SemiAnalysis 的 Jeremie Eliahou Ontiveros：数据中心与电力全景

Jeremie, how’s it going?

Hey, man. Good to see you again. Thanks for having me.

Look, you’ve got an open invite. Your last pod was—I never reveal stats—one of the most popular podcasts we’ve put out this year. The YouTube stats were great. I don’t know what it says that the audio stats were even better. I don’t know if that speaks to my face, your face, or whatever it is, but both sets of stats were great, and the audio in particular was off the charts.

The people loved it, so I’m super excited to have you back. I want to dive into all things SemiAnalysis. We were just wrapping earlier today, and obviously I think our conversation is going to cover everything—semis, power, and everything—but I’d love to start in the semis and power space. What’s real? What’s the next big trade in the semis and power space? I guess that’s the headliner.

Yeah. One of the interesting things you’re seeing is the surge in on-site gas. A lot of people talk about deploying CCGTs to power data centers, but there’s an issue with those: the lead time. Everyone knows there’s a multiyear lead time to get CCGTs. What people might not realize is that you can actually deploy at very large scale much faster with smaller, more modular units.

You have 2 types of systems. You have the aero-derivatives, which are turbines mostly sold by GE Vernova and Caterpillar, and you also have basically huge car engines. They’re called RICE engines—reciprocating engines—supplied by folks like VoltaGrid, as well as Caterpillar and a few others.

In both cases, we’re increasingly seeing solid, reliable data center projects that I believe are going to be operational in 2026 and 2027, serving some very large end users like AI labs and hyperscalers. I think they’re going to be deploying those technologies.

If you think about the market for on-site gas for data centers, in 2024 it was pretty much Elon Musk. It was pretty much the Memphis site. In 2025, Elon Musk again is a driving force. In 2024, you had around 300 megawatts of on-site turbines—or 250, whatever. In 2025, he’s probably adding another 500 megawatts on-site.

But you have other sites. Another one that’s well publicized is the Stargate Abilene, Texas, site, which is going to deploy 300 and 160 megawatts of those GE Vernova aero-derivatives. But there’s more, actually.

That’s a fantastic pitch. I was just reading some stuff to prepare for this podcast, and SemiAnalysis is huge. If you haven’t subscribed, my personal favorite is that you guys have the drone shots of all of these things, especially the Oracle co-located one, where you’re like, “It’s Oracle here.” It’s just really interesting. You get a really deep knowledge of natural-gas turbines.

You said it, and I think, as a dumb-dumb journalist, the thing that worries me is that GE Vernova, which provides a lot of these turbines, has been a screamer. I believe the spin-off from GE happened in 2024, and it’s like a six-bagger since the spin. The stock is up 1.5 times this year. I’m not saying I’m scared to buy a stock that’s gone up, although that is always a little worrying.

I guess the thing with GE Vernova is that what I’ve heard is, hey, these gas turbines are complicated, right? But pricing is going through the roof because demand is crazy. What I worry about as a generalist is 2-fold. First, if demand pauses—and I do want to talk a lot about AI capex—but second, I believe pricing is also going up because GE and the other manufacturers are in something of an oligopoly. The other people who make them are saying, “We don’t know if we can trust this demand,” so they’ve held back on building new plants and bringing new supply online.

They’ve got this beautiful situation where demand is through the roof and they’re not bringing supply online. How long can that really hold? At some point, somebody is going to say, “I need to bring a little bit more supply online,” and once that happens, the floodgates break. You worry you’re buying into a supercycle, and there are 2 methods for it to break. I know I threw a lot out to you, but I’d love to get your overall thoughts on that.

Yeah, sure. Gas turbines are through the roof, as you said. However, I would say there are 2 types of demand here.

One is the big turbines—500 megawatts per unit. Those are mostly serving the grid, with a few select, very large data centers, like Meta, for example, talking about building on-site power plants. But mostly, they’re serving the grid, so it’s still related to data center growth, but it’s for a broader use case.

Now you have another use case, which is deploying turbines on-site for time-to-market purposes. What’s interesting is that the dollars flowing to these turbines are, to some extent, not net-new dollars. They’re taking market share from something else, which is the diesel-generator market.

What you could see, and what you’re starting to see, if we take the Abilene, Texas, data center as an example, is that I look at satellite pictures and I don’t see any diesel generators. What I understand is that they’re going to be deploying those turbines, maybe at first for on-site power, but then, as grid power comes online, they’re going to use those turbines for backup.

I think you’re seeing that same pattern with the Memphis data center, where Elon brought online those smaller turbines, which are much faster to scale up and can be manufactured at scale. They’re factory-made. He first uses those turbines for on-site power, and then, as grid power is built and they have a substation and such, they use them as backup.

Let me ask a question on that. You’re the data center expert; I’m not. You had a great piece on intermittent-power issues at data centers and their impact on the grid. I was fascinated by it, and we can talk about that later.

But when you’re building these big gas turbines and backup systems, you’re making tens and hundreds of millions of dollars of investments in these natural-gas turbines, right? They initially serve as primary power and then eventually serve as secondary power.

And this might blend nicely into our capex discussion at some point. But when I hear that, I say, “Oh, you have a rush right now, right?” But when you're spending hundreds of millions of dollars on something that's going to end up being your backup power once the grid is kind of online, that doesn't seem sustainable, right? Yes, it's sustainable when you need to get stuff up now, but when you're planning your 2027 or 2028 data center, at that point, aren't you kind of looking and saying, “Hey, we're going to be connected to the grid. Maybe we don't need to spend hundreds of millions of dollars on these giant turbines”?

And also, as batteries get better, maybe you need fewer backup sources. It just strikes me that if these are eventually going to be backup sources, it sounds great in the near term—and obviously, it sounds like diesel generators are really up a creek—but the medium- to longer-term outlook for these strikes me as a little bit murkier.

