Infineon key visual in RGB

The server rack evolution toward 3-phase PSUs

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 23, 2026

Podcast

In this episode of Podcast4Engineers' "We Power AI" series, host Kelsey Markl sits down with Dr. Diogo Varajao, Head of Power Supply and Battery Backup Systems Group at Infineon, to explore how power supply design is rapidly evolving to keep pace with the soaring energy demands of AI infrastructure. The conversation dives deep into the dramatic rise in GPU and server rack power consumption, from under 1 kilowatt per GPU just a few years ago to projections exceeding 600 kilowatts per rack in the near future, and what that means for data center architecture. Diogo walks through the generational shift in rack design, as well as novel energy buffer concepts, advanced PFC topologies, and the strategic use of WBG technologies to maximize efficiency and help hyperscalers drive down their Power Usage Effectiveness (PUE). 

Listen on Spotify Watch on Youtube

Speakers

In this episode of Podcast4Engineers, host Kelsey Markl speaks with Dr. Diogo Varajao, Head of the Power Supply and Battery Backup Systems Group 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.

Diogo Varajao

Guest:

Dr. Diogo Varajao is Senior Director and Head of the Power Supply & Battery Backup Systems Group at Infineon Technologies, leading a global team of 20+ engineers focused on application strategy, reference designs, and product roadmaps for datacenter, telecom, and satellite communications. He holds a Ph.D. in Power Electronics from the University of Porto and received the SEMIKRON Young Engineer Award 2018 for his patented ACDC CUBE technology. A Senior Member of IEEE and APEC speaker, his work at Infineon focuses on powering AI data centers from grid to core, leveraging silicon, silicon carbide, and GaN technologies to advance next-generation high-efficiency power systems.

Additional information

More episodes

Guest: The GPU power is increasing a lot. A few years ago, a GPU was consuming less than 1 kW. Today it's about 1.8, and next year will be 3.6. It's increasing very fast. And then it's also increasing the total power demand of the rack. A few years ago, we were talking about 30 to 60 kW per rack. Today, between 150 kW to 250. And very soon it will be more than 600 kW per rack. At Infineon, we developed a quite novel energy buffer concept that we can integrate inside the PSU between the PFC and the LLC stage. And with this energy buffer, so it's an active circuit, we can basically make better utilization of the energy that is stored in the DC link capacitors. And in this way, we can buffer these big loads from the GPU.

 

Host: Welcome to the Podcast4Engineers and to another episode in our We Power AI series. My name is Kelsey Markl, and today we're diving into a big shift happening behind the scenes of AI infrastructure: how power supply design is evolving to meet soaring compute demands. And to walk us through it, I'm joined by my colleague, Dr. Diogo Varajao. He's the head of Power Supply and Battery Backup Systems Group. Diego, thanks so much for being here.

 

Guest: Thanks for having me.

 

Host: Okay, Diego, let's get started. Let's get right into it. We have talked in previous episodes with some of our colleagues about the big picture, how we're dealing with the increasing power demands of data centers at a bigger scope. But let's focus today on the server rack. What challenges are we seeing here in the server rack?

 

Guest: Yeah. The GPU power is increasing a lot. A few years ago, a GPU was consuming less than 1 kW. Today it is about 1.8 and next year it will be 3.6. It's increasing very fast. And then this is also increasing the total power demand of the rack. A few years ago we were talking about 30 to 60 kW per rack. Today between 150 kW to 250. And very soon it will be more than 600 kW. This is a big change in terms of power. And another challenge is that the GPUs need to be put closer as much as possible in order to increase the computing density inside the rack. And this will allow us to reduce the cabling length of the rack in order to achieve high bandwidth within the GPU cluster of the rack. And actually, there is also an evolution. Traditionally, the GPUs and the CPUs in the motherboard were mounted in a compute tray, horizontal place.

 

Host: Okay. Something likes to pull out.

 

Guest: Exactly. It is easy to pull out horizontal. And actually, now there is in the next generation, they will be mounted as a blade. These will allow us to actually pack more GPUs inside the rack and in the end increase the computing density inside the rack to have many more GPUs there. And of course, if you have more GPUs, you also have more power, which means in the end we'll have also more current flowing in the busbars that are supplying all these GPUs inside the rack. And as a consequence, it means that you need to have busbars that are thicker and in the end, they will require more copper.

 

Host: Okay. Copper is pretty expensive, I hear.

