# A Safe, Efficient, and Scalable Approach to 800 VDC Architecture (Presented by Eaton)

## Abstract

The growing demand for high‑performance computing and dense GPU deployments is accelerating adoption of DC power distribution in next‑generation data centers. By eliminating rack‑level AC conversion, DC architectures improve efficiency, reduce space overhead, and enable higher‑voltage distribution (800 VDC), increasing overall system scalability and lowering losses. Within this landscape, Eaton plays an active leadership role in the Open Compute Project (OCP) community, and in the development of Underwriters Laboratories (UL) and International Electrotechnical Commission (IEC) standards, helping shape industry frameworks that enable the safe deployment of high‑voltage DC architectures at scale. This standards‑driven approach ensures that emerging DC ecosystems are interoperable, fault‑tolerant, and aligned with real‑world operational requirements for AI and HPC environments. Building on its deep power management expertise, Eaton focuses on enabling higher power densities safely. By combining active standards leadership, and a laser focus on safety, Eaton is helping establish the foundational power architectures required for the next generation of AI and high‑performance computing data centers.

## AI Summary

- The presentation highlighted the rapid advancement in data center technology comparing a 3KW rack from the 1990s to a 17KW rack by 2020 and the transition to liquid cooling and co-packaged optics
- The focus shifted to the modular data center approach emphasizing the medium voltage solid state transformer as a key solution for high-density racks reducing costs size and maintenance needs
- The presentation discussed the importance of direct current (DC) distribution in data centers including the potential for up to 30% reduction in power requirements and the need for different distribution methods (busway vs DCRPP) based on the specific sub-segment of the data center industry

