For those steeped in the world of AI development, the major differences between tech giants’ chip strategies are easy enough to see: the speed of Google’s TPUs, the affordability of Amazon’s Trainium, and Nvidia’s market dominance. But those same companies’ respective energy strategies often fly under the radar.
As data center buildouts surge across the U.S., the tech industry is hitting a massive wall: a power grid that can’t move fast enough to support them. In this episode, Shayle sits down with Jeremie Eliahou Ontiveros, who leads coverage of infrastructure and power at SemiAnalysis, to lift the curtain on the hyper-competitive world of AI energy procurement.
They cover topics including:
The hyperscaler leaderboard: Why Google remains the most energy-sophisticated tech giant, as others continue to innovate
- Why power has effectively become revenue for frontier labs
- How Google uses its massive balance sheet to provide financial backstops for Anthropic’s data center buildouts
- Why the interconnection queue is forcing AI labs like OpenAI and Anthropic to bring their own generation
- The sudden market frenzy over modular speed-to-power options—ranging from aeroderivative gas turbines and massive reciprocating engines to billions of dollars of fuel cells
Resources
- Catalyst: How data centers are complicating transmission expansion
- Catalyst: Live from Transition-AI 2026: Inside Google’s massive AI capex
- Catalyst: The rise of flexible data centers
- Catalyst: AI scaling pathways: On grid, on edge, off grid, off planet
- Open Circuit: The new reality for data centers: No easy answers
- Open Circuit: Can data centers regain their social license?
- Open Circuit: Grid utilization vs expansion: The 100GW debate
- Latitude Media: FERC to grid operators: Connect large loads to transmission faster
- Latitude Media: The rise of the data center power exchange
Credits: Hosted by Shayle Kann. Produced and edited by Max Savage Levenson. Original music and engineering by Sean Marquand. Stephen Lacey is our executive editor.
This episode of Catalyst is brought to you by ENGIE, the smarter energy supplier. ENGIE doesn’t just provide the power to run your business — they supply the energy to move it forward, with reliable, flexible solutions built for what’s next. Learn more at engieresources.com.
Catalyst is brought to you by EnergyHub. Peak season puts every grid to the test — and the utilities that pass are the ones that built flexible capacity before they needed it. EnergyHub works with more than 170 utilities to coordinate 2.5 million devices and 3.4 gigawatts of dispatchable flexibility through a single platform designed to perform when it counts most. See what that looks like at EnergyHub.com.
Catalyst is brought to you by Bloom Energy. Bloom Energy fuel cells deliver affordable, ultra-reliable onsite power for hospitals, utilities, and data centers – at speed and at scale. Learn more by visiting BloomEnergy.com.
Transcript
Shayle Kann: I’m Shayle Kann. I invest in early stage companies at Energy Impact Partners. Welcome to Catalyst. So let me make two observations. First, if you’re paying close attention to the AI world, I’m guessing that you can articulate pretty clearly the major differences amongst the hyperscalers, maybe even also the frontier labs in terms of their chip strategy. You know about Google’s TPUs, you know about Trainium coming from Amazon, you know what Microsoft and Meta are doing. You probably know about the partnerships Nvidia has set up across the value chain, who’s vertically integrated, who’s not. But do you have the same level of knowledge down to the individual company level about their power strategy? I would guess no. Here’s the second point. You’ve heard, I’m sure, about the war for AI research talent going on amongst all these companies. TBPN and its ilk reports every single time a top researcher switches teams.
But do you know the same thing is happening on the energy teams? I think there might even be more teams switching there lately, which is all to say I think that the topic of exactly what strategy all the companies building out the AI infrastructure are employing with regard to power is poorly covered and not well understood. So let’s fix that. For this one, I brought on Jeremie Eliahou Ontiveros, who leads coverage of infrastructure and power at SemiAnalysis. He’s coming up next.
///
Shayle Kann: Jeremie, welcome.
Jeremie Eliahou Ontiveros: Yeah, thanks for having me.
Shayle Kann: You have a GPU sitting right behind you. What is it?
Jeremie Eliahou Ontiveros: Which version is it? That’s a Hopper H100. I run a lot of agents on my local computer now, just joking. I don’t have the GPU actually. I just did a visit of the factory of a neocloud that does a lot of pretty cool engineering stuff and so they gave me this token of appreciation.
Shayle Kann: Good swag. I haven’t been given a GPU yet, just hint hint to everybody in the audience. All right, I want to talk about the power strategies of the companies who are building, owning and operating data centers. Here’s a first question for you. We’re going to talk about the different categories, but let’s focus on the hyperscalers for a second first. When you think of the hyperscalers and how they approach power, provisioning power, getting enough power to build the infrastructure that they want for AI, do you think of them as being fairly monolithic and they are all approaching basically, do they all have the same strategy in your mind and they’re just in a land grab or do you see meaningful differences within that group?