Data centers have been deploying backup forever: diesel generators. And if you think about it, most of the time those are a stranded expense, right? They never actually serve. If you normalize the cost of power in terms of how many hours they run per year for these diesel generators, it's absolutely high, probably in the thousands per megawatt-hour.

So those are not economical purchases. They serve when you have a blackout. For example, in Spain a few months ago, there was this horrible blackout. The data center industry was, to some extent, proud—I don't know if that's the right word—but they were actually able to navigate through that event.

You've got to be very careful, because you're the data center industry. You're like, “Don't worry, guys. We had 100% uptime through the blackout,” and then you're like, “Hey, the hospital down the street was offline. Granny was sitting in a 100-degree apartment.”

I think that answered it well. I do have some more questions there, but let's go to the overall AI context. You guys published a piece. Oh, did you have something else? Please.

Yeah, I want to flag something else, which is that today it's all about time to market. But as the market matures, maybe think of it this way: It's obviously technical, so right now it's the upcycle; time to market matters more than everything. But then, as we get into the downcycle—which for sure is going to happen at some point—you're going to see companies turning to cash-saving mode.

When they do that, they're not going to be contracting new cloud capacity, or much less. They're going to be using their existing footprint, which basically means taking training GPUs to serve inference. So what if you build a training data center with no backup at all, because uptime doesn't matter so much for training data service, and then suddenly you have a downcycle and you want to save money?

Your inference service—I mean, you want it to be high availability. That's also why it matters to have backup. Again, that's a decade-old rule for the data center industry: to have backup. I just think you shift from diesel to gas, basically, because it serves an initial purpose with gas. Sorry, go ahead.

No, no. Hey, that's fantastic to think about. And obviously, when I asked the question, I wasn't framing it as, “Hey, these guys always need backup power.” But I guess my push to you would be: right now, they're doing it with the natural-gas turbines because you do the natural-gas turbine, it serves as your primary, and then you can use it as your secondary once you're plugged into the grid.

If I was planning in 2028—let's say I had something in 2028 and I had a hookup to the power grid in some way when this thing comes on in 2028—would someone be planning that 2028 power plant with the natural-gas turbines? Or would they go back and say, “Hey, obviously I need backup. Nobody's debating that”? But would they go back and say, “Hey, in 2028 batteries are much better, so maybe I can store a lot”?

Even if I don't want to rely purely on battery, diesel is, as you said, always the backup. It is so much cheaper than gas backup. Should I just use diesel plus battery as my backup instead of buying these giant, super-expensive natural-gas systems, because I don't need a primary anymore?

So, first of all, with the diesel cost, I actually think it's pretty similar. Natural gas.

Oh, is that right? I thought diesel was significantly cheaper.

If you look at the diesel engines, they're actually very similar to the natural-gas engines. For example, CAT has both a diesel engine and a natural-gas engine. They're pretty similar, with roughly the same pricing.

Now, again, as I said earlier, you have basically 2 types of on-site gas power for these deployments. You either have what's called reciprocating engines, which are basically giant car engines—3 megawatts or 4 megawatts per unit—and then you have the turbines, the aeroderivatives, which are like 30, 15, or 40 megawatts per unit. You're seeing both being used today. I'm not exactly sure what the share is going to be, but both seem to be getting a lot of traction, maybe slightly more on the turbine side.

In both cases, actually, both the turbines and the reciprocating engines have a roughly similar cost structure. It's expensive, to be clear, but it's not like a huge difference compared with diesel. So, in terms of cost, it's roughly the same.

The difference is really about site selection. Basically, you need to be near a pipeline. That's the only difference. You didn't need that before; you could just build a tank on-site for diesel, whereas now a site-selection criterion is having that gas access.

Perfect. And I actually do want to come back to site selection clearly. Well, let me back up a second. Capex—you guys published this great piece that looked at the 5 hyperscalers, basically, right? If I remember correctly, Oracle, Amazon, Google, Microsoft, and I think you guys had CoreWeave thrown in there.

You said, “Hey, we are forecasting capex.” I thought the 2 interesting pieces were that your capex estimates for the rest of 2025, and especially 2026 and 2027, were way above Street estimates, right? So the party ain't stopping for capex.

And then the other interesting point—actually, I'll just pause there. What are you seeing that suggests to you that the capex estimates for the Street—which, by the way, is pretty bullish on capex, and I think capex estimates have been going up a lot as the year has gone on—are still not just low, but way too low, in your opinion?

I just want to say, first of all, that the report you're mentioning was sent on Core Research. It's only for institutional clients. It wasn't on the broad newsletter website, so sorry, guys, it's not accessible for free.

We can still talk about it, though. We can still talk about it.

Talk about it, for sure. Of course. And we're going to do it.

Anyway, why is the Street's capex estimate so low? First of all, you look at sell-side capex estimates, and they're normally modeling growth versus 2025. In 2026 and 2027, they have some growth—maybe high single digits, maybe low double digits—but it's not a lot, right?

I guess we go back to something we discussed last time, which is: How do cycles typically play out? Is it generally a slow, steady trend, or is it actually strong, steady up and then down? I think, again, 50 years of history back me up in saying cycles tend to be very strong toward the upside. Anyway, this is just overall theory.

Now, what we're seeing more on the ground—using our data center model, for example—is that construction starts for self-built data centers are surging. We're seeing that the leasing rates—pre-leasing, I should say—of hyperscale data centers are extremely high today, which, again, means that when you pre-lease it, the data center is going to be operational next year. So that's an indication of next year's capex.

Most of the forward-looking data center indicators point to very sharp growth in the high double digits, which tracks pretty well with the numbers that you saw: high-double-digit growth for hyperscalers and way ahead of the Street. Those signals clearly don't show capex being stabilized at high levels. They show capex going up, up, up.