## Transcript

**[00:00:04 – 00:00:08]** Hello everyone. A little bit of a technical difficulty.

**[00:00:08 – 00:00:13]** It's a good thing I brought my pinnacle of technology of

**[00:00:13 – 00:00:18]** the 1980s, which is out in the 90s.

**[00:00:18 – 00:00:21]** If you know what this is, please raise your hand.

**[00:00:21 – 00:00:24]** If you used it, raise your hand.

**[00:00:24 – 00:00:28]** So you guys probably can also use a stick shift or

**[00:00:28 – 00:00:30]** a manual transmission.

**[00:00:30 – 00:00:32]** All right.

**[00:00:33 – 00:00:38]** Now, luckily, I brought my laptop or my iPad as a backup.

**[00:00:38 – 00:00:40]** But I also have a couple other things.

**[00:00:40 – 00:00:46]** We used to transfer data on one of these 3.5 floppies.

**[00:00:46 – 00:00:50]** But today, when I plugged in,

**[00:00:51 – 00:00:57]** I have my multi-port, solid-state transformer, my DC distribution

**[00:00:57 – 00:01:02]** cable, and luckily I brought my iPad because now I have

**[00:01:02 – 00:01:03]** a display to go on.

**[00:01:03 – 00:01:06]** The show always has to go on, everyone.

**[00:01:06 – 00:01:10]** So the difference between this technology here that was in the

**[00:01:10 – 00:01:14]** 1990s to my iPad today is 30 years.

**[00:01:15 – 00:01:19]** That 30 years of innovation is fantastic.

**[00:01:19 – 00:01:22]** What we're about to do in the industry and have been doing

**[00:01:22 – 00:01:29]** is going from an AC input device that takes 110 or 240, 50 or 60

**[00:01:29 – 00:01:37]** Hertz to a DC device in three years in the time it took us to go 30.

**[00:01:40 – 00:01:42]** So we're gonna talk about the approach to the direct

**[00:01:42 – 00:01:46]** current and where we're going with this in the industry.

**[00:01:46 – 00:01:49]** So the workloads have been increasing more and more and have

**[00:01:49 – 00:01:52]** been less and less fault tolerant.

**[00:01:52 – 00:01:56]** Most of this first principle design comes back to the network.

**[00:01:56 – 00:02:02]** That network is what drives all data center design.

**[00:02:02 – 00:02:07]** And what's been happening in the last five years is packet loss.

**[00:02:08 – 00:02:11]** And that packet loss has driven a convergence.

**[00:02:11 – 00:02:15]** That convergence started back in 1990 with my wonderful

**[00:02:15 – 00:02:18]** laptop that I have now.

**[00:02:18 – 00:02:20]** With a 3KW rack.

**[00:02:20 – 00:02:24]** That 3KW rack per tile, 600 millimeters wide, 1200 millimeters

**[00:02:24 – 00:02:30]** deep, 48 RU high, was 3KW back in the 90s after we standardized

**[00:02:30 – 00:02:33]** the TCP IP protocols in the 80s.

**[00:02:33 – 00:02:37]** That existed and lasted quite a bit, all the way to about 2020.

**[00:02:37 – 00:02:39]** That 2020, we had a 17KW rack, H100, maybe an H200,

**[00:02:39 – 00:02:42]** it was beautiful.

**[00:02:44 – 00:02:50]** You had about 20 megawatts to get you to about 10,000 GPUs.

**[00:02:50 – 00:02:54]** Those 10,000 GPUs would monetize about 3.4 million a week.

**[00:02:54 – 00:02:58]** What we learned earlier in the keynote is talking about

**[00:02:58 – 00:03:01]** the need for the convergence.

**[00:03:01 – 00:03:05]** That convergence started to begin in about 2020 to 2022.

**[00:03:05 – 00:03:07]** It's something called liquid cooling.

**[00:03:07 – 00:03:13]** Liquid cooling transition occurred not because of a thermal problem,

**[00:03:13 – 00:03:14]** But because of a network problem.

**[00:03:14 – 00:03:18]** Jensen talked about it quite a bit, about the MVL72 link,

**[00:03:18 – 00:03:24]** condensing those 72 GPUs, or nine H100s, into one rack to mitigate

**[00:03:24 – 00:03:26]** and be within the copper domain.

**[00:03:26 – 00:03:28]** Now that we have co-packaged optics, that's still going

**[00:03:28 – 00:03:32]** to support, but now we have a power convergence.

**[00:03:32 – 00:03:36]** So, would anybody go back to the 1990s technology and love

**[00:03:36 – 00:03:39]** to carry it around with me today?

**[00:03:39 – 00:03:42]** I got one, all right, good.

**[00:03:44 – 00:03:49]** So, there's a form factor benefit from going to removing the AC to DC

**[00:03:49 – 00:03:56]** conversion device out of your tower desktops to bring it into a device

**[00:03:56 – 00:04:00]** that's outside of the form factor.

