Ep. 039 - [EMERGENCY] Our Numbers Are "Insane." Here's the Proof. (Datacenters) | Jeremie Ontiveros, Reyk Knuhtsen, Jordan Nanos

Reyk, show me that bed again. It looks like if you lie down, you’re gonna kick your feet through the wall.

Like, dude.

Yeah, if I have a night terror, my foot goes right through the fucking wall.

Yeah, take a look at the pole in the middle of the room.

Right.

Like—

Hell yeah.

Right there, man.

Reyk’s gonna be dreaming about Brad Gerstner tonight, just fighting the wall and kicking the wall.

Yeah, put his picture in there. My room is fine. It’s just that I got two single beds, which was kind of weird. But besides that, the room is okay.

That room is—look at that carpet. You know that room’s nice, yeah.

They just said, “Yeah, give Reyk the fucking shitty one.”

All right, Reyk.

I know. Talk to Michelle.

Yeah, I gotta talk to someone.

Okay. You guys ready to rip into it? You guys ready to start?

Yeah.

You know.

Everybody get in bed for the cold open.

Let me jump in.

Hello, everyone. Welcome back to another episode of “Semianalysis Weekly.” In this episode, I’m joined by Jeremie and Reyk, our experts on the data center model. We’re gonna talk about their article, “Data Centers 2027: The Year of Acceleration, 50 Gigawatts in a Single Year.” This one comes against a bit of a backdrop. Just the other week, Brad Gerstner from Altimeter was showing our data on the All-In Podcast and then in his BG2 interview with Jensen. They were talking about how our forecast of 43 gigawatts in 2027 is too aggressive.

We’re gonna dig into why that’s not the case, and we’re actually increasing the forecast. When people think that only about 25 gigawatts are gonna get built, we’re gonna go through exactly what the FUD is about. People talk about moratoriums and other political restrictions. They talk about power generation and all sorts of reasons why this might not be the case. But guys, I don’t wanna walk through it. Let’s talk it through with you. Welcome.

Man, thank you. Thank you.

Thanks for having us.

Yeah, we got a lot to talk about today. This is a good one.

Okay, before we get into the actual content of the article, let’s lay the land. Can you guys give me a little bit of history on your track record with the data center model? Jeremie, can you tell me a little bit about the history of this model, when you launched it, and some of the forecasts you’ve made in the past and how that’s played out in terms of the fidelity of the model?

Sure. The model launched nearly 3 years ago now. We published an article in March 2024 to discuss some of the results we got from building the first iteration of it. That article is still available today. It’s “The AI Energy Dilemma: Race for Data Center Space.”

We start that article by showing other market forecasts out there and then putting our own forecast up, which was materially ahead of many other firms like McKinsey or whoever else was making forecasts at the time. That has largely, of course, been proven right. We’ve increased the forecast a bit since then, but overall, the orders of magnitude are there. Back in early 2024, we were expecting a material acceleration in 2025 and then in 2026.

The way we do this is by combining several data points. We use data from the chip side, the data center side, and the energy side. We triangulate as much as we can through everything we can follow. Another thing I could point you to is our article from a few months ago called “Stop Saying Half of U.S. Data Centers Are Delayed,” where we show this chart of our 2027 number and how it has been revised over the different versions of our data center model. You realize it has barely moved, right? So we were always expecting—

Sorry, it was the 2026 one, though.

Yeah, whatever. 2026, 2027. People can go look at it. Same thing.

It’s all right, Jeremie. It doesn’t matter which year it is.

Yeah, you can pick any year. It doesn’t matter.

All right. It’s 2026. Sure. Yeah, exactly. It’s been pretty close. We’ve been doing this for nearly 3 years now, so the track record has been there.

The methodology is actually not that complicated. You just gotta be thorough and comprehensive and make sure you’ve got everything. We spent a lot of time and built a whole team to ensure that. That’s what got us there. Now we have extremely strong signals pointing to 2027 being a gigantic year, with many different subcomponents that we can talk about in depth.

So there we are, after dear Brad called us out publicly and cut our forecast in half. We’re like, “Damn, this is actually at least some number we made up.” There’s a lot of thought and work that went into it, and we’re firmly behind it and stand by it.

