Five years ago Starline was powering data center racks in the 10 kW range. Today it’s speccing racks at 142 to 150 kW, and customers are already asking what 200 kW looks like next year. Over the same stretch the busway itself went from 250 to 400 A as a normal order to 1,000 and 1,200 A as a routine one.

Chris Osian works in product management at Starline, the overhead busway brand at Legrand, where he’s spent about 14 years, a little over a decade of it as an applications engineer sitting with customers and turning their specs into a power chain. His job now is to keep busway ahead of whatever the next generation of racks asks for.

The number that makes that job hard isn’t 150 kW. It’s 18 to 36 months, the length of a typical data center build, set against GPU and rack deployments that now iterate every one to two years. The building takes longer than the spec it was designed around. Osian calls the result design instability, and it shows up as late-stage changes to a power architecture that used to sit untouched for the life of a project.

The Data Center Engineer sat down with Osian to talk about why customers are deliberately buying more busway than their specs call for, what open channel bus does that a remote power panel can’t, the arc flash math that isn’t getting enough airtime, and where the 800 V DC conversion is going to land.

The following is our conversation, lightly edited for length and clarity.

Rack densities are climbing exponentially. How is that changing power distribution inside modern data centers?

We’re seeing requests in the next year above 200 kW, which is kind of insane to think about.

Osian: It’s changing pretty drastically. Five plus years ago we were seeing rack densities roughly in that 10 kW range and sub-20 kW range, and within that timeframe we weren’t seeing a lot of growth in power density. It looked more linear than anything, not exponential like we’re seeing today.

Fast-forward to roughly three years ago and we were seeing anywhere between 20, 40, 60 kW. Today we’re speccing in and powering racks that are 142 to 150 kW. You can see from that trend it’s not linear. It’s turning into a very exponential type of increase in power. And even looking forward, that’s not going to stop. On our roadmap, and in what we’re seeing from customers, we’re seeing requests in the next year above 200 kW, which is kind of insane to think about. Because again, 5 to 10 years ago we were one order of magnitude less than that.

To put that further into context, at least with the busway and the power chain, for a long time we were distributing 250 to 400 A busway at 415 V. Since the AI boom we’ve more than doubled that. Today we’re distributing 1,000 A and 1,200 A in a very common space, and we’re also deploying bus duct, where that can climb up to the 5,000 A range. So it’s safe to say AI rack density is really changing the power architecture we’re seeing in the space. And not only busway. Upstream and downstream we’re seeing the same things happen, an increase in power both at the PDU strip level and upstream at the PDU level.

The one thing I’ll say, and I don’t know if it’s talked about enough, is this design instability, you’ll call it. The typical data center build is roughly 18 to 36 months, give or take, depending on what type of deployment it is. But the one thing that’s changed is that we’re seeing these new GPU deployments and new rack deployments come up every one to two years now in terms of iteration. That throws the whole data center build timeline off a bit, where you’ll see customers either over-engineering at some point or changing the design late in the game because there’s a new design coming down the pipe in the next year. So we’re seeing a lot more late-stage changes in the power architecture. I’ve seen it in the busway and other places in the power chain as well, where in the past it was a very stable specification and it lived that out through the life of the project build.

Are there design choices that keep the power architecture flexible enough to avoid major rework?

I’m calling it over-engineering or over-speccing, just to mitigate some of the risk.

Osian: Yes and yes. It really depends. For hyperscalers, they have a longer build cycle and a little bit more flexibility and resources to handle some of the changes, or maybe even wait an iteration. For some of the shorter builds, flexibility is key.

We have customers today that have a power spec, but what they’re asking to practically build is over what they need, because of what they don’t know. So I’ll take busway, for instance. If the spec calls for an 800 A busway at 480 V, we’re seeing customers spec in a higher amperage, like 1,000 or 1,200 A, or whatever the highest is offered, to help mitigate some of the unknown design specs that are coming, or some of the known ones like the rising rack densities. They want to prep for that. So there are spaces where, again, I’m calling it over-engineering or over-speccing, just to mitigate some of the risk in future-proofing whatever design they’re doing.

