just around design the data center, we also show the ability to simulate the data center with different types of simulation engines. We have a CFD engine called ITD CFD, we have a great power simulator called ETAP, and then we also have a fluid flow simulation called a Viva process simulation. All those kind of can run native, dispatched to through Omniverse. But the other key is then what it operates is that now I've got this sort of digital twin in my environment. Now, how do you now turn it into operations? And that's where we kind of flip to the fact that okay, if you're at a large gigawatt scale center, you might want to use an element like well, we have an Aviva piece of software, which is a fantastic system with OSI Pi system that allows us to get millions and millions of data sets, data points, and kind of see it coherently. But if you're a smaller data set, maybe that's overkill. Maybe we might have another kind of ecostructure building system that allows that space. So no matter what, the idea is I think it all starts with the digital twin to make sure the design certainty is there. But then don't forget about the operations phase in the monitoring and also in the in ensuring that it stays as tuned as possible in that space.

SPEAKER_00 15:29

Yeah, I mean, like I said, you're all your boots on the ground talking with clients. I mean, this is what they're asking for right now, these digital twins.

Liquid Cooling Gets Real

SPEAKER_01 15:35

Yeah, absolutely. And you know, what what the beginning is is that they think they can do it in their existing data center. So we do data center readiness assessments very high. And what that prevents is an NVIDIA block showing up or pod showing up without it anywhere to go. And we've both seen that the last year, where a stroke of the pen is made and it's like, where are you gonna put this thing? We don't know how we're gonna do it because a CEO or a chairman decided to sign off for it. Yeah, yeah, yeah. So now they want more predictability, and that predictability comes with the digital twinning. These are not small investments. I mean,$20 million to build a basic data center, probably$50 to build a more advanced one. I those are those are round numbers. But I think ideally, before you make that spend, the better you're equipped with the detail of how it's going to go, including operationalizing it, you're gonna make sure that's better. The other thing we're seeing is that people are concrete may not be the way to build the data center. We have some that are doing modular, which is a faster way to market. You can cut maybe six to nine months off time if you're trying to get to first token by using modular. So in that design, that you might have, it might say, well, great, we could go do this in a physical data center and build it out, but your time to market means it might be modular. Well, then you flip that to a modular design. Yeah.

SPEAKER_00 16:42

Yeah. Let's let's start to unpack cooling here a little bit because you mentioned it um a bit earlier. Just just to kind of level set, because I feel like the the more I talk to people, there's still some confusion around exactly what the technology is. Is it liquid? Is it not? Just a high-level overview there.

SPEAKER_02 16:58

Yeah, so you said a very high level. There's always this kind of this thought of like, hey, is is liquid cooling or air cooling which one's better? And the answer is it's not really a better or worse. It really depends on what is what type of compute there is. There's no doubt when you start packing anything more than say 500 watts per package, it becomes much more effective and efficient to use liquid cooling. And because that's the dominant scheme being deployed for AI today, that's why you saw the uptick in liquid cooling in the space, because it's the most effective way to remove heat from very small packages. You can do with air and you can push that, but it it gets very big and you're starting to it's not as efficient at that space. At a rack level, you see that change around 30 kilowatts per rack, but it really starts at the GPU level as well. So all that stuff is, and again, this is something that the world's been doing for a while, but at small scale. What's happened though is that now all of the server vendors, all the people that are delivering the compute, have all now kind of you know adapted their supply chains and their design models to be director chip, liquid cooled, yeah, cold plate designs with a standardized, you know, tap-off, you know, leakless quick disconnects. And that's really what's allowing the industry, you know, to scale. There still is some holdover for some classic networking equipment, and uh even some power supplies are still air-cooled, which is fine. It works that way. So every native today, you know, even uh even an NBL 72 rack today, you know, as of now, it's probably 85% liquid cooled, but there still is 15% that skills are rejected in air that has to be captured somewhere else. That might change in the I'm sure it will change in the future. There'll be like 100% liquid cooled in that space, but they both will coexist. On enterprise, it gets even a little bit more monky because uh that's a technical thing. Yeah, yeah, yeah. Which is the fact that, you know, while you still might get a liquid-cooled server, you might have air-cooled servers. You probably will never have a fully, you know, liquid-cooled environment. And there are paths forward to allow liquid and air to exist, even in enterprise, that do not have to be crazy disruptive in the in that space today.

SPEAKER_01 19:04

Okay. Yeah, and you know, what Schneider's done, they acquired Motive Air last year, which helped with reardoor heat exchangers. And so as the need for cooling has advanced, you can not only move to reardoor heat exchange before you go to direct liquid chip, but it that combination also allows you to have some of that air-cooled and the liquid cooled. So it's a it's a great combination to be able to add as well.

