In The Arena by TechArena - Schneider Electric’s Jim Simonelli on Breaking AI’s Power Wall

Episode Date: October 12, 2026

At Yotta 2026 in Las Vegas, Allyson Klein sat down with Jim Simonelli, SVP and CTO of data center business at Schneider Electric,Simonelli explains why IT moving on six-month cycles and power access running on six- to 10-year cycles makes the data center a systems problem. He covers why rack densities of 600 kilowatts to 1 megawatt drive the move to 800 VDC, and how Schneider’s new Software-Defined Medium Voltage switchgear separates the physical build from protection programming to speed time to power.

Welcome to Tech Arena, featuring authentic discussions between tech's leading innovators and our host, Alison Klein. Now, let's step into the arena. Good afternoon. My name's Allison Klein. I'm the CEO and founder of Tech Arena, and I've got Jim Simonelli with me. He's CTO of Data Center business at Schneider Electric. I have been looking forward to this interview all week. Schneider has been everywhere at Yada and everywhere in the data. to center conversations that we've been having all year. It's so wonderful for you to be here with me. Thanks so much for coming. Thanks, Alice. That's very good to be here as well. Very exciting show, too. So, Jim, you've been at Schneider for a long time. Tell me about what you do with the company and how this

informs what we're seeing from Schneider this week at Yada. Yeah, so interesting enough, my role of Schneider, the title is CTO, right, chief technology officer. And largely, I'd like to say my higher and fire responsibility is to make sure that Schneider is in a good place to support the industry in two to three years, make sure we're making the right investments across the full portfolio, software, hardware, services in that space. Clearly, I'm still employed right now, so I've been doing something right for the past few years in that space, but I can say that the job that I have right now has never been harder because thinking about what really the day is and is not just today, but in the next

six months is very hard to predict and align in that space. but it is one of the things that I do enjoy most about my role. And certainly with Schneider is because, at least in the data set of context, I do like to think about the fact that we can bring a full portfolio. We'll use the word like, rid the chip and chip the chipper plus the software and services unpassive it. So it gives you a nice big sandbox to play and create and have an impact. Now, everybody is talking about the power wall.

I think that there's a reason why there's thousands of folks in Las Vegas right now congregating to talk about power delivery and data centers. Why are we at this challenge? Talk to me a little bit about the landscape that you're operating in. And why is a company like Schneider really in the center of the conversation between utilities, data center operators, and even the ship companies who are trying to power the semiconductors that actually deliver all of this cool compute? Yeah. So it's interesting. I might make it down to a few different sections there. So first of all, why are we here and why are we talking largely around power and power access? It's not like that wasn't a conversation before, but really some of the

dynamics that have happened in the industry have been, I would say, at least twofold. The first fold is that the pace of change for compute has rapidly accelerated. And the ability to build or even draw power is anywhere from six years to 10 years in that type of a cycle. If you think about just getting access to energy now, in the cycle life of IT is moving at six months, those two cycles just are not in step. That puts a little bit of a challenge in the system overall. The second is the scale. The notion of thinking about annual increments of gigawatts just in the U.S. back in 2020, 2021, maybe we're talking about plus five gigawatts. Now we're talking over five years of easily over 170 gigawatts. So the scale is another factor. So that's kind of the

challenge is that everyone's facing why it is such a common conversation today. And there are other factors for sure, Greek constraints and all those areas as well. But what Schneider brings to the table and the way that I like to think about this problem is a systems problem. It isn't just a power delivery problem. It is just a construction problem. It isn't all just around how to work in regulatory environments of utilities. It is actually about think about as a whole system. And that is about thinking about a data center as a system that has to interoperate with the grid, has to be focus towards producing the most tokens in the cost of AI, most efficiently and effectively in that planet.

And that takes a system point of view, not just on power delivery, but also on heat rejection as well. And staying up to date and being in a company that thinks about not just as a system, but also about its full lifecycle. And we'll talk about this from a thing about data centers having to be thought of from a design side to ensure you're simulating digitally to building it, operating and maintaining it, that whole life cycle and the whole depth of what makes a data center a system is really where Schneider comes into play in some of the aspects that we think about.