And this is up-up, right? From 2023 to 2024, capex was up 50%; from 2024 to 2025, another 50%. You guys have, I think, only 35% to 40% growth in 2026 and 25% to 30% growth in 2027. At that point, these companies are eating the world, right? Just those 5 companies, I think you guys are saying, are going to have over half a trillion dollars of capex in 2026. Unless you have anything to add there, I have some questions on that, but if you want to add anything, I can pause there.

Yeah, I mean, again, all the signals go up. We're just saying these companies are committed to investing. If you think about the drivers, maybe we can talk about that afterward. There are a couple of ways to frame what's driving the markets.

I would just say that we don't see anything today that would suggest they're slowing down. I also want to emphasize that our report was actually sent before earnings. The early indications that you saw—the best one was from Meta, where they basically said, “We're probably going to add roughly the same dollar amount of capex,” which suggests an additional $40 billion year on year.

So they're spending $65 billion to $70 billion in 2025. They're basically saying they're going to be close to $100 billion by 2026. Soft guidance, right? That's the first indication from management that we've had, and it clearly goes toward what we have, which is that, in terms of dollar amounts, the year-on-year growth is going to be roughly similar.

But you've got accelerating spend. These companies are spending at levels that, as a percentage of the economy, I don't think we've seen since the railroad buildouts in the late 1800s. I could be wrong; it's not like I've got my fingers on all of that. But it's going to be over 1% of the economy really quickly.

So, at a high level, obviously it's not accelerating, because we're going from 50% to 35% growth. But I think no one would fault me if I said, “Hey, it's not exactly diminishing when you're talking about numbers this big.” Are there any worries about signs of diminishing ROI yet?

Yes and no. If you take a simple view of the balance sheets of these companies, you're definitely seeing the revenue-to-assets ratios going down for folks like Amazon, for example. And if you compare revenue to assets for these companies versus CoreWeave, for example, you're seeing that the GPU cloud business model clearly has lower asset turnover, which is, I guess, a way to say returns are going down.

But that's not fair, though. You're going from Microsoft, where they were selling capex-free software licenses, right, to this heavy business. Yes, it's going to be lower ROIC and lower whatever it is than the software licensing business, but it doesn't mean the returns on capital aren't incredible.

The returns on capital, from the estimates that we've done, are decent. They're not as good as a typical Azure business, for example, which was very high-margin, but they're pretty good regardless. One of the big questions is: What's going to be the useful life of these chips?

Actually, that's my next question. Do you mind if I frame it? I did a lot of work on this. If you talk to bears, they will tell you Jim Chanos—and I hate to pick on him, but he's got 2 prominent tweets in the last 5 days on it.

The one that really jumped out to me was a tweet that said, “Hey, Meta is depreciating their entire useful life at 11 to 12 years, and a lot of that is GPUs, which they're depreciating, I believe, at 3 years.” There's a lot of debate over whether you should be depreciating GPUs faster.

I think he kind of missed the point, but he's basically saying, “Hey, if Meta is building all these things and depreciating them at—his best guess was—11 to 12 years, he thinks it might be 20 years if you do something.” It's basically saying—not accounting fraud—but are they overstating the returns on investment because you should depreciate these at 5 instead of 10? So you're actually way overstating the investment. I'd love to get your thoughts on the useful life for depreciation that you just teed me up to ask about.

Yeah. Off the top of my head, the accounting depreciation that we have today is about 6 years. Meta is 5.5, if I remember correctly. CoreWeave is 6.4 years for servers. The big risk is: What if from 6 years you go to 4 years?

There are a couple of ways to think about this. First of all, we can look at empirical evidence. The best empirical evidence is the HPC world—high-performance computing. Some of that stuff is pretty public, right? If you think about the TOP500 supercomputers, some of those have been running for over 6 years. Typically, what you hear from the HPC community is 5 to 6 years of useful life.

So I think when you use empirical data—what has happened historically—it's tough to make a case that 6 years, or 5 to 5.5 years, whatever, is not an accurate measure. That's the first thing I would say: It makes sense based on the signals we've seen historically.

Now, you could also argue that we're running these systems at max power these days. We're just doing more crazy stuff, I guess, with these GPUs. Obviously, we're trying to sweat every single watt out of these GPUs—the whole point is to utilize them as much as possible. As such, you could imagine maybe they're going to die faster. Honestly, it's anyone's guess at this point. I just cannot know what's going to happen in the future.

Again, just looking at the historicals, 5 to 6 years seems okay. But I do think it's a risk. If it's 5 instead of 6 years, or 4 years instead of 6 years, then I think what is going to happen is that they're going to have a massive write-down in a few years.

We had a post on Oracle showing that the very large OpenAI-type contracts, if you look at how they work, actually have very high margins. We estimated about a 40% EBIT margin, just on a project basis, to be clear. That doesn't include all the structural costs and such, but on a project basis, this is a very high-margin project. Now, of course, if the useful life is actually 4 years, they're going to book a massive write-down in a few years, and they may have 1 quarter with a $10 billion accounting loss.

What strikes me is, like you said, the useful life is about 6 years. If you go from 6 to 4 years, maybe you overinvested a little bit, but it's not the end of the world. I think the differentiation is, if he's right and they're using a useful life of 12 or 20 years, that is the crux of the bear tweet.

If you used a useful life of 12 or 20 years and then wrote it down to 4, well, then it was a disaster. But what you're saying is, if it's at 6 and it goes to 4, I don't think anyone's sitting here saying, “This giant $100 billion data center build is going to be completely useless within 5 years.” So I'm kind of with you now.

There's still the question of return on investment. Let me just ask that: You said the returns on capital are good. How do you guys measure the returns on capital here? To me, you do have the issue—there's the famous thing with bank financials: The scariest thing in a bank is a fast-growing bank.

If it makes a dollar of loans today, then $500 of loans 2 years from now, and then $50,000 of loans 3 years from now, its metrics are going to look great, but that first-dollar loan might be terrible. You don't know it until you level out, and then you say, “Oh my God, we've just been fooled because all the new loans aren't paying off.”

