Infineon key visual in RGB

Powering the gigawatt era: The shift to DC grids and solid-state infrastructure in AI data centers

We’re bringing you the latest from the world of semiconductors – straight to your ears! From quick takes on trending applications to deep dives on product innovations, our experts give you their take on the tech behind the tech.

Podcast

Jun 02, 2026

Podcast

In this episode, host Kelsey Markl welcomes Head of Systems at Infineon’s Green Industrial Power Division Mehrdad Baghaie Yazdi back to the show, where they discuss the massive power demands of gigawatt-scale AI data centers and the necessary shift toward efficient DC grid architectures. They explore how solid-state transformers and advanced semiconductors are replacing traditional infrastructure to reduce energy losses and improve grid resilience. 

Listen on Spotify Watch on Youtube

Speakers

In this episode of Podcast4Engineers, host Kelsey Markl speaks with Mehrdad Baghaie Yazdi, Head of Systems of the Green Industrial Power division at Infineon.

Kelsey Markl

Host:

Kelsey Markl is Head of Strategic Marketing Communications and Content at Infineon with more than a decade of experience in communications, content strategy, and storytelling. Not being an engineer is part of what makes the journey interesting. Kelsey approaches technical topics from the perspective of a curious outsider, asking the questions that help make complex ideas more accessible and relevant to engineers and broader audiences alike. Passionate about innovation and the people behind it, Kelsey enjoys exploring the trends, challenges, and breakthroughs shaping the future of engineering.

Mehrdad Baghaie Yazdi

Guest:

Mehrdad Baghaie Yazdi is

Additional information

More episodes

Guest: There's a lot of power in one of these GPUs. And you got about 6 of them there. And these are resistive loads. If something happens, you can destabilize the grid. The way the fault propagates in your system needs to be stabilized, right? One of our real unique assets— there is barely any semiconductor company out there, and I would say actually none other than us, that will cover from the bipolar technology all the way to the low-medium voltage technology.

 

Host: This is the Podcast4Engineers, the show where we translate today's engineering challenges into tomorrow's systems. When we talk about AI data centers, we often say they need a lot of power. But at today's scale, that phrase no longer does the challenge justice. Some AI facilities are approaching gigawatt-level demand, putting them in a class of infrastructure few engineers have ever designed for. In this episode of the We Power AI series on the Podcast4Engineers, I'm joined by Mehrdad Baghaie Yazdi, who's the Vice President of Systems, Applications, and Solutions at Infineon. We're going to explore what high-power architecture really means and how semiconductors and systems will enable the future hyperscaler infrastructure. Mehrdad, thank you so much for being here.

 

Guest: It's my pleasure, Kelsey. Thank you.

 

Host: Okay. Can you maybe help me put this in context? When we talk about power for AI data centers, we're really talking about a lot of power.

 

Guest: Correct. Yeah. Imagine the following, right? One of the most modern car manufacturing factories today, a Daimler factory, it typically utilizes 10 to 15 megawatts. If you look at one of the most power-demanding industrial applications today is aluminum smelting. The world energy utilization is between 100 to 150 gigawatts. If you look at the city of Munich, it's projected to utilize about a gigawatt of power by 2030s. And now you just said in your intro, we're talking about 1-gigawatt data centers. That sort of should give you an idea of the scale of power we're talking about, right? 10 data centers would be 10% of the global aluminum smelting capactity.

 

Host: So a data center outside of a city could actually require the same amount of energy as the whole city next to it? So what's important when we're thinking about how we generate all of the power required for these data centers?

 

Guest: Absolutely, yes. That's a question a lot of people are thinking about right now. How do we generate, distribute, and stabilize this enormous power demand? It clearly will be some sort of a heterogeneous power generation and power supply. People are thinking about building dedicated nuclear reactors, right?

 

Host: Just for the data center?

 

Guest: Just for data center. I mean, today's most powerful nuclear reactors can generate 1.6 gigawatts, so it's barely enough. You need to connect to the grid as a backup. So you will have centralized and decentralized energy generation, plus a lot of infrastructure around it.

