Hello and good morning everyone. Thank you for being here with us today. I'm Giordano Albertazzi from Vertiv. I'll talk you through what really keeps everything AI alive. What enables AI to actually work and exist, if you will. It is the physical infrastructure. Clearly, there is no need to explain the importance of AI, the volume, the demand behind that, but certainly at least to some audiences it's important to remind how important it is to have an AI infrastructure that can support and make the actual AI that is used by businesses and individual consumers exist. So it's not about proven AI, it is about maximizing the useful AI as it is generated by an AI factory.
The world is changing. What we thought in terms of the physical infrastructure, the traditional data center, may not and in most circumstances is no longer adequate for AI infrastructure. Certainly capacity and compute are constraints right now. But the constraint of compute is the AI factory. So the speed at which an AI factory is deployed, the efficiency at which the available power is turned into tokens, is really the key and the solutions to the problem. It's all about talking in terms of AI output per megawatt. It is tokens per megawatt. For those who come from the physical infrastructure industry, in the past we used to talk in terms of PUE as a metric of efficiency. We are beyond that. Think about an AI data center and AI factory as a factory. It's about how much output you extract and how do you maximize that output. So it is a constrained world when it comes to compute capacity. It's a constrained world when it comes to available power. It's how we maximize that available power.
So there are five forces at play in this moment in the industry. One is density. Density of data center is changing dramatically. Density at rack level is changing dramatically from traditional kind of rack density of power in the tens of kilowatts. We are now going to 100, 600, in the future one megawatt per rack. That changes everything. Everything in terms of the power and the thermal architecture, in terms also of how you operate a data center. Think about how much power and current you can have in a data center. The speed at which you make that capacity available typically defines the success of a certain AI player. How fast are you in making your AI available to the market is absolutely critical. So how fast you are, the time to token for a data center and AI data center, is absolutely important. Scale is orders of magnitude bigger than it used to be in the past. So forget about bespoke data center design. It's all about scale. It's all about optimization. It's all about rapid deployment. Complexity is increasing. If you really want to be super efficient and fast, you have to architect everything at one. And don't forget that whatever everyone in the IT and data layers are doing, they're creating a lot of challenging loads from a current and thermal standpoint that need managing. So all things that were, if not existing at all, but certainly much less acute in the past, are now central.
So what changes is really the very economics of a data center and infrastructure. It's really about maximizing tokens per second per megawatt and look at it in terms of efficiency, productivity. It's really about how fast you are deploying capacity and how good you are at extracting compute out of every single watt. So traditionally, data centers are built a piece at a time. You have a power train that is made of a switch gear, a UPS, etc. All thought through as individual optimized components, but the system is not working in an ideal manner. What we call the seams between the various pieces are not optimized. If these seams are not optimized, the data center is not optimized. Those token economics are not working ideally. So we got to change that paradigm completely. We have to really make sure that there are no capacity leakages in the system at all. That means that the architecture of an AI factory, of a data center, needs to be engineered, re-engineered, rethought through holistically with all parts not like discrete elements of a system but like intrinsic components of one big seamless system across the powertrain, the thermal chain, everything controls, and seamless to the compute to maximize the tokens per watt per second we were talking about earlier.
So just a quick walk through the traditional way of building data centers versus what we see becoming mainstream. As Vertiv, we have been in the industry for decades and we are absolutely proud of the technology we bring about with every component, be it a UPS, a chiller, a cooling distribution unit for liquid cooling, you name all the components, a switch gear, everything. But we know that no matter how optimized the individual component is, there is a limit, there are diminishing returns to the optimization of the system. When you just think about the data center in a traditional way of building, designing, and engineering as somewhat discrete pieces, we see an opportunity that we call an opportunity for convergence. When you look at a data center as a whole, it means that the individual components of the data center are not only optimized as such but are designed and built exactly to realize that minimization of losses or inefficiencies at the level of the seams that we are addressing. So we now think in terms, and the whole industry and certainly the thought leadership is around optimization of the whole power train. It is a power train that is grid interactive. The power constraints on the power side are such that not only we see a lot of AI factories and data centers becoming, let's say, autonomous when it comes to power, so behind the meter power generation, but also become very interactive with the grid, so helping the grid in many circumstances as well and making the behind the meter power generation also more economically viable.
