第018期 - 别再说2026年美国数据中心容量有一半被取消了（数据中心、能源）| Jeremie Eliahou Ontiveros, Reyk Knuhtsen, Ellie Holbrook, Jordan Nanos

Hello, everyone. Welcome back to SemiAnalysis Weekly. I'm here this week with the data center energy and industrials team. That's Jeremie, Ellie, and Reyk. We're going to talk about a couple of articles they put out recently. The first is “Stop Saying Half of 2026 US Data Center Capacity Is Canceled.” Nice title. We'll clarify some things there.

Good title.

Yeah, pretty descriptive. The second one is about behind-the-meter data center power generation: 40 gigawatts by 2028. Guys, welcome to the show.

Thank you.

Thank you, Jordan.

Good to see you again.

Yeah, thanks for having us.

We're covering 3 continents on this one. We got Reyk in Singapore.

The latency.

We'll strut on to—

The sun never sets on SemiAnalysis, bro.

Normal day, yeah.

1. The Cancellation Claim Collapses

Okay, let's dig in. The first article had a great title, and obviously the conclusion didn't bury the lead at all. I've been seeing this all over the media, where everybody keeps saying that half of data center capacity is canceled. You walked through some reasons why people are saying this, and then clarified what's actually happening. Can you give us a lay of the land as to what's reality and what's fake?

Yeah, I can start with the why. I think Bloomberg started this with the big headline, “Half of US Data Center Capacity Is Delayed.” Everyone else started piling in and citing the same number. Most of these articles point to the same underlying source, which is a report that's available out there.

When you look at the report, it says there were 12 gigawatts of data center capacity scheduled to come online in the US in 2026, and only 5 gigawatts are under construction. For us, when we saw this, we were like, “Bro, why is everyone talking about this? It's just not possible.”

You don't even need to do anything fancy. You can disprove this data so easily. Amazon announced publicly that they built 4 gigawatts in 2025. Is that going up or not in 2026? Obviously, it's going up. Amazon alone is probably going to add 5 gigawatts or more. That's basically all they think is going live—one company.

Obviously, you add all of the hyperscalers. CoreWeave is going to add a gigawatt in 2026, and all of that is under construction. So you're telling me CoreWeave's 1 gigawatt under construction is a fifth of the market? Come on. There's a massive issue in the denominator. You're just wrong. Just don't publish that.

That's the high-level take. Reyk, if you want to add more details, but for me, that was the trigger. It's just looking at the underlying source and saying, “Oh, guys, you're just off. It's just wrong.”

Yeah. We had seen this getting paraded around quite a bit. Everybody would go viral just reposting it, and it was mind-numbing to see every time on my Twitter timeline.

Obviously, our clients are pretty smart. They're going to take every data source they can and compare them against each other. We also had to help every client on that front, and it was like, “Why don't we just write a newsletter about this and say, ‘This is just really fake. We don't have to go into this as much anymore. Please just read the newsletter.’”

A lot of the baseline here isn't really worth talking about sometimes. That's why we ended up writing it.

Yeah. In addition to that, it's not even that the spirit of the article was fake, in a sense. You can imagine or forgive somebody getting the exact numbers wrong, but if I look at this chart that you put out around the outlook you forecasted—from roughly April of last year to May of this year—it hasn't even changed by 5%. I'll put this chart up on screen.

Maybe you can talk through what the reality of some of these forecasts is: some things get delayed, but very few cancellations are actually showing up in the market, right?

I would disagree. I think there's a pretty substantial amount of cancellations. What we keep saying is that these cancellations are early-stage projects.

I'm not sure if you have this one on the screen, but there's a map of the US where we show the large load requests, which is slightly different but makes the same point. As of today, you have over a terawatt of data center load requested by operators in the US alone. This chart is from 6 months ago. It's more than doubled now.

We have over a terawatt. Obviously, the whole US system right now has a peak load of 750 gigawatts. You're not going to double it in just a couple of years. It's not possible. It's not a reality. So, obviously, this is fake.

There's a lot of early-stage projects, and it makes sense. I think the data center market started booming toward the end of 2023. That's when people first started to realize there was going to be a big constraint on this. The first big deals started getting signed at that moment.

Since then, there's been this massive search for power. Everyone was trying to find where power was available, and that leads to certain behaviors. Because everyone is doing it, if you want to be successful, you also have to be aggressive and plan multiple options.