You know, this gets into compute density and the amount of compute that's being pushed into racks. You've been in this space for a long time. Did you ever think that you were going to see us transcend a megawatt per rack and even look higher than that? Yeah, it's not a. I sometimes have to check myself because for years, I've been in the industry for 34 plus years, I won't go deeper than that, but is that we've been talking about things of 10, 12 kilowatts per rack. Max was like a typical 20 to 30 kilowatts per rack in sort of a build style for rack densities. And be clear, it was only five years ago where we talked about the fact that the rack densities were growing.

We thought, what do we do with 100 kilowatt per rack or 120 kilowatt per rack, which was the original spec for a DB200. And now we consider that normal. So as you think about taking this forward, and certainly when you get above, say, 250 kilowatts per rack, the whole system changes a bit. That's where you start thinking about different forms of delivery, 80s DC, larger and larger CDUs with liquid cooling dimensions come into play. And I think those are solvable problems, but the fact that they accelerated on the planet very quickly and now sort of is where our mindset is we've re-normalized to what rack density is to having something that was unthinkable. four or five years ago to being, oh, we know how to go do that. And now we're focused on that ultra high density compute, which, you know, is certainly thought up today, solvable maybe in that

six to one megawatt per rack. Who knows if it's going to keep going or not? I never bet against compute technology in that space, but it's a reality for the next three, four years at least. Now, Schneider earlier this year and it may be before then, but this is when I started following the plot, made a big bet on 800-bolt DC as, being a technology that could shepherd us through. Tell me why a movement to DC and add 800 volts is so critical. Yeah, so it falls along with the previous conversation that we just talked about. It's not, no one's building with 880 volts DC because they like the number or they like

DC versus AC. It is really to fulfill a purpose. That purpose is, hey, if the compute side can deliver the most efficient compute, if it's AI, it's around tokens, bypassing. packing the GPUs very dense in a single 19 inch or 1 in 1 inch rack format, then it's our job to figure out how do we deliver that? And that's where the understanding of, hey, to deliver power to that rack, if you deliver that at, say, 480 volts AC, thinking in the U.S., the number of power cables and copper that come into that rack, just starts taking up space away from compute.

It becomes not necessarily inefficient, but ineffective in the fact that you've taken up critical space where you really need that to be for compute. And by raising that voltage, almost doubling that voltage, it gives much more freedom to drive compute. And we knew from a while that if that density happened, that there was a need to switch to that. So we've been planning a portfolio, again, for probably three years or so, to cover all different dimensions of not just power delivery,

but conversion, distribution, and protection around all those schemes as well. But they're not here right now, but you can start seeing shooting in the next six to nine months the first deployments of those racks, which are requiring age of volts DC as well. Now, I know you talk to a lot of operators and probably the who's who of operators that are here gathered at this event. When you think about what you just talked about, is this something that you talked about in the keynote about reserving room for revenue generating compute? Yeah. Is that a metric of efficiency in your design? Or how do you think about that from the

system perspective you talked about earlier? Yeah. So look, I think we always try to keep our eyes on the prize of our customers. And really what that means is that as much as I love copper steel structures and cool data center architectures, that actually is not what data centers are built for. They're not built to heat. They're built to drive compute. So therefore, keeping a metric on what is the value of that data center to deliver for every watt that comes in or every electron that comes in a data center gets transmitted into valuable compute. And also any heat molecule that gets generated gets transmitted out with zero loss is our job in that space. And the way you value that is not just on the capital cost of the structure, but it is time to compute and time to allow more tokens to be

delivered in that space. And that means thinking creatively around things like ensuring that you are using your redundancy levels to the maximum. So if you have spare capacity for redundancy, maybe there's times where you want to maximize that to deliver more tokens. The revenue of a token, is significantly higher than, say, the cost of saving 10% on an infrastructure architecture architecture space. We always keep an eye on that. And in the other dimension is speed. And this gets back into the fact of, sure, there are potentially supply constraints that exist, but we want to double that. We don't want to work against that. And that is also thinking creatively around how we deploy our systems and how the industry needs deploy our systems. That's what