But really, if you look at the red area between the optimized system level, whole data center system converges, that's the gain that you have thinking, designing every component, the whole data center, and then the delivery of a data center as a converged system. So it's end to end, it's pre-engineered. This re-engineering happens with the ecosystem partners, the silicon players, the data center developers, the hyperscalers. A very important and easy to grasp example is what happens in the white space. Typically you build a data center, you have a big white space, and then you deploy compute. But before deploying compute, you have to fit out with everything: liquid cooling, power distribution, cabling, and whatnot. A very complicated activity inside the white space. So rethink it completely, and that's where the industry is moving, where we're leading the industry towards, is to optimize the design of everything inside the white space and deploy very rapidly through modularization. So again, not discrete or independently designed components in the white space, but everything exactly designed and deployed at speed, reducing the deployment time inside the white space by an order of magnitude, which means a much faster time to revenue, time to token for the data center developers, and certainly much earlier compute capacity available. So this is now becoming mainstream in the industry and demonstrates not only the power of modularization but also the efficiency that comes in terms of total cost of ownership and deployment time from a converged infrastructure at white space level inside the data hall.
So what you see here combines cabling, power distribution, liquid cooling, and you see it even here. So from discrete field integrated components to an entire infrastructure that is prefabricated, pre-engineered, and deployed at very high speed. So tokens per second, tokens per euro, tokens per watt, time to first token, absolutely optimized at levels that were unthinkable simply a couple of years ago. But then we cannot just stop here. Think about the entire infrastructure. Now we know that traditionally there are many ways of building compute, making compute available. We have seen and we are driving in the industry an evolution towards increasingly large levels of optimization, prefabrication, and modularization. Of course, all what I said combined with optimized engineering, and very often this optimized engineering, for example, designs a data center model, reference design, exactly to a certain silicon model so that you have the certainty that running over the years that type of silicon you can do that in the most optimized way. So think about this wave of optimization and modularization combined to extreme engineering optimization of the infrastructure that allows us to go all the way to the next level.
Just like when we were talking about white space, we were talking about a deployment of a white space infrastructure that can really be seen as one SKU, one SKU that is rapidly deployed and one SKU that is manufactured. So instead of having also the quality variability that you experience on site, let alone the difficulty to find and to have access to trained and capable labor on site as you build an infrastructure data center and ultimately compute, in this case you can have everything done in a factory, much more control, much more repeatable. So not only optimized and not only very importantly rapidly deployed, but take this concept of moving from building a data center to manufacturing a data center. And then you have the concept of design and optimize the entire data center as one product. One product made of various modules that are deployed and they can be deployed over time as the need for compute capacity expands or as more power becomes available in a certain site, and those modules can then scale all the way to hundreds of megawatts or gigawatts in a seamless way. Again, think in terms of deploying one SKU, deploying one product. So it's a dramatic shift in the way things are defined. And it's a dramatic shift because it enables the industry to completely cut the time that it takes to engineer, design a data center. These designs, these reference designs I'm talking about, are already optimized on the leading silicon and that can be deployed at high speed.
When you think in terms of a fully modular converged infrastructure, then we can really think in terms of an infrastructure that can be deployed in approximately half the time, that can be deployed on a smaller footprint, and very importantly, optimized from an engineering standpoint on the specific silicon or more flexible depending on the type of business model that our customers have, and that can dramatically lower the total cost of ownership. But again, don't take total cost of ownership lightly. When you think about data centers that run half a gigawatt or a gigawatt, then that TCO translates dramatically in terms of cost of running the infrastructure. If that translates dramatically favorably in terms of cost of running the infrastructure, that then translates into a dramatically improved cost per token, and that is an enormous lever for competitiveness in the AI world. So a lot of what we're doing has absolutely this type of concept behind it. It's not just about moving in an isolated manner in the industry. I've been in the industry for a few years, and in the past, the various players would move independently, and anyway the stability of the technical specifications that we're addressing such that it didn't matter. Things move so fast now that absolutely you need to cooperate very closely with the silicon providers, very closely with the hyperscalers, with the deployers, because it is really a teamwork. The design and the technology that we deploy today will host tomorrow, two years, three years down the line the new silicon, and we need to be ready. So it is about working with partners. It's about enabling and having the capacity as Vertiv to really serve this type of model at scale and globally.
It is about having a service organization because again, when you deploy this massive capacity, then it all starts because then your asset will run for 15, 20 years, and those 20 years better be very efficient. But also, we enable an industry that relied a lot on different opinions, technical flavors, etc., towards speed of deployment, really pre-optimized design, and off you go. Industry, here is your physical infrastructure. Go ahead with building compute capacity, compute power. And as Vertiv, certainly one of the most, we like to think and we believe we have the broadest portfolio of physical infrastructure technologies, really powertrain, thermal chain, from the grid all the way, follow the power all the way into the rack, whether it's a traditional powertrain or if it is a new powertrain at 100 volts, and the same thing on the cooling side and all the infrastructure around that. So the world is changing. It is about working together with our partners, and especially it's about pre-engineering, optimizing, and deploying at scale. So with that, thank you very much everyone.