Jeremie Eliahou Ontiveros: I think there’s pretty meaningful differences. Company by company, you see really varying degrees of first of all, USA versus international appetite to sort of behind the meter versus grid connection. Sort of location of data center, how close to the end user versus sort of middle of nowhere big campuses. So I would say pretty different overall. Also with regards to the way they negotiate with utilities. Generally speaking, I think it’s fair to say that Google is the most sophisticated company. And on the energy side, they have sort of the biggest trading desks. They’ve stroked some pretty big deals with utilities as you probably know for load flexibility kind of stuff. So they’re definitely sort of at the frontier of innovating on the energy side. Another way you’d see this is when you look at the minutes of the conversations with officials in PJM or ERCOT, you always see Google’s name.
You generally see them more than others. So I would say probably the most sophisticated company is Google. But other companies have different strategies. For example, I would say Meta was probably the first among the four big guys to adopt behind-the-meter at a bigger scale.
Shayle Kann: Is that the Louisiana project or which is… There’s like a five gigawatt behind the meter project that they’re doing somewhere?
Jeremie Eliahou Ontiveros: It’s actually in Ohio, in Columbus, Ohio. So as of today, they’ve announced two major training clusters. There’s a bunch of others that are on their way. But the two majors they’ve announced and the sort of closest to delivery, there’s one in Columbus, Ohio and the other one is in Louisiana. The one in Ohio is really interesting because they actually rolled out a new type of data center design that is significantly faster to deploy. We call it the tent because when you look at it, it literally looks like a tent. The reason, speed, and they’ve done a whole lot of crazy stuff in Columbus, Ohio. I think probably one of the craziest examples is you could see a picture where basically there’s a piece of land that they own. There’s a piece of land nearby that is from another company from whom they’ve leased the data center.
And so they have these two substations and then they own another piece of land in between. This is where they build the tents, the one in between. They didn’t have the time to build a new substation. So essentially what they did is they took the medium voltage current from sort of the two already established utility connected substations. And so they’ve built a new medium voltage line to feed their new data center, which means that you get the idea they’re not adding new data generation, they’re using the same transmission line. So you could imagine in peak summer days, they’re going to have to actually curtail that data center. So that was sort of a temporary solution and now they’re building onsite power plants with, if I remember correctly, solar titan turbines from CADs to deploy to get power on that side.
Shayle Kann: Yeah. From my observation, I’d say both things that you said so far ring true. One being Google is the most sophisticated. I think actually Google was best positioned because they have long been the most sophisticated on energy. No offense to my friends at Microsoft who I think have been kind of second behind them historically, but Google was always pretty deep in the energy world. They were early to PPA, offsite PPAs, VPPAs. They were the first to make the twenty four seven clean commitment and that required a lot of sophistication. They were doing carbon aware flexibility long before they were doing this version of flexibility. And so they were like shifting workloads. This is pre AI, right? But they were doing shifting workloads geographically in order to minimize carbon impact. So I think they had the benefit of all that. And the other thing I would say is that this is speculative, but anecdotally I’m pretty sure it’s true.
I think Google’s energy team is probably a multiple of the size of the next hyperscaler, just the sheer number of people they have dedicated to it. So that seems true to me. I would have said for a long time, Microsoft was a clear second there, but they’ve actually lost… This is another thing I’ve noticed, which is that the small group of people who have spent a long time focused on provisioning energy for data centers are hot commodities in the same way that AI researchers seem to be. And so there’s a lot of people changing roles. I think Microsoft has lost a bunch of people in that in recent days. And so now it’s not clear to me who’s kind of second in terms of sophistication.
Jeremie Eliahou Ontiveros: Yeah, 100% agree. I don’t think I would call Microsoft now the second. I probably would rank them fourth now. I think you’ve seen Amazon step up pretty dramatically and you’ve seen them at a frontier of a bunch of unique types of deals. For example, obviously Talen Energy was a pretty flagship one more recently, Vistra Comanche Peak. So they’ve done also pretty large scale PPAs with gas power plants, with nuclear power plants in addition to of course renewables. So I would view Amazon as fairly sophisticated as well. And I think Meta is probably doing the most unusual stuff of all the hyperscalers. And in some sense, that’s sort of where you see Meta being different because they don’t have this cloud business. They sort of have less use cases. They’re actually much more of a first party company. And that’s a pretty big difference when you compare this to Amazon and Google, which build their infrastructure for actual customers.
