The SpaceX IPO, Fable 5, AI Capex Update & Market Check w/ Gavin Baker, Andrew Fox & Clark Tang | Edited Transcript

A copyedited transcript of BG2 on the SpaceX IPO, xAI compute, Fable 5, frontier models, and AI capex.

Chapter Timestamps

00:02 Opening thesis: why SpaceX is framed as both a space and AI compute bet

00:40 Show setup: SpaceX IPO, Mythos launch, Taiwan and Computex takeaways

03:05 xAI compute economics: Google and Anthropic deals, operating profit per gigawatt, and Colossus IRR

13:28 “Elon Web Services”: SpaceX becomes an AI hyperscaler inside the IPO model

19:08 Data centers as engineered systems: power, speed, behind-the-meter design, and supplier advantage

26:01 Orbital compute economics: launch cost, space power, and why timing is still debated

29:25 Cursor, xAI, and model data: why proprietary coding data changes the SpaceX AI story

35:05 Bull and bear case: whether SpaceX can 8x revenue and how to size post-IPO risk

37:00 Post-IPO trading: drawdowns, lockups, position sizing, and the psychology of “must-own”

43:56 Fable 5 and Mythos: long-running agents, benchmark limits, and model-release implications

51:00 Frontier vs. open source: why revenue may concentrate at the frontier even if open models catch up

1:04:11 AI capex math: $1.5T in spend, $300B in revenue, depreciation, rates, and power constraints

1:18:13 The next trillion: SpaceX, Anthropic, OpenAI, market-cap creation, and closing reflections

Made with: The Transcript Desk Chrome Extension

Full video:

Brad Gerstner sits down with Gavin Baker and Andrew Fox of Atreides Management, alongside Altimeter partner Clark Tang, to break down how investors should think about the SpaceX IPO, xAI compute economics, orbital data centers, Fable 5 and Mythos, frontier-model revenue, and the latest AI capex math.

Transcript

00:02-02:30

Gavin Baker: I think we are all pretty “AI-D.” If you’re “AI-D-L,” it means you believe we need to build a lot more compute than the world currently expects and that these models are going to be far more valuable than people realize. When you combine that with their core business, I don’t know of another entrepreneur or another company that represents a better bet on the future than SpaceX. For most institutional investors, it’s a “must-buy” and a “must-own.” It’s a “set it and forget it” investment if you want a real stake in both the future of space and AI.

Brad Gerstner: All right, here we go. It’s an early morning in Silicon Valley and BG2 is back. We’re chopping it up on all things tech and markets. To do that, I have none other than GB in the house—Gavin Baker from Atreides. He’s brought his main guy, Andrew Fox, and of course, I had to draft my partner, Clark Tang, into the mix to discuss some of the big questions of the day.

Specifically, how should we be thinking about the SpaceX IPO? What are the big levers? There are some massive numbers floating around regarding what might happen over the next few years, so let’s break that down and simplify it for everyone.

Also, Mythos launched yesterday. I want to talk about who’s up and who’s down in the race for superintelligence. Where do we stand, and what did we learn from the Mythos launch? Additionally, Clark was in Taiwan last week with Jensen Huang at Computex and GTC. What were our takeaways from that? What’s the status of GPUs and memory? Where are the bottlenecks, and where do we go from here?

Brad Gerstner: To kick things off, Gavin, let’s talk about the SpaceX IPO. It’s happening in two days. You’re a major shareholder—congratulations. We are shareholders as well and expect to be buying more in the IPO. The Wall Street Journal and Goldman Sachs are both reporting projections of $160 billion in revenue by 2028. We know the IPO is priced at $135 a share, implying a $1.77 trillion valuation. When we consider the primary drivers, there are so many moving parts in this deal. Nobody is better than you at breaking this down.

Brad Gerstner: Nobody is better than you at breaking things down and simplifying them. What are the key levers we should be thinking about over the next few years?

Gavin Baker: It’s great to be here. Thank you for having me. I thought we were going to call the show “BGGB,” but we’ll stick with BG2—I’m in your house, after all.

Brad Gerstner: Hey, that’s subject to revision!

02:30-04:46

Gavin Baker: That’s okay. I think there are two big levers or variables that people should focus on. I won’t comment on where I think those variables are headed, but one of them involves a chart you guys have. Did you post this on X?

Brad Gerstner: I did. I posted it previously, and we’ve also included a new version that incorporates xAI’s latest deals.