That's actually nothing new. Hyperscalers have always had multiple options when evaluating projects. It's even more critical under these constraints, and so you would sometimes see hyperscalers with maybe 10 different options for a single final investment decision.

That would be 10 projects where they talk to the local counties, talk to the utilities, and so on. These are not all realistic. They're testing the field: Which county can enable me to build my large-scale project? Where do I have the workforce? Where do I have the supply chains? Where can I build it?

There's an oversupply of very early-stage projects. Our point—and that was Reyk's brilliant phrasing—is the “cloud-coded projects.” I just love that. This is amazing. That's the kind of thing that AI is, I guess, not very good at filtering.

When you have humans in the loop, it's pretty obvious that some of these announcements are way too aggressive. We gave a bunch of examples in the article of people who announce a 10-gigawatt project and say the first tranche of 500 megawatts is going to be available next year. Then you click on the website and see “Contact us,” with nothing more. You start digging into the permits and whatnot, and you see nothing.

At some point, you're just like, “Okay, these people probably have a lot of land in Texas or something like that, but they don't yet have a real project.” Any human judgment would filter that. But I guess for AI, it's still pretty hard these days.

If you want to build a forecast of the industry, you have to base it on realistic forecasts. Extensive triangulation has always been our playbook—not just for data centers, but also for chips, because obviously this has downstream implications for Nvidia and upstream implications for the AI labs, their revenue, and all of that. It all connects to each other.

That's the SemiAnalysis flywheel: covering every single one of these industries and building a cohesive view. Others don't do that and easily struggle to find the ability to put these side outputs together.

Makes sense. In the article, you cover 3 different types of data center delays. One is an aggressive announcement by a newer data center developer. The second is an advanced project with overly optimistic construction timelines that might just be delayed or pushed out. The third is something facing permitting and local opposition issues.

Ellie, can you come in here and explain one of those examples where we look into it and find something in the permits?

2. Oracle New Mexico Faces Delays

Yeah, sure. The project facing quite a lot of local pushback—not just local opposition in New Mexico, but also from the jurisdiction and regulatory authorities there—is Oracle's Project Jupiter in New Mexico.

Essentially, they're trying to construct a pipeline to feed the behind-the-meter data center in Doña Ana province. This pipeline hasn't been built, and the route they want to build on hasn't been approved. They haven't got a preferred route; they have a secondary approved route.

Essentially, none of the ways to get gas to the site seem very feasible.

So, looking at the pipeline issue they're trying to make happen, every FERC filing that I see come through on the docket shows that they haven't made much progress, or any progress at all, because essentially it's defaulted to a type of regulatory process for which there is no precedent for it being done sooner than 2 years. Given the local opposition as well, it's very unlikely that it'll be sped up. There are other options for getting gas to the site, like shipping trucked CNG and LNG to a facility, as we point out. However, at this scale, I think the only proven delivery of CNG is at around 200 MW—so, a maximum of 200 MW.

I think maybe even that's optimistic. There aren't many producers in the local area, and there aren't enough trucks to facilitate that. So essentially, there's not only a pipeline that doesn't exist and has no viable route, but getting gas to the site via CNG and LNG is even harder. It's just not viable.

It demonstrates an interesting bottleneck that could occur with more behind-the-meter facilities: building gas pipeline infrastructure, especially in territories like New Mexico, which is not an extremely friendly area toward that kind of infrastructure. That's how we basically look through the filings—we look through all of the filings. We already had it on our radar because they had issues with the initial turbines that they were trying to use at the site, so they ended up switching to Bloom fuel cells.

Through that research, I spotted this timeline, and I was just like, “This doesn't make sense. Why is there all this pushback, all these local comments?” And then FERC and the other regulators are still thinking about it.

So, is this an example of a Claude Code data center project?

So, I think what this speaks to is that we have to rethink the way we build data centers in order to meet this demand. But it's not like other people had built 1-gigawatt sites before this, right? We're still in the early innings of bringing the first gigawatt-scale data centers to market. We had to try a whole lot of different things. No one has experience doing this—maybe today, yes, but as of 1 year ago or 2 years ago, no one had experience.