Customers are going to run their infrastructure and they need to have a whole bunch of different services so they build for maximum fungibility. And Google is kind of increasingly in between because they have more and more services. They used to be, I think, much more first party where they basically just had Google search, but now they also have Google Cloud. They have a whole bunch of other workloads. So they kind of sit in between. And that’s in Meta because they’re sort of fully their own workloads, their own cloud, their own infrastructure to support their first party workloads. It’s kind of easier for them to try new things. As I said before, they’ve been the first to move in a bold direction and behind the meter. Another interesting thing they’ve done, for example, is more of the data center level than at the energy level, but no backup.
You look at any one of the modern Meta dataset, they don’t have any gen sets or they have 10 megawatts for a hundred megawatt data center.
Shayle Kann: Yeah. I wanted to talk about that and that’s actually a really interesting point. So this gets to the behind the meter stuff, right? Which is I think people, I don’t know, people who are paying a little bit of attention to it, here are these huge numbers for we’re building mostly gas behind the meter. Occasionally it’ll be something else, but why? None of these are actually off grid or none of these are intended to be off grid long term, let’s say that at least not yet. You could tell me if you know of an exception to that that has been publicly announced. I know of some that haven’t been announced yet. Mostly what they’re doing is they are using behind the meter generation to bridge to a grid connection that is going to come at some point in the next few years. But in the meantime, if they are acting as a bridge and you are running off of those gas gen sets or gas turbines, then you’re not getting the four nines of reliability that you normally would provision for a data center.
Now, I think you make an interesting point. Maybe Meta doesn’t need the four nines in the same way that the others do. They don’t have customers who are relying upon their uptime. So maybe that gives them the flexibility to do, as you said, behind the meter, gas with no backup power, which is wild.
Jeremie Eliahou Ontiveros: Yeah. And look, that’s where I would actually think that the most important question is not the one you asked first. I would slightly change it. I think right now, if you want to understand the world of power, you don’t have to ask about hyperscalers, you need to ask about AI labs. These guys are the ones that are driving sort of the demand in the market. And so increasingly, you basically look at Amazon or Microsoft, how many gigawatts are building every quarter, half of that is going to either both of them combined, right? OpenAI and Anthropic. These guys are basically proxies increasingly for open AI and Anthropic. So maybe right now these guys don’t have the financial strength to do everything themselves, but they have a clear view on how to get there. And look, I’ll also answer your first question, which is, why do you go behind the meter?
And I think this is where there’s maybe kind of a misconception. I think just folks don’t understand the order of magnitude we’re talking about here. Just look at how much generation is being added on the grid. It’s just not enough. We’re actually publishing a report right now as we speak. I’m not sure if it’s live, it is going to be live probably the next couple hours. So when you look at the data center build out, if you’re looking at tens of gigawatts per year, the trend is increasing every year by high double digits. Let’s say 50% per year. Currently, no signs whatsoever of our slowdown, so things just keep going up. Now compare that against how much generation is being added on the system. How much gas is going to be brought online next year? Five gigawatts, six gigawatts, how much solar and battery, like 20, 25 nameplates, but as you adjust for ELCC values and whatnot, you just sort of add it all up.
There’s just not enough being added on the grid to support the build out. Now, how does that actually show up? You’re talking about generation constraints essentially. How does that show up? Basically you talk to any data center developer in the US, they’re all going to tell you the same story. They’re all going to tell you about this time where they talk to utility, they were promised a gigawatt or half a gigawatt by 2027 and then two months later, the utility tells them, actually sorry, it’s going to be 2029, it’s going to be a hundred megawatts and I’m going to give you the gigawatt by 2042. There’s a broader conversation with regards to incentives. Obviously have no penalties whatsoever if they fail on their commitment. So that’s probably one issue. But the bigger issue is that if you’re an AI lab, power is the lifeload of your business.
Power is revenue for Anthropic. Power is huge revenue in the form of training for Anthropic. So they need power more than anything else. If they don’t have power, their business doesn’t exist. So they basically cannot make a multi-billion dollar investment decision. If you’re talking about a gigawatt, as you probably know, CapEx is like $50 billion or if it’s a contract because someone else pays the CapEx 60, 70 billion, whatever, you cannot make that kind of investment decision if there’s uncertainty on the timeline. So at some point you got to think first principles, how am I sure that I’m going to meet with my roadmap? And let’s talk about roadmap, Anthropic. One and a half gigs of capacity end of 2025, by ‘27 they want over 10 gigs. So they’re going to build basically more than eight gigawatts in two years. That’s the size of Google today.