Gavin Baker: Clark, whom I’ve known for many years, did a great analysis here. He shows that xAI’s deal with Google for cloud computing generates more operating profit per gigawatt than Anthropic, Meta, Google, or OpenAI. Their deal with Anthropic also likely generates more operating profit than anyone else.

Gavin Baker: Your colleague at Altimeter, Frieda, also calculated a 55% IRR on Colossus 1. If you can borrow money at 6%, 7%, or 8% and invest it in something with a 55% IRR—well, I’m not the most sophisticated thinker, but that math definitely “maths.”

Gavin Baker: So, I think one of the two most important variables is how quickly they can bring terrestrial data centers online. We know from Jensen Huang that Elon Musk brings data centers up faster than anyone—122 days. Speed is literally cost, because every day you’re paying electricians and plumbers, that adds to your expenses. And now, they are monetizing them.

Gavin Baker: They are now monetizing these models at arguably the highest rate. Everyone should run their own math on that, but it is a massive variable.

Andrew Fox: It’s a truly massive variable. We have a chart here that is already wildly out of date, which is amazing. We made this chart only 10 days ago, but in the 12 days since, we’ve seen the Pareto curves for OpenAI’s Codex and Anthropic’s Opus 4.7. Since then, Opus 4.8 has been released, and now we have Fable and Mythos. It’s freaking wild—in just 10 days, we would have had to update this chart twice.

04:46-07:14

Gavin Baker: What the Pareto curve shows is how much intelligence you can get for a given cost. I believe all frontier model revenue will accrue to the Pareto curve. This specific one is for coding. What’s so impressive is that you can see Composer 2 was Pareto dominant. Even at the lowest level of intelligence with very little training, this reflects the quality of the data. I know you know Cursor a lot better than I do—a vast amount better, actually.

Andrew Fox: My understanding is that Cursor and Anthropic have more tokens of proprietary coding data than anyone else—more than even exists on the public internet. Cursor used their own private data, performed reinforcement learning (RL) and supervised fine-tuning, and developed a really good model. Then, they spent three weeks on the Colossus 2 cluster and produced a model that, as of 12 days ago, was Pareto dominant with Composer 2.5. Now, that’s based on their own benchmark, CursorBench, so perhaps take it with a grain of salt. However, I think this suggests that the Cursor data is incredibly valuable for coding. When it is trained to be Chinchilla-optimal...

Gavin Baker: If you train a model to be Chinchilla-optimal or beyond, especially using reinforcement learning, I think it suggests that xAI and SpaceX AI have a real shot at being major players in the coding space.

Brad Gerstner: I think one of the most interesting things is how you just answered that question. We didn’t even talk about launch, Starlink, or communications. Up until about six months ago, those segments *were* the business.

Gavin Baker: Exactly.

Brad Gerstner: Right. And then we merged in xAI and Cursor, and then we announced these deals where it became very clear he was building “Elon Web Services” right under our noses. But I want to go to Fox now. Give us the breakdown. We have three big lines of business, right?

Brad Gerstner: We have the communications and Starlink launch business, the AI compute business, and then I want to circle back to the xAI piece you were just touching on. But if we look at the core business, what do we have to assume goes right with both launch and Starlink to hit the numbers being projected?

Andrew Fox: Yeah, sure. Look, I think the foundation for everything is the launch business.

07:14-09:22

Brad Gerstner: Right.

Andrew Fox: This is the crown jewel of SpaceX. It’s something no one else really has—notably reusability, and soon, rapid reusability. That is what you need to believe in to reach the AI economics that make orbital compute economically attractive, especially given that we are currently facing shortages in both power and chips.

Andrew Fox: So, I think rapid reusability is the main thing we are watching, and I think most people should be watching it too. Elon talks about it a lot—getting these rockets to fly at a cadence comparable to an airline. Gavin has used this analogy before...

Andrew Fox: Gavin has used this analogy before, but the old rocket industry was like boarding a plane, flying to California, and having the plane explode right after you get off. What SpaceX is ultimately trying to achieve is to have both stages of Starship—not just the booster—fly 30, 40, or 50 times before needing a retrofit. When you do that, you’re amortizing the cost of the vehicle over many flights, which is what brings the cost down significantly.

Brad Gerstner: But that’s a really hard problem to solve.

Andrew Fox: Extremely difficult. The company has been loud and clear that they’re going to attempt to bring back the second stage of Starship later this year and then make it reusable—actually reflying that second stage next year. From there, they’ll ramp up the cadence. At the end of the day, driving down the cost of launch is what enables all of these other businesses and makes them so attractive relative to the incumbents.