Obviously, when you try all these new things, there are going to be errors. That's why, when we analyze big projects like this, we always try to think about what is a reasonable timeline. I think it's pretty common in the industry that you have folks who pitch more aggressive timelines, and that's valid for every single industry. Contractors—I think anyone who's done construction will tell you they tend to be always overoptimistic. Suppliers are always overoptimistic. It's not like this is anything new.

Especially when you try something much bigger at that scale, with new designs and so on and so forth, in new locations, new counties, and new municipalities handling it, then obviously you're going to face a few challenges. So, that's a great fit for a category of projects that are already well advanced because they already got financing. They got a lease from Oracle for OpenAI—or, sorry, Oracle contracted it to OpenAI—so there's already a massive contract on this, but they're facing execution challenges now.

3. Behind The Meter Takes Off

Yeah, makes sense. Maybe we could talk about behind-the-meter now. You guys published some numbers on this in the second article. It was the first Y-axis on one of your charts in a little while there, Jeremie, which was pretty cool. So, when we talk about the size and scale of some of these projects, I guess the behind-the-meter net additions, when compared with the available grid capacity, are a significant difference.

You specifically forecasted over 40 GW of net additions of data center capacity behind the meter by 2028. That number effectively rounds to 0 right now, right? There are a few behind-the-meter projects, but people are just getting going, ordering turbines, and the supply chain is ramping up. What does it actually take to get there? And do you expect behind-the-meter projects to be more at risk of cancellation based on some of the reasons Ellie just described, or less at risk of cancellation because they don't depend on the grid?

Excellent question, man. I guess the first thing is that this forecast mostly depicts our analysis of the US grid and the gap. For context, we're not adding enough generation on the grid to meet that demand, which is going to be in the tens of gigawatts per year and just keeps increasing every single year based on all of the signals that we keep seeing, right? So, is it realistic to assume there are going to be 40 GW of new behind-the-meter data centers added by 2028? You have to look at what is being planned right now.

I can tell you that in the last 2 weeks, 3 GW of data center deals were signed for behind-the-meter purposes, so deals are happening. Site selection, I think, happened beforehand. I think you saw a massive move in part of 2024, mostly 2025, where folks really started to have access to a gas pipeline as one of the main site-selection criteria. I think you started seeing diversions where maybe some of the traditional data center operators were more focused on, “Let's be in a tier-1 market like Northern Virginia. Let's find grid-connected sites because we need the five nines.”

Then you saw newer operators, maybe more AI-native, that tried to foresee this trend. I think Crusoe is a great example. They've been quite ahead of the curve on this, and they've been able to sign massive deals, like 672 MW in Abilene with Microsoft, announced publicly by Crusoe in Q1. That's behind the meter, for example.

In terms of whether the supply is ready for it, there are many developers that have secured sites. In terms of manufacturers, Ellie can probably do the extremely detailed rundown. In terms of the probability of delays, I think the probability of delays is lower with behind-the-meter because of the way these grid constraints play out. As a developer, you talk to a utility, and they have no binding obligation to abide by the schedule that they provided you.

With the anecdotes we hear from developers, they say, “I thought I was going to get 500 MW by 2027. I had a handshake with the utility or whatever.” Then, a couple of months later, they tell me, “Actually, sorry, bro, I can't do it. It's going to be 100 MW by 2028. Your 500 MW will be by 2032 because I have to do bigger network upgrades because, hey, actually, I didn't consider in my analysis that another guy also wants it. I don't have enough generation coming,” and so on and so forth.

Utilities are realizing they're faced with more delays than they thought. In many cases, if your strategy is only grid, you're extremely likely to be disappointed. If anything, that's more so on the power side. Behind-the-meter adds a new risk, which is more at the execution layer. Can you build a power plant on time? Can you get the permitting? That adds a new set of complexities.

We're going to see some high-profile delays. Obviously, the New Mexico one, based on our analysis, is the highest-profile, right? Basically, it's a $100 billion deal for OpenAI with Oracle, so it's the highest type of delay that you can have. I hope they solve it on time; wish them the best. I think you're going to see more of that, but I think in terms of raw volumes, because there are so many options now for developers, you're going to see a lot of success stories as well.

Colossus 1 and 2 are proof that you can do this at scale. Now, I don't know if they're the best examples because of the issues they have with regard to permitting, but they're demonstrating that it can be done, and there are others doing it. Crusoe is another example.