**[00:04:02 – 00:04:07]** That form factor occurred because of the convergence for the network.

**[00:04:07 – 00:04:10]** What we learned today and what went public was the co-packaged

**[00:04:10 – 00:04:13]** optics, which is the next convergence point.

**[00:04:13 – 00:04:17]** The next convergence point associated with it is in the

**[00:04:17 – 00:04:21]** near future, and we learned about it this morning.

**[00:04:21 – 00:04:24]** Today what we're talking about is the first portion of that, and

**[00:04:24 – 00:04:28]** it's the shift into direct current.

**[00:04:29 – 00:04:31]** That shift into direct current comes into stages.

**[00:04:31 – 00:04:36]** The modular data centers are needed into the five different blocks.

**[00:04:36 – 00:04:40]** Block one being the grid interconnect to medium voltage

**[00:04:40 – 00:04:43]** with onsite generation, medium voltage to low voltage

**[00:04:43 – 00:04:48]** with power quality, the IT module, where you take power, water,

**[00:04:48 – 00:04:52]** cooling, and fiber, and you bring it together in a control plane,

**[00:04:52 – 00:04:53]** and then you need to cool it.

**[00:04:53 – 00:04:56]** Software, of course, to optimize.

**[00:04:56 – 00:05:00]** Today we're talking about that conversion point of the

**[00:05:00 – 00:05:04]** MV to LV with direct current.

**[00:05:05 – 00:05:09]** And that shift is coming in a few different phases.

**[00:05:09 – 00:05:11]** In particular, that shift of the medium voltage solid

**[00:05:11 – 00:05:16]** state transformer in order to support the high density racks.

**[00:05:16 – 00:05:20]** Those high density racks are significant in the fact that

**[00:05:20 – 00:05:26]** you transform that technology and support it in order to

**[00:05:26 – 00:05:29]** be able to take the device that you plug into the wall

**[00:05:29 – 00:05:35]** and scale it to a power factor that you can use in the data centers.

**[00:05:35 – 00:05:39]** So in the data center applications, it's about how do you apply

**[00:05:39 – 00:05:44]** and produce that direct current at scale.

**[00:05:44 – 00:05:47]** So thinking back to the first block associated with it, the medium

**[00:05:47 – 00:05:51]** voltage, the low voltage transition point, currently we take in

**[00:05:51 – 00:05:59]** anywhere from 9kV all the way up to 500kV, transform that down into the

**[00:05:59 – 00:06:04]** 9kV to 35 and into medium voltage and transition into low voltage.

**[00:06:04 – 00:06:08]** We're in a point where we're transitioning, and I think

**[00:06:08 – 00:06:12]** we're probably year two into this transition, of going

**[00:06:12 – 00:06:16]** to a direct current distribution to support the compute module.

**[00:06:16 – 00:06:19]** The medium voltage solid state transformer is one of our

**[00:06:19 – 00:06:25]** solutions that we use to transform that in a safe, effective manner.

**[00:06:25 – 00:06:27]** The benefit of the medium voltage solid state transformer

**[00:06:27 – 00:06:31]** is you're taking a lot of devices that we used to have

**[00:06:31 – 00:06:35]** disparated across different portions of the value chain.

**[00:06:35 – 00:06:39]** So you used to have a pad-mounted transformer that would transform,

**[00:06:40 – 00:06:44]** then you have the medium main switchboard, a UPS,

**[00:06:44 – 00:06:46]** low-voltage distribution.

**[00:06:46 – 00:06:49]** The benefit of going to a medium-voltage solid-state

**[00:06:49 – 00:06:52]** transport technology is we can combine that technology

**[00:06:52 – 00:06:56]** into one unit, in particular, into the pad-mounted or the

**[00:06:56 – 00:06:59]** modular application.

**[00:06:59 – 00:07:05]** Pad Mount and Modular Application will allow us to move that.

**[00:07:05 – 00:07:08]** In a form factor, it will fit in the back of a truck.

**[00:07:08 – 00:07:12]** Oftentimes, building a data center comes down to speed.

**[00:07:12 – 00:07:17]** Shovel to revenue is currently the focus and the need of the

**[00:07:17 – 00:07:21]** industry, and being able to have that in a modular fashion matters.

**[00:07:21 – 00:07:25]** Also, you have a lot of human capital constraints.

**[00:07:25 – 00:07:29]** The human capital constraints, in particular, comes to

**[00:07:29 – 00:07:30]** on-site technicians.