Brad Gerstner: Starship 3 just launched. How many launches do you think the consensus is assuming two or three years from now? What is the launch cadence? Are we launching one of these every day, every week, or every month? Where are we in terms of expectations?

09:22-11:41

Andrew Fox: Right now, expectations are going from roughly 160 to 165 launches last year up into the high hundreds over the next several years. We’ll likely get into the thousands of launches in the three years following that.

Brad Gerstner: Okay.

Andrew Fox: I think the company has aspirations—

Brad Gerstner: With thousands of launches, you’re doing two or three launches a day, right?

Brad Gerstner: Right. Talk to us a little bit about what this enables. Obviously, I’m here in Silicon Valley and I can’t even keep a call connected on Sand Hill Road, two decades into the mobile revolution. It’s the craziest thing; it’s like a third-world country.

Gavin Baker: It’s a major business problem when you’re freaking out here.

Brad Gerstner: It’s crazy. Right by the Rosewood dead zone, I’m thinking, “How can this possibly be?” It’s almost like a joke. It’s the epicenter of technology in America and you can’t maintain a call. So, we’re all going to switch to Starlink Mobile when it comes along because I don’t want to lose that call on Sand Hill Road. Walk me through this at a high level. It’s a big portion of the revenue growth expected in the business over the next two to three years. My hunch is a lot of this is driven by direct-to-cell connectivity. Walk me through those economics.

Andrew Fox: Yeah, it’s actually interesting. The broadband business is still in its very early stages when you think about the percentage of households that have been penetrated to date. If you look at the percentage of global households with Starlink, it’s less than 1%. That’s the broadband side, where you have a base terminal at your house, on your car, on your boat, and now in airlines as well. I actually think broadband can scale to hundreds of millions of terminals.

Brad Gerstner: Hundreds of millions of users. And today the subscriber base is...

Gavin Baker: It could reach hundreds of millions if they achieve rapid reusability of Starship, which is really hard. If there isn’t significant competition, hundreds of millions is possible, but maybe...

Brad Gerstner: I always say around here—and it’s funny seeing a PM and an analyst in this situation, because it’s exactly what I do with Clark—someone will say something and I’ll say, “The future is a distribution of unknown probabilities.” It’s either more likely or...

Brad Gerstner: What are the probabilities? Is it more likely or less likely? Give me the distribution. Are we talking 20% or 30%? It’s hilarious.

11:41-13:34

Gavin Baker: Well, no, it’s 100% the same thing. I’ve watched Elon do many hard things, and this is a really hard thing. I think it’s reasonable to believe they’re going to succeed with rapid reusability, but I think it’s important to acknowledge—

Andrew Fox: That orbital compute, Starlink V3, and Starlink Direct-to-Cell require first reusability for Starship V3, and then rapid reusability. That is what unlocks a lot of this, right?

Brad Gerstner: Right. When I see the models the banks are putting out there—the ones reported in the Wall Street Journal that have been widely leaked—they largely have the revenue on connectivity. Let’s call it Starlink Direct-to-Cell, etc., going from roughly $10 billion to $50 billion by 2028. I’m not asking you guys to give me your specific numbers, but when I’m talking to Clark, all I’m trying to size up is the order of magnitude. Do we think we can 5x the business over the next three years? Is there enough TAM in both broadband and direct-to-consumer? I think the answer to that is yes.

Gavin Baker: Yeah. I’ll just say this very simply: I travel with Starlink. I’m a big video gamer, and very consistently, wherever I am in the world, Starlink provides the best connection.

Brad Gerstner: Yes.

Gavin Baker: It’s the fastest and has the lowest latency. I think once they achieve rapid reusability, they’re also going to have the cheapest cost per gigabyte or megabyte delivered. “Better, faster, cheaper” has always been a winning formula. So, $50 billion represents about 0.3% penetration of the global telecom market. Now, maybe there’s some deflation with Starlink pricing, but that’s how I’d frame it.

Brad Gerstner: I like betting on better, faster, and cheaper.

Brad Gerstner: I like betting on better, faster, and cheaper. Clark, I would say the biggest surprise of the last six weeks is that Elon—you know, we talked about it on the All-In Podcast and called it EWS, or “Elon Web Services”—has struck these massive deals with Anthropic and Google.

13:46-15:31

Brad Gerstner: I don’t think people were even considering SpaceX as a player in the AI compute game. If you looked at the models a few months ago, it was all about connectivity via Starlink and then xAI as the model. But this entire category of taking all this compute—which he is uniquely good at standing up—and reselling it in a highly profitable way was not in many people’s forecasts. Now, it’s a major component. You and I did that podcast with Jensen Huang where Jensen said Elon is an “N of one.”