I was just going to add to that: we're seeing the OEMs really reacting to the behind-the-meter story as well. We're seeing, obviously, the major OEMs—the established players, like Siemens, GE, Lenovo, all those guys—expanding manufacturing capacity. But then, throughout the stack, you're seeing new players, new entrants, new types of technology, and recycled old types of technology, which we've written about in several different notes.

I think that's a good tell for how this market is developing at speed. On permitting, that's something that could cause delays, since we've seen companies switch to Bloom fuel cells in order to get speed to power, possibly because of Bloom fuel cells' manufacturing footprint, but also because of their lower NOx and SOx emissions.

I think the Nebius New Jersey facility, for example, switched to Bloom. I think it had Bergen turbines previously, in order to combat the permitting issues there. We might see technology shifting around, maybe relocating more to other states, like Texas rather than the East Coast. But that's the story going on in the BTM market right now: full growth.

Makes sense. Jeremie, maybe we could go back to one thing you said related to the size of some of this. At this point, when you say that you know about 3 gigawatts signed in the last week, and we're talking about, in the context of the article, 40 gigawatts of behind-the-meter power generation by 2028, can you just, at a high level for the general audience that listens to this podcast, describe a rough timeline for when something gets signed, when construction starts, when you can power the first tranche of GPUs, and when a large site, like a gigawatt-scale site, might actually be completed?

Is 3 gigawatts signed in mid-2026 actually going to have anything by 2028? Is it a 2029 or 2030 story? Is this a year and a half, 2 and a half years, 3, 4, or 5?

I would say it varies a lot depending on what is actually signed. There's a bunch of deals out there. The bulk of the volumes these days would be turnkey leases, where, let's say, a company like Digital Realty or QTS builds a data center for Microsoft. They sign a firm lease, a turnkey lease, where QTS takes on everything and Microsoft just rents.

You have powered shells, which are a bit different. That's a lower burden on the developer and more CapEx from the tenant. You have just power deals, where it's basically a PPA—for example, Oracle with VoltaGrid at the Shackelford County, Texas, site. Essentially, that's a PPA; there's no data center involved. So, depending on the deals, the timelines can vary.

The rule of thumb would be that the deal is signed, and 18 months later you have capacity for the first tranche. Then, depending on how fast you can build it, I think it's pretty clear that these days the expectation is that the ramp from first capacity—the first phase—to full ramp is expected to be faster and faster.

The point we make in that article is that, from a buyer's perspective, behind-the-meter is now becoming much more attractive than the grid because you're in control of your destiny. You know you're going to have gigawatts by a certain time. Now, obviously, it has to be permitted and so on, but at least from a power standpoint, you know exactly what you're going to have on-site and when, provided that there aren't delays from suppliers and so on.

The idea is, if you have all of this power, then you also need to build a data center, because otherwise it's useless. The expectation is that ramping up these 3 gigawatts isn't going to take 5 years. It's going to be much faster than that. Generally, what you observe in the marketplace is that when there's a gigawatt-scale deal being signed—let's say Oracle with STACK in New Mexico or with Vantage in Texas—you see financing for the whole project take place shortly after.

We saw DigitalBridge raise money for Vantage maybe a month after the lease was signed, and we're talking about $25 billion of financing. The whole amount is already secured. Once you have that deal, you secure all the money, and then you just try to build as fast as possible. It's not going to take 5 years. It's going to be much quicker than that.

Yeah. So, not 1, not 5—somewhere in between. Makes sense.

4. Hyperscalers Drive The Buildout

Okay, I want to show you guys one chart and get your reaction here because I thought this one was fascinating in terms of understanding the market share. I think a lot of the people who buy the model are using the data that you guys provide to make decisions for companies. On the legend here, AWS is in orange, Google in blue, Meta in this fuchsia-pink color, let's call it, and Microsoft in green. These are the hyperscalers, and they make up a large percentage of the market in terms of how many megawatts they have under construction.

The market share that these guys represent going forward is really interesting as we see them compete with each other. Reyk, what's your take on the relationship between the hyperscalers competing with each other, and then hyperscalers versus everyone else in the market, in terms of just how big everybody's going right now?