Let’s walk through where some of this opposition comes from. Reyk, when you hear people take our number and cut it in half based on a hunch, what’s the hunch actually coming from? Is this purely moratoriums? Is it the supply chain? Is it capital, labor?

Yeah, it’s a number of things. When you read a lot of the news articles or sell-side reports, they all hinge around the same fear-inducing topic: moratoriums are going to get rid of all the data center buildout in the U.S. It’s all over, they’re canceled, right? Then you have other reports coming out where people aren’t covering every data center in the U.S., but are taking headline numbers and saying, “Oh, well, this headline number was there before, but now this got canceled because of a moratorium.” There’s the energy side of things, where there are no grid connections anymore. These data centers are never gonna get power. There are a number of other things—labor, the actual buildout itself, a million things here.

The issue is that nobody digs as granularly into it as we do. Jeremie brought up that it’s granular, but what that literally means is we go site by site, building by building—not even at the campus level, but building by building, permit by permit, power allocation by power allocation. We buy satellite photos. We estimate starts of construction, see how the project is progressing, and figure out when it should come online. It’s hard to refute any of this. When I’m physically looking at all of these data centers under construction, I don’t know how I could haircut it. These are the numbers that add up when you count all of this.

We have that, and then we also address the same problems people are worried about. With the moratoriums, we said, “Okay, then let’s count all of them and map all of them.” We looked at over 400 moratoriums, combined them with every data center we had mapped into the region or city—whatever jurisdiction may apply—and found that 10% of them were actually affected. Energy is the same story: you look at grid headroom, and then you get to behind-the-meter solutions, where maybe you don’t have to have grid headroom in the first place. I can just build an on-site generation project. There are a number of ways around this.

The point is that these topics are broadly true, but they don’t apply to every project. When you dig into the appropriate projects, the ones that are under construction, they all pass this bar. All of them pass the bar. That’s why our methodology can give us these numbers and why we have such confidence in them.

Yeah, take me through the individual projects. One of the great charts in the article was “From Dirt to Data Halls,” where you’re tracking 3 individual projects from when they broke ground back in 2021 or 2022 to when they’re fully built out and the data halls are available. Clearly, these projects we’re talking about are targeted for 2027. Maybe moratoriums cover cancellations and, to a smaller extent, delays. Can you talk to me about the chance that half of these construction projects slip by a year, or two quarters, enough to push them into 2028 or something like that? What’s the risk of small delays resulting in a change from 2027 to 2028?

Looking at all the data centers we have before us—we have thousands in the model, and we’ve reviewed them over and over—you get a sense of what looks like a delay or what might constitute a delay. Maybe there’s a winter storm in the northeastern U.S. during the TeraWulf buildout or Nebius’s buildout. Maybe there’s an equipment problem, like the transformer blowing up at the Core Scientific site. There are a number of things you can see or gauge from public reports, as well as overly aggressive timelines. You start to understand what the buildout actually looks like from a timeline perspective. In our view, 12-month timelines are pretty aggressive, especially when you compare them to the historical industry.

These are quick timelines. When anybody is guiding to sooner than a 12-month timeline, it already rings a red flag. And if it’s a 12-month timeline from a developer that just came out of nowhere, and some of the permits are obfuscated, the power source is obfuscated, maybe we can’t identify it perfectly, and maybe nobody has signed on or is the tenant of the project, all of it raises enough questions that I don’t totally know how certain it is.

It’s not to say that these things don’t catch us off guard. They can catch you off guard. I didn’t predict a winter storm; it’s not my job. But the point is, we can get close enough that we have enough of a filter to not let these longer-duration projects become 12-month projects in our methodology.

Yeah. You’re already considering the potential for this.

Yeah. Just to build on that super quick, this is also why it’s somewhat counterintuitive to apply a haircut to our numbers. We’re admittedly being a bit conservative on some things. There are sites under construction right now, and they don’t add all their megawatts in 2027, or maybe they don’t add megawatts in 2027 at all, as with some international instances. We’re not being overly aggressive either; we’re taking all of these factors into account and using a reasonable timeline.