These guys are accelerating capex so fast. How are they measuring their current return on spend? They're spending $350 billion this year, and that's not even coming online for 12 months. How do they know that they're getting this great spend? It's not just, “Oh, yeah, the first $10 billion was great, but the next $350 billion was just crazy.”

Yeah. It's basically the CoreWeave business model, right? The bears are looking at that and saying this is the worst business model ever. The analysis that we've done—which we actually posted in a report—was pretty positive on CoreWeave. To be clear, we don't provide financial advice, but we just looked at the trends, and everything to us looked much more viable than many people were suggesting.

The way we do it is we build a project-by-project analysis. We actually built a comprehensive model about a year ago called the AI Cloud TCO Model, which we built for someone building a GPU cloud, literally. We feel very good about all the estimates we have here.

What we did is estimate all the different capex-related costs and all the different opex. We went extremely deep down the rabbit hole to model every single line item.

What we're seeing is that there are a few things that impact your returns. Of course, you want to have low data center costs. Cost of capital matters a lot as well, because many people in the neocloud industry have a very high cost of capital. A year ago, it wasn't unusual to see a 15% to 20% cost of debt; cost of equity was probably also above 20%.

I think it's going to come down because now these companies are getting more and more mature, especially CoreWeave. People are much more comfortable lending to CoreWeave at much better rates.

There are a lot of assumptions baked into the business model. If you're able to optimize your opex and your cost of capital, and you can also optimize your capex to some extent—which is something we like about what Oracle is doing with networking—you can improve the returns. Oracle has a really good networking configuration to serve very large clusters, enabling them to have lower capex than others.

All of these optimizations taken together can take a business that, for a random neocloud, may be—I’m making it up—5% to 10% ARR and turn it into a 25% to 30% AR business. The point being, to answer your question, they’re thinking about this on a project basis, just like we're doing. At least that's my understanding; I might be wrong.

I think the best way to think about those is that, because there's so much upfront capex, you just want to know what your likely returns are. That's also why you see a lot of these big hyperscalers signing much longer contracts than the average market rate.

From CoreWeave disclosures, and from leaks in The Information and Reuters about OpenAI, it's pretty clear that they tend to sign 4- to 5-year contracts. When you sign a 5-year contract, you basically have guaranteed ARR.

Unless something goes wrong, of course, or AI fails and whatnot.

Just on CoreWeave, you said you published a report that was positive. I wish I had really looked at that IPO for a while. I was like, “Man, this IPO—you get an IPO that launches with a low float.” We’ve seen this a few times with Arm, especially with Arm. I really looked at it, and then the stock was up 3× a month later. I was like, “Gosh damn it, Andrew.”

You sold.

But I do want to ask you—and we’re not making financial advice; I’m not even talking about the stock price here—but on CoreWeave, you heard lots of bears, especially around the IPO. Even now, a lot of people will say, “Hey, the stock price is completely inflated by a thin float. There are lots of pod monkeys trading around with borrow rates and everything.”

Just on the business, it jumps out at me that this business was really spun up in large part because Microsoft was so desperate for capacity that they entered a huge deal with CoreWeave. Microsoft and OpenAI have said, “Hey, we’re not doing that deal again. We’re trying to do everything on our own,” right?

Look at this company. Their capex spend, even at their level, is a pittance compared to their larger peers. Their larger peers obviously have other business models where they can subsidize. They can go build a new data center knowing, “Hey, most of it is going to be taken by our core internal processes, and then we can fill the books with third parties if necessary.” CoreWeave doesn’t have any of that.

I’d love to ask—not about the stock, but just about the core business—why do you think it makes sense, and why are you guys still positive on the business versus these giants that are playing in a similar field with other advantages?

Yeah. Basically, if you think about the market for very large contracts—OpenAI, hyperscalers, Anthropic, these guys—it’s basically a commodity. What these clouds are providing is bare-metal infrastructure. You have to build a data center, put some machines in there, and assemble the machines through networking. There’s no software layer on top of it. There’s no technical moat or technical differentiation. There is some to some extent, but it’s really much lower compared to what we’re used to seeing from AWS, Azure, and others.

Barriers to entry are much lower. If you think about it, okay, it’s a commodity. How can you win? There are 2 ways to win. One is, long term, having the best cost structure. But actually, that’s not what matters today, because what matters today is speed.

That’s really the thing that matters, because that’s what these big end users want. That’s what OpenAI wants. That’s what Anthropic wants. They want the clusters as fast as possible, so you have to optimize everything for speed.

This is where you look at CoreWeave’s strategy, and they’ve done some things that enabled them to beat everyone else on speed. They contracted over 2 gigawatts of power in roughly 2 years. They had to take on, of course, material financial risk and such to do that.

For example, they’ve been working with those crypto miners, which no one was considering back then. CoreWeave was among the first to sign with a miner. These guys have the power right there, they have the substation, and everything is there. You just need to trust that the miner has good enough contractors to actually build a data center, but the time to market is unrivaled.

If you think about the challenges of building a greenfield data center, CoreWeave said, “No, I just want to go brownfield.” They have a couple of other partners, like Chirisa Technology Parks, and a couple of others that build brownfield data centers at an extreme speed.

In terms of cost, honestly, it’s probably not the best cost structure, and I think they’re trying to optimize that increasingly as of today because now they’re scaling and they’ve gained that trust from many partners. That’s why they acquired Core Scientific: that way, they’re more vertical. They can own the infrastructure and such. But they really scaled at an impressive pace in 2023 and 2024.

If you think about the time to build a data center, ask most people who have been in the industry for a while. They’re going to tell you those are 2-, 3-, or 4-year projects. That’s the time it takes to build a data center.