 

Host: Getting that power from one, you said either at the source, so right at the data center or from the grid, from sources spread out around the country is one thing. But how do you convert that power from the grid into actual— into the power, the electricity you need in the data center for the HVAC, for the servers, for the GPUs, for everything else?

 

Guest: That's an excellent question. It's a really different scale and a very different utilization of power we're talking about. Classical, you know, factories that I mentioned before, like car factories, a lot of it is about rotating masses, about robots and belts and so on. Here, the majority of the power consumption is within the GPUs. Yeah, these GPUs are transistors, transistor-based, right? Like, like our semiconductors, just to fulfill a different function. And there's something called resistive loads, and they need cooling. That the building needs cooling. You will have HVAC, as you mentioned. You will have motors that are pumping fluid to cool the GPUs and to cool the electronics. You have a series of rotating bodies which are smaller, HVACs and pumps, and the majority of the power is going into these GPUs. Now, what you do today, or what you're doing right now, you asked about the power conversion, is you coming from the grid, and the grid is an AC grid, right? Alternating current. The famous discussion Tesla versus Edison and so on, right? So the AC grid kind of became our standard, and you enter the facility and then you have transformers because that grid is at a very high voltage level, 110 kV, 35 kV, and so on. And then you have something called the medium voltage transformer. You typically come to that transformer with about 35 kV if you're in Europe or in the US, and about a little bit more than 10 kV if you're in China. And you enter these medium voltage transformers, they're still AC. After that, you have again AC, and now you start with a series of energy conversion or electric conversion. You go from AC into DC, and the first step down brings you to about a 400 V— you go from this AC, you go to 400 V AC, and then you take that 400 V AC and you start converting into DC. That DC then is driving your power supplies that are then driving your GPUs in multiple step-downs, right? You need to go all the way from your 400 V input down to 1 V or 0.5 V at the GPU, and you need a couple of hundred volts for your HVACs and so on, either 3-phase or single-phase. You have all these electric conversions going on, and you can imagine what that means at one gigawatt utilization of the GPUs plus all the infrastructure around it.

 

Host: Yeah, because conversion means loss, right? Every time you convert, you lose some energy. So maybe a simple question then, how do we make this process more efficient?

 

Guest: And that's the million-dollar question. We need to do something, right? It's not only that, it's that when you have such a high load with these many DC processes and your end load, your most important power drain is a GPU, which is a DC load. And just to put that into context, one of these GPUs dissipates about 1.4 kW. That is similar to your cooking plate, right? When you're boiling your plate on an electric stove, right?

 

Host: Not the induction, but like the coils. The coil ones, right?

 

Guest: That's about— when you put it all up, it's about 1.4 kW, 2 kW. There's a lot of power in one of these GPUs. And you got about 6 of them there. And these are resistive loads. If something happens, you can destabilize the grid. Even if you have a local energy generation, the way the fault propagates in your system needs to be stabilized, right? You look at all of these and you're like, what can I do to improve efficiency but to also make my grid more resilient? And one of the answers is you move towards DC grids, right? Because inherently your loads are DC. So why not make your entire grid natively DC? And that's one step.

 

Host: Mm-hmm. Okay. But this seems like a drastic step. This seems like a complete shift in the way that we're designing these systems, right?

 

Guest: Correct. And this is the exciting part. A DC grid will require you to change a couple of things, but it will come with huge benefits. You talked about efficiency. We talk about efficiency at multiple levels. You have converter efficiency, but you have also losses in conduction cables. Yeah. If you look at an AC grid, you typically have 3 phases, which means you have 3 cables, yeah? If you look at a DC grid, you most likely will have 2 cables. You have some with 3 cable configurations, but typically you have 2 cables. Now, the other thing is, if you go to a DC grid, you can also up the voltage. Meaning you're not at the 400 V for 3-phase, but you go to an 800 V system. And that means, to put some simple formulas, you know, it's for engineers, right? So I'm going to use some formulas, right? We say power is equal to voltage times the current, So if you up the voltage, you can down the current. And if you down the current, you need less, you know, circumference in your— diameter, sorry, in your wires.

