Ep. 029 - Modular Data Centers Cut Build Time to 12 Months (Datacenter)

We are here to talk about the wild, wild west of LEGO data centers. A couple of weeks ago, we put out an article on modular data centers and construction, and I've got Eric, Nigel, and Nico. We're going to dig into what modular means, what the benefits are, how the bottleneck has moved from construction time to labor, and all sorts of other things: the trade-offs between modular construction and traditional construction, what sort of risks exist, and who the big players are that are pursuing it. I'm going to pass the ball—

Great to see you guys.

Yeah, exactly. Nico and Nigel are returning, and Eric is attending for the first time. This is a nice cross-section across SemiAnalysis. We've got Eric from the consulting team, Nigel from Core Research, and Nico from the data center energy and industrials model team, so you get a little taste of everybody who touches modular data centers and contributes to newsletter articles like this one.

Yeah, excited to be here.

Awesome, guys. Okay, so let's start by passing to Nico to give us an overview of what a modular data center is, what it means, and what makes it different from the traditional way to build data centers.

1. What Modular Data Centers Mean

Sure. Yeah, let's start with that. I think that is a single definition for what a modular data center is, and that was the single reason why we decided to put out this article: everyone has their own definition of what modularization or prefabrication—sometimes called that—means. Essentially, we put out that note, a big part of which is a whole taxonomy of what we actually mean or are talking about when we talk about modular, to answer that question.

But just to start off with a pretty simple idea, modularization or prefabrication—and the reason why we named the article LEGO Data Centers is because—instead of building your data center in a traditional way, where you essentially bring all the equipment to the site and then bring all the workforce, all 7,000 workers at peak, to the middle of Abilene, Texas, and start building the data center, you're going to build it like your old LEGO, piece by piece.

Of course, those pieces can vary a lot. Think of it as, instead of bringing all that equipment and assembling it and doing connections on site, you're going to do part of the work off-site. You're going to do some modular sections because you're actually using modules, or you can think of a metal enclosure with equipment inside. You're going to do that in the factory, then ship it to the site and stack different modules, different LEGO pieces of that data center. Then you're essentially going to connect them all. It's supposed to be plug-and-play. We can talk about why I say “supposed,” but essentially it's a plug-and-play build of a data center. That's, in super-simple terms, what we mean by modularization or prefabrication.

Yeah, and maybe there are a few different types of modular data centers when you talk about the different options for describing how construction happens with a modular data center.

Yeah.

On screen, I've got this chart describing stick-build versus the baseline today, MVP heavy skids, full modular build, or containerized, where it's literally a shipping container. They can range from up to 2 years down to as little as 12 months when it comes to getting stuff on site, from site prep and groundworks in gray to the yellow structure shell, then the blue mechanical and green commissioning, for those who are looking at the chart.

Yeah.

Eric, maybe you can talk about some of the benefits you've seen from people who are going with the modular approach and how this speed-up is actually helping them.

2. Time To Power Drives Adoption

Yeah. I think there are 2 of the biggest benefits that we have seen on the market. The first one is probably labor. We have seen multiple people come to us asking about modular designs because the labor market is super strong right now. Labor rates are super high right now, and people can't find enough labor to support the immense data center build-out. They're really turning to modular data centers as an alternative solution for the labor bottleneck we're seeing on the market. That's the biggest one.

I think the second one, more importantly, is certainty in modular designs. I think that's related to financing and capital, because when we look at a lot of data center build-outs, certainty is a big thing. Even a lot of smaller providers don't have all the necessary equipment to build an entire data center, and therefore they turn to modular as a potential alternative solution to build a data center, make it faster for time to power, and also have certainty about when they will get power on site.

We see a lot of different players trying to look for different modular solutions based on where their constraints are, both labor and economic constraints, as well as their constraints around certainty. By bringing modular solutions, it can give both investors and themselves a quicker time to power, and also more certainty about when that power can come online, based on bringing a completed package as a whole.

Yeah. And this isn't just random crypto miners that are doing this. Obviously, this became popular when we put out the article on Meta.