If you look at the listed companies—think Digital Realty and others—that’s what you see on their statements: when they start building, it becomes real and generates revenue roughly 2 years later, or 18 months later—sometimes slightly more, sometimes slightly less. At CoreWeave, if you look at their deal with Core Scientific, things are moving faster. They’re doing that stuff in less than a year in some cases. That speed—again, speed is really what enabled them to gain those contracts.

So, really, I think they took some pretty innovative approaches to infrastructure that are much more optimized for AI, whereas everyone else was still in the cloud mindset. I’ll give you another example, which is what Oracle did—

Can we pause on Oracle? I actually do want to come back to Oracle. Do you mind if I ask one more question on CoreWeave?

I heard a lot of interesting things there, but if I were a bear, I would say, “Hey, what Jeremie just described was that they took on risks that no one else was willing to take, and they really emphasized speed.” In an upcycle, that’s everything you want, right? But one thing Jeremie said is that there is always a downcycle.

Back in 2022, when people were just first starting to talk about this, everyone said, “Hey, even NVIDIA—there’s a downcycle.” All the NVIDIA long-timers might have missed the big move because they were like, “Things are starting to look good. That’s when you sell, because there’s a downcycle coming,” and they missed it.

If CoreWeave optimized for an upcycle—if that’s basically what they’re optimized for—aren’t they going to get crushed the moment a downcycle comes? All of a sudden, they’re committed to, “Hey, all they’ve done is accelerate and take as much as they can, and they’re delivering it as fast as possible,” and then a downcycle, even if it’s a blip for 6 months, comes and they’re just completely stuffed. Am I crazy to think that?

You’re not crazy, and actually, I think that applies also to Oracle, which we can talk about in a bit. Yes, 100%, it’s a lot of risk.

The simplest way to frame the risk is this: your longest GPU contract is going to be 5 years, while your data center deal is going to be 15 years. So, if you cannot replace those GPUs and find a new contract, you’re stuck with 10 years of paying rent to a data center operator without any revenue. That’s the simplest way to frame the risk, which I agree 100% exists.

You can think of it as a bet on the underlying companies. Are these companies likely to need capacity in the future? That’s the way the bet is framed, right?

If you think about it, I’m exposed to OpenAI through both CoreWeave and Oracle. The way you think about it is, OpenAI is not going to fail. OpenAI is going to need GPUs forever because they have a business model that structurally requires GPUs on the inference side, and they always have big training requirements.

That’s the way you can frame it: I’m betting on the AI industry to be a long-term consumer of GPUs and a long-term consumer of power. As I secure power and build data centers, I can renew those contracts over time. I actually have some more questions on power, but let’s go to Oracle.

Yeah, you wrote a piece on Oracle, and I’m sure you didn’t see it, but for my last book club, I read Larry Ellison’s biography from 2002.

As I was reading it, in the back of my mind—I mean, Oracle stock has been on a heck of a run this year, right?—I was kind of thinking, “Hey, Larry Ellison somehow managed, as a man in his late 70s or early 80s—I can’t remember if he’s in his 70s or 80s—to catch the AI wave.”

He wasn’t crazy early, but in 2023-ish, he saw where it was going. He made a big bet, and boom, this man, who’s called so many trends, calls this one again. Oracle’s really benefiting right now.

I just want to throw that background out and ask: what is Oracle’s AI strategy, and why is it working out for them?

It’s CoreWeave’s path. Oracle’s strategy is, “I’m going to use my balance sheet to get the largest contracts.” It’s using its investment-grade signature to get, call it, normal contracts.

Let me explain. Basically, one of CoreWeave’s issues was that, because they’re not investment grade and they’re not a reliable hyperscaler and such, they struggle to get capacity from reliable data center operators. Sorry, when I say reliable, I mean experienced guys—Digital Realty and all of these people that have been building forever, that have a lot of land and can deliver power.

These guys love working with hyperscalers. Oracle fits in the hyperscaler category, so Oracle can easily get capacity from folks like Digital Realty and others. That’s one advantage that they have, which all hyperscalers share, to be clear.

But what Oracle did on the side was basically go, “CoreWeave, I’m going to make a massive bet on this company.” By then, if you look on paper, who is Crusoe? Crusoe is a crypto miner that never built a data center for uptime-related, traditional data center purposes. They built mines, and I do think they had amazing engineering teams and such.

They're a great company, but back then, if you looked at it on paper, betting on Crusoe was the same as betting on other crypto miners today. Oracle took that bet. They signed this contract. And if you think about it, it's the same situation: they're getting a 5-year deal with OpenAI, but the deal they signed with Crusoe—I think the leaked numbers were around 15 years, right?—was a 15-year deal, probably $15 billion to $20 billion over those 15 years.

So if the contract fails after 5 years and they cannot renew it, they're stuck with 10 years of paying $1 billion a year or more. It's the same thing: it's taking massive financial risk, betting on the success of these companies. If something were to happen to OpenAI or to the overall AI industry growth, you could frame a future where Oracle isn't very happy.

I guess it's easier because even in 2023, this was a $300 billion–$400 billion EV company. They're making a big bet, but CoreWeave was existential: this works, or the company goes to zero. It's interesting that Oracle made that bet and it was not purely existential. Oracle would exist if this bet had gone up in flames, but at the same time, CoreWeave, as many people pointed out, took the optionality bet: this works and we're 100x; this doesn't work and we're at zero.

If it went up, they'd benefit, as they are benefiting—the stock is up probably 50% this year—but if it didn't work, they were on the hook for all those payments. And then they're sticking around and saying, "Okay, we're the largest data center people in the world. We've got all this excess capacity, and we're going to have to lease it for a song." Anything else?

Yeah, I was going to say that I guess you could frame a difference with Oracle, which is that they already have the cloud business. There's always the hope that by signing a gigantic deal with OpenAI, you also incentivize OpenAI to use your CPU-based infrastructure. I'm sure the spending that OpenAI does on Azure for traditional cloud services is fairly high. Again, to manage 700 million weekly users, it's not just about models; it's also about data storage and traditional CPU-based front-end processing and whatnot.