drove Schneider significantly towards a prefabricated approach to ensuring that we're putting more and more of the trades in factory environments to deliver power and cooling skids and equipment. So the commissioning time, the trades on site become much less. And the other brand new dimension for us right now is we're also seeing there is a value to something we would call software-defined or software-defined protection in that scheme. And the reason why that's in part of this conversation is because what that enables, and this is a brand new announcement that came out from us today, as a matter of fact, but something we're working on for a while has been, truly we think about software-defined

and software-defined protection, medium-voltage switch gear in this case, is that it gives this notion of flexibility. After things have been put in place, you can actually change your protection scheme, adjust it maybe in real time. It's been done on the utility side for a while. But the real reason of why the importance to the data center in that speed and that time to power, that time to power means earlier token production. And the way that you enable that is by separating the build of the physical switch gear. All of the current carrying parts, the cubicles can be built without being programmed and wired for the protection scheme. And you put that in software.

They basically can run in parallel. So it's not just efficiency. It also is how you design to allow more power to get access. to produce more tokens, but also packaging delivery for prefab using more software-defined, to separate out the physical build from the software or things that can be done in data bits versus atoms to build physical things. Now, I wanted to unpack this, so I'm glad that you brought it up. Software-defined is something that I always associate with compute.

Software-defined, composable architecture. When you extended into energy systems, are you seeing completely? posability of energy delivery within this model dynamically, or are we even there yet? Yeah. So look, I think it's coming in. Let me give a concrete example. So softivine networking. That is good of the inspiration of it. Typically, you can't reconfigure large copper wires or aluminum bus bars. But what you can do is there are elements that you can put more in a software layer that have typically been in a hardware or a very controlled networking space. So there's one layer that It's inspiration from the scene, the fact that, again, classically, protection schemes of where

systems need to trip off, how they have to switch have been hard programmed into devices locally. You can imagine that in a software-dify networking scheme is that the equivalent would be having an operator come up and reroute wires or have to go up and plug a USB in to change the routing scheme on a router. That is no longer how things are done. That same thinking is being brought to at first the protection scheme. But the very next step, and we're very close this as an industry, and I will talk more about this in the Schneider context,

is that when you think about the energy in the energy composability of where your energy comes from, you think about a microgrid, or we might use the word in our context, an AI energy park, is that look at those don't need to be static program device following rules-based controls. They actually can also be mission-based,

which is that, oh, I would like to have my data center today or in the next few hours be carbon seeking, minimizing carbon and you would operate your energy storage systems and your grid systems and maybe you're complementary of a gas system a little bit differently in that scheme. So that's another aspect of software to find by having things be more emission based. Same thing with cooling today. Cooling killers are usually pre-programmed or rules-based for what chiller goes on when and how they're commissioned. Those are hard program. Those could be more software to find in those spaces as well. So that Your insights from learning from the digital world as to how to flex and give new targets and goals,

the infrastructure is the same thing that's happening to the power space. In a data center space, it's probably the first to adopt because it enables that speed, flexibility, in some cases, more efficient use of the infrastructure. I was talking to one of the major neocloud operators recently, and one of the things he was talking about is he wanted a control plane that could work across computers. and power and cooling and have an aligned view across those spaces. Is your moves calling for software-defined energy a move in that direction, do you think? It is.

Again, the operative word here is collaboration, working together towards the goal of ensuring that we can maximize whatever compute is there for whatever workloads. And in the context of AI workloads, right, if you really want to maximize that, the idea there is to share insights in the context of, of your neocloud example is that, look, there's only so much that we can do on the infrastructure side, assuming the current use of energy that we can measure that's going into the compute,

imagine that what could be done if there was even cross-insights in terms of what workforce are most important, what their predictability was in the future. It's a better align, not just from the demarcation of power to the rack or to the server, but actually right down to the CPU and GPU. Those are real world things that can be done today

with very bespoke designs, but I would anticipate that in the near future, and near future still might be three to five years away. Sure. That inside of being bespoke for one very motivated customer that becomes now at least nearly table stakes, but just that's the way you do it. You wouldn't buy a router today with software-defined networking. That's a very sophisticated piece. You almost can't buy it without it. I expect that's where the data center land is moving as well. I like that. Now one final piece on that is that You talked about that the vision includes openness. Tell me why open innovation is important in this space. Look, I think part of this is the fact that we don't know everything from every dimension.