Clark Tang: What they achieved is singular; it’s never been done before. To put it in perspective, a 100,000-GPU cluster is easily the fastest supercomputer on the planet. Normally, a supercomputer of that scale would take three years to plan, and once the equipment is delivered, another year to get it all working.

Brad Gerstner: Yes.

Clark Tang: We’re talking about 19 days.

Brad Gerstner: Wow. “N of one” is right. Elon’s ability to secure supply, stand it up, and deploy it in a way that is coherent and effective—both for himself and now for others—is incredible. Walk us through that, because it looks to me like this is now a major component of the revenue story.

Clark Tang: Totally. We were all at the Memphis data center, and the amount of engineering that went into building those sites was very evident. People always talk about Google’s ability to build a TPU and sell it to Anthropic to generate AI revenue. I think it’s a pretty...

15:32-17:58

Clark Tang: I think we’re seeing a similar dynamic here with Elon. He is able to secure power and build these sites faster than anyone else, and he can now monetize that for this massive AI market ahead of us. If you look at the relationships he’s forged with his suppliers—whether it’s Jensen Huang or the various site owners who want xAI as a tenant—his ability to finance these deals at very attractive rates relative to other players is a major advantage. These advantages compound over time.

When you’ve built the credibility to stand up these sites and monetize them at this level, it becomes a very attractive proposition for everyone involved. In fact, if you look at these deals in particular, Gavin, as you pointed out, they are actually monetizing their infrastructure perhaps better than any other player in the space.

Brad Gerstner: The margins are a lot higher. Google is obviously paying SpaceX a huge premium for this compute. Fox, you said something earlier that I thought was really important: it may very well be that in order to get first in line for space-based compute—which Google certainly wants—they are willing to pay a premium for their terrestrial compute today. To me, that’s how you square the circle on why they’re paying that premium. What are your thoughts?

Andrew Fox: Yeah, look, I think there’s some of that embedded there. But at the end of the day, SpaceX can stand up compute quickly, they can do it coherently, and they can put a massive amount of it in one place and have it readily available. I think that accounts for most of the premium. But beyond that, certainly—

Gavin Baker: People are going to space eventually.

Brad Gerstner: So you pay a little call option to get first in line for space.

Andrew Fox: There you go. Exactly.

Gavin Baker: We’ve all been investing in the NeoCloud space, so there is a fundamental...

Brad Gerstner: There is a fundamental belief around this table that we lack the compute needed to continue pushing the frontier of intelligence. Consequently, we have to build a lot of compute. Currently, there is a competition unfolding. On one end, you have the hyperscalers building out that capability. Then you have AI-dedicated clouds building it out. Now, literally in a matter of weeks, a giant has emerged in this category: SpaceX.

17:58-20:10

Brad Gerstner: The question for you, Gavin, is: can they consolidate this market? When I look at the marketplace, Elon has a unique ability to secure supply and a unique ability to cut deals on the other side. Nobody can stand up infrastructure as fast as he can. I think we might see a real consolidation in the AI compute market where you have the hyperscalers on one hand, and on the other, he emerges as the largest and strongest player in the space.

Gavin Baker: Yeah. Are they the number four or number five hyperscaler today? After the Google deal, they will be number four.

Brad Gerstner: It’s kind of wild, right? In 30 days, we went from not being an AI hyperscaler to being number four. We passed a lot of companies, including Oracle.

Gavin Baker: Coreweave is a huge business that we are investors in, but there are many other players like the Nebius’s and the IRONS of the world. I would say there are probably 50 “Neo-labs” being funded in Silicon Valley right now because of the compute shortage. So, achieving that in 30 days is just extraordinary.

Gavin Baker: What I would say is that there is a common belief that these data centers are commodities. I do not share that belief, and I don’t think anyone around this table does either.

Andrew Fox: I don’t think anyone around this table shares that belief. In the same way that Elon was able to re-engineer a rocket from first principles and make it reusable, he engineered an electric car from first principles. Everyone else was trying to build an electric car like an internal combustion engine vehicle, but he thought about it differently. I think he looked at data center design from first principles as well, and he designed something fundamentally different.

Andrew Fox: I actually told the team, “Hey guys, maybe be a little less public about things that are obvious to you regarding data center design, but are revelations to everyone else.” I think what they’re doing is more differentiated than they realize. It’s so logical to them, but it’s not logical to everyone else. That’s how he was able to get it done in 122 days.