 

Host: Okay. So smaller cables, basically.

 

Guest: Smaller cables, thinner cables, which is an advantage. Copper prices are high. At the same time, it means fewer thermal losses in the cables, due to electric conduction. You're reducing these losses because you go to higher voltages. At DC, you reduce the current need, you reduce the cable need. So, the cabling of your building becomes simpler. You don't need 3 wires anymore, you use 2 wires. At the same time, you don't need all these stages of converters anymore because now you're going into a DC grid and then you don't need to have intermediate converters anymore. Because even today you have a lot of DC sources. If you have solar close by, fuel cells, we hear a lot about fuel cell fuel cell backup generators in these power centers. All of that today has intermediate converters. Also, your power supplies at the REC, they need to go again, you know, you do AC/DC, then you go again into another stage of conversion. And here you lose stuff. You maintain the whole thing at a DC, and you have a net gain and imagine, recall I said the car factory has about 10 to 15 megawatts. If I have an efficiency improvement of only 1% at 1 gigawatt, that's 10 megawatts.

 

Host: That's a whole factory's worth of savings.

 

Guest: That's a whole factory worth of savings.

 

Host: Exactly. Okay. Yeah. So, but the way you were explaining it also, it's not just about energy losses, but also the material costs, right? You mentioned the copper. You're simplifying the architecture with this new. You also mentioned that this architecture would require a lot of steps. Are we already seeing this right now in the designs we're seeing for hyperscalers, for AI data centers? Are people already implementing these new architectures?

 

Guest: Yeah, to be frank, if you talk about the scale of these powers, it is really a global project. It's a global project with a large complexity and it is really an infrastructure paradigm shift, right? So we're talking, we're talking about something that will not heterogeneously be driven right away everywhere. There will be different stages, and different stages will require changes in the way we— certain systems that we use for power distribution, power protection. And yes, we do see a big push towards this 800 V architecture. Some of the leading AI companies, some of the leading providers of AI computation, have 800 V on their roadmaps talking about DC. It is a reality. It is not a theory from, from engineer, passionate engineers. It's, it's a reality and it will come and it might be in different incarnations, and it might require one or two additional intermittent steps, but it will come and it will require specific solutions which are all semiconductor-based to facilitate that 800 V move.

 

Host: Okay, so you describe basically a journey. We're on a journey as a, as a society to deal with this issue or this challenge of all of this power demand. What are we doing in the meantime? I mean, what do our data centers look like today? What's essential today and what's going to be essential in the future.

 

Guest: Yeah, that brings me to the systems and to the semiconductors. How do we look today? I said today we are mostly in a classical common state, sort of a little bit on steroids, I would say. So what we do is we take things like bunch of gas turbines and, and we add them to the grid connection and so on, because it's not easy to get a 1-gigawatt grid connector. Connected. Yeah, it takes 7 to 10 years if you go by, uh, standard acquisition paths to, to get an official grid connection. What we do today is we have classical transformers, we have all these AC/DC converters, we put gas turbines to, to stabilize the local energy generation. Um, some people are adding renewables, but you would need a lot of renewable areas, right? Solar, you would need huge area to cover that to get these powers. And we utilize things like STATCOMs, which are electronic grid stabilization, to sort of make sure that failures in the data center do not back affect the grid or problems with the grid don't affect the data center. These are today. And you asked me about the future, and that's really the exciting part because we have an obligation to make sure that the grid is stable infrastructure. And at the same time, we have the obligation to make sure that there is as little load on this infrastructure as possible because many things depend on the grid, right?

 

Host: We all do. Our lights.