Yeah.

A few months ago, we talked about their tents. Then, in this article, we described how AWS is doing this as well. Maybe you guys can talk a little bit about the different players that are exploring this.

Hmm. Yeah, absolutely.

Yeah, absolutely.

Yeah. No, I mean, I was just going to say that maybe before actually addressing the players, I wanted to go into what Eric mentioned more. He mentioned the labor shortage, which of course is a huge topic. We can spend an hour just talking about labor. Also, predictability and quality.

But essentially, one thing that's clear after all the conversations we've had with different data center builders and developers is that there's no clear pecking order. Some folks are doing modularization for one reason: they actually like this kind of predictability of knowing that when they're going to plug in their modules, they're going to function as promised. Then some others are doing it because they cannot get enough electricians or welders to come to the site.

But there's also a third reason that I would say is the clear number 1 today, which is essentially time to power, or building time. Essentially, it's shrinking the time it takes to build a data center from the day you get your heavy machines onto the site to do the site work to the day you're ready for sale, or ready to energize. That's a common practice.

When we think of the players that are doing this, or the different strategies these players we're going to talk about now are pursuing, essentially it all comes down to how I can build this as fast as I can. How can I shrink these 2 years—18 to 24 months—that it usually took to build a data center, sometimes over 3 years, to 1 year?

Because now everything could translate to tokens, and those 12 months of opportunity cost of not being able to generate 12 months' worth of tokens, of course, we've talked about recently in our newsletter. Before, we were talking about $11 million to $15 million per megawatt; now we're seeing compute deals much higher than that—$40 million per megawatt, with Anthropic generating some configurations over $100 million per megawatt. Think of how much 12 months of token production time is worth. So, with that, we can go back to the miners.

Yeah. Yeah, I mean, specifically, to make the point and bring this chart back on screen, when you look at site prep and groundworks on this chart, the difference between 5 months and 4.5 months doesn't really change much.

Hmm.

This is for a containerized build versus a baseline.

Yeah.

But the part that compresses so much is mechanical and electrical fit-out. You're doing 3 months of work instead of up to 9 months of electrical work on the system.

Yeah.

Can you guys talk about what this means specifically for the biggest companies in the world that are building the biggest data centers? I thought this chart at the top was incredible, where you show the rapid increase in the cost of labor at the biggest sites in the country or the world, such as Stargate Abilene with Crusoe in Texas, as well as Microsoft's Fairwater site in Wisconsin.

This chart really ties to one of the points, which is the labor shortage, or the difficulty of finding 7,000 workers to come to the middle of nowhere, essentially. It's evidence of that in the sense that, yes, we hear stories of, “Oh, it's really hard to get the electricians on site.”

But then we actually asked, how can we get some empirical evidence of this? We looked into it: How are the wages of new hires—these new electricians who are being hired—trending?

And we came across this chart, which essentially shows that, in these states and counties where the biggest campuses are being built, the actual campuses demand between 3 and 4 workers at peak. We saw that the wages for these new hires were essentially doubling, sometimes tripling, in some key states.

Nigel, what’s your take on the labor shortage? I assume you hear about this from investors and maybe many different companies across the industry, modular and otherwise.

3. Labor Becomes A Local Bottleneck

Yeah, I think it’s a great question. I think the first thing to note—and one that is increasingly being appreciated across customers, data center operators, investors, and so on—is that, of the different bottlenecks out there, there’s power and there’s compute. There was once compute as a bottleneck, and there was memory. There was the topic of interest, and then labor.

Labor is one of the constraints that is really difficult to resolve because the supply isn’t very elastic at all. So even when you talk about power, I think the economics of AI—meaning the huge premiums that are there for the taking if you can bring a unit of compute online quicker—have brought in a lot of your marginal power sources and power formats.

Again, with the energy team and here at Core Research, we’ve talked about how we’ve written a lot about reciprocating engines being one of the marginal formats that have come in, and everyone kind of knows about fuel cells as well. But if you think about labor, it’s really difficult to solve that, because these aren’t your garden-variety electricians who can just come in and stand up a data center with all the specialized electrical wiring that is needed.