20:10-22:31

Clark Tang: Yeah, to that point, Brad, we were meeting with one of our portfolio companies yesterday and discussing “behind-the-meter” solutions. When you really think about it, there are only two or three players right now who can reliably engineer a behind-the-meter data center. There is significant engineering work involved in all of this.

Brad Gerstner: Think about it this way: if you’re a company like GE Vernova and you only have a certain number of gas combustion engines, who are you going to sell them to? Are you going to sell them to xAI, or to one of these startup “NeoClouds”?

Gavin Baker: Well, there’s another dynamic at play. Everyone starts making more money the moment those GPUs are energized and sold. Speed is everything.

Brad Gerstner: Exactly. Speed is literally money for all the suppliers—power, land, and turbines. We’ll see how it plays out.

Brad Gerstner: Right. But this is just terrestrial we’re talking about. I want to hit on this, and then you can flip it back to me. Let’s assume they continue to build out the terrestrial landscape and continue to find buyers for that. Walk us through what this unlocks and how it relates to space data centers. Once you start talking about terawatt capacity and beyond—we’re talking a thousand gigawatts—and this year we’re only doing 25 or 30 gigawatts, just to put it all in perspective.

Gavin Baker: Yeah, 20 to 25 gigawatts.

Brad Gerstner: Okay. So once we start scaling up, walk us through this: do we have to have space data centers in order to get excited about buying the IPO? There’s obviously a debate about the timeline. I heard Jeff Bezos say he thinks it’s more like six years, but Elon is going to say three—because if he says six, it will take even longer. So he says three, and we may get it in four or five. Are space data centers integral and essential to the IPO? What do you think the timeline is?

Gavin Baker: I don’t think so. I think if you consider the variables around what Cursor could mean for xAI...

Brad Gerstner: Yeah, and we have proof of existence that once you really reach that Pareto frontier...

Gavin Baker: Revenue can scale rapidly. It’s called Anthropic. There seems to be a lot of demand for coding. Amjad Masad, the founder of Replit, posted something very interesting.

Brad Gerstner: The founder of Replit.

22:31-24:52

Gavin Baker: He called it “Bitter Lesson adjacent”—that coding may be the fastest path to AGI and ASI. If a model is truly excellent at coding, it can write code to do anything.

Brad Gerstner: Correct.

Gavin Baker: I think that’s a profound point, and I believe coding will continue to be...

Gavin Baker: I believe coding will continue to be a critical factor. If you consider that variable, along with Starlink’s direct-to-cell capabilities enabled by Starlink V3, and the speed at which they can deploy terrestrial compute, orbital compute isn’t strictly necessary to justify the IPO valuation. However, it is certainly an important piece of the puzzle.

Brad Gerstner: Perhaps another way to look at it is that we might reach Artificial Super Intelligence (ASI) faster than we achieve full-scale orbital compute. That could take us from a 300 IQ level to 400, 500, and beyond, with the ability to scale up until it consumes 10% of global GDP. But maybe that’s where we should head next in this conversation.

Gavin Baker: Actually, on the topic of orbital compute, it would be great to have Foxy or Clark lay out the math from first principles. Clark has an excellent chart detailing the costs per gigawatt and the overall dollar-per-gigawatt breakdown.

Brad Gerstner: Great. Walk us through the economic case.

Clark Tang: To the point of whether orbital compute is key to investing here—I don’t think it is. First, let’s look at the implied monetization rates based on current expectations for the AI business. There is a $160 billion valuation figure that has been leaked and is being widely discussed.

Andrew Fox: The implied monetization rate on that number is roughly $14 billion per gigawatt, per year, for the AI business.

Brad Gerstner: And they just signed Anthropic at $22 to $23 billion, and Google at $50 billion.

Clark Tang: Exactly. So, you can invest in the terrestrial AI business and still be very excited about it. But regarding orbital—it’s an important point. Most investors have an easier time wrapping their heads around how SpaceX wins on the ground. They understand the strategy: can they acquire the land, secure the power, and build it out?

Brad Gerstner: Can they secure the land, power, and chips? The answer to that is a high-probability yes. What we’re saying is that the rate at which they are monetizing already gets you to the numbers being leaked, even before you take the leap of faith that they’ll extend their lead with orbital data centers. But take us through that part as well.

24:52-27:22

Andrew Fox: Sure. With orbital compute, I think the key factor is two-stage reusability—and beyond that, rapid two-stage reusability.

Brad Gerstner: Right.

Andrew Fox: Today, with Starship, they’ve shown they can successfully land the booster. As for the second stage, we’ll see what happens later this year. I think they’re attempting to bring that back and make it reusable by next year. The reason two-stage reusability is so critical for the economics of orbital compute is that the cost per kilogram drops significantly.