**[00:07:30 – 00:07:33]** Those on-site technicians are constrained, and this is one

**[00:07:33 – 00:07:38]** of the solutions we think that'll help transition that industry.

**[00:07:41 – 00:07:44]** As you move down that stack, it's going to help you reduce the

**[00:07:44 – 00:07:50]** cost, the size, for you to be able to have lower amount of maintenance

**[00:07:50 – 00:07:53]** and a higher reliability associated with it because there's less

**[00:07:53 – 00:07:58]** equipment associated with that for the install.

**[00:08:01 – 00:08:03]** Thinking about this is important, because then you have to

**[00:08:03 – 00:08:05]** think about the distribution and the protection.

**[00:08:06 – 00:08:08]** These are the things we've been focusing on at Eaton,

**[00:08:08 – 00:08:12]** also through the Open Compute Project, the power distribution.

**[00:08:12 – 00:08:16]** That paper's about to be released in the next two weeks for

**[00:08:16 – 00:08:21]** the first alphabet and taxonomy we have for the industry.

**[00:08:21 – 00:08:24]** Think about it as first the conversion point,

**[00:08:24 – 00:08:27]** then the distribution and protection is important.

**[00:08:27 – 00:08:30]** Those things are actively being built

**[00:08:30 – 00:08:34]** and delivered in the industry.

**[00:08:35 – 00:08:37]** After you go around for the distribution point, you need

**[00:08:37 – 00:08:42]** something, some sort of solid state sidecar or a DC RPP.

**[00:08:42 – 00:08:48]** Once that energy is distributed, you also have to have the ability

**[00:08:48 – 00:08:53]** to swap that out in a fungible manner as the loads change.

**[00:08:53 – 00:08:56]** Currently, rack densities are around 150 kW, looking to go

**[00:08:56 – 00:09:02]** to 250 kW, but we still look like we're going to go to 1.5 megawatts.

**[00:09:03 – 00:09:06]** And having a fungible device that can do that within the breaker

**[00:09:06 – 00:09:09]** and fault domains is important.

**[00:09:14 – 00:09:16]** The key here is you have to think about the different

**[00:09:16 – 00:09:20]** aspects of the distribution.

**[00:09:20 – 00:09:23]** Busway versus DCRPP.

**[00:09:23 – 00:09:26]** And that depends on which segment of the industry you're in.

**[00:09:26 – 00:09:28]** Most people think about the data center industry as a

**[00:09:28 – 00:09:30]** monolithic industry.

**[00:09:30 – 00:09:33]** In reality, there's about nine sub-segments in the data

**[00:09:33 – 00:09:38]** center industry, and the form factor that's appropriate for you

**[00:09:38 – 00:09:40]** Needs to be thought through well, right?

**[00:09:40 – 00:09:44]** So an AI training cluster versus an inference cluster,

**[00:09:44 – 00:09:48]** a co-location versus a cloud hyperscale is different.

**[00:09:48 – 00:09:52]** Understanding that path to navigate is important.

**[00:09:52 – 00:09:56]** And this is some of the thoughts that we have associated with it.

**[00:10:03 – 00:10:06]** The other portion of this is if you've heard about the

**[00:10:06 – 00:10:07]** AI waveform mitigations.

**[00:10:07 – 00:10:11]** Show of hands who's heard about Senate Bill 6?

**[00:10:11 – 00:10:14]** So Senate Bill 6 is much like here.

**[00:10:14 – 00:10:18]** We have about thousands of people in this conference

**[00:10:18 – 00:10:19]** right now together.

**[00:10:19 – 00:10:22]** We're all talking about something different.

**[00:10:22 – 00:10:25]** We did that, and that's something called power diversity.

**[00:10:25 – 00:10:29]** So it's something we've been doing for the last 30 years in the data

**[00:10:29 – 00:10:32]** center industry, where we're all talking a different rack load, a

**[00:10:32 – 00:10:34]** different program being ran on it.

**[00:10:34 – 00:10:37]** With the advent of training AI, it brings everybody to

**[00:10:37 – 00:10:41]** be singing the same sheet of music in the same area.

**[00:10:41 – 00:10:44]** So if we were doing that in this room, the loud voices

**[00:10:44 – 00:10:47]** would go up and down together.

**[00:10:47 – 00:10:49]** This causes

**[00:10:50 – 00:10:54]** AI waveform causes oscillations into the grid, and those

**[00:10:54 – 00:10:58]** oscillations into the grid are something that can damage

**[00:10:58 – 00:10:59]** the grid as well.