This is very critical, mission-critical work that is being done, and you need to have these guys go out in the middle of nowhere. The term for this is field labor. As you see from that chart, the wage rates of these guys are just soaring.

Because of that, I think it’s being appreciated that one way to solve this is modular. As we’ve kind of pointed out in a tweet that went out today, this actually shows up in the share prices when you look at the entire data center industrial space and how the construction, engineering, procurement, and construction guys—the EPC guys—have traded this year relative to the equipment OEMs. You see the construction guys outperforming.

And yeah, it goes to show what my colleagues Nico and Eric have been talking about on labor.

Definitely. I think maybe just to put a finer point on this, we hear a lot about bottlenecks, and the term “bottleneck” gets thrown around a lot. There are 2 different types of bottlenecks in my mind. There are industry-wide, high-level constraints, like how much CoWs can CSMC produce and how much memory the big 3 memory vendors can produce. This is a constraint on the whole industry: how many chips you can produce.

Then there are local bottlenecks that would restrict a given project from coming online in a certain number of months or weeks that they care about. These are 2 different things because, in some ways, individual projects that have multiple phases may have their allocation of chips from the global supply chain that just can’t be increased above a certain amount, but they have their allocation.

It’s waiting for them to come and install, and their bottleneck is things like the labor that you’re describing here. It’s things like reciprocating engines or whatever. But if we look at the industry as a whole, some of these individual bottlenecks that show up at individual sites are not necessarily representative of what’s going on globally when we look at the massive amount of progress of chips being brought online over time.

Hmm.

Nico, maybe you can comment a little more on high-level, data center-wide, industry-wide tracking and differentiate between the local bottleneck of labor and the industry-wide demand that we keep seeing.

Hmm.

Yeah. No, and Jordan, I think this is an extremely important point because absolutely, labor is a big challenge and one of the main constraints that operators name when we talk to them: actually getting the labor on site.

But we shouldn’t take from that the message that we don’t have enough electricians in the US. It’s just a really complex optimization problem. In the US as a whole, at least in the next 24 months, with the capacity that is expected to be built, you probably have enough labor.

But the thing is, how much of that labor is not only capable of working on a mission-critical and very complex project, but also reachable in these areas? Labor is more of a logistics problem than it is an actual structural problem because of labor being inelastic.

It’s a really complex one. What we are seeing now is that, other than the solutions that really contribute on a net-positive basis to reducing the amount of man-hours required, like modularization and what we covered in our article, there are other strategies. It’s about finding the right incentives and the right solution to this optimization problem: how we actually reallocate all of the available labor in the right locations.

The good thing at SemiAnalysis is that, because of the work that the data center team has done over 3 years, we actually know exactly where all of these sites are going to be built. When we talk about 15 gigawatts or 30 gigawatts being built, we actually have the breakdown by specific county.

We can take all that and say, “Okay, let’s take this capacity, these megawatts, and convert that into labor demand.” Then we can say, “Okay, in this specific county, we need this many labor hours, and we don’t actually have them. How do we manage to bring in the people, the electricians, who are living 300 miles away?”

This is our conversation. This is what is actually happening in the industry today. Of course, wages are a big part of this, and this is why we included the previous chart. But we’ve also heard about some other things, like per diems and everything, all the way to taco trucks in a parking lot, just because they’re demanded by the actual workers who are going to the site.

It’s absolutely a local problem, and the solution is not easy.

Yeah. Well, let’s talk about the specifics. Eric, maybe I can pass it to you to talk a little bit about specifically what some of these people are doing and what specifically makes a modular data center different.

You guys had a great taxonomy for this in the article: systems, shell, and site—the 3 layers of the stuff that’s getting built. Maybe you can break down a little bit about what these 3 components are and, when we look at a modular site and compare it to the traditional building process, what actually makes it different across system, shell, and site.