That's also, I guess, the hope that Oracle has: by signing those GPU deals, they can also grow their Oracle Cloud Infrastructure business, which was more than 10x smaller than the other rival hyperscalers. I think Oracle, if you look at the history, made this very bold bet on cloud in 2016 or 2017, and it turned out okay, but not that good. They were still lagging way behind the other hyperscalers, growing decently, but not at a triple-digit pace that was enough to catch up with the others.

It also makes sense from a strategic perspective to increase the size of the cloud business and hopefully upsell some services. What we've said in the report is that we don't see a lot of evidence that this is happening today. We don't think it fundamentally has to happen, because it's very easy to connect GPU infrastructure with another cloud provider.

We're yet to see a meaningful benefit of colocating—or having on the same cloud—GPUs and CPUs. But it's always a possibility, right? So they have this option. I don't know if it's going to happen or not. I don't think so, but it could happen, and if it does, it's of course very positive for their purposes.

And I think the thing that's coolest about Oracle, just at a high level, as someone who's not a specialist in this, is Larry Ellison. Again, he's in his—let's just call it—70s; I can't remember his exact age. Oracle is late to the cloud computing party. You just said they started in 2016 or 2017. AWS starts becoming the driver of Amazon in, what, 2012? Microsoft with Azure. They're late, they miss the boat, and they're basically an also-ran there.

Larry Ellison, in his 70s, says, "Hey, we missed the boat in cloud computing." Hyperscaling starts taking off in 2023, and there were plenty of people at the time saying, "Hey, this is a bubble. Too much capacity is already getting built. Where are the returns?" All that sort of stuff. And Ellison says—they instantly go all in.

When I read that Oracle autobiography, it fits totally with his personality. He goes all in, and he hits it again in his 70s. It's just crazy to have that mental flexibility and that forecast of the future. I want to ask you a few things about power, but is there anything else on Oracle?

No, I think the way you frame it is great. Definitely, that was a big, bold bet, paying off pretty nicely now.

Power. It still strikes me that—for the most part, and we had this discussion in our first conversation—all these data centers are going up. I don't want to say only domestically, but the US is the hot spot for data centers. You can tell me if I'm wrong. There are big data centers in Asia. Stargate 2 is in Saudi Arabia, I think—I can't remember—but for the most part, they're going up in the US.

They're going up not in New York City, but pretty close to urban areas because, obviously, connection latency and all that sort of stuff is a factor. Tell me if I'm wrong on any of that, but when can we start seeing data centers getting built in crazy places? I think there was an Elon Musk tweet about putting a data center up in space. I doubt we're getting one in space, but you could see the appeal, right? You don't have to worry about power as much because you don't have any heating components—any heating worries—if you're in space.

Alaska, right? We don't have to go build in Antarctica or somewhere with— But there's infrastructure in Alaska. It's really cold up there, and there's a lot of access. We talked about natural gas; there's a lot of it. When do we start seeing a big data center get built out in Alaska, where land is cheaper and labor is cheaper? I'm asking—I see you laughing because it's clearly a silly question—but why is it a silly question?

No, I don't think it's actually silly. I think it could happen. Generally, the problem that you see in more remote locations is labor. Can you actually get thousands of people on site? Are the logistics good enough to handle a project of that scale? That can be one of the big issues.

But overall, I think what you said in the beginning is inaccurate, because I do think we're already seeing very large data centers pretty far away from population centers. West Texas, I think, is already growing. Look at Abilene. You could say Abilene is a city—it's not too far away from Dallas—but I think we're going to see a few massive data centers in West Texas in the next 2–3 years, in more remote locations.

You could also say Ellendale, North Dakota—sorry, Applied Digital's site—but that's really far away from everything.

Ellendale's the Applied Digital site, right?

Yeah, correct.

The logistics to go there are pretty insane. It's pretty tough to get there, but it's still possible. I do think we're already seeing a move away from those metros. A lot of these data centers are being built in areas that are just more remote. But even there, it's still in the domestic US, right? I haven't looked at the map, but I'm sure getting out to Ellendale involves a pretty long car ride and a flight into a very small airport. But Alaska—I don't think I said Alaska because my first thought was super-far-north Canada. Super-far-north Canada has nothing built out there, right? So you'd have to build out the gas pipeline and a fiber connection. Alaska has all that right now.

I'm not saying it has all the infrastructure you need for a $500 billion Stargate or something, but it's got a lot of that. And if it's got a lot of that, you can lay some more fiber and stuff. I'm just surprised about Alaska. I know there were a few people talking about the Nordic regions, and I know there are some smaller ones out there, but again, I'm an outsider. I'm surprised that you haven't seen something huge getting built somewhere a little more out of the way.

You start saying, "Hey, we're spending $6 billion on this. Let's pay the data scientists an extra $5,000 to fly out there when we need them," or something.

You might have seen an announcement, I think 2 or 3 weeks ago, from Crusoe saying they're going to build a massive data center in Wyoming. It's still, I guess, close to Cheyenne, but still pretty far away, I would say.

Overall, I definitely expect to see data centers going to more remote locations. It's already happening today, and I think it's going to happen more. I haven't looked at Alaska specifically, so I can't really tell you.

I was just throwing that out there.

Actually, one thing we can touch on is a question I saw on Twitter asking about the Permian Basin. I would say—it's just a thought; I haven't looked at Alaska again—but the issue with the Permian Basin is that it doesn't have grid infrastructure.

Yes. Ideally, you want to build grid infrastructure because it has lower electricity costs and higher uptime.

A grid is actually the best power you can have on-site. Ideally, you want to have a grid connection. Fully islanded data centers are pretty expensive. It works out, again, as we discussed before, for a fast time to market, but then this equipment serves as backup, and you want to have primary power coming from the grid.