**[00:10:59 – 00:11:03]** So, when we're thinking about that, we have technologies

**[00:11:03 – 00:11:06]** and we've released a paper, Good Grid Citizens, since

**[00:11:06 – 00:11:09]** I only have 15 minutes to talk to you today, is a paper

**[00:11:09 – 00:11:11]** that we discussed through there.

**[00:11:11 – 00:11:14]** That's how we think that it would be a good way to go

**[00:11:14 – 00:11:17]** through and mitigate those technologies with some technologies

**[00:11:17 – 00:11:19]** we have called Energy Aware.

**[00:11:20 – 00:11:21]** The last portion, so we got a little bit of a late start,

**[00:11:21 – 00:11:24]** so I need to rush ahead.

**[00:11:24 – 00:11:27]** I want to leave with you and think about it this way.

**[00:11:27 – 00:11:31]** Right now, the top of the illustration is where our

**[00:11:31 – 00:11:32]** current state is.

**[00:11:32 – 00:11:36]** We currently have a conversion point with a pad-mounted

**[00:11:36 – 00:11:38]** transformer that will convert from medium voltage to low

**[00:11:38 – 00:11:42]** voltage using a UPS, which I highly recommend because it helps

**[00:11:42 – 00:11:45]** with the AI waveform mitigation.

**[00:11:45 – 00:11:48]** And then it goes to the rack itself.

**[00:11:48 – 00:11:50]** That's the classic example that we have.

**[00:11:50 – 00:11:55]** There's been some value engineering for some data centers that

**[00:11:55 – 00:12:02]** remove the UPS, which then exposes the racks to raw utility power.

**[00:12:02 – 00:12:05]** The next step or phase, we see the sidecar, right?

**[00:12:05 – 00:12:09]** So if you go out to the vendor wall with NVIDIA, you can

**[00:12:09 – 00:12:11]** see our sidecar that we have.

**[00:12:11 – 00:12:14]** That's the next phase in this development that we see, or

**[00:12:14 – 00:12:19]** stage gate, in which if you have a existing deployed capital and

**[00:12:19 – 00:12:23]** an existing data center, this is one solution you can use to support

**[00:12:23 – 00:12:25]** the new direct current loads.

**[00:12:25 – 00:12:30]** As you transition, we feel that going into a medium-voltage

**[00:12:30 – 00:12:33]** solid-state transformer is the next step.

**[00:12:33 – 00:12:37]** So if you're thinking about current greenfield designs,

**[00:12:37 – 00:12:40]** I would really think about and consider how would you

**[00:12:40 – 00:12:44]** use a medium-voltage solid-state transformer for DC distribution.

**[00:12:44 – 00:12:47]** In the OCP white paper, we talk about it happening in

**[00:12:47 – 00:12:48]** a couple different stages.

**[00:12:48 – 00:12:53]** First, power to the actual IT device itself.

**[00:12:53 – 00:12:57]** Next to the overall IT module, which inside of that IT

**[00:12:57 – 00:12:59]** module, you can think of it as like the white space.

**[00:12:59 – 00:13:03]** So if the CDU's in there, anything that would be powered

**[00:13:03 – 00:13:06]** would be on that DC.

**[00:13:06 – 00:13:08]** 70% of the world is already direct current.

**[00:13:08 – 00:13:10]** So this is a transition.

**[00:13:10 – 00:13:14]** If you think about your variable frequency solid state, variable

**[00:13:14 – 00:13:17]** frequency drives that you use in your cooling applications,

**[00:13:17 – 00:13:21]** it takes it from AC, goes back to DC, to step DC.

**[00:13:21 – 00:13:23]** The last step is the full facility.

**[00:13:23 – 00:13:25]** And this is where the real benefits come in.

**[00:13:25 – 00:13:31]** We talked about an issue associated with power constraint.

**[00:13:31 – 00:13:35]** In some of the DC applications, you can reduce your overall

**[00:13:35 – 00:13:40]** power requirements by up to 30% by shifting to direct current.

**[00:13:40 – 00:13:43]** A lot of that has to do with how the pumps are utilized

**[00:13:43 – 00:13:45]** and the braking resistors.

**[00:13:45 – 00:13:49]** But we've had in manufacturing applications, particularly

**[00:13:49 – 00:13:53]** in the automotive field, we've seen empirics 30% reductions.

**[00:13:53 – 00:13:55]** And as you do that transition, that's something else to see when

**[00:13:55 – 00:13:58]** we do that shift to direct current.

**[00:14:00 – 00:14:04]** All right, thank you so much for your time, I appreciate it.