In terms of competing with each other, I think when we take a look at the capacity race, you're going to notice that it's mostly driven by the respective AI labs. Most of the time, an AWS capacity build-out will be fueled by Anthropic, right? Meta will typically be for MSL. Microsoft is typically for OpenAI as well. We know that there are some Anthropic deals going on there as well.

These are, in any case, the drivers of the data center build-out right now, because they're the ones with the capital able to actually build out the data center capacity that's needed for the AI labs. They're the ones with the investment-grade financing able to start these projects and get them moving forward.

Most projects struggle with actually getting the financing. If you're a neocloud—which, if you're listening to this later on, maybe you've already read the other article—it's a bit hard to get the financing to start, get a data center, buy the GPUs, and spend the capital for all of this build-out, because it's probably going to come out to billions of dollars.

In essence, you're mainly left with hyperscalers driving the AI build-out, or very well-funded or well-capitalized neoclouds. If we look at any of the Blackstone-backed guys or any of the KKR-backed guys, they've got plenty of money coming in from those firms to fund their own build-out. In any case, I forgot your second question, actually. So, if you could say that again.

No problem. Yeah, yeah.

It was interesting. Maybe let me ask a different one because I'm thinking of it now, listening to you talk. At the very beginning of this article, we were commenting on the public research that has driven all of these articles in the media about how half of 2026 U.S. data center capacity is canceled.

But that chart that I put on screen, and roughly what you're saying about the hyperscalers building for the AI labs, says that even if you just take 2 of those 4 hyperscalers that have more than 5 gigawatts under construction, and then there's the whole rest of the market—the other 2 hyperscalers and everybody else—you can see that there's a massive amount of capacity.

Can you opine on where the gap is in terms of people trying to do research on data center capacity under construction and just missing entire gigawatts' worth of capacity that's a real thing for next year?

Yeah, yeah.

Luckily, Jeremie was able to explain the whole article in 3 minutes earlier, so I can expand on what he said. He's a bit too efficient with his communication.

Basically, when you have a model like ours, we do everything in a bottom-up way. Each individual data center is, for example, a row in an Excel file for us. You can get the forecast for each individual building, the tenant, and, if the end user is OpenAI or Anthropic, you already have that information for that specific building. Then you can go ahead and find the timeline for the building.

A lot of people don't have the expertise to judge timelines on these things. When we do our data center research, I like to joke that Jeremie and I have probably seen 10,000 to 20,000 satellite images. Not a lot of people can say that they understand what that single gray pixel on a brown floor means. It's like, “Oh, that's actually the concrete pad coming in. That's going to be a good sign.”

Hopefully, we see vertical construction, which is the darker gray pixel in the corner over there that most people wouldn't get. The satellite images are super useful for us, but that's just one that's really fun to talk about. The other stuff is a bit more boring: we go through every permit portal for every county, city, state, and country in the world. We like to talk about how much we use Claude, as you see in the article: 170K in a week.

The Claude part—everybody loves the Claude part. Anyway, we’ve got our lovely chart on the front of the page: “$170,000 spent in a week,” right? We’re token-maxing, you know; we’re token-mogging Meta, or whatever you want to call it—more tokens per employee than Meta.

But in any case, we tend to use Claude. We’ve built out basically an entire harness and workflow for our agents—we love our agents—that go ahead and scan and scrape every permit portal basically around the world, down to the state, municipality, city, and county level. With this, we get a ton of public filing information that allows us to basically estimate timelines. Without going into too much more detail, that’s 1 part of basically 3 in our methodology, and for the other 3, you can also buy the model to figure that out.

Yeah. And so, Reyk, I put the Claude Code spend chart on screen here. To be clear, the color coding is a per-user breakdown, I believe.

Yes.

People who are looking can see the really big—

Yes.

—the yellow block at the bottom.

Yeah, yeah.

Yeah, the culprit may be on this. The culprit is on the phone right now, yeah.

I’m 90% sure that’s Jeremie.

Good forecast.

That’s Jeremie just melting Anthropic’s GPUs at the bottom there.

Yeah, yeah. The ones that he’s looking at from Blue Sky are also cooking up some of his requests there.

Yeah, exactly. They saw Jeremie spend, and they went and bought out the SpaceX capacity right after. So thanks to him.

Good forecast.

Yeah, we’re moving markets now like that. Uh, go ahead.

Ellie, what are you thinking about over there?