The Permian Basin doesn't have a large electric-transmission infrastructure. There's a project going on that's going to be 5 years out or something. So maybe that's one of the reasons why Alaska isn't considered today.

Let me ask a general question on power. It strikes me that power is the limiting factor. It is the constraint. Obviously, there's time to build these things, but it's been a race for power, and you've seen this in the stocks of all the power players—they've gone up. I think we touched on this in the first episode, but I want to follow up again. If I look at the history of things, power efficiency tends to improve over time, particularly when it's a limiting factor.

I would point to the biggest improvements in car fuel efficiency coming after big oil-price spikes. Airline fuel efficiency, I believe jets today are, depending on your source, 50% to 75% more fuel-efficient than they were 30 or 40 years ago. Homes today consume a lot more power because we use a heck of a lot more power, but they are much more power-efficient.

I just want to ask: at what point—right now, we're in the rush—but could you see a world where there's still a rush for GPU capacity, but 2 years from now they say, “All right, demand is still growing, but it's not growing exponentially. Let's optimize for power”? Then all of a sudden you have a world where a lot of these data centers—not that they're stranded assets, but you were building 1-gigawatt data centers, 2-gigawatt data centers—and you say, “Hey, now that we've optimized for power, it turns out we need 50% less power.” All these data centers are sitting here saying, “Oh my gosh, there's just no need for us,” because even though demand's growing, we just took our cost down because it's the bottleneck, and physical bottlenecks tend to get solved over time. Does that question make sense?

Yeah, of course. It makes a ton of sense. It's always a big fear that we're doing a massive oversupply and that AI demand is going to go away, which is completely possible, to be fair. It's a scenario.

We can look at the cloud-computing world as a proxy, where you've seen electricity costs as a share of revenue for these huge cloud providers go down over time. There seems to be a moment where it doesn't really go down that much. One thing you can do is just look at Azure's revenue growth and what they've published in their ESG reports, for example, and you see that it actually tracks pretty well. They're growing revenue a lot. They're growing electricity consumption a lot.

Basically, my point is that even if the unit of compute itself is improving—the CPUs in this case, especially the GPUs, are getting more and more powerful—this is kind of the Jevons paradox, I guess, which is that you're just going to use more of those GPUs.

Generally, when you look at GPUs over time, if you look at Nvidia's roadmap, of course pricing goes up for their GPUs, but power costs also go up for the GPUs. Actually, you see power and price tracking and correlating pretty well. That's because, if you look at Blackwell, the reason Blackwell is more powerful—there are a couple of things, but one obvious thing is that it's actually 2 compute dies, not 1 compute die.

When you have 2 compute dies, you don't double the power, but it's actually close to it. Then you have system-level improvements and whatnot. But the point is, at the hardware level, you actually have a pretty good correlation between the price and power of the system.

Can I slightly push back on that? It strikes me that for the past, let's say, 50 years—since the dawn of the computer industry—power has never really been a constraint, right? And by power, I'm using electricity. I'm basically saying the cost of electricity—the cost of power. I'm sure people would have loved to get more power into the GPUs and everything, but you could build with the assumption that you were basically treating your electricity and power costs as free, right? Obviously, they were not free, but they were so low.

Now power is actually the bottleneck. I think we would have a lot more data centers and a lot more capacity right now if people could just get access to the power and electricity. As we go into a world where power has been the bottleneck since mid-2023, for 2 years, do people start asking whether Nvidia's next product—what they're releasing in 2026—optimizes for this power-constrained world? Do GPUs go up, but maybe a little bit less, while power usage goes way down? Does that make sense? Is there any roadmap there?

If you look at Nvidia's roadmap, what they're doing is, as I said, price and power track pretty well, but actually the throughput of the chip goes up a lot. In terms of throughput per unit of power, energy efficiency goes up a ton, to be clear. If you think about the system-level improvements, it's all about having more GPUs working together and whatnot to deliver output with roughly similar power. So yes, Nvidia is very clearly trying to push power efficiency.

The reason they're doing it is mostly because you can use those additional tokens to generate more intelligence. That's where we go back to the debate of Jevons' paradox: what are you going to do with your extra compute power? Are you going to use it to save on costs or to increase intelligence?

The path that the leading AI labs are choosing is to increase intelligence, because I think what they're saying today is that the best way to monetize LLMs is to have the single best models. Very simply, Anthropic has the single best model for coding, and you're seeing everyone use Anthropic models.

If you go down the stack and start looking at the market for cheaper models that have a pretty good price-to-intelligence ratio, actually building those models is much cheaper because you can use techniques like distillation. You use 1 model to generate synthetic data and such, to distill intelligence into a smaller model. All the AI labs are doing that.

This is where you get into much more competition. If you think about the market for these mid-level models, there's a lot of competition from the Chinese, the open-source firms, and the big labs. But the real moneymaker is the frontier model. That's why we think, because of this specific structure, there is an incentive to always use the extra computing power that you get from NVIDIA to increase the intelligence of your model.

You can frame it this way: in terms of physical constraints, yes, the price per GPU goes up, and power goes up as well—it's pretty linear—but then the compute power you get from NVIDIA goes much higher, and the intelligence you get out of the model also goes way, way up, on a more exponential curve, I guess. Does that make sense? Any pushback?

It does make sense. It makes total sense. There are some follow-up questions I want to ask, but I am aware of time, and I want to ask one on a completely unrelated subject.

You guys had a report on robotics that I thought was very interesting. I'd love to just pause here, and you can give overall thoughts on robotics, and then I had some specific questions.

Yeah. Basically, the idea behind the robotics piece was that we just wanted to provide a framework to help people understand robotics markets. We added this classification of levels of autonomy, which we've seen in automotive.

I loved it because you instantly knew, right? You had level 0 to—I think yours went up to level 4 versus level 5 for driving. But as soon as I saw it, I was like, “Oh, I can equate it to vehicles.” I really like that framing.