 

Guest: Our lights, our hospitals, and so on. And we want to be as sustainable as possible with efficiency. The move to 800 V will potentially go in steps. But one of the exciting things that's happening already today with some of our customers is we are utilizing new high-power Semiconductor solutions, 6.5 kV, IGBT technology, bipolar technology, all in the Infineon portfolio. One of our real unique assets, there is barely any semiconductor company out there, and I would say actually none other than us, that will cover from bipolar technology all the way to the low-medium voltage technology at the motherboard. And exactly at that high-power part, we're already developing with some of our partners something called eSTATCOMs, which are the stabilization electronic components, infrastructure components, which are extended by energy storage backup. So not usually batteries, but very fast-reacting supercapacitors and they can react a lot faster to destabilization of the grid. Yeah, we have UPS systems at the data center side to make sure that we have uninterrupted power. Those are another set of, um, backup power supplies with energy storage capabilities. We're seeing a large growth in these segments, and these power supplies, their conversion part is semiconductor-based. And we're seeing a move of these semiconductors from classical silicon-based semiconductors to wide-bandgap-based semiconductors like silicon carbide, which improves their efficiency. And this is sort of like the intermediate step, right? We're moving to three-phase, more power, eSTATCOMs, more efficient UPSs, uninterrupted power supplies. All of these things make sure both the data center and the grid are in balance if something happens, while around it we have gas turbines and whatnot. If we look into the next stages, a lot more exciting things are going to happen as we move towards that DC architecture. Our grid protection is going to substantially change. We're going to go towards something called solid-state circuit breakers. Now they are the evolutionary step from your classical mechanical breaker.

 

Host: Like what you have in the house when you blow a fuse, correct?

 

Guest: Yeah, okay, that thing you go and kind of flip back on. Exactly. Now just a little bit bigger, right? And you— they, they have, you know, there are different levels on the factory floor. Even in your house there are different levels, you just don't see it. You usually know the fuse in your little box, but there are also fuses at the house level, and then there are fuses at the, at the district level, and so on. Same in a factory, right? You got different levels of breakers. And you will see that as we move towards grids that are more— need to react faster, grids that have a certain fault, need to have certain fault resilience, you will move towards solid-state circuit breakers.

 

Host: Because they're faster, right?

 

Guest: Absolutely. Because they're faster and smarter. They're inherently semiconductor-based. That means you're not using an electromechanical component anymore or a vacuum breaker or an air gap breaker, but you're using a semiconductor. And here, Infineon provides state-of-the-art silicon carbide-based JFET, a brand new developed technology. There's none parallel to that on the market, particularly in the breadth of power levels and packages that we're bringing. This will facilitate building these solid-state breakers. And as you correctly said, they can react faster. They react at the speed of semiconductors and not at the speed of mechanical relays. They don't arc because there is no mechanical contact to arc. And if you go to DC grids, you will have them throughout your building to react at different stages. They can be bidirectional, meaning they can facilitate power movement in bidirectional, protect in bidirections. They can be intelligent monitoring power. Monitoring failure is faster, can be centralized remotely. You can even use AI to look at your power usage and so on. So practically, the infrastructure they are protecting and enabling can be used to optimize it or similar infrastructure. And then that's a component throughout the building and even at the entrance of the building. And then I talked about these transformers, right?

 

Host: Yeah, these big metal things.

 

Guest: Beautiful description. These big metal things, right? Today there are practically a bunch of coils. Magnetic cores sitting in an oil bath and some connection.

 

Host: And you see them outside of cities, like the big, big transformers.

 

Guest: Exactly. You see them at, you know, where you have this big power lines connect to a to the town part or to the factory. You see them outside of factories. Today, if you look at a data center, you will see them at regular steps in front of these, at the wall of these data centers where power comes in. And the reason is wherever they come in, then the power goes into the building and is then distributed along all these conversion chains. Yeah. In a classical AC grid. If we look into the into the near future, one of the ambitions is not only to move to DC grid where you will need the solid-state circuit breakers, but in order to generate the DC grid, you want to sort of merge a lot of these conversion stages into something we call a solid-state transformer. The solid-state transformer practically means semiconductor-based medium voltage to whatever grid you need transformer. And it does not rely anymore on this big bulky metal box that you so nicely described, but it operates at a higher switching frequency. Hence, you have high frequency smaller transformers there, and it directly goes from the AC to the DC that you need for your entire building floor. And One of the visions is that you kind of directly come out with cables from these Solid-State Transformer boxes, and you can directly connect to your server rack. What's important for these guys? Real estate area.