So you’re building a Google in two years, you need to have certainty if you want to do that to ensure that your revenue is going to grow as expected and so on and so forth. That’s not going to happen with the grid. There’s just no way. No one is building generation fast enough and no one can give you the guarantees that they’re going to give you that power on time, except in some very specific cases. So essentially, if you want to be in control of your destiny, the best option from a buyer’s perspective, some of the folks that actually deploy power is to simply bring your own generation. You have your equipment, you handle your own permitting and then you know when your thing is going to be online. At least there’s always uncertainty, but it’s sort of lower than if you talk to utility that has obvious generation constraints and cannot sort of promise you anything and doesn’t have any binding sort of reason to comply with this.
Shayle Kann: I would offer one… I’ll make a bet with you, which is I think we’ll add way more generation than you are saying that we will. Set aside behind the meter, like exclude that. I think we’re going to add a lot more generation. Now that doesn’t entirely make your point wrong because the constraint in my mind is less generation and more transmission distribution, right? And like you can add generation, but you still need to upgrade a substation and you need a high voltage transformer and that takes three years to order whatever. There are many things that come up the ability to provision power in the hundreds of megawatts or gigawatt scale, but I actually think from a generation perspective, there is a wave coming that is not behind the meter as well. There’s clearly a wave coming behind the meter too. But I just think it’s important to distinguish between a generation or a capacity constraint, which exists, but I think won’t be as big as a lot of people think there will and a T&D constraint, which I think is as big as people think it is or maybe is bigger.
Jeremie Eliahou Ontiveros: Well, I think it’s both. I don’t know. Let me ask you this. How much generation obviously adjusted for like ELCC, UCAP and whatnot, how much do you think is coming online in say ‘27 or ‘28?
Shayle Kann: Oh, let’s say 10 to 15 gigawatt. I’m just giving you numbers off the top of my head. I think we’ll do 10 to 15 gigawatts of gas and I think we’ll do, let’s say ELCC adjusted 20 gigawatts of solar and batteries. So,say, 35 total, something like that.
Jeremie Eliahou Ontiveros: Yeah. On our numbers, the ELCC adjusted is lower for battery and gas and solar gas, I agree with you. In both cases, rookie numbers, when you compare that to data centers, right?
Shayle Kann: If we’re trying to build 50 gigawatts of data centers every year, plus all the other load growth that we’re going to see, that’s tough, right? Yeah, I get that point. You talked about the frontier labs. I’m interested in how you think about their energy strategy such as it is, because there’s one extent to which, okay, mainly they’re just buying capacity from those who are actually building the infrastructure, which is the hyperscalers, but that’s not entirely true, right? They are building their own capacity as well and back to the point of the energy teams, both OpenAI and Anthropic have started to higher up energy teams. They’re small but mighty at this point compared to the hyperscalers, but what is your perspective on what OpenAI and Anthropic are doing from an energy perspective themselves?
Jeremie Eliahou Ontiveros: Yeah. So you’ve seen them work at different layers. I think in a lot of time they’re looking at sites themselves. And you could argue, for example, Stargate was kind of a two-way street between OpenAI and Oracle where they were both involved in the decision making process to get this done. I think both of them evaluate a lot of powered land sites. They look at their different options. They hire a bunch of people internationally as well to look at what do these markets look like and then once they sort of find sites that they like, they can bring in partners, whether it be Microsoft or Oracle or CoreV and so on and so forth. I think overall their biggest problem is that they just need a lot. They just need a lot and they have a financing constraint in the sense that they’re obviously not investment grade and a lot of this is very capital intensive and it’s upfront CapEx that they just can’t afford.
So that sort of slows them down in their ambition and desire to be more vertically integrated, which creates a very large market opportunity for hyperscalers. And again, as I was saying before, hyperscalers are essentially half of their business, Amazon, Microsoft in terms of like megawatts built, goes to Anthropic and OpenAI. So these companies are the number one sort of folks that deliver capacity. Now the second thing is that there’s the Silicon Wars that are at play and the Silicon Walls are fascinating and Google struck first, surprising doesn’t happen very often, but Google struck first and on a very large scale. The way they did it is that they sort of understood very well this issue with regards to financing. And so they were very innovative in the sense that they sort of invented, you could argue at scale, the concept of backstops. And so they also saw the tremendous value in selling their hardware externally.
I think they understood very well that when their biggest customer is Anthropic and that customer is basically becoming bigger than Google from a gigawatt perspective, there’s just no way they’re going to be always reliant on Google cloud, renting from Google. They want to go virtual. It’s just absolutely normal. It’s table stakes. And so I think Google sort of saw the opportunity as a way to sell their hardware externally and compete straight against Nvidia. And from an energy point of view, that is interesting because again, like that led Google to tell Anthropic, you can build your own capacity. I’m going to support you financially if you buy PUs essentially. And Anthropic loves TPUs. So it’s a two-way street, right? Anthropic also wants to deploy TPUs because it’s just a great piece of hardware.