SPEAKER_02 19:25

Yeah, that was one of the things that it would that one of the motivations for the motive air acquisition was so there was a catalog that was there, but it was actually the experience. They've been doing this for well over 10 years in the in hypocrisy. So they know how to do rear-door heat exchangers, not just that you know fans and coils, but how to do that well. Same thing with the CDUs for direct-to-chip liquid cooling, and also how to do liquid-to-air heat exchangers, so which allows for even a uh a server that might have direct-to-chip liquid cooling, but you actually can put it right next to uh a heat exchanger or even an in-Rack CDU that allows you to kind of pre-plumb that and make it be an easy drop-it system for even the enterprise in that space. That was the that was really the kind of the magic that we saw with Motive Air. It was surely a catalog is robust and good, but it's a know-how that goes into it. Believe it or not, it is not just pumps, fans, and coils. There's more that goes into that. If you want to do it with certainty, uh, in particular you've got to think about the cost of the asset you're protecting right now, it's it's kind of a high stakes right now.

GPUs Are Reshaping Everything

SPEAKER_00 20:25

Yeah. You know, recognizing that every year here at GTC is constantly pushing the envelope. Jensen's just, you know, innovating at a rapid clip, and everybody at NVIDIA for that matter. What are you seeing here this week that you know is gonna force data center design to be more and more innovative? It could be, you know, along that flexibility, it could be more in the power and cooling space. What what type of innovation needs to happen in your area to keep up with where Jets is going?

SPEAKER_02 20:50

So I think a few things. So one is you heard Jensen speak, you know, about a few devices. One is that you know, while it's possible, it's likely, it's unlikely there'll be a monolithic AI factor, right? You uh you they talked quite a bit around the adoption of the IP from rock for inferencing in that space. So you'll get sort of a data set that has different types of workloads and different types of dimensions that began adding to the diversity factor in that space. The other is so that data center, it still is flexible, but it maybe it's not as high density as a pure trading data. So that's that's a that's a bit of a thing that the industry has to adapt to. Understanding that you know, if you were just kind of thinking about AI factories are all the same, they are not, they're becoming a more even more diversified because the workloads are changing. You're also there was a statistic that he put up where I think he talked about, I think I remember it correctly, it was like a 60 40 split from hyperscaler versus enterprise in 28. And it's about a 90 10 split today, which means that to get to that growth in two years, guess what's happening? All enterprise GPUs are going more towards enterprise. That's the other shift is I do think they're going to see a lot more enterprise owned GPUs, which now We have to figure out okay, if there was a challenge to stand up the gigawatts of data center, how do we index to help that scale grow at maybe smaller on-premise or or uh or I was you know edge or close to the consumer type. I was trying to avoid the word edge, but uh you know, uh in that in that's in that in that space is also putting another gear change in how we and how we work together at the at that space. You also in another instance of that was you heard the you know the the open claw dimension that so you can just hear the usages of AI is going certainly from the training and the localized scientific use to the more broad-based use. And and that does mean that I think we will start seeing even at the hyperscale adeo cloud level a little more heterog heterogeneous workloads and also more types of workloads that are driven by compute that is maybe not owned by hyperscale as a deal clouds, but also more for the enterprise, you think differently about those designs.

SPEAKER_00 22:56

As I mean, if if the enterprise is gonna start to own more, more of the GPU or is gonna start to balance itself out, Chris. I mean, what do enterprises have to do to get ready for that? Is it about mostly what we've talked about, or are there additional considerations that need to come into play?

SPEAKER_01 23:09

Well, A, there's a few things. If they're gonna choose something that's cutting edge from an NVIDIA standpoint, let's just say it's Ruben Vera, they're gonna have to build from scratch. So there's a bridge they have to build to be able to use that. It may put in a co-location provider, maybe they'll rent it from a neo cloud, but then they ultimately want to build it themselves. And so they have ground, they're breaking ground, probably working with us in Schneider, but it's gonna take two to three years to build that out just due to the supply chain. Yeah. So we're really talking about what are you gonna do with that workload? How much demand do you really need now? So one of the things we've seen happen is that the enterprises that are buying into the clouds are buying big chunks and using training. They might spin up the workload for six weeks and then not use the cluster for two and then come back, but they just want access to the chips. So I think initially there's gonna be a big demand for access to the chips through providers of others, and then as they build it themselves, it's gonna be well, what do we really need? Do we need to train here on-prem? Do we need to just run inferencing? Right. And I think the training could be done in the cloud. The inferencing can all be done on-prem. So we're still in the early stages. The other thing I'll note is that the new AI factory is now a gigawatt. Right. And that's a different SU than we've been working with for a while. So it changes what we are supposed to be doing as providers and what the expectation is because that that SU is very, very large. Yeah. And we have to scale for that.

AI Factories vs The Power Grid

SPEAKER_00 24:31

Yeah. Articulate a little bit of that pivot that he was just talking about from your end. What are you guys doing differently or going to be doing differently to handle that that gig AI factory?

SPEAKER_02 24:41

So look, I think the gig eye factory is just, it's, it's, it's, it's, it's, we could see it coming. You know, they're already kind of on the books and building.

SPEAKER_00 24:48

Yeah.