Adding a haircut on top of that is doing a bit of double counting, or doing the same work over again when we’ve already applied it. And so you end up with, I guess, not our number. But we like to think our number’s pretty good.

Look, the other thing is we’re obviously on the ground quite a lot. We travel the world, going to conferences and talking to all of the folks actually building these facilities. So we also get on-the-ground insights from the people who are doing it. For the most part, they’ll tell you these days that it’s very challenging to build these things, that they’re facing a lot of community pushback, that they’re struggling to get labor, and so on and so forth.

But on the other hand, the timelines they’re targeting are getting shorter and shorter. In many cases, folks tell us, “Hey, I’m expecting to have this ready by mid-2027.” In our number, we purposely add delays of 1 to 3 quarters, depending on what we’re observing, because we expect stuff to be delayed by default because of all of these challenges.

Yeah. Okay. My favorite chart from the article actually went site by site across 40 sites. It’s really impressive. It’s clear from this that these are massive sites with a ton of buildings. They’re all under construction and they’re all counting toward 2027, maybe not all of their megawatts in 2027, as you’re saying, but they are contributing to the total forecast.

A couple of people have commented on a few of these sites. Maybe first, just give me your reaction to the size and scope of the forecasts you’re referring to, and how even these 40 sites are still not the full complement of data centers under construction worldwide right now. It’s still under 50% of the data centers globally that are shown on screen here. But there are individual sites that some people know better than others that maybe have some high-profile delays. You can talk through the difference between an outright cancellation and a delay, and why things are still going to come in in 2027 and not be around 25 gigawatts, but closer to 50, as you guys are saying.

Yeah, there are two things. One is: when is the first phase of a site going to be delivered? So it’s about how fast the site will ramp to full capacity. For some of the projects we show in this picture, we assume much more conservative ramps than for others. For example, Oracle, New Mexico: we were the first in the market to talk about all the issues they’re facing on that site, and obviously that’s been proven right. All the media is all over it now, and they’re calling force majeure and whatnot.

They’re obviously facing tremendous challenges getting power to the facility, and that’s why we assume a much slower ramp. If you look at the picture, you can see that essentially the whole campus is under construction. There’s over a gigawatt of IT capacity under construction right now, but our forecast is more conservative than that because of the challenges we’ve identified. So the share of megawatts that goes into our 2027 number for this specific campus is actually fairly low.

Another one: some Twitter accounts commented on our ramp on Crusoe in Childress, saying that we were assuming a whole gigawatt was included in our 2027 forecast. That’s wrong. We have under 200 megawatts as of the latest version published. We assume these things ramp over time because you’re not going to go from zero to 300 megawatts. Some of these buildings are 300 megawatts, 200, 400. You don’t go from zero to 400. Stuff gets ramped over time, and if you’ve ever been to a data center, you know that—which we do a lot, so we actually know how that stuff works.

Yeah, makes sense. Talk to me a little bit about the leading indicators, because this isn’t a forecast you just created 2 weeks ago. It’s been pretty consistent for a number of months, if not years, at this point, considering that the model is years old and you’ve been tracking 2027 as it comes to fruition. If we’re forecasting for 2032 right now, that depends on leasing and pre-leasing activity happening now, and we expect it to increase, not decrease, because there’s only so much you can track at that early stage.

But right now, if something’s going to be built in 2027, there is construction activity happening. Talk me through how you started a forecast for 2027 a couple of years ago, and how you’ve tracked it through to how it’s playing out right now.

And I want to say that’s also the reason why we have different models to track the whole ecosystem. Around the data center topic, for example, we have the energy model and the Industrials Model. There are plenty of suppliers exposed to projects at different phases, and we try to track every single one of these companies. We’ve been extremely successful in predicting the strength in orders for some of them.

One example: we published a big deep dive on our Industrials Model about a company called Sterling. Sterling does early-stage construction: groundbreaking and site work. That obviously happens first. It’s the first thing that happens when one of these projects starts getting built. We can track this company at the project level, essentially, and we were good at predicting that their orders were going to surge, which is exactly what happened.