Shayle Kann: You’re saying Google has enabled Anthropic to be more vertically integrated and actually build their own infrastructure, which means dealing with energy directly rather than indirectly.
Jeremie Eliahou Ontiveros: Yeah, absolutely. And they do that via financial support in their format of backstops. The first way was through front party developers. So you saw Terrawolf, Hut 8, Cipher Mining. These guys signed deals with Fluidstack and Anthropic in the end for about a gigawatt of capacity all backstop by Google. Google enables these sites to bring to fruition through sort of, again, their credit signature stepping in case fluid stock or Anthropic fails. And now they’re doing this at very large scale. So if you look at how much capital they have sort of indirectly deployed, it’s already $50 billion of applications that they have on their balance sheets that are solely in sort of the goal of supporting Anthropics data center build outs. And it’s actually brilliant because these $50 billion, which are, let’s call it four to five gigawatts, that’s going to be then capacity that’s going to yield tremendous revenue for Google because Anthropic is going to be buying the TPUs at $20 billion a gigawatt.
So that’s actually 50 billion that generated $100 billion in revenue.
Shayle Kann: It’s kind of equivalent at a single customer larger scale of what Nvidia has been doing with all the neoclouds. It seems like Nvidia has been just like one by one trying to stand up more neoclouds because they’re going to buy GPUs essentially. It’s kind of like Google is doing with TPUs and Anthropic.
Jeremie Eliahou Ontiveros: It is, but what Google did is actually more aggressive. What NVIDIA did initially was just investing in these companies. So they invested in Coreweave, they invested in Nabias, in Lambda, and a bunch of others. So they supported them financially just through equity investment, but not really much more than that. And obviously co-design and so on and so forth, but the burden of sort of securing data center capacity, securing financing for the GPUs was still on the sort of books of the New York clouds themselves. So actually Google is much more aggressive because they’re directly supporting the build out. And I think what you’ll observe is that for a company like Coreweave, if they wanted to build a gigawatt of data center capacity or most likely to lease it, that would be more likely for them. They’re just not going to be able to do that.
They don’t have the credit profile to do that, right? Because if you’ve seen a gigawatt of capacity, then the data center operator needs to get that financed typically very high loan to cost. And it’s typical real estate. You need a construction loan that’s not going to go through if you don’t have an investment grade signature.
Shayle Kann: Yeah. Related to that actually, I’ve been curious, I mean, speaking of the neoclouds, they’re not investment grade, but in addition to all the other challenges that presents from an energy perspective, if you want to go buy turbines right now, you need to put down pretty big deposits. If you want to do a large load interconnection, the deposits are getting larger and larger. The dynamic in the market is that because there is so much of a supply constraint in providing power or generation or whatever, the supplier can demand more out of the customer. So yeah, is that becoming a challenge for the neoclouds? Is it putting them at a competitive disadvantage in being able to build capacity relative to the hyperscalers who obviously have big balance sheets?
Jeremie Eliahou Ontiveros: Yeah. And a pretty massive one, you can just look at the numbers. CoreWeave, they have three and a half gigawatts of contracted power. Contracted for them means signed leases for the most part. Some self-bail mostly signed leases with third parties. And what you saw was that this number was about, if I remember correctly, 1.3 gigawatts in Q4 2024. So they’ve scaled that up pretty fast, but since Q3 25, they haven’t really been able to secure more and that sort of coincided with the overall tightening and financial conditions where you saw a pretty massive bond sell off, which impacted the likes of CoreWeave of Oracle and many of these guys and suddenly sort of the high yield market froze to some extent. And that’s also obviously related to the fact that this market is not that big and basically they massively increase the supply in that market.
Anyways, we’ll sort of get to where we are today, which is that it’s getting pretty tough for these companies to get the financing for all of these parts. And they’re all, as you said, more and more capital intensive. Utilities now are asking multi-billion dollar commitments for gigawatts of power, turbines and so on and so forth, the same thing. So yes, pretty massive disadvantage. And again, like I go back to what I said earlier, that’s essentially what Google sold for by providing their balance sheet as support and it’s actually becoming a pretty existential risk for NVIDIA because if they don’t support the neoclouds, they’re at risk of basically seeing Google, Amazon and others take tremendous market share on the silicon side. So you would see Trainium take more share, TPUs take more share. NVIDIA shares go down just as a function of deploying more capital to get that power secured and power is revenue.