Overhead busway versus traditional cable whip distribution. What are the tradeoffs?

Osian: Between RPP and busway there are a couple of major differences. The most obvious one is that with open channel bus you can add and remove circuits under an energized situation, and you don’t need an electrician to do it. You just need somebody who’s qualified and trained to do it.

With an RPP, the remote power panel, you typically have to shut down, which results in downtime, and you have to use an electrician to wire up the new circuit and energize the new load. So in terms of operational effect, on a day two situation where your data center is operating and energized, it takes more time or it takes downtime to add a circuit in the RPP space versus busway, where you can simply order it and have your qualified personnel add it without any shutdown. The flexibility there is really good. If there’s a new rack coming down the line, or if you’ve already reserved it, you don’t need the downtime to add that circuit. You can just add it live.

You literally plug it into the bus, plug it to your load, and you’re done. It takes like two minutes to install.

The other big thing worth pointing out is that the install time is a lot different between the two. With RPPs you have to route the wire and land each of those terminations, strip them, land them on each location, route them to their load. And you can imagine, with data centers there are hundreds and thousands of these circuits that you have to land. With busway, on day one you’re literally just installing what we like to call somewhat of a Lego set. It’s a very preset, limited set of straights and feeds, what we use to power it, and then plug-in devices where we pull the power.

The nice thing there is that everything comes in preassembled at the factory, and all you have to do is tie these three or four pieces together. Then once the open channel is set overhead and in place over the server row, the plug-ins themselves, you don’t need to land any wiring. They come preassembled from the factory. We do factory checks on the way out, but you literally plug it into the bus, plug it to your load, and you’re done. It takes like two minutes to install. So from that point of view it’s just a lot quicker and less time-intensive than RPP. And the result of that is lower cost as well.

How important is flexibility now that workload demands are less predictable?

Osian: To expand on what I said earlier about that day two install and over-engineering, flexibility right now is super important. We’re seeing these fast iterations of rack designs being deployed so quickly that if you don’t plan to be flexible or mobile within your system, especially in the life of the build, you’ll be stuck retrofitting, in my opinion, at that point. So there’s a lot of value in thinking about future-proofing your design.

What I will say is that in terms of flexibility, the one thing we’re seeing is that people are now more open to different types of infrastructure within the white space than they were before. Because rack densities are getting so ridiculously high, we’re seeing customers now entertain using bus duct versus open channel on the floor, because it has a higher amperage ceiling. That traditionally has not been done with data centers. But we’re also seeing fuses being implemented into the white space, which again really wasn’t a common thing five years ago or before that.

The arc flash risk goes up as your energy goes up.

What the fuses do is knock down the fault super quickly, which affects the let-through, limiting it to something very small compared to what breakers do. So we’re seeing a trend to fuses because of that. Not only does it help with the power infrastructure and the short circuit withstands and let-throughs of the power chain in general, but because it does that, it also keeps your arc flash risk lower as well.

With this rising power trend, what’s hopefully being talked about among customers in the field and the market is that the arc flash risk goes up as your energy goes up. So you need to be thinking of things like putting in fuses to help bring that arc flash risk back down. If you get something that reacts super quickly, that affects the let-through that’s going into the busway, which then affects the arc flash risk, and eventually that makes it a little bit safer for your workers to be around the infrastructure. It’s something that maybe isn’t top of mind, but it’s something I think the field should think about as we move forward into this high-power world.

What else should engineers prioritize to hold reliability and safety as densities climb? Is it just fuses?