SPEAKER_02 24:49

One thing you think about an AI factory, right, is that at that scale, even at the you know, 200 megawatt level, is they are no longer things that just can be, you know, Greek takers. Okay. So the idea is that there is a lot more, we talk about flexibility in compute and design. There's also flexibility in an asset that is a gigawatt that needs to interact with the grid much more better. That's why you see a lot of new standards and regulations that is causing some challenges with regard to Matt. Think about a data center that is a gigawatt data center that let's say it is moderally inferencing for three weeks. It's nice and calm and steady. It's a you know, grid friendly device, not too disruptive. And it goes to a week of training, holistic training. It becomes spike, you know, spiky, bumpy, lots of things that can cause power challenges for the grid. And as a as a designer and a provider, we have to accommodate both those cases. So there's a lot more going into thoughtfulness in terms of you know, how do we absorb the the power variation that exists in an AI data when it's training, but it may not always be that way, so you don't over provision for it. There's also new requirements with regard to ensuring that data centers are are stable during just grid faults. If there's a grid fault upstream, even for a very short duration, they don't want 31 gigawatt data to all disappear and go on to battery or go on to some generation because now the grid just lost 20 gigawatts in an instance. Right. And that can drive the whole grid instability. So when you the gigawatt scale, I think we're getting our hands on the block size and deployment. However, as you start thinking about how does it match with the bigger picture of the TSOs and the network interconnection, that is becoming much more much more of a requirement and proper than it we've ever seen in the past. Now, Schneider, we have solutions to this to help with AI smoothing. Everyone is being called ride-through grid compliant, but I expect more and more to come in that space when you're dealing with a gigawatt scale, not to mention at least some of the gonna have on-prem power as well, which also changes that space, you know, it as well.

SPEAKER_00 26:48

Uh so we're running short on on time here, and the two of you have been generous with your time. I know it's an incredibly busy week out here. Just real quick, Chris. I mean, for us, if we're sitting here next year, what do what does the enterprise need to get right this year so that next year we're feeling good about hitting the ground running?

SPEAKER_01 27:06

Yeah, I think they have to have expectations over how their use cases are developed and how they're gonna do it. You can't boil the ocean right now, there's just not enough power and and ships available. So you might start with a more measured approach. So you're saying we're gonna use a workload that takes a certain amount of power, maybe it's a little bit more, but you're not gonna make a huge commitment. It's just not available. Yeah. So I would think a year from now the power question will be coming more into focus. And I'd be curious if you have anything on microgrids or anything, because I do think alternative power is another thing in the next few years we'll be talking about. But until that comes, we're gonna be talking about well, how how do we manage this chip short, the availability of chips? It's not a shortage, can't say that, but availability.

SPEAKER_00 27:49

Yeah. And from your end, I mean, what what has to happen in the data center provider space? What do you need to tackle or take care of so that you know everybody else is in good position next year?

unknown 27:59

Yeah.

SPEAKER_02 27:59

So I think yeah, again, a few things. I do think on the enterprise side is that is about having a strategy. You know, Chris talked about this, is making sure that look, you might have to use a different part of your PL to buy tokens less CapEx, but then you be aware that you might have to put a CapEx face and then bring things on-prem or build your own data center. Those are decisions that make sure there is a plan there to be successful. Because if you over-index to all on-prem, you may not work well for you, but you have to balance those types of things. What I'm really watching is if there is sort of an on-prem or enterprise-owned type space, is just what Chris talked about, is make sure that we can reliably power and cool that in a different environment. And that is the last thing I want to be is a barrier to deployment. So I have to make certainty of power. If it means that I have to use a microgrid to get it on that space, we will do that. Remember, my goal and really the industry's infrastructure goal is to make sure that every electron that gets generated goes to compute. I want 100% efficiency on the electron delivery. And if you generate the heat, I want to make sure that I don't have to use any energy to put that out. That same thing we're getting very good at at the gigawatt data center scale. I have to take that same philosophy and bring it down to the space, particularly when the stakes are high. In other words, I can free up just another 3% of power to a data center. They can deploy that for tokens, which is far more valuable to them. So that's kind of how we think about make sure that power and cooling is never a barrier for on-prem deployments, which we're gonna start seeing more and more in this space.

SPEAKER_00 29:28

Yeah. Well, exciting times ahead. Lots of work to do, but uh certainly lots of opportunity uh for everybody here. Uh Jim, Chris, thanks so much for the time here today. Hopefully uh the rest of your week goes well, and hopefully we have you on again soon.

SPEAKER_02 29:40

Yeah, great. Thank you so much for thanks for the discussion. Very good. Thanks for having us, Brian. Jim, thanks for joining. Yeah, thanks for it. Always good to connect. Yeah, yeah, it's always good. Okay, great. All right, cool. Thank you.

SPEAKER_00 29:49

Okay, thanks to Jim and Chris for joining today. The lesson here is that AI success will be decided by how early leaders turn workload strategy into infrastructure strategy with the power, cooling, flexibility, and deployment certainty to get to production without delay. This episode of the AI Proven Ground Podcast was co-produced by Nas Baker, Kara Kuhn, Sarah Chiadini, and Addison Ingler. Our audio and video engineer is John Knoblock. My name is Brian Felt. We'll see you next time.