So it’s becoming a pretty strategic angle and I expect NVIDIA is already reacting pretty strongly and expect to see pretty big announcements pretty soon. Our institutional clients know all about this already. We talk about this all the time. NVIDIA’s big moves on the data center market and on the power market. It’s a pretty big topic for them, but I think they’ve been a bit late to the game, but I expect second half of 26, you’ll see NeoCloud growth be unlocked much more through the actions of NVIDIA as they provide their balance sheet of support.
Shayle Kann: I want to talk through some individual energy generation technologies and I want to hear from you how you think about them and particularly like who amongst all these players really leaning in because it does seem to differ starting with, okay, so everybody as far as I could tell, is trying to get their hands on gas turbines and everybody’s buying from GE Mitsubishi Siemens if they can. Let’s set that aside because that’s kind of the incumbent thing. Just focused on gas though, this has been a time of the rise of reciprocating engines, air derivative turbines, all these other things. Who in your mind is doing the most sort of interesting, innovative stuff around gas that’s not just buying from the big three?
Jeremie Eliahou Ontiveros: Yeah, Bloom Energy, man. Fuel cells.
Shayle Kann: Well yeah, fuel cells was my next one. I was going to talk about that. Let’s set full cells aside for a second.
Jeremie Eliahou Ontiveros: Okay. So look, I mean the way I kind of see it is like there is like basically tiers and obviously the most premium is disturbance and then you sort of start going down and maybe then the second best thing that you like, and maybe even like it better than H class turbines would be error derivatives and then you start going down and down into reciprocating engines, high speed, medium speed and so on and so forth and then fuel cells. So basically it depends on the appetite of players with regards to behind the meter. And look, the name of the game here is generally speed as well. So you have to think of like what enables fast deployment. XAI as usual sort of showed the way and everyone followed. So what did XAI do? They deployed these aeroderivatives or actually IGTs. IGTs in the order of 20, 30 megawatts per unit.
And then you started seeing other players follow. So that was ‘24. In ‘25, you saw OpenAI being sort of the next one through Crusoe and Oracle in Abilene, Texas using again, Aero derivatives. Then you saw Meta essentially doing the same thing in Columbus, Ohio. Again, aeroderivatives, IGTs, a lot of this. And I guess like you started really getting into very large scale reciprocating engine in 2020, actually 25 as well, second half of 25 with OpenAI again. They signed this gigantic deal with Oracle in Shackleford County, Texas, where they’re going to deploy 2.3 gigawatts of four megawatt reciprocating engines. So you’re going to see a whole lot of reciprocating engines. The way to handle the transients is also pretty interesting because of the kind of deploy flywheels. So there’s going to be things running all day sort of spinning. That’s a pretty interesting one. And essentially it’s kind of like companies one by one are sort of folding and realizing that their best way to keep scaling is sort of going in that direction.
Now I would say the two companies that have avoided this to some extent so far are Amazon and Google because they’re the most aggressive in putting deposits to every single utility in the US. When you see these gigantic numbers of like contracted loads by all of these utilities, AEP or Dominion and the bulk of that – Xcel for Google. Excel. The bulk of these contracted loads, I think as of end of 25, we had like 150 gigawatts. The PPL in Pennsylvania, that’s mostly Google and Amazon. They’re the ones sort of scouting all of the utility markets and putting deposits and so on and so forth. So they’ve been more aggressive on that side and folks like Meta have been more aggressive on the BTM side and also I would say what you’ve generally seen is hyperscalers being generally speaking a bit slow to react to that trend and being a bit late to the game and realizing that, hey, I’m going to face generation constraints and I might not be able to scale as fast as possible.
And that has led to a tremendous rise of third party data centers and the general share of sort of data center build out, you had a period in time in like 2024 and part of ‘25 where self-build was the bulk of it. And now you’re seeing the leasing markets that really go up tremendously through the likes of Oracle who really don’t sell build, but now even the Metas of the world, the Microsoft of the world, they’re essentially scaling through leasing. And when they lease, they lead through operators that oftentimes for gigawatt scale sites, use behind the meter and that can be reciprocating engines, either high speed through like Jenbacher or medium speed for Bergen. All of this is sort of being adopted at scale now. Essentially the name of the game is, do you have it soon enough?
Shayle Kann: Well, so speaking of speed then, let’s talk about fuel cells because it’s been a wild ride for Bloom. Bloom’s, as of this recording, I don’t know, I haven’t looked today, but it’s like a $90 billion market cap company somehow and the fuel cells are, they were not designed for this scale, right? They’re similarly like few megawatt individual units. They’re also higher CapEx, generally speaking, though CapEx on the gas turbines is going up too, so you can tell me where you think the comparison lies today, but historically higher CapEx, somewhat higher efficiency. The main thing seems to be availability, which is like Bloom was not sold out till 2031 and so they were able to take advantage and particularly with Oracle it seems, but like how do you think about fuel cells in that cascading chain that you described before?