No, I was just going to make a joke about Jeremie’s thesis about making Anthropic go to the moon. But, yeah, it’s going to be interesting to see how much of this is going to be built behind the meter by the main AI labs as well.

And I think behind-the-meter is 1 term. There are various different definitions of moving energy on site, not taking all of it from the grid. I mean, you can co-locate your energy, or you can have a net-metering solution, which is more of an interactive relationship with the grid. You can have a fully islanded, off-grid energy supply for your data center.

But, yeah, I think that’s going to be interesting to observe in the future as more of them are built, seeing what’s preferred. I mean, which state: fully islanded, or a bit of grid connection? Maybe that might hedge some interconnection in the future. Also, what happens to these assets—these turbines—once maybe they do connect to the grid? They might end up being part of the grid themselves.

A lot of these turbines could be used as peakers, gas peakers, and help solve the actual lack of power on the grid itself. I mean, obviously, there’s transmission to sort out, but I think it’s more symbiotic than is portrayed by a lot of other outlets. I don’t think it’s 1 or the other forever.

5. Power Suppliers Expand Capacity

Makes sense. Let me throw this chart on screen that I thought was really awesome from the second article about behind-the-meter power generation, based on what you’re talking about. This here is a breakdown of the supply ceiling by OEM per year, so we can see the increase of some of these suppliers that would be responsible for power generation in behind-the-meter gas settings and how much capacity they’re increasing in terms of how much they can produce.

Two things jump to mind when I look at this. First of all, it’s just how much it’s increasing, so it seems to be going up at roughly the rate at which we’re going to increase the power-generation requirement. That’s a good thing. The second thing is just how many different suppliers are available.

We’ve looked at constraints in the market, whether this is GPUs, wafers, memory, or anything else. Generally speaking, when something’s a constraint, it depends on 1, 2, or 3 suppliers. But when you have this diversity, where the top 8 that are on screen and color-coded make up just more than half, let’s say, and then your footnote at the bottom of this chart shows 25 more names that can possibly produce power, we’ve talked about this previously. Maybe you guys can walk through just how diverse the options available for power generation are and what sort of shared supply chain benefits people who are looking for alternatives when they need power and aren’t sure where to get it for behind the meter.

Yeah, sure. As I was saying, there are established players who have a large supply chain, and they’ve been in the business for years and years and years. They have manufacturing capacity globally. They quite often say in earnings that the bottlenecks are the castings or the blades for whatever kind of turbine, aeroderivative, or reciprocating solution they’re offering.

But it seems that this is not slowing anything down, because they’re still announcing capacity increases, and they’re still saying that these capacity increases are driven by AI demand growth. It seems to be a bit of a narrative talking about the constraints to castings and upstream constraints to building these turbines—a narrative from last year.

I think this year, the large OEMs are realizing that many more players are coming to the market, and they’re trying to keep their market share and not lose it. I think that literally was in a Siemens pre-close call this morning—or yesterday. They’re very much aware of this growing market, and they can’t play the scarcity game and try to get everyone to freak out. People did not freak out; they saw an opportunity and are converting old parts of plane technology into engines and then boilers. Obviously, there are aeroderivative options as well.

We’re seeing all of these different types come into the market, some with more of an advantage in terms of permitting and some with less. I think there’s also something I’ve been noticing: more of the car industry is having a larger role in the data center industry, in the sense that EVs are not as popular. I mean, they’re quite popular in California and on the East Coast, but in lots of the U.S., EV growth has been quite slow.

A lot of that manufacturing capacity, I think, is going toward this market. We have Tesla as a good example, and then we saw Ford as well providing BESS—battery energy storage system solutions—to the data center industry. So we’re seeing these existing players switch to the data center industry as well. I hope that answers your question.

That’s cool. Makes sense. Jeremie, what’s jumping to mind here when you’re hearing us talk about all the different suppliers that you can use for behind-the-meter power generation?

Yeah. I think it was always to be expected. Last year, the narrative was that behind-the-meter was not possible because there were these 3 manufacturers that were so slow. Everyone was throwing around the famous chart of gas orders in the early 2000s, where you had this massive wave, and everyone was saying GE Vernova and Siemens are so scarred by that era because they invested massively in capacity and are going to be very conservative, right?