Andrew Fox: We’re talking about going from roughly $1,500 per kilogram on Falcon 9 to somewhere around $250 per kilogram—or even lower. The more you can reuse the rocket, the more that price drops.

Brad Gerstner: Right, because you’re just depreciating the cost of the launch. Eventually, you asymptote to the cost of the fuel.

Andrew Fox: Exactly. That assumes you can use a rocket forever, which will take us a very long time to truly achieve. But at that point, we’re talking about a cost well south of $250 per kilogram.

Brad Gerstner: So then you look at the specs of these AI satellites. Elon did a great job explaining this.

Andrew Fox: Yeah, that podcast where he laid out the satellite specs the other day was incredible. It was really great because...

Andrew Fox: It was really impressive because they are finally showing people how you could viably design one of these satellites. They’ve detailed the weight of the satellite and how many you could fit into a Starship launch. When you back into the numbers, you get something like five megawatts of capacity per Starship launch. Since there are 100 metric tons in one of those Starships, you can calculate the cost per gigawatt to launch this compute into space. The math you arrive at—before accounting for things like faulty GPUs or satellite failures, which will inevitably happen—is roughly $5 billion in CapEx per gigawatt to put these in space.

Brad Gerstner: For comparison, let’s look at terrestrial costs. When you talk about switchgears, generators, transformers, the shell, and securing the power, that currently costs about $20 to $25 billion per gigawatt. So, we’re talking about a 5x reduction in cost for half of your bill of materials for the data center. That is a massive number.

27:22-29:26

Andrew Fox: Exactly. To put it simply, it costs about $60 billion to put a gigawatt on the ground today. We’ll say $35 billion of that goes toward the GPUs and the silicon doing the training and inference, while $25 billion covers the land, the shell, the power, and the cooling. I would hypothesize that those terrestrial elements are likely to be inflationary, so that $25 billion figure may not decrease. However, in space, power and cooling are effectively free. And when I say “space,” I mean the physical room—there’s no “land” in space, but there is plenty of room. You’re talking about putting a gigawatt into space for a total of $30 billion.

Andrew Fox: You could put a gigawatt into space for $30 billion and have lower operating costs. Compare that to $60 billion on the ground, which is inflationary. That $30 billion—or even $5 billion—could become deflationary over time. However, we have to consider reliability and maintenance.

Gavin Baker: Exactly. Everyone can do the math, but as long as these satellites aren’t failing at an astronomical rate, the math works out.

Andrew Fox: By the way, we know GPUs melt and lasers fail. We see this in data centers, particularly during massive training runs. So, as long as the reliability and maintenance costs aren’t dramatically worse, the logic holds—especially once we achieve reusability and rapid turnaround with Starship V3.

Brad Gerstner: When we look at this, we’ve already walked through Starlink and agreed that direct-to-cell is a reasonable assumption you can wrap your head around. Then, regarding terrestrial data centers, it’s not a stretch to think—based on the deals Elon has already done—that SpaceX is going to build a much larger business there. On top of that, you have this “call option” on space-based compute that could drop the price even further.

29:26-31:56

Brad Gerstner: The one thing we haven’t discussed yet is their model. I find this surprising: six months ago, xAI was competing and doing well, but they’ve done something dramatic over the last couple of months by acquiring Cursor. Cursor has 700 to 800 people and was already doing incredibly well from a revenue perspective. Our own projections suggested they could exit this year with up to $10 billion in revenue. They were growing very fast as one of the leading coding agents, but they also brought in an incredible team.

Clark Tang: They also had this incredible team with the potential to really build a frontier-level model, but they were compute-constrained. Suddenly, they get acquired by xAI, which has massive compute they can now train on. When I look at the AI revenue line item in these models—going from $10 billion to $150 billion—a lot of that will be the Coreweave-type business they have. But the real question is: how much of that will be the core xAI business, powered by the new team from Cursor? Any thoughts on that, Gavin?

Gavin Baker: Right now, Composer 2.5 was Pareto-dominant 12 days ago. It was trained on the Kimi K2.5 base model. Now, the Grok 4.3, 1.5-trillion-parameter model is training. One would hypothesize, based on scaling laws, that it might be a better base model. Furthermore, the Cursor data is being injected into the pre-training process, not just reinforcement learning. We’ll see, but I think that is going to be a very important data point when it comes out.

Clark Tang: I just think everyone should keep in mind that once you are at multiple places on that Pareto curve, if you have compute, you can scale really rapidly.