Jeremie Eliahou Ontiveros: Yeah. I think the biggest disadvantage that fuel cells have not really cost, it matters but not so much these days. I can explain why, but generally not. I would say as a bridge power solution, it’s really bad. Because Bloom Energy fuel cells, they have to run extremely hot. And so if you want to use them as backup, this basically takes two days to go from zero to a hundred. Whereas aeroderivatives, as you know, can scale up fairly fast, reciprocating engines can scale up fairly fast. And a lot of folks, the usual hope of behind-the-meter was that it’s all going to be bridge power, is I’m going to deploy these power plants for a year, two years, maybe three years and then the good is going to come and hey, maybe I’m going to use this as backup. In many cases, you see folks starting with sort of lower redundancy, no diesel gen sites, like that.
And so in some sense, Bloom is like the ultimate play on power constraints because it’s to play in islanded data centers. And if you do bloom, you’re basically islanded for life. Either that or maybe you get grid at some point and then you move your fuel cells to some other location, but you can’t use them as backup. It’s not a very efficient solution for backup purposes.
Shayle Kann: One thing I think has been interesting to see some projects, I don’t know if any fuel cell projects are doing this or not, but certainly some of the recent projects and so on are their bridge power, which then upon the grid connection coming transfers to the utility. So the utility then gets to operate, own and operate them. So it’s not like they have to get moved to a different location. They don’t sit there as backup. They do provide value to the grid on an ongoing basis. You just kind of like shift them so they’re no longer behind the meter, which I think is kind of clever. Yeah. So fuel cells, interesting story. I think the perfect sequence of events for Bloom. We’ll have to see how it evolves over time.
Jeremie Eliahou Ontiveros: I was just going to say, I think one of the lead times that is generally maybe a little sort of underestimated is not just the equipment, it’s the actual deployment. And as you know very well, CCGTs take forever to install a couple years or something. Bloom energy fuels are really fast to deploy. So that’s a massive advantage for them is as you’re desperate and for some reason you can’t get your power. The two high profile examples we’ve seen so far are permitting challenges, which I think are going to be happening more and more as well. Bloom is sort of the natural option because you can just plug them in. It’s pretty fast. And so that makes them an amazing play. And again, as the data center market sort of grows tens of gigawatts, 50% year growth per year essentially. Basically, if you keep that direction by 2030, you’re talking about 100 gigawatts per year being added.
Then at some point you got to think about like what can be deployed fast enough. I guess the other thing is who can expand capacity fast enough and who has sort of the incentive to do so? And Bloom score is really well because obviously their high capex also means that the payback period on a new factory is pretty low. As opposed to turbo manufacturers, you could argue now prices go up so much that buyback is shortening, but still it’s like much more modular. You can sort of build Bloom capacity faster. So if you assume the super AGI scenario of half a terawatt by 2030 or whatever per year, you could assume that Bloom could become this gigantic company because they’re sort of one of the best at scaling and meeting sort of the demand to scale.
Shayle Kann: Let’s talk about solar and I guess wind. All of the hyperscalers for a long time have been signing VPPAs for renewables. Let’s just set that aside, assume they continue to do that. As far as behind-the-meter goes, clearly you have Google. Google acquired Intersect. Intersect was already in the process of developing a bunch of these kinds of hybrid sites that have a lot of capacity for data centers and a bunch of behind the meter stuff including solar and batteries and some gas as well. So clearly Google has at least a play in behind the meter renewables. Do you see anybody else active there?
Jeremie Eliahou Ontiveros: Yeah. I think the action is going to concentrate in Texas. I think the way big data centers, big data center campuses are going to look like in a few years from now. Again, like obviously the scenario is we keep adding tens of gigawatts per year that keeps growing. So we’re in this AI scenario. I’m not religious. I’m not saying it’s going to be a terabot per year. We’ll see. But just if you sort of keep on that trajectory, I think at some point the only path is going to be to essentially build these massive campuses in West Texas. It’s already sort of going up pretty tremendously in West Texas and I think it’s going to keep going up. And for solar, that is actually a strong positive. Obviously because there’s a ton of land in West Texas, having solar on site can reduce your energy costs, sort of shield you from, I guess variations and sort of power prices and whatnot.