There are many other companies down the supply chain. We also track all the electrical equipment and cooling equipment very closely; that happens afterwards. That can impact companies both ways. Sometimes they’re delivering ahead of schedule, so these companies see strength and you see their revenue come through strongly. We’ve also had times when, with Industrials Model clients, we talked about potential supply chain delays, meaning some of these sites are late. We’ve flagged several sites in our model that have been delayed by equipment-shipping challenges and supply chain challenges.

Energy is also, of course, subject to very long lead times. Everyone knows gas turbines are multiple years out. Now there are a bunch of other solutions, notably reciprocating engines, that have shorter lead times, and that’s another indicator we can track to understand the timeline of these projects. Essentially, there’s this waterfall of things we can track that are more or less binding, and we look at all of that to ensure our timelines are accurate.

The two we presented most strongly in our article are self-build construction starts, which are pretty trivial: you just have to track every single building and when construction starts. You can see how the number goes up quarter after quarter. If you understand the timeline at which this stuff is getting built, that’s a strong indicator of how large 2027 is going to be relative to 2026.

The other one is pre-leasing, which means hyperscalers, neoclouds, and AI labs signing lease contracts with data center operators.

This one is actually even earlier because typically you sign a lease, then you raise debt. The timing can vary, too, because leases can be signed at different moments in the life cycle of a data center project. But generally speaking, it’s a bit earlier. We track that extremely closely, and every quarter we forecast these numbers before they’re published by others. We’re generally extremely accurate, which helps us understand ahead of time what’s going to get built under these binding contracts that cannot be canceled.

Yeah. And this is just a really big number, Jeremie. Just double-click on the over $1 trillion of data center leasing that’s been signed. That drives a lot of the forecasts from the credit and capital markets model that Dan and the guys are working on. We just talked about it a week ago, and we’re going to be doing more of this, but this has been a pretty big story in the news recently: just how big the scale of some of this is.

Yeah. Look, there are a bunch of indicators you’ve got to track at the building level to get to that number: things like pricing, contract term, escalator, that kind of stuff. Some of these deals are public, so if you look at all of the crypto miners, for example, companies like Applied Digital and Core Scientific, you can see how quickly their backlog has been increasing. It’s now in the tens of billions of dollars.

But there are plenty of these companies that are private, right? The public markets don’t have a good sense of all the contracts being signed between companies like QTS, EdgeConneX, Aligned, and Vantage, and all of these leading hyperscalers or AI labs. That’s stuff we track extremely closely. By having a good sense of all these private companies and tracking every single one, we can gather the data and put out numbers that maybe other folks are missing, but that I think are a crystal-clear indication of the strength of the build-out we’re going to see.

1. Modular Builds Move Faster

Awesome. Okay, Reyk, tell me about logistics. When we talk about modular data centers, we wrote a whole article on LEGO data centers and how these things go up. You’re the guy who stares at a lot of satellite images and watches these things. The trend seems to be that the changes in the satellite images are happening faster, not slower, right?

Yeah.

I brought in one of the charts that shows Google’s self-builds and how they’ve changed their design for the electrical side—

Yeah.

—to get them up faster for the modular sites. Can you talk me through how this is going faster, not slower over time?

Yeah. I think this is one part of the build-out that’s a bit overlooked, to be honest. With traditional data centers, maybe labor would be called into question, maybe the timeline would be called into question. But in our minds, the larger-scale utilization of modular deployment is the reason this works and the reason this happens.

When you look at traditional timelines of 18 months compared with, say, DayOne’s timelines, which can go down to about nine months for hundreds of megawatts, you can also look at BrightRay down there in Southeast Asia, which can build in about six months. You can look at the Anthropic reference designs, too, which are going to employ modular. Modular is becoming more and more widespread across the industry, and you’re going to see build times drop. You’re going to look at the Meta tents and the new Amazon buildings, and you’re going to see these buildings basically pop up out of nowhere.