What we found in our research—and maybe some people are going to nitpick on some things, because you cannot always make a level that's going to make everyone happy; there's always some nuance—is that overall, we can frame it in a way that's easy to understand.

If you think about our levels, level 0 is a rigid robotic arm. Level 1 is a slightly more flexible pick-and-place arm. Level 2 adds mobility, so it's a robot dog. Level 3 is sort of a weak humanoid, and level 4 is a strong humanoid. Typically, again, that's an oversimplification, and I hope robotics guys aren't going to kill me when they listen to that.

It's kind of the simple way to think about it: there are these different capabilities that add up over time. It's a simple way to frame it, I guess is the point, and it's easy for people to understand.

What we wanted to do with that piece was help people understand where we are. If we think about level 4, which is those super-humanoids, we're still pretty far away. We're still in the research phase, but we're actually already seeing level 2. People don't really talk about that, but all of these mobile quadrupeds, like the robot dog...

You're seeing that actually in early production phases. That might be a trend that could be interesting in the next 1 to 2 years.

That was the overall question I had on robotics: How quickly do you think we accelerate in robotics? I mean, with AI, maybe people are starting to get disappointed with ChatGPT-5 versus ChatGPT-4, but if you were sitting here 4 years ago and talking about where we are with AI, I think most people would have their minds blown. ChatGPT wasn't even out at that point.

How quickly does robotics start accelerating? It does strike me that, with Tesla—controversial—the robotics are out on the field, but you're starting to see more of the robotics get out into the field. I saw there was a robot-fighting league. You're starting to see more, and once you start to see more, it tends to accelerate, especially with the AI. So if you and I were talking here in 4 years, do you think we're seeing level 3 out there? If we're talking in 15 years, how quickly do you think this is accelerating?

Yeah, I'd say currently we don't have any evidence that it's going to happen as fast as we've seen with LLMs. One of the big reasons—

Robotics are just harder than anything else. Yeah.

Correct. Right. And in terms of data, there's always the data issue. With text, we have internet data—trillions of tokens. It's pretty hard to generate high-quality data for robotics. So that's one of the big bottlenecks in increasing capabilities. This is currently being solved, and I do think progress is starting to accelerate, but I struggle to see a world where it gets as fast as we've seen with LLMs.

If you think about it, LLM leaders love to say that today you have some models that are nearly as smart as advanced university students, and 2 or 3 years ago it was like a dumb, dumb kid.

I like when you say it's nearly as smart as an advanced university student, because I've given ChatGPT research projects and had it spit out brilliant insights. It's digested scientific papers in half a second and spit them out in language that—I like to say, “Hey, I'm someone who hasn't taken a science class since high school. Please put it out to me in a language I can understand,” and it does that.

So you get these brilliant insights. Then yesterday I asked it how many Bs are in “blueberry,” and it said, “I would bet my life there are 3 Bs in blueberry.” You're just like, man, on one hand it can understand science at an advanced-university-student level; on the other hand, it might be as dumb as my kid. It's so funny.

Yeah, I know, for sure. But I'm definitely with you. I'm a huge user of Deep Research and those kinds of tools. I think it's incredible.

In terms of robotics, anyway, to go back there, there are a few additional challenges. I think it's going to accelerate, but not to the same extent. I'm not the main robotics analyst, so I don't want to make a crazy prediction. Let's say level 4 in the next few years seems unlikely right now.

Wait, it's funny when you look at this, and this gets more into a dystopian future. Maybe we need to have a science-fiction writer on. If they're as smart as university students, and you're already hearing lots of—I think it's overblown, but you're already hearing lots of consulting firms struggle with AI taking consulting jobs—then you're saying, “Hey, if level 4 is here, I mean, how long until it starts taking low-wage, manual-labor jobs?”

You're like, “Jeremie, what are we going to be doing in 6 years?” We can't use our hands, and it's smarter than us. What are we going to be doing?

Yeah. I mean, there's this great analogy. I think it's interesting when you think about accountants before Excel, before spreadsheets. I've talked to people who told me that back then people were thinking spreadsheets were going to kill accounting jobs, right? And actually, you've seen accountants go up a lot. They're just doing new types of work. They don't have to do all the manual calculations and such. They can let the machine do that, and they can use their brains in some other ways. So, yeah, I guess that's the framework I would use.

Which—the one people point to now is law firms, right? You say, “Hey, this is going to take the work of a lot of junior analysts,” but at the top, we have no lack of legal work, and legal cases are accelerating.

But you do wonder. Or there's the famous—what is it? They tried to ban the sewing machine because it was going to replace all the seamstress jobs in Britain. It's a little scary, right? I tend to be pretty optimistic and say, “Hey, there's going to be jobs, and it's going to create really new, interesting ones.”

But when you start getting to, “Hey, you've never had something that can replace human thinking before,” you realize that you still needed a human. While it could automate basic tasks, you still needed a human to go flip the French fries over and hand them out to the customer. And if on one hand robots take over the French-fry machine and on the other hand computers think better than a McKinsey consultant, it's like, what is left for us to do?

Yeah, I don't know. I guess maybe more relationships. The value of social relationships probably goes up over time because at some point you trust humans more than machines for certain things. So maybe, yeah, maybe that's the future. You've got to be sure to be well connected, to have a good network.

That's why it's a good thing we're so handsome. That's the one thing robots can't take. You and I can hop on and go be fashion models. Robots can't take that—for now.

Jeremie, this has been great. The first one was a hit. I think this is going to be a hit. Maybe we'll have to have you on before the end of the year or something to do outlooks for 2026, or something. Talk about how high that capex spend is going in 2026.

10 trillion dollars. No.

10 trillion. Wow. I said it wasn't accelerating. We're getting real acceleration. This has been great. I'll include a link to semi analysis in the show notes and uh we'll go from there.

All right, man. Always a pleasure. Thanks for having me.