But that leaves a market opportunity. One question we like to ask management teams at these power companies is, how AI-pilled are you, essentially? How much do you believe? How much risk are you willing to take? There’s also a function of how easy it is for you to take risk, in the sense of what your economics are on building more capacity and new factories. There are a bunch of companies that score very well.

The one that we've been flagging for a while—we especially doubled down at the end of 2025 when we did our big deep dive—is Bloom Energy. I think clearly the management team is very AI-pilled. I think they have economics that enable them to build capacity faster than others.

The behind-the-meter conversation is really interesting because I think it's company by company, just adjusting their mindsets as they adapt to the new reality, and also solution by solution, you keep adjusting your expectation based on the constraints. So what I'm saying is that initially everyone was like, “Okay, this is going to be bridge power. This is going to be a matter of 1–2 years. I run off-grid, and then I'm going to have my grid come in, and maybe it's going to be backup.” So you would only consider systems that are good at backup.

But that's the ultimate disadvantage for something like Bloom: it's not very good at backup. When you run it extremely hot, it's 15,000 degrees Celsius. It takes 2 days, as far as I know, to go from 0 to 100. So for backup, it's really not that good a system.

But hey, if you have no other alternative, then maybe that's it. Maybe you just have to go for it, right? If your other option is “I'm not going to have power,” then you're screwed, and you're not going to be competitive in the marketplace.

Other issues are power costs. I think when you look at the recent SpaceX deals, it's pretty clear that the revenue per megawatt that folks are making on the cloud side, also on the lab side and on the model side, goes up at every layer, I guess. As that increases, it means that power costs are increasingly irrelevant. Power costs for solutions like Bloom aren't that expensive anyway, relative to what we have in Europe, for example, with the grid.

Well, certainly not as a percentage of the total construction costs of the project or when you consider the GPUs and stuff. People can pay double for the turbines they want compared to somebody else competing in the market for them and not really affect the total CapEx of the project, right?

Yeah. The other thing is also estimating the bill of materials, which I think has proven quite complicated for many of these vendors, especially as labor rates are going to the moon. You thought you were going to have local labor, but actually everyone in Texas is already occupied building data centers, so you have to call people from Denver or Ohio, and obviously it's much more expensive.

This also favors solutions that are fast to install, fast to deploy, because your BoP is more predictable. Your timelines are more predictable, both on the labor side and on the full-deployment side. So again, analyzing timelines and ease of deployment is key, and I think some solutions score extremely well.

I also think we're going to keep seeing more and more new entrants, because everyone looks for capacity earlier and the standards are dropping, right? There are a few new companies that entered recently that are basically coming from the automotive industry—not themselves; they source engines from the automotive industry.

When you think of this, you're like, “Wow, automotive.” We're talking about 100 gigawatts-plus per year of production. That's an industry everyone knows runs at very low utilization rates—50%, 40%. These factories are not doing too well. So the capacity and the incentive that these guys have to sell their engines to the data center market are also tremendously high.

Yeah. Yeah.

That unlocks gigantic capacity.

Absolutely. At a minimum, it's diversification for their business. The one thing that we haven't talked about necessarily yet is alternatives to gas, let's say. I'm sure a lot of people who are new to this space and are listening are thinking about renewables: What about solar, wind, hydro, nuclear? Who's considering that, and what's going on?

I'll throw 2 charts on the screen, and then hopefully you guys can comment just on the reality of being able to use renewables plus batteries or any renewables, let's say.

Yeah.

Yeah, go for it.

So yeah, this is grid-connected. The point of this, obviously, is that nameplate additions are overall growing, and we're talking about 60 gigawatts, 50 gigawatts per year, so it's a lot. But as everyone knows, 100 gigawatts of solar is not a true gigawatt for the grid because it's intermittent and only turns on at certain hours and so on and so forth, right? Pretty simple.

It depends on the area. You see an ELCC value, adjusted for the actual capacity value it brings to the grid. Depending on the area, it can be 10% of it, 20%, sometimes below 10%, sometimes over 20%. It depends; it's all a very complex system-level calculation.

The problem is that the grid is also facing a lot of these transmission issues. It's intrinsically slow. There are giant interconnection queues. So I think the real question for solar and batteries is: Can you do it behind the meter? Can people build these things on-site? I think it's going to happen.