This is the Project Houdini program, which I think has a stated goal of 6 gigawatts a year for the next 2 or 3 years, announced maybe 6 months ago. These are the reason this build-out works. Without this, you’d have a much more labor-intensive build-out. We did the math in one part of the article: if you have your MEP, the mechanical/electrical side of the fit-out within the data hall, more modularized—for example, if I take the power hall and imagine a literal shipping container that I slot into the shell of the data center and plug a few things into—it demands much less labor than having 20 electricians in there trying to coordinate everything and “stick-build” the entire hall. You can imagine this on a much grander scale. I think we’re working with the industrials team to put together a comprehensive newsletter on labor for the 2027 build-out. So far, everything points to modular being what’s making this work, and you can see it across the industry.

Yeah. I would add that if you actually visit data centers—the big ones, some of the 2- or 3-year-old ones, and some of the smaller ones—you can clearly see the difference in how many of these new builds are fully designed and optimized to go as fast as possible on the build time, ramp time, and commissioning time. Absolutely everything is thought through for an accelerated timeline.

Many folks tell us they’re going to pull this off in 6 months. We purposefully haircut a lot of these assumptions, but it’s still extremely impressive when you visit these sites and see what folks are doing to accelerate while keeping the bar high on quality and key delivery requirements.

Yeah. I think the point here is that somebody might ask how you measure modular adoption. We have it on the industrials tracker, but the other thing is that if you don’t follow suit with the modular build-out or this kind of design strategy, you probably won’t be allocated the labor you need, nor will you be competitive in the build-out market.

If I can offer you a data center in 12 to 14 months, I’m really not worth as much as the guy who can offer it to you in 8 to 9 months. That’s going to start forcing people’s hands in this direction, and we’re already watching people do interesting things. For example, BrightRay is talking to multiple neoclouds right now about selling entire data center modules—not the 1- to 2-megawatt crates, but the actual hall that you can ship and install into your core and shell, or with the core and shell from BrightRay. People are going to keep going to these folks who can provide this equipment, and they’re going to keep buying it. I think this build-out is exceedingly led by the modular aspect.

Yeah, makes sense. How about the balance? We put this behind the paywall, to be clear, but it seems like generally speaking, people are trying to call whether we’ll have accelerators and an overbuild of data centers, be constrained on data centers and have an oversupply of accelerators, or have a perfect match.

One comment we can make is just how many people overdeployed GPUs and then used power capping. It seems like power is still a constraint for a lot of these, whether they’re grid-connected or behind the meter. There’s only so much power you can bring on, even if you have the physical space and the chips available, so you’ll be power-capping GPUs to fit within the envelope.

With construction timelines reducing so much at modular sites, maybe you guys can just say at a high level: do you think there are going to be a lot more data centers built in the next couple of years?

I mean, look, it’s obviously tough. Generally speaking, our forecast for data centers is slightly easier, in the sense that it’s based on infrastructure signals you can track ahead of time. Chips can be slightly more volatile, but overall, we also have a lot of very strong indicators pointing to what’s going to happen in 2026 and 2027.

I would say we’re a bit more worried about the risk of oversupply of data centers relative to chips than the other way around. We’re already in a sort of oversupply of chips today, as we speak. One always has to be a slightly bigger constraint than the other. If you want to deploy something right now, it’s easier to get GPUs than data center megawatts.

But when we compare how these different supply chains are expanding into next year, and especially into 2028, we do see a risk of that flipping in 2028 and of potential data center oversupply. That’s what we’ll watch most closely: how fast the chip supply chain can expand.

That’s obviously a strong positive for things like CapEx, including investing in new semiconductor fabs and buying new equipment. That’s the type of thing we’ve been flagging for our clients for several months already.

2. Responding to the Skeptics

Let’s wrap on this one, guys. Great job taking us through all of this; it’s been super helpful. We’ve covered all sorts of aspects, from the moratoriums and labor to electrical equipment, cooling equipment, and actual construction progress. There are all sorts of things that play into a data center arriving on time, and I think we’ve covered them in great detail.

Let me get your direct reaction to people like Brad Gerstner showing Jensen our chart with an old forecast, which basically says, “You’re going to build too many chips and not have enough places to put them.” Jensen kind of acknowledged it—“Yeah, yeah”—and moved on with the conversation. Where do you think that sentiment comes from? Maybe you can respond directly to that.