Yeah, absolutely. Before that, I just want to comment on what Nico just said earlier. I feel like there’s always an upper bound to how much labor can be available. You can’t produce humans overnight. It’s even harder for you to train a human to be, I guess, data center-proficient and get to a level of being site-ready.

There’s a mismatch between where a data center wants to have power online today and when humans can come in to actually build a data center in the future. That mismatch creates a huge demand for modular. Modular is not just about costs and timing; it’s more about the availability of things and how we get power online—the sooner, the better.

Going into the specifics of how modular works, Jordan, do you want to pull up that chart real quick?

Yep, got it. Coming.

4. The Three Layers Of Construction

Yeah. Essentially, there are 3 layers of data center build-outs that we have seen: system, shell, and site. The first layer is the site. It’s more the land grading and foundations. That part is not something that can be modularized because you have to physically grade the site and lay the foundations for it to be able to build the data center.

That’s one of the biggest bottlenecks right now as well: site construction. The part where we do get modularization is the shell of the system. The shell is more of the outer bound—obviously the roof, the structure, the skin, and everything else that weatherproofs the enclosure.

The biggest example that we give is that a Meta shell—a Meta tent—is literally tens of tents that stand up on the ground. I think we have a picture later in the session. Right there, yes. These are tents that are put on the grounds for Prometheus in New Albany, where Meta puts a very soft enclosure tent on top of the foundation.

It’s quick, it’s a fast turnaround, and it’s quick to process, but it doesn’t stand up to very hard construction work. It does the job in that it can get tied to power sooner or later. Therefore, it’s very fast to build out a construction. That’s the shell taxonomy for it.

If you go back, the second part is more about equipment manufacturing. If you look at traditional versus prefab, we basically see that prefab is almost 29% faster than traditional. Where that savings comes from is parallel construction: the equipment is built at a factory while the site itself is being constructed.

Once the equipment is built, it can essentially be shipped to the site and plugged and played, instead of following a traditional sequential process. The work we’ve seen from industry takes about 15 weeks end to end, where you have to do the basic construction first, and then you do the equipment and plug it together. With parallel processing, you do both: one in the factory and one on site. Once everything’s done, you put it together and plug it in on site for commissioning. That’s where we see the savings from modularization.

5. The Risks Of Modular Builds

Awesome. My mind’s going into a few different places right now, but specifically, when I’m working on ClusterMAX, a big thing that some of the buyers think about when it comes to modular is that it may introduce some risk.

Sure, you can go faster, but with traditional data center construction, you have some guarantees around redundant power, redundant cooling, the internet connection, and the size of the system that you can get on the scale-out network. Can you talk a little bit about who owns the risk in these projects?

Is it the day-two operations team that’s taking on the site and expecting to have the exact same uptime as a traditional site, while managing all the risk of changes in the design? Does it fall to the OEM to change power and cooling requirements or keep up with the site design as the racks get denser? Is it up to the EPCs that are being contracted to actually do this stuff? Nico, maybe you can talk about how people are thinking about the risks and trade-offs here.

Absolutely. We can go through some of the different risks and challenges that we’ve heard—actual stories that we’ve heard about what happened when deploying the systems. Many of them have to do with the fact that we’re at really early stages when it comes to actually deploying this product, so it’s understandable that we run into some problems. Some margin of failure is acceptable.

One of them, as I mentioned at the beginning, is pure failure of these modules: when they’re supposed to be plug and play, you plug them in and they don’t play. Something doesn’t work as expected. Then you need to bring all your service guys to the site. You need to wait for them to go to the site to understand what’s wrong—why, I don’t know, the UPS isn’t turning on, whatever.

We’ve heard stories like this from MEP contractors specifically saying, “These modules that are sold off the shelf—now we’re hearing stories that they call us in the middle of the night, and we need to go there just to figure out what’s happening.” In my view, in that case, the responsibility is on the OEM side. If we think of OEMs—the Vertivs, the Schneiders, Flex, all of these companies selling this package—they need to make sure that their systems actually work.

