Five forces are intensifying AI infrastructure complexity, and the biggest value lives in the seams between components. Standardizing those interfaces is what unlocks operability.
In the first episode of the Vertiv DCD Management & Operations Innovation Day 2026, Martin Olsen, Vice President of Segment Strategy and Deployment at Vertiv, speaks with Stephen Worn, CTO and Managing Director for NAM at DCD. They cover how repeatable architectures, coordinated commissioning, and lifecycle services help data centers operate as intended.
Stephen Worn, CTO and Managing Director, NAM, DCD: What does designing data centers for operability mean, and what is driving the need right now?
Martin Olsen, VP, Segment Strategy and Deployment, Vertiv:
There are five intensifying forces that exacerbate what is going on: The first, and arguably one of the most important, is speed. You're dealing with billions of dollars of compute that need to be deployed.
Number two is density. We are at 140 kilowatts (kW) in a rack, moving to 250, 600, and we are already designing for megawatt racks. Those racks are probably no longer the racks we know today.
The third is scale. These are large operations. An AI factory is getting so enormous that getting it stood up is a massive undertaking, and operating and managing it becomes a task in and of itself.
Number four is complexity. Complexity specifically points to interfaces and seams. Seams are all the different connections and interfaces between components. That is where the biggest value lies today. There are a lot of good products out there, but it is where they all interface and connect.
And then lastly, most of these AI workloads are very dynamic. They will spike up and oscillate pretty dramatically, and that puts a big toll on the system.
You got to design for operability. There's day zero, the design, where we spend a lot of time getting all different interfaces figured out. Day one is commissioning the deployment, where we're handing over the keys. The third is day 2, which is the operations phase, but it's rarely well connected to the other two phases. So here, the operator is inheriting this system and does not look for the project to be a list of bills of materials. This is a system that needs to function as a system.
We call it V - R = 0, where V means the virtual world or design, and R is your reality. What did you actually deploy? If you subtract those, any differences between the two should equal 0 because you want exactly what you have designed in the field. The five forces are intensifying that separation between all these and becomes much harder to sustain.
Stephen: What do we need to focus on to solve these challenges?
Martin:
We need systems-level thinking. At the end of the day, the customer has one problem that requires one solution, not a plethora of solutions. Seams are a critical phase in design, deployment, commissioning, and operations. They are basically an interface, a connection. It is pipes that meet, it is electrical connections, it is digital connections, it is structural connections, how it all fits together. It can be organizational connections. Did that design in fact transfer to training the people that are going to run it?
There are economic interfaces as well. What trade-offs are you making to get up and running faster or with greater density? These seams become serious gaps when putting a system together, because simple things like components not being right sized across, make you lose capacity.
Stephen: What else are customers dealing with in interfaces and seams? How do you fix it?
Martin:
Hyperscaler and cloud operators run things a certain way. You have your tier one colocation operators, tier two and three, enterprise, and neo clouds. They all have different designs in the way certain functionalities are centralized or distributed.
If we can standardize the seams and the interfaces, we can bring that interoperability. It is truly in the seam and the interface where standardization needs to happen. So you understand exactly what the component prior to it, component after it, how it operates, the capacity of it, with the telemetry, metrics, KPIs, and APIs that flow between. That's what we focus on. We call it converging the physical infrastructure.
Stephen: How do you standardize when you have so many different architectures? How do you deal with speed, the talent gap, and the demand for a system level approach?
Martin:
One of the key things I work with our R&D groups on is I will ask them, name the seam. You have to be able to name, measure, and put value to it. Because we are talking about seams, interfaces, and connections that are abstract.
But that is where we get all the way down to name the seam, which could be the electrical connection from the busway to the top of rack power distribution unit (PDU). That is the seam. Name it, then identify the risks, costs, and time it takes to connect these things.
Some seams are more important than others, so it is prioritizing that. Once you have the system down, that is the opportunity. Take that telemetry, all of that information that comes back from the different components, and improve it. Identify new seams that we did not know existed before.
Stephen: 800 VDC is being talked about everywhere. How do you get scalability, repeatability, and operational excellence when technologies are changing this fast?
Martin:
800 VDC is probably the biggest level of complexity being introduced, but not necessarily new in and of itself. It's pretty standard in the EV or electric vehicles world. We have deployed variations, like the ±400 VDC, in data centers in Greater China for the past two decades.
But the real challenge is the safety element. The arc flash for DC power is very challenging. It is not an oscillation of energy coming through, so you never cross zero, which means if you were to put your hand on a DC bus bar, it would be incredibly hard to let go and incredibly dangerous. That needs to be figured out from a regulatory standpoint and standards within how you service it.
We are working closely with chip manufacturers and regulatory bodies in the US, Canada, Europe, and Asia. It takes an ecosystem, because 800 VDC to the rack is one thing, but what about the rest of the support infrastructure? Do we keep running that on AC or do we also switch to DC? It is a question of optionality and building a platform that can come together in a variety of different ways.
Stephen: Can we really get to standardization? Do we all need to go modular to integrate the power systems and the cooling systems?
Martin:
Modular is an excellent way of driving that standardization. There is a speed element to it, taking a lot of the work out of the field and bringing it into a factory. But right now everybody wants to modularize and manufacture, and it is primarily optimized for speed. It does not necessarily solve a lot from a system standpoint. Parts improve linearly. They are only going to improve incrementally as you put these together. System performance compounds. You get more performance out of the whole system when you look at it all the way across.
From the compute side, they have standard rack scale build standards as well as commissioning standards. L11 and L12 are two terms that are being used quite frequently. L11 being it is ready to go, built and cabling done. L12 is fully commissioned and delivered. What is missing is L13 and L14. Think of L13 as the cluster within the AI factory, and L14 as the entire AI factory. Multiples of the clusters going into this AI factory, which in many ways is, multiple L11s go into the final L12 rack scale compute commissioned.
That does not exist today, but that drives standardization in the interfaces and predictability in the performance of the system. Now you can introduce things like predictive maintenance more effectively.
Stephen: Pre-commissioning, pretesting, and manufacturing work are becoming more important, isn't it?
Martin:
Yes. You can achieve that with modular manufactured systems, because you bring a lot of this into a factory and alleviate the field work. When you bring it into a factory, there is the opportunity for convergence across the physical infrastructure. That allows you to have an intimate knowledge of how this system will actually operate before you make any connections on site.
That allows pre-commissioning, because commissioning can take months. Add the component prior to that, which is the digital twin, building it and simulating it in a virtual environment where we can iterate 100,000 or a million different scenarios in hours or days. That has cut commissioning time for some of the things we do for our customers by as much as 80%.
We have got to flip the script. In the past we spent 20% of the time on design and 80% in the field manipulating big boxes. We have to spend the time where we have not cut any steel yet, simulating that piece of it. That three months has now reduced down to a couple of weeks.
Stephen: What are the key takeaways?
Martin:
Really focus in on these interfaces and get a good grasp for what that means for you. Can I observe it? Can I commission it? Can I diagnose it? Can I maintain it? Name the seam so that you can take that to your provider and ask those questions. That will make a massive impact. A resilient system is one that just works on day one. It is one whose design intent can be preserved, observed, and adapted throughout its operating life. That is where you want to have an eye toward being future-ready.
Watch the full discussion: Designing for operability - How system-level infrastructure improves lifecycle resilience