 

Guest: Yes, copper is quite expensive, right? At some point you simply hit a wall that you cannot go beyond a certain power level, right? So today, air-cooled busbar can allow to flow roughly 300 kW with a 50 V distribution, right? And in order to push it further, you can use liquid cooling busbars to bring it beyond 600 to 700 kW. But then it's really the limit, right? And of course, these liquid-cooled busbars are also expensive, right? These are some of the challenges that we are facing today inside the rack architecture.

 

Host: Okay, so before we start looking at what's coming next, what are we looking at today? Can you explain a little bit more what a 50 V server rack looks like?

 

Guest: Yeah, for sure. The current rack with the 50 V, basically we have all the systems inside a single cabinet. We have compute trays with the GPUs and CPUs inside. We have network switches for communication inside the rack. We have also then the power shelves with the PSUs. And here could be single-phase PSU with 5.5 kW or 12 kW. And we see now an evolution towards 3-phase power supplies in these traditional racks with roughly 18 kW. The PSU is then responsible for establishing this 50 V inside the rack and powering the busbars. And in the end, if you also want to have, for example, battery backup units inside the rack, then you have even less space for computing, for the computer trays. Right. And therefore, space is a constraint now in order to scale computing capabilities. And the other topic is, as I mentioned before, power losses inside the rack in order to power all the GPUs. We simply hit a limit. We need to change the architecture of this rack.

 

Host: Okay. Yeah. From the way that you described it, it sounds like it's getting really crowded here in the server rack. What does this— what needs to change? You mentioned the architecture. What does that actually look like?

 

Guest: Yeah. In the end, we, in the next generation, what we call Gen 2, will basically disaggregate computing from power, right? We'll have a dedicated rack, what we call IT rack, where we'll have all the computing trays and also the network switches inside. that truck. And then we have the power supplies and the battery backup unit in a separate truck that we call power sidecar. In that truck, we have 3-phase power supplies with 25 kW each or more. And these power supplies will then establish the high voltage DC bus inside the power sidecar. This is also a change. We move from 50 V towards 800 V or plus minus 400. So, there are these 2 variants, right?

 

Host: Okay. The racks still look the same. It's just in one you said you have all of the IT, the GPUs. And then the other, we're putting all of the power supplies.

 

Guest: Correct. So, we have in the end 2 racks in this Gen 2 architecture. And by moving towards high voltage DC, it will allow us to reduce a lot the currents that flow on the busbars. And in the end it will allow us to save copper. There are some studies that we can, for example, save roughly 45% of copper amount compared with Gen 1 architecture.

 

Host: Okay.

 

Guest: This is one advantage. Another advantage is also that we can have much better thermal management inside the power sidecar. That is another benefit. And another aspect I think is quite important to highlight is that GPUs consume power with huge spikes.

 

Host: Okay.

 

Guest: The GPU power can go up instantaneously 180% of the nominal power. Right. And if we don't have, in case we don't have anything to filter this, then this will propagate towards the grid and create a lot of disturbances in the grid. We need to have also in this case the so-called capacitor bank unit, CBU, inside the sidecar that can in the end buffer all these spikes and ensure that we are not disturbing the grid. This is also another change because with increase of power, then we really need to take care of these big loads of the GPUs.

 

Host: Okay. Yeah, that's a conversation that we've had a lot is how we stabilize the grid and protect the grid with all this big rise in data centers. You mentioned, okay, so now that we have the power supply in a separate sidecar, I guess we can really optimize things for power. Does that mean that we can also look at different topologies?

 

Guest: Yeah, absolutely. The change also to 3-phase PSUs means we need to use 3-phase PFCs. And here the most commonly used topology is the Vienna rectifier. It's well known, it was used for several years in EV charging modules for charging stations, for example. It's quite proven robust. This was the first topology to be adopted by the ODMs, but also with the increase of efficiency requirements, there is a need to change and to use also other topologies like the T-type PFC or, for example, 3-level flying cap PFC or even active neutral point clamp or ANPC, so-called ANPC. So, with these topologies, we can really push efficiency further. And for example, my team has been developing several reference designs, and we've proven, for example, T-type PFC can increase efficiency by 0.2% compared with a Vienna rectifier. And for example, a 5-level ANPC can even increase efficiency by 0.4%. And, you know, for the hyperscalers, every 0.1% matters because this is reflected directly in the utility bill, right? And their OPEX. That's why we see interest from our customers also in using our products and in particular these solutions from our reference designs in order to increase efficiency.