Then, other than quality issues, I’d go to logistics. That’s a risk because when we think about shipping these modules—these containers that are sometimes huge containers—it’s not an easy task from a logistics point of view. What we’re seeing is that part of the value proposition from a supply chain point of view, whether it’s OEMs or system integrators assembling and shipping these modules, is that they’re trying to be close to the sites and have a footprint close to where the data centers are actually being built.

That’s going to save a lot of headaches compared to shipping these huge containers with all of the expensive equipment around the world. We’ve heard stories, of course, of a module that included some really precious electrical equipment tipping over in the middle of the road, and there you go: a UPS just completely broke down. In terms of logistics, that’s a huge challenge—one of the main challenges that operators are finding in real life.

It seems that it’s going to be just as easy as shipping these kinds of modules. In this case, again, I’d say the responsibility is in the hands of the OEMs and the integrators that are shipping this to the customer.

These are some of the biggest companies in the world that are pursuing this right now: AWS, Meta, Crusoe with the Stargate sites, Hut 8, and Nebius. These are the biggest neoclouds and data center operators, right? Compass, QTS, and Aligned—you guys talked about all of them pursuing modular in different flavors.

Yeah.

Eric, you’ve got some horror stories from this, right?

Yeah. I can’t say the vendor, but we’ve heard about North Virginia, where they were shipping modular data center equipment, and the truck carrying it basically flipped over onto its side on the highway because of how heavy it is and how bumpy the highway is. It caused a big issue for the provider.

The consequence of that is that a lot of insurers are unwilling to insure these data centers because of how expensive the modular equipment is. If a truck flips over onto the side of the highway, a module is a huge mess to clean up. There are also huge penalties for them to pay up on the insurance price.

What we see from a lot of providers we talk to is that they have an issue trying to find an insurance company to insure them from a logistics perspective. Therefore, we’re seeing more modular factories getting closer to the site itself to remove that logistics constraint.

There are all sorts of logistics, right?

Further on the logistics point, when you talk about supply chain issues, if you think about where the industry is going with modularization more broadly, you could see a world where your OEMs—the Etons and Schneider Electrics of the world—become predominantly modular, or at least where the modular part of the business is much bigger than before.

In such a scenario, these companies are more exposed to supply chain risks because there are many components that go into a skid or a module that you will ship to a customer. Missing one component now means that you can’t ship an entire skid out the door, which has a much higher average selling price compared to the status quo, where you could continue to ship the other stuff that isn’t affected by supply chain issues.

That’s one. I think the other thing that we’ve been seeing, which makes a ton of sense given where we know the industry is going, is that OEMs and integrators are all investing significant sums to prepare themselves for the modular capacity that is needed to meet the demand.

From that alone, I think that introduces some level of risk, whether they’re able to get share at an important customer or whether the customer would bring in a second source or a third source, et cetera.

So I think just the entire build-out itself introduces—or, I guess, brings along with that—some risks to the OEMs or the modular integrators.

6. The Modular Supply Chain Takes Shape

And you guys maybe have done a bit of analysis on these OEMs and these different companies in the supply chain that you think are more aggressively pursuing modular, seeing where the trend is going so far.

Yeah, I think—sorry, sorry, Nico—but as you alluded to there, with modular, it's very important to see the suppliers that have a strong position with the customers because it is really these customers who are setting the pace, so to speak, of modular development and modular capacity build-out.

One company that everyone knows is at the forefront of this is Comfort Systems. About 2 months ago, a little more than a month ago, we went out with a note on this, and our analysis started with a very simple question: We know that these guys are at the forefront, or at least one of the early movers, and we know that they're in a good position with Google and Meta. Let's see where that shows up in the numbers.

Just to share our screen a little bit, we looked at how much they have invested in capacity that is capable of producing modular. To do that analysis, we used a peer that isn't really known for its modular progress, at least to date, as a baseline. You can see the scale of the CapEx that they have been engaged in for the purpose of modular.

I think this is a very good visual representation of the sums being invested in modular, the opportunity ahead of these guys, and what they're really gunning to capture.

Yeah.