And also that there's no such thing as a single supplier or a single off taker in that space. So if we really want to advance the compute or the value of compute across the board, there is a degree of openness that has to happen, probably in an exchange layer at the very least. But, Well, when I talk about software-defined and software-defined, certainly at the layer up beyond the median voltage switch gear, which is an element of an asset, it needs to come up a layer, and an upper layer has to be cross-platform, cross-vender.

And yes, it does take some conversation. We have these conversations that are ongoing. In the U.S. DCFlex is an activity program run by APRI to help on the energy side. There's lots of activity happening with Nvidia partners at OCP, to try to open up that bridge in a more standardized way. way. And it's not just open for openness. It is the fact that if you really believe that the data center, the value of compute for GDP, for economics, choose what of a value would like to get out of it,

is that it is not a single-threaded, single-customer area, something that has to spread out across the area, across the board and be multi-ventor with finding the right interfaces. Ironically, what's interesting is we're actually learning a lot from AI. So by able to more simply mapped and we're doing this with our software-defined energy and our energy intelligence platforms, which are vendor agnostic, but they're up a layer of abstraction that used to be very hard to create an ontology, what an asset is, what's connected to what lots of times with AI and actually just a little bit of information. Can I form that very quick? And once you have that, it is a lot easy to then also think about what can I do with that and define, once you have

an ontology, the description of something, there's only so many ways to describe a UPS or a killer that it does enable actors right above to more simply create software-defined assets and applications, which is really where the industry needs to be, not around, oh, my BMS is better than your BMS or my EPMS has been to yours. It should be more about, sure, that's something that has to be there, but the real action happens at that layer above. And that's where there's a lot a convergence going on right now. No, one thing that I can see is operator control gets better, but there's another aspect of this that you've talked about before, which is the data center moving from being more of a passive player in the grid to more of an active player and rounding

out their use of power to not see these crazy spikes. Can you talk a little bit about that? Yeah, so look, data centers themselves have been large off takers of the grid, just like the rack densities, large as a matter of perspective. Like if we were having this conversation five years ago, We'd talk about a large data center of 30 megawatts that's seen as almost a trivial bump. You put 30 megawatts on a grid. You don't have to think as much about that. You put a cluster, you put one gigawatt or half a gigawatt site that's impactful for the grid. And you put several of those in a region.

It's impactful for the full transmission system. So it is incumbents upon the data center industry in particular and those that are owning and operating to make sure that those assets, however they deployed, aren't disrupting the grid that, is being used for many other things other than just compute. And that dimension of ensuring that there is buffering of any type of volatility of powerful that could disrupt the grid are captured somewhere in the data center, inside the data center boundaries. We have a belief that a lot of that should be captured inside the white space, but there's a piece that should be captured at maybe a UPS or ACD-DC converter level or maybe

even at the site level with an on-premise best. But it really is that dimension that how we architect now data centers to be ensuring that nothing leaks out that can disturb the grid or if it's on-premise power in that space is now becoming compulsory. The first instance we saw of this was actually ironically the other way around, which was ensuring that data says as such large off-takers were well-behaved in the event of grid falls. So we can talk more about the story at PGM, but largely the artificial. fact was a single small outage on the grid with data centers that were tuned and fine tuned to switch them off the grid for several seconds or sometimes minutes or hours. Look at 30 megawatts going off the grid, no big deal. But if you take off three, four, five gigawatts off of a grid, you can destabilize it.

So it's become the notion of in those types of requirements are now not just nice to do. They're compulsory. And the same thing of going the other way of making sure that no, no, no, this barrier is now, we recognize you're not just a power off taker of the grid or a turbine, is that now there are requirements with the amount of fluctuation and powerful that can be put at that space. It hasn't yet become a hard requirement. It become a de facto requirement for any putting on-prem power because the fact of the matter is if you don't do that, you run the risk of quickly aging your very expensive assets or breaking

them in that space. So our architectures have evolved and Schneider has a portfolio that is aligned to solve both of those problems as well. All this has happened, accelerated crazy quick just over the last two to three years and a lot has happened just in the last 18 months as well. As we talk about this, you're driving incredible innovation. We've talked about it through the conversation here. But the amount of demand for moving quickly, I've never seen anything like this in the industry. What do you think is the biggest challenge ahead for the industry to navigate this hockey system of deployments for AI? and what makes you most optimistic that the industry is going to work together to actually get to the other side of this?