 

Host: Mm-hmm. And when we're— so we've covered topologies, but what about materials? I know that when we're talking about these high-power dense applications, often we turn towards silicon carbide. Is that what we're looking at here in these 3-level?

 

Guest: Yeah. In the 3-phase PSU, actually the 1200 V and 650 V silicon carbide MOSFETs are well-suited technology to be used there because silicon carbide offers a very high thermal conductivity when compared with silicon and GaN. That's why we see a lot of demand for our silicon carbide products in 3-phase PSUs. But of course, it's a bit of a blend, so other technologies can also be used. We see silicon being used, for example, in the 3-phase 50-volt PSU for the Gen 1 architecture, for example. And now there is also a lot of interest in GaN to use it, for example, in the back-to-back switches of the Vienna rectifier or the T-type, right? Because With our GaN BDS, since it is a monolithic device, we can replace 2 or 4 MOSFETs for these back-to-back switches, which is then translated in PCB area savings, also allowing us to increase the power density of the power supply and to save overall costs. Right. We have less devices there in the power supply. This is one, one key use case for GaN in the PSU. And, in the traditional, let's say, 3-phase 50 V PSU, we also see now a lot of interest from customers to use 80 V GaN into the synchronous rectifier because GaN allows us to have to reach lower RDS, and RDS is super critical in the SR. because it's directly impacting the losses. And another aspect is also GaN has better gate charge parameter. So in that case, we can also reduce further the losses, right? In the end, for 3-phase PSUs, for sure silicon carbide is very well suited, but I think we can also use silicon and GaN in specific places inside the topologies.

 

Host: Yeah, it sounds like there's an opportunity for everything. For sure. But I assume you're talking mostly about discretes from—

 

Guest: Yes. In the current designs, mainly discretes are used. For example, bottom-side cooling devices like the TOLL, it's quite popular. And if we talk about top-side cooling packages, the Q-DPAK is also getting now a lot of interest from our customers. In particular, Q-DPAK can dissipate more power per package. This is also super interesting as we scale the power of the PSU. Yeah. So discretes at the moment are the most preferred solutions.

 

Host: But as we're talking about going to more and more power and more and more power density, is there a tipping point where we would say a module makes more sense?

 

Guest: Yeah. We see also some projects for liquid cooling PSU and we have then some customers asking about power modules, right, to be used in that liquid cooling PSUs because power modules offer also a higher level of integration. Right. And in this case, these projects, we are talking about more than 50 kW in a PSU, right? So power is really increasing in that sense. So that's why there is some interest. Of course, it can also be implemented with bottom-side cooling or top-side cooling discrete devices. But in the end, it's a decision of the R&D team of the OEMs to decide which package to use. Right. And here at Infineon, actually, we can offer a very complete portfolio to our customers. We really enable them to choose what best suits their requirements.

 

Host: I mean, it sounds like that's an innovative or forward-thinking opportunity to use modules. Is there any other innovation happening in that space?

 

Guest: Yes. As I mentioned before, the GPU consumes power with these IP clothes, right? And actually, at Infineon, we developed quite novel energy buffer concept. that we can integrate inside the PSU between the PFC and the LLC stage. And with this energy buffer, so it's an active circuit, we can basically make better utilization of the energy that is stored in the DC link capacitors. And in this way, we can buffer these peak loads from the GPU. And this offers the opportunity, for example, to replace the CPU. To not have a CPU in the sidecar by having this buffering feature implemented in the PSU. And in this way, we can actually ensure that we are not disturbing the grid. And if we compare, for example, with the alternative solution, that would be to have simply more capacitors inside the PSU, right, in order to do this buffering. And that is a solution that was implemented by some ODMs for the 18 kW, 3-phase, 50 V PSU for the Gen 1. We can, for example, save up to 50% the amount of capacitors inside that PSU. In the end, we're talking about, first of all, saving cost and saving also space. Meaning that we can reach higher power density with this solution. We have implemented it also in our 18 kW reference design and we see a lot of interest from OEMs basically to use our solution in their designs.

 

Host: Okay, that's really exciting. We talked about what we have today, let's say Gen 1, and then this 3-phase PSU would be Gen 2. Is there— is this the end of innovation in the space? Will there be a Gen 3?