Brad Gerstner: To me, that is the one piece being lost in the story. It’s easy for everyone to get excited about the deals with Anthropic because you can wrap your head around how much revenue that generates. I see debates about the 90-day termination clauses, how long those deals last, and what multiple to put on those revenues. But the thing getting lost is that they’ve dramatically advanced their capability when it comes to building a frontier model. People outside Silicon Valley may not know Michael and the team at Cursor as well, but this is an extraordinary team that was just downloaded right into SpaceX. xAI was already building good models, and what they have now is a way to monetize compute that gives them...

31:57-34:14

Brad Gerstner: You can monetize compute in a way that gives you a call option. You can pull all that compute in-house to train a model and then run it. If there is an upside surprise, I suspect—if we went around the table—that this is the area getting the least amount of attention while offering the biggest potential upside. Clark, do you have any thoughts on what is being overlooked or misunderstood about the business today?

Clark Tang: I would say the last few weeks have proven that Elon and his team can stand up all this compute. If you look back just a year and a half ago, they were behind in the race to stand up compute; they didn’t have many H100s. Then they brought Colossus online, followed by Colossus 2, at a scale much larger than anyone else. Now, as we gear up for the Vera Rubin chips, my conversations suggest they have secured perhaps up to 20% of that capacity. Especially in the early days when these chips are incredibly scarce, they are going to have a lead because people believe they can stand up this compute better than anyone else.

What the last few weeks have shown is that Elon will take his shot at hitting the frontier, but if for whatever reason they have over-provisioned capacity, they own a very scarce asset. They’ve demonstrated they can monetize that asset with best-in-class margins and payback periods.

Brad Gerstner: The irony is—and you and I have been doing this long enough to know—that is exactly why Jeff Bezos built AWS.

Clark Tang: Right, he had to build capacity for Black Friday.

Brad Gerstner: Exactly. But for the rest of the year, he was sitting on all this excess capacity they had built, and he figured out an incredible way to monetize it. And by the way, investors at the time, in 2009 and 2010...

Brad Gerstner: Back in 2009 and 2010, when Jeff Bezos was building out the capabilities for AWS, investors hated it because he was consuming all that free cash flow. Meanwhile, he was digging the biggest gold mine in the history of the world.

Gavin Baker: It was certainly one of the biggest.

Brad Gerstner: Among them at the time, it was probably the biggest.

Gavin Baker: Yeah, although Google Search might want to have a word about that.

34:14-35:34

Andrew Fox: By the way, I do think this is important: Grok 4.3. If they acquire Cursor, that may end up being very significant. As of about 10 or 12 days ago—and these things move fast—Grok 4.3 was on the Pareto frontier. It was the most intelligent 500-billion-parameter model in the world. They were right there on the frontier. There are really only four companies on that frontier: xAI, SpaceX, Google (with Gemini 1.5 Pro), and then the rest of the space is dominated by Anthropic and OpenAI. But xAI was on the Pareto frontier. Now, we’ll see what they do with Cursor.

Brad Gerstner: Yeah.

Andrew Fox: I want to come back to that in a second. But Gavin, I want to ask you: what do you think? Do you think the biggest source of potential upside is the model?

Gavin Baker: Yes.

Andrew Fox: What do you think, Clark?

Clark Tang: I think that is the part that is least talked about.

Brad Gerstner: It’s the least talked about, right? Listen, when I look at the bull and bear case for the IPO, the bears are looking at last year’s revenue—say it was $18 billion—and comparing it to the bank forecasts of $160 billion three years from now. They’re saying, “Look, not many companies in history have 8xed their revenue over a three-to-four-year period.” That’s why people get nervous about the valuation. But when you break it down from a first-principles analyst perspective, part by part—which is what I’ve tried to do here—and you look at Starlink, it looks totally doable. When I look at...

35:36-37:54

Brad Gerstner: It looks totally doable. When I look at what they’re building in AI compute terrestrially, it seems completely achievable over the next three years. Looking at the model itself following the acquisition of Cursor—and combining that with the compute they already have—that looks like a potential upside to me.

I think when we look back three years from now, there’s a decent chance everyone will say, “Oh my god, that was super obvious,” even though today all of these things carry risk. To go back to where we started: none of us are here to pump the IPO at a $1.77 trillion valuation. We’re really just here to break it down the way we do inside our own shop and ask: what is the distribution of future probabilities? What is the probability that it goes higher from here?