I think it’s easier to feel what’s happening right now than to know how that feeling is going to evolve by the end of next year, 12 or 15 months from now. Right now, it’s crystal clear that there’s more supply of chips than data centers. Anyone who wants to deploy AI infrastructure can get access to GPUs, but they can’t get access to land, power, or a shell in the next three to six months.

The constraint right now is crystal clear, so obviously we want to make sure the situation gets solved. I think that’s why Jensen absolutely wants to make sure this doesn’t happen. He has a long history of being extremely smart about the supply chain, and he can be aggressive as well. We’ve seen him secure tremendous amounts of supply, like in 2021, when he secured GPUs for the crypto-mining build-outs, which obviously crashed.

There can be a bit of a cyclical aspect, but overall, it makes sense. It perhaps makes the company slightly cyclical, but the whole semiconductor sector is cyclical anyway, so it doesn’t change too much. His company is in extremely good financial health, and he’s writing these massive checks to secure supply from memory and logic suppliers. He’s really good at securing supply from these guys, and he wants to make sure he can build as much as he can without constraints. So, yeah, it makes sense to pressure everyone in the supply chain to build more.

The other thing I’d add is that a lot of it comes from people operating without the most solid foundation of information. It’s one thing to understand that moratoriums are slowing data center build-out. Grid connections are, too. So are labor and equipment supply chains. All of that is true, but being able to measure it accurately is a different story.

Yeah.

I don’t know if there’s anybody else out there who’s really— Actually, as an aside, Jeremie, did you say we were looking at oversupply in 2028 before I continue?

I didn’t say that formally. I’d say it’s the main risk as we get into 2028.

I was going to start talking about that, but let’s just move on. I’ll finish my thought: a lot of people are operating without the most solid foundation of information when they come to these conclusions.

Yeah. Look, all of Wall Street sort of tells us, “You guys are plain wrong.” Some of the great investors out there, like Brad, say it, but he’s obviously not alone. He just did it on a big podcast, but many other folks tell us, “No, you guys are way too high.”

There are a lot of data sources out there with data center forecasts that are materially smaller than what we project. We’re familiar with a bunch of them. A lot of these Claude-coded forecasts track data centers, and people just prompt Claude and say, “Track all of these sites.” There are three, four, or five of them with numbers that are 2× ours. Some of our clients tell us, “Hey, these guys are lower,” and we tell them, “We didn’t prompt it,” or whatever. There’s more that goes into it.

Generally speaking, a lot of sources probably aren’t based on accurate information, unfortunately. Every single CEO is also telling folks that the biggest constraint right now is land and power, which goes back to the sentiment driven by what’s happening right now.

Maybe I could say something directly: what set me off on this one is that a month ago, the guy got on stage and presented to a YouTube audience of 330,000 people. Then, in this one from four days ago, to 265,000 people and counting, he cites our data and then just says it’s wrong.

He says, “Yeah, I don’t think so.” Like this.

Citing other stuff and saying it’s different is fine. Cite our data and provide specific critiques; great. We can always learn. An offhand “This is the forecast; we think it’s wrong” seems a little strange. I’m glad you guys could respond directly to some of that in this article, and I look forward to walking through ’27 and then continuing to see if the track record on your data center model holds true.

Just to be clear, the goal wasn’t to respond to one person. It’s a huge topic that people ask us about all the time. Our institutional clients have been aware of how we do things and what we’ve been projecting, and I think we wanted to give a piece of that to our broader audience, put the content out there, and see how folks would react. But, yeah, the point is—

Yeah.

—we’re extremely bullish on how big the infrastructure build-out will be in 2027, and we think we have rock-solid indicators to back that up.

It’s not the first time we’ve gotten the feedback that our numbers are insane: “We don’t know what to do with these numbers. Nobody agrees with them.” So we wanted to address that directly.

Awesome, guys. Great job today. I appreciate you taking the time to educate everybody, as always, on what you spend your days doing and how accurate it is. Appreciate it.

Cheers, Jordan.