 

Guest: Yeah, good question. Yeah. In the end, a 3-phase PSU is for sure a key landmark in this architectural evolution, right? We have PSU in the first generation, 50 V. Now we have that also in the Gen 2 in the power sidecar, but we are now talking about Gen 3. And in Gen 3, it means that we'll have a centralized system at the data center infrastructure level that will generate these high voltage DC buses. And that 800 V will be brought directly into the white space, into the data center room to supply the racks directly with these 800 V or ±400 V, right? In that way, it means that in Gen 3, there will be no longer these 3-phase PSUs inside the data center room. Right. And they will be replaced by these centralized systems.

 

Host: Okay. So maybe no more sidecars, but the idea of the sidecar would then be centralized and would feed the racks directly. What does that look like? What is a centralized system?

 

Guest: Good question. I think there are many implementation approaches. So one approach is to have the so-called medium voltage rectifier. It's still a low frequency transformer plus an active front end. And this active front end is then generating the high voltage DC. Another alternative is to use a solid-state transformer, and this solid-state transformer is implemented with high frequency transformers plus a lot of power electronics. Right. And this allows to reduce the space and the amount of materials to implement the solid-state transformer compared with a traditional low frequency transformer. These are, let's say, 2 of the many options to implement this so-called Gen 3 architecture.

 

Host: Okay. The centralized solution seems pretty nice on paper, or as you described it. Why not go directly to Gen 3? Why this step with Gen 2?

 

Guest: Yeah, very good question. I think there are some regional aspects that we can highlight. For example, today in China, their data centers, they already run with DC voltage. Okay. 240 V DC. I believe China will be faster to implement Gen 3 solutions. And actually, we see, for example, some pilots of medium voltage rectifier being implemented in some of the hyperscalers in China. This is already happening now. But in the end, there are also other aspects that we need to take into account, right? For example, SST is still a technology that is in development, is not yet fully mature. It will take still a couple of years. And this is just one aspect. Another aspect I think is super important is about standardization, because running a data center with 50 V has completely different safety requirements compared with running it with 800 V or ±400 V. Which means that we need to have standards. to define how this is implemented. And another aspect is also to train the personnel that will have to be inside the data center room to operate it, right? Because they need to be trained, they need to be certified for safety reasons. And this has also to happen, right? There are multiple aspects that are still on the way to implement or to jump directly to Gen 3. That's why in the end, we need to scale the power today now. In the end, we have to use proven solutions, which means single-phase and 3-phase power supplies allow us to implement this, being it with the Gen 1 at 50 V or then with Gen 2 high voltage DC architectures. And then as we go, then there will be for sure a transition towards the Gen 3 with a centralized system.

 

Host: Okay. So, we've really been painting a picture today of Gen 1, which is already here, Gen 2, which is coming soon, coming now, and Gen 3, which is right on the horizon. But what else is coming next?

 

Guest: Yeah. The deployment of massive data center hubs and the so-called AI gigafactories, right, with power consumption beyond 1 gigawatt will basically mean that we have something equivalent to a medium-sized town. Which means we need to have very efficient data centers in order to save energy and resources in the end, right? So, there is one factor that is the so-called PUE, Power Usage Effectiveness, is a ratio that measures, for example, how much power is delivered to the data center to operate the data center compared with the consumption by computing the GPUs and the CPUs inside the data center. That is in the end what is useful for the data center, right? So today the average PoE is between 1.6 and 1.8, right? The hyperscalers are really pushing the industry to offer solutions in order to bring the PoE below 1.2. So, in the end to increase efficiency, right? So that's why it's super important to use very high efficiency power supplies in order to contribute to improving the PoE. And so also in the end with our system solutions and our products, we can for sure help to contribute to reduce the PoE.

 

Host: Okay. It sounds like an ambitious goal, but I think we're on the right track. You mentioned several times a reference design and some other materials that I think might be interesting for our listeners. If they want to find out more or check into, look into this reference design, where can they find that information?

 

Guest: Oh yeah, it's very easy. They can just go to our website, www.infinion.com/aipsu. It's very easy to remember, and then they can find out all the content about our reference designs.

 

Host: Okay, thank you. Thank you so much, Diogo, for being here.

 

Guest: Thanks for having me.

 

Host: And to our listeners, thanks for listening to the Podcast4Engineers and our WepowerAI series. If this episode on 3-phase PSUs has sparked your interest, tell a fellow engineer and check out the rest of our series on more for AI power demand, architectures, efficiency, and the future of AI power at scale. See you in the next episode.