Obviously, the high-level challenge is that the land required to build a massive amount of solar is just tremendous. Logistically, these projects are fairly complex. If you're talking about, “Hey, I need 20,000 acres just to throw solar panels at it,” and it's going to be for a gigawatt data center or something like that, then buying all that land can get complicated.

Sometimes what happens during the course of this land buyout is that the landholders realize, “Hey, if this guy wants to build a $100 billion project, maybe I'm going to sell my land 10× more expensive than what he thought I was going to get,” right? That sort of stuff happens as well.

The complexity of building these massive projects is fairly elevated, which removes the time-to-power angle to some extent for solar and batteries. But as folks get bigger and bigger and start planning multiple years ahead, I think there are already a lot of big projects like this underway.

The other issue is obviously SMRs and nuclear. I'm not going to reinvent the wheel. This is slow stuff, right? Everyone knows building nukes takes time. So again, it's not a time-to-power option.

Right now, you have a lot of non-binding LOIs in the market, or non-binding deals that are contingent on milestones, execution, getting the regulatory approvals, and so on and so forth. I think in the 2030s we're going to see a gigantic diversification of energy sources to power data centers. I think for the 2020s, we're going to be very much in the gas world.

Makes sense. Okay, as we move to wrap here, guys, anything left unsaid? Ellie, what you got?

I was just going to say that we might see, as we move on, more carbon capture, utilization, and storage facilities being co-located with BTM sites and with data centers as well. I think that's something interesting to observe going forward.

Good to know. Reyk, how about you? What's jumping to mind, man?

I don't know if I have too much here. I think we covered what the listener needs to know, I'm not going to lie. No, I don't know. I think the BTM movement is going to be very, very big, but maybe Jeremie touched on it already. It's interesting to see what we call “peak turbine” in 2026, where there was a huge overbuying or a huge over-purchasing of turbines in 2026, and not everybody knew what to do with them.

Not everybody could figure out how to get the permits, or not everybody could figure out how to build the data center to actually use the turbines. So you ended up seeing a lot of these turbines go underutilized or, later on, start to hit the secondary market, for example.

I think this is going to read pretty bearish to people at first, and people might freak out about it. But long term, we're still very, very pro behind the meter. I just think it'll be interesting to watch how people interpret this, though, because it'll probably keep picking up, right?

You'll probably see your favorite project's turbines start to hit the market at some point and you're going to freak out, and then the question is: Is it over, right? We don't think it's over. We still quite like BTM, but I think that's going to be a fun dynamic to watch play out.

Okay, we've got to end by going around the horn here, Reyk. What's your favorite project?

My favorite project? Yeah, actually, I might have one.

I might have one. I would go with the Nscale Norway project. I love this one, actually. Love this one.

Okay, Jeremie.

Big Jeremie’s one.

Jeremie, what’s your favorite project? Or top 3. You don’t have to choose.

I mean, Colossus II, bro. Colossus II: fast, efficient, scale, super-high revenue. W.

Ellie, what’s yours?

I would say the Nebius—yeah, switch to Nebius in New Jersey, so it’s the Nscale, right?

Why not the—

Hey, Reyk, I got a question for you.

... why not Oracle New Mexico? All right, go ahead.

I don’t know. Oracle New Mexico, possibly. Maybe, yeah, I guess that’s my only thought. That’s the one I—

That’s lovely. Ellie spent the most time with that one, I think.

Reyk, when are we gonna get robots building data centers?

Yeah, Jer. Yeah, yeah, yeah.

We need robots building data centers. Hey, they’re coming. They’re coming. Hey, we should, uh, you guys should watch the other podcast. You should watch the other podcast. We talk about this.

Give me a year.

I don’t even think we did.

What year are we gonna see a data center fully built out by robots?

Fully built by robots. Wow. That’s a lot of man-hours.

End to end. No more electricians.

A lot of man-hours. What is it, like 8.5 million man-hours? Dude.

It’s a lot, man.

2029 at the earliest. 2030. 2030.

Reyk’s going to back-calculate actuator constraints in the supply chain to be able to just figure it out.

Yeah. Hang on, let me think about how many permanent magnets are being manufactured right now. Man, that’s a tough one. That’s a tough one, man.

Yeah, the long pole. Oh, man. All right. Thanks for joining, guys. Good job.

All right, thanks.

All right. Cheers. Bye. Bye, Dylan.