 

Host: Yeah. Okay. So you can make the whole, I don't know, project smaller.

 

Guest: Precisely. You can improve the space that you have, the footprint that you have for for server racks, which is the brain and the moneymaker of the data center, while the conversion is brought into a box that can be put into the ground floor or outside the building or even next to the rack and directly go and connect to the rack.

 

Host: Yeah. Okay. And with this, we're eliminating, as you mentioned, all of those conversion stages. We're simplifying the architecture.

 

Guest: Correct. You gain about 1.5 to 1.5% of efficiency by reducing the number of conversion stages. Yeah, a thought fallacy would be to compare the classical transformer to the solid-state transformer because the solid-state transformer contains a lot of these conversion steps. And if you look at the total chain from a classical transformer to the REC power supply versus what the solid-state transformer does, you're gaining 1-1.5% efficiency. That's your factory there.

 

Host: Yeah, that 1.5% of the 1 gigawatt is quite a bit.

 

Guest: Exactly. And at the same time, you're buying yourself more floor space for servers. You're allowing yourself to build the architecture differently. And now you also can natively connect your DC sources, your fuel cells, your solar. If you happen to be able to build a big solar farm, if you happen to have wind

 

Guest: It also sort of changes the way you build UPSs. You don't really— if you integrate the ESS with the SST, it practically intrinsically becomes a UPS.

 

Host: Okay.

 

Guest: So excellent point. It completely changes your flexibility in architecture and the amount of things that you need to build it.

 

Host: So what are you most excited about seeing? When are we going to see this? What's your, what's your take?

 

Guest: Well, there are many things that we're excited about. All of these things are revolutionary. In academia, they've been around for a long time, but it's one of these things that people said, why would we need them? The transformer works perfectly.

 

Host: Why would you ever need 1 gigawatt of energy in one place?

 

Guest: Exactly. Yes. Why would you need 1 gigawatt at one place? Why do you need your circuit breakers to work perfectly. Why do you need that? So it's exciting to see a paradigm shift in the way we build grids, in the way we transport power. At the same time, it is exciting to see that our state-of-the-art technology is what's needed to facilitate that.

 

Host: Semiconductors.

 

Guest: Precisely.

 

Host: So semiconductors are really driving, or they will be the foundation then for these new shifts in architecture.

 

Guest: Semiconductors are the enabling components for these architectures. We're not talking about metal bending anymore. We're really talking about high-tech control algorithms. We're talking about state-of-the-art semiconductors. We're talking about innovation in package and isolation that enables these things. And if you look at a solid-state transformer, if you look at ESS, if you look at solid-state circuit breakers, solar, wind, hydrogen fuel cells, If you look at the power, the high voltage power conversion at an intermediate bus converter of a future 800 V system, all of these steps, they're all enabled by semiconductors. If you look at an eSTATCOM, it's enabled by semiconductors. And this is really an exciting part. It's the shift from bending metal and connecting coils to radically shifting into utilizing state-of-the-art semiconductors. And here is where innovation will lead the pack. And that's the strength of Infineon.

 

Host: Well, as you mentioned in the beginning, this is a journey and it sounds like it's an exciting journey to be on. Thank you so much for sharing your expertise on this topic. It was really nice talking to you about what these future hyperscaler projects could look like with our semiconductors.

 

Guest: Yeah, thank you very much for having me. It was a pleasure.

 

Host: And thank you to our listeners. As we've heard, stabilizing the grid and efficiently delivering power at gigawatt scale will become more and more important, and that depends heavily on advanced semiconductor solutions. To our audience, thanks for listening to the Podcast4Engineers, and stay tuned for our next episode on the We Power AI series, where we will continue our journey.