Yeah, so I probably tier this as a few tier in terms of layering of not crying. I might cry sometimes too, but layer this a bit in the fact that one solvent problem is just capacity, capacity of physical things, switch gear, batteries, compute, GPUs. That's a capital problem and it's a capital speed deployment. we saw that very hard before all this ramp during COVID. It's been a lot of response and a lot of basically as the market has grown, a lot of actors have come in to help support that. I think we're working ourselves into or out of the ability to supply, the demand. It still is a challenge just to be clear.

I don't want to say that that's gone, but I feel that that is moving in the right direction. It still takes a long time to get turbines and generators, but it used to be not just as a lot of space. The power access is a problem for many, many different reasons. And therefore, the industry has, again, getting very creative. There's much more, I wish I could quote a fancy statistic here, but the behind the meter power for getting access to energy was fractional five years ago. We were getting better at that and showing that it can be done as an asset for the community as well in that space.

Not just an off taker that's going to have negative impacts, not perfect, but things are moving in that right direction. I think there are potentially some regulatory constraints that I don't know my way out of. or the industry doesn't know our way out of, but I can tell you that our ability to solve those problems creatively is fast evolving because we started to deploy things that used to be only notional a few years ago, having just even a convert or a UPS, saw fall right through and filtering of these bumps was not even in the radar.

We spun that up in a matter of six to eight months in that space. So liquid cooling was a big constraint. Now we have those things. It's coming out of the way. I believe that we might run into more hurdles on the access to power side. It's a complicated problem, as I mentioned earlier, that we have weird time scales going on where compute wants to move at the 18-month cycle and utilities can only move at the 10-year cycle. Infrastructure can maybe move at the 2-3-year cycle.

That's not going away, and that's causing some congestion and just our ability to deploy quickly. But my hope actually comes on the compute side. I've never bet against new technologies, optical networks, photonics, new ways, particularly be specialized towards compute inferencing versus training. I do believe that the productivity of a data center for the use of a single watt to produce effective and efficient tokens based on software, hardware, new silicon, and new technology, we've only just seen the beginning of that. And that actually is kind of what gives me hope in the fact that, while I do believe,

that the demands for the compute, that digitization is not going away, our ability to do that more efficiently from a watt to token or walk to flop dimension is going to get better because there's a lot of smart people on the planet working to crack that code and there's a lot of interesting things coming up in that space that I think is a very interesting thing to watch that I think gives me a lot of hope to still solve that digitization problem without assuming a hypersonic or parabolic rise of power demands. I agree with you. I think that the industry is working very hard to drive up tokens per watt, and I can't wait to see how we accomplish that. But the work that you're doing is so foundational to even getting things powered on to deliver that capability.

I loved our conversation today. I'm sure others do too. Where would you send them to find out more about the technologies that we talked about today and engage your team? Yeah, so a great place to start with is certainly at Schneider Electric or s.c.com. But the key there is to focus on the data center base because there's so much in that area on AI and really track what we're doing with reference designs with our thought leadership on liquid cooling on these due dimensions of AI factories, what's going on on the power. Really think about when you look at that AI factory or that data center side of our web page,

it's the best way to do. Look at our white papers. Lots of good information. And by the way, I've written many white papers. I love a challenge. Any of your audience that reads something and says, oh, that's not right or I really like that, reach out to me directly and we'll try to also engage in that because we learn by sparring. And it's a great place to start.

Well, you heard it here first, guys. Go ahead and challenge him with anything that you see in their content and their technology direction. Thank you so much for being with you today. Thank you for having me. It's a great conversation. Thanks for joining Tech Arena. Subscribe and engage at our website, Techorina.com. All content is copyright by tuckering.