Aside from Comfort Systems, there are also several other players that we at SemiAnalysis are keeping a close watch on, but I'll let Nico and maybe Eric touch on them.

No, Nigel, I mean, I think those are the key players in the space. My point would be that, in the article, we tried to be very clear. We've been discussing what AWS, Meta, and all the COVs are doing in the modular space. That's from a demand point of view, but also from a supply-chain point of view: Who's actually supplying these modules?

It helps to have a taxonomy for our audience to understand because we've been talking about Vertiv, Schneider, and Eaton, but then Comfort, Sterling, and Quanta. These are all EPC companies. We can split this into 2 groups that are actually selling these modules to the customer.

First of all, you have the OEMs: Vertiv and Schneider. These guys are already selling the UPSs, switchgear, transformers, and rectifiers to the customers. They realize—and Vertiv was the first company to realize—that they can put all of their equipment in a container, into their metal enclosure, and ship it. It's not only going to be useful for their customers because of all the reasons we've been discussing on this call, but for them, it's much better, too. They have the opportunity to expand their addressable content per megawatt simply by including Vertiv, Schneider, or Eaton content in that enclosure. That's the OEM group.

Then there's also a second group that we call EPC integrators: companies that aren't selling the UPSs and switchgear to customers but have the expertise to design and engineer these mechanical and electrical systems. Nigel was talking about Comfort Systems. They are one of the main MEP contractors in North America, working with the biggest hyperscalers. They have that expertise, and what they are doing is working with these players—not supplying their own UPSs or their own switchgear, but integrating, assembling, and helping their customers design and engineer the systems.

Those are the 2 groups that are integrating and selling the modules to the customers.

Yeah, and proof is in the pudding. It's working, right? These sites are up. There are people using them. It's not theoretical that they should maybe do this in the future; they're actually building them right now.

It seems to me like, if you think about those 3 different approaches—where the OEM is leading it, versus the system integrator is leading it, versus the operator is actually leading it, like the self-build, operator-led approach, where AWS or Meta just says, "I'm gonna do this"—that's obviously the way to make everybody figure it out in the supply chain.

Yeah.

Maybe the conclusion is that people who weren't considering modular last year or 2 years ago are now going to get the benefits of this whole supply chain being built out when they go to build a site in the future. All the suppliers have already had to serve the biggest customers, and now they're going to have to serve the new guys doing new stuff, too, right?

Maybe you can talk a little bit about the process. One thing I'm fascinated by is how factory testing works. You tell the story about the shipping container rolling off a truck. Many racks have been in trucks all the time, getting shipped to data centers, and generally what happens is you have to do a whole bunch of integration work on-site in the data center.

As much as you try to test stuff in the factory, it's always going to have to be retested and commissioned on-site. Do you see this meaningfully changing now, where there are hubs that can produce all of this equipment near where a lot of these sites are being built? Is it going to de-risk the quality—in other words, the time it takes to actually bring the systems online—meaning the IT systems, as opposed to just the power and electrical equipment?

7. Commissioning Remains A Bottleneck

I think, of course, talking about commissioning and how prefabrication helps with commissioning is also a big question, a complex topic, in the sense that there's a lot of misunderstanding out there. We actually hear stories that, "Oh, they are running commissioning fully—the 5 levels of commissioning—at the factory, and you don't need to do anything on-site." That's not true.

There's so much of the commissioning process that you can do at the factory. But all the way through level 5 of commissioning, you need to do some level of commissioning on-site. That's one point.

Second, the actual commissioning time is really dependent. We try to show our midpoint, or our guidance, on how much the construction timing can be shortened from stick build to a full containerized data center. But bear in mind that commissioning is a part of the process that's super dependent on the project, and it can run anywhere from 3 months to 8 months.

It's not easy to give a general answer: "By doing this, using this kind of module, you can directly avoid commissioning or squeeze the commissioning period from 5 months to 3 months."

The last point I would make is that we've heard of operators doing different kinds of commissioning. Maybe they are just avoiding some steps or less trusting the process and saying, "Okay, I'm not going to do commissioning. I'm going to turn on my GPUs, and then if something happens, well, my bad." We've heard of that, and that's not really a full consequence of modularization. That was happening already.