And then you can put a bunch of batteries and actually, it’s actually quite profitable to do arbitrage and so on and so forth. Anyways, so it makes sense and the answer is yes. There’s other companies that are doing this. If you want to talk about like the ones that have said it publicly, I think Lancium has been pretty open about this. If you don’t know them, they’re the ones owning the land and developing the power in Abilene, Texas for Oracle and OpenAI. They have a bunch of other sites that are pretty massive in West Texas and there’s other companies that are thinking about this as well. Cruso is another one. They have a really interesting site in Armstrong County what they reported to me with Google. So that’s another interesting one where they’re connected with a wind farm across the meter as they say. I think we’re going to see more of that towards the end of the decade.
More of those campuses with potentially tens of thousands of acres, gigantic campuses that are going to have gas, solar, batteries, maybe wind, but I think that’s going to be a pretty big recipe as you get into that gig of scale. And one last thing I would say is I think the market for like five gigawatt campuses is really changing now actually in the sense that over the last two years, you’ve seen a bunch of announcements of most of the time companies that no one really knows, you look at their websites, you see contact us, what? I know of very reliable developers now that are planning five gigawatt data centers in West Texas. So I think it’s really changing. We’re at an inflection point and I think you’re going to see some pretty gigantic announcements. And as you said correctly, Google sort of paved the way for others and I think folks are really all looking at this now.
Shayle Kann: Alright. Just rounding out the technologies to talk about, let’s bucket together the clean firm stuff. So in nuclear and geothermal. So in nuclear, Meta, Google, Amazon have all made pretty big announcements, reasonably sized investments into different nuclear reactor technology companies into projects. They’ve bought power from upgrades or restarts of existing reactors. So there’s like a fair amount happening in nuclear amongst the hyperscalers. And then in geothermal, I think it’s a shorter list, at least off the top of my head. Google obviously is a big partner of Fervo, which just went public. Meta signed a PPA with XGS, which is kind of a next generation geothermal developer. So rather than running through these all one by one, I guess how do you think about this world of like the clean firm category from the hyperscalers perspective?
Jeremie Eliahou Ontiveros: I mean, I think the big difference with other types of technologies is just the type of contracts that have been signed so far are for the most part non-binding subject to milestones, nuclear being the best example. If you don’t get your approvals and so on and so forth, the contract obviously is not executed. So I think right now it’s more of an option for hyperscalers. If it works and is economical and so on and so forth, I think it’s going to be massive. So it’s more on the execution side for all of these companies to actually be able to deliver right now in that more speculative phase where I think many of these, especially these next gen technologies, SMRs for the most part or next gen geothermal, I think it’s still like overall fairly early stage and at the non-binding stage currently.
Shayle Kann: I mean, maybe the way to put it is that all the other technologies we’ve talked about are there for the purpose of speed to power. That’s why they’re getting directly involved, right? It’s like, we’re going to do a bridge power, whatever, we’re going to do fuel cells, we’re going to do something. Nuclear and geothermal are not speed to power today. Long term, they could be massively scalable capacity, which the hyperscalers believe that they will need, but they’re serving a different purpose. It’s long term capacity, not speed to power.
Jeremie Eliahou Ontiveros: Yeah, 100% agree. And you know what, I will double down on that. So we have these thesis currently that we’ve shared to our clients. We think gas turbine orders are going to pick this year, which I think is a pretty bold sort of idea because everyone is sort of like GEV, Infinity and whatnot. But the reality is that I think the bulk of the orders have been driven by utilities indirectly data centers, but a lot of that is utilities, the Dukes, the AEPs of the world. And I think those guys are not going to be ordering so much because now the sort of window is really concentrated on like ‘27, late ‘27, mostly ‘28. You can’t really get turbines for 28, they’re all sold out. So really the rush right now is to secure whatever is available and works. And that’s really a favorable environment for fuel cells to some extent, for reciprocating engines for the most part and to some extent, new types of turbines, like you’ve seen PROENERGY and some of these new guys that have good solutions.
And so yes, 100% agree in the sense that I think things that are not speed to power are deprioritized today over speed to power and still think we could see like more nuclear deals and geothermal deals, but those are going to remain more sort of options rather than binding contracts.
Shayle Kann: Jeremie, this was a lot of fun. Thank you for doing this. Jeremie Eliahou Ontiveros leads data center and energy infrastructure research at SemiAnalysis. This show is a production of Latitude Media. You can head over to latitudemedia.com for links to today’s topics. This episode is produced by Max Savage Levenson, mixing and theme song by Sean Marquand. Anne Bailey edits the video version of the show. Stephen Lacey is our executive editor. All of our episodes are now on YouTube. You can subscribe to Latitude Media for episodes of this podcast and also Open Circuit, their other podcasts. You can also find the audio version of the show anywhere you get your audio podcasts. I’m Shayle Kann and this is Catalyst.