Osian: Fuses is a big one. But we’re seeing new technology, especially in our part of the power chain. We’re seeing LSI breakers being implemented to help control some of that reaction time to faults, and to be a little bit more mobile there. There are also things you can do in terms of safety, like adding an IR window to whatever power infrastructure you have. For us it would be our end feed. That allows you to catch any hotspots early. It’s more of a preventative maintenance thing. Now that we’re dealing with more than double the power, implementing things like IR windows, or even power quality meters or meters in general to monitor the health of the system, that can be done to identify hotspots but also anomalies in the system before it becomes a real problem. It either can result in downtime, which is one of the worst things for data centers, or even worse than that, an injury to personnel.

One of the biggest things is safety education. Again, it’s probably something less talked about.

But I think one of the biggest things is safety education. Again, it’s probably something less talked about, but there’s a standard out there called NFPA 70E, which is a national standard giving guidance to engineers, contractors, and techs, anybody on site. It sets up a system of education to educate the tech that’s working on the switchgear, educate the tech that’s working on the busway, to make sure that they know what they’re doing but also have the equipment that they need. Are you wearing the right PPE? Are you using the right tools? Things like that. I think it’s worth bringing up because we’re seeing this drastic increase in power that we hadn’t seen before, and it’s something less talked about that you really need to pair it with. Are you doing things to mitigate the risk for your personnel on site as well?

As operators move to bus duct and other high-amperage options, what separates a good design from one that creates a headache later?

If you have to retrofit at a later date, that’s horrible for you.

Osian: I’m going to restate what I said earlier. If you have to retrofit at a later date, that’s horrible for you. When you’re thinking about these designs, I think you need to make a decision on whether you’re going to stick with your plan of deployment, or whether you’re going to leave some space, either a reserve space or something in the design that allows you to be a little bit more flexible.

The last thing you want is to retrofit in the end because a new iteration comes out and now you’re stuck, because you’ve either stranded power somewhere else, which we’ve seen before, or you under-delivered on the power on day one and now you can’t power whatever is coming down the pipe in the future.

As AI compute pushes facilities toward higher loads, how does power distribution architecture evolve over the next several years?

The next two to three years, we’re seeing specs that are in the 400 to 500 kW range.

Osian: I don’t know if it’s a secret, I don’t think it is, but the big topic on the power side of things in the data center is 800 V DC distribution. I talked a little bit earlier about how we’re seeing specs this year and next year of around 200 kW or more. The next two to three years, we’re seeing specs that are in the 400 to 500 kW range, which is more than doubling again what we’re seeing today and into next year. So to get there, 800 V DC distribution is kind of the popular thought, the topic of discussion. It’s even roadmapped in some vendors, in some GPUs. I don’t know if I can say names here, so I’ll keep it generic. But 800 V is out there, and I think that’s going to be a huge topic.

The other is sidecar deployment. A lot of the power today is handled within the rack, and again I’m talking about busway distribution to the rack. We’re seeing in-rack conversion today from AC to DC. We’re seeing now sidecar deployments where that conversion happens next to the rack, eating up a little bit of rack space next to it, but being able to handle this large conversion of power from the bus to the rack.

This whole 800 V power distribution conversation, one, it helps us get to a higher power to the rack using the 800 V distribution. But I think there’s still trying to iron out where that conversion’s going to happen, whether it’s going to be upstream away from the rack or in a sidecar situation. So I think that’s still a bit unknown today. But that 800 V, I think in the future, is going to be one of the keys in keeping up with the rack densities that we’re going to see in three, four, five years.

We’re obviously trying to gear up for it, and I think the market and other vendors are as well. I mentioned bus duct earlier and open channel busway. On our roadmap we’re looking to keep up with that as well and making sure we’re equipped to handle the 800 V. And not only handle the 800 V, but continually trying to climb that power amperage range to make sure we’re positioned pretty well to distribute power to the racks.

Is 800 V the ceiling, or does that number keep climbing too?

Osian: I didn’t mention it earlier, but 1,500 might be a thing as well in the future, so who knows. The future is unknown, but we know one direction. It’s up for power, for sure.