Whether an operator is doing the full commissioning process or just skipping some steps—and we can discuss why operators are choosing to skip some steps—again, many times it all comes down to time to power. If there's a long lead time for liquid load banks, I'm not going to wait for that. I will skip the commissioning of my CDUs.

Makes sense. Yeah.

That does happen. The consequence of that is that quality issues then arise. But again, this is not a full, direct consequence of doing these skids or modularization.

Yeah. I think what I'm going to add is that, when we talk to vendors, it was surprising to us, too, that L5 commissioning is becoming one of the biggest constraints of the modular story right now.

If you think at a high level, there are 2 parts of testing. There's factory testing and commissioning testing. Factory testing is what you do in the factory, where you prove each module works independently.

But really, on site, you have to prove that the entire system works together under different workloads and different failure conditions, right? So, let's say utility power drops, for example: do the rest of the UPS units line up together in the right sequence? What we have heard is that even if something is tested and works 100% in the factory, it might not work on site. So the operational risk is really high, and therefore L5 commissioning has been super difficult for a lot of people. That's one of the things we have heard from people: it's something that people need to overcome as a result of the entire process.

Makes sense. As we move to wrap, the article came out a few weeks ago. It's not necessarily breaking news the way we've been covering some things recently on the podcast, but it is a total, full review of the whole industry. I'm sharing on screen the modular vendor map that the guys put in this article, which shows the vast number of players that these guys are tracking across the entire industry. 116 products, according to this chart, and a lot of names that you might recognize from other businesses that are just getting into data centers now because of the great opportunity that modular data centers provide. So that's what I'd like to leave it on. Do any of you guys have comments or thoughts that we can leave the audience with as we close up?

Yeah, I think the one thing that I'll leave is that a lot of people came to me asking, "Hey, is modular really about cost?" I think the argument is that it's not just about cost. We actually see the cost benefits as not as much as what people think. It's about 8%, from what we have seen. The bigger benefits are what we talked about throughout this podcast: labor and time-to-power availability. What's driving the whole modularization is the entire industry that's shaping it.

Good one. My take on the vendor map—the whole state of the supply chain—would be that one thing to look at in the coming 12 months is probably lead times and actual manufacturing capacity. Yes, this is great: we're saving labor; we're taking all the construction to the factories. But that means that we actually need footprint. We need manufacturing plants. We are industrializing the construction process of a data center, and that takes footprint and production lines. It's completely changing the model of what it takes to build a data center.

What we are seeing, even though we are still in the early days of this transition, is that some of the companies at the leading edge of the supply of these modules are already running into supply constraints. They don't have the space. They may not have factory workers. They have a lot of electricians in the construction companies, but they don't have the factory workers. So this complete transition is going to take some friction, and we at SemiAnalysis are closely looking at the state of the supply chain.

What if we run into not electrician shortages, but shortages of factory workers who are actually working on the production lines? You still need electricians to assemble these modules, even if it's on site. Yes, there are a lot of productivity gains, but you are going to see that the labor problem here is not fully solved just by moving everyone into a factory.

To produce more people.

No, it's true. We are seeing companies—some of the leaders—with lead times for these products that are over a year. If you want to tweak just a little bit of the module that they are selling off the shelf, then the lead time is 18 months. We are already talking about the lower range of the time that it takes to build a data center in the traditional way. So, if you are looking at a 12- to 18-month lead time for these modules, then you need to start thinking, "Am I actually getting anything from these?"

That is why, in our vendor map, we were showing 160 players in this space. When the top companies—when, let's say, the leading companies—are fully booked, the industry is going to go to these other new entrants, these integrators that have capacity and expertise. It is a huge opportunity for a lot of new players, probably coming from the construction space, that have the integration expertise to come into the modularization and prefabrication space. It's a huge opportunity for them.

All right, Nigel, anything from you as we close?

Just more to come from the industrials team here at SemiAnalysis.