I think we’re all pretty aligned. If you’re long AI, it means we have to build a lot more compute than the world currently thinks, and these models are going to be far more valuable than people realize. When you combine that with their core business, I don’t know another entrepreneur or another company that represents a better bet on the future than SpaceX. For most institutional investors, it’s a “must-buy” and a “must-own.” It’s a “set it and forget it” type of investment if you want a real bet on both the future of space and AI.

Gavin Baker: From your lips to God’s ears.

Brad Gerstner: I mean, listen, I think you’re going to have to wait. We saw that chart that came out last week—everyone was sharing it on Twitter, and it was conveniently timed. It showed the average maximum drawdown post-IPO for about 20 major companies, including Facebook, Twitter, Alibaba, and Shopify, was over 50%.

So, maybe we’ll end this section here. Gavin, you and I have been doing this a long time. We know it’s going to be bouncy around the IPO. As a manager, how do you try to...

Brad Gerstner: How do you, as a manager, try to handle that? Do you try to trade around the IPO, or do you just set it and forget it? From an Altimeter perspective, we tend to take a base position that we set and forget, and then we might size up or down depending on how the market reacts in a particular moment. But what are your thoughts on this chart, or how are you guys thinking about it specifically? You obviously own a lot going into this.

37:54-40:19

Gavin Baker: First, I agree with absolutely everything you said. I think about it the same way: set it and forget it. You’ve talked about having ballast that you move around—moving it to one side of the ship when you want to lean into the wind to go faster, and moving it to the other side when you don’t want the ship to tip over. I think that’s a great analogy.

Andrew Fox: I think about all the important companies in the portfolio the same way, so I 100% agree. I mean, this chart is a bit of a bummer.

Brad Gerstner: Yeah, it is.

Gavin Baker: What I would say is, while this is the data on IPOs, we are in a truly unprecedented situation. We’ve never had an IPO this big.

Andrew Fox: We’ve never had an IPO that is going to enter an index this quickly. We simply don’t know how much selling there will be from investors. I would hazard a guess—though I don’t know for sure—that Elon doesn’t need the liquidity. Foxy, what does he own? About 50%?

Andrew Fox: Roughly 50%.

Brad Gerstner: And by the way, he’s locked up for 365 or 366 days. So we know he isn’t selling.

Gavin Baker: Exactly. It’s an unprecedented situation, and the honest answer is that I don’t know what’s going to happen in the short term.

Brad Gerstner: The right approach, which I would encourage every investor to consider when making their own decision, is to think exactly the way you do.

Brad Gerstner: The best approach is to think exactly the way you articulated it. We have these different levers and variables. You have to look at each one of them from first principles and make your own decision.

Gavin Baker: Do your own due diligence and be thoughtful, but there are a lot of variables at play here. It’s a little funny to me that people were focused on it being valued at a hundred times trailing twelve-month revenue. Well, after the deals they just signed, I think that multiple is down to 39 times.

Brad Gerstner: That can change fast.

Gavin Baker: They added $29 billion in a single month.

Brad Gerstner: Exactly. Have you ever seen anything like that happen before?

Gavin Baker: Never. It just goes to show that Elon is not only a great engineer, but he, Gwynne Shotwell, and the rest of the team are incredible at business. They understand exactly what needs to be done to raise the capital required to reach the next phase. They have a long-term mission for the business.

To me, looking at what we’ve seen over the last few weeks—with Cursor and these new deals they’ve cut—I don’t know if any of the “Magnificent Seven” companies could have moved that quickly to adjust their business. It is exceptionally entrepreneurial at a massive scale, which is something we very rarely see. I would just add two other things.

Brad Gerstner: Can I give you a hug, Gavin?

40:19-42:43

Gavin Baker: There are two other points I’d like to make. First, people talk a lot about the total amount of capital being raised. If you add up the capital here—what Anthropic might raise, what OpenAI might raise, and what SpaceX is raising—let’s call it $250 billion. That is only 1% of the total market cap of the Magnificent Seven.

To me, that is a very reasonable bet on a future we all believe in. If you ask where we are out of consensus or what our “variant perception” is, it’s that we actually think this is going to be bigger and happen faster than people realize. We’ve felt that way for a couple of years now. So, first, it’s only 1% of the Mag 7 market cap. And then, as you referenced, the amount of...

Brad Gerstner: You referenced the amount of selling. I have a chart we’ll post here showing the staggered share release for SpaceX shareholders. There isn’t much that can be released until after the first earnings report. We saw a version of this with the Cerebras IPO. I think the banks have been thoughtful here, knowing this is a very large IPO. I’m not saying the stock won’t trade down—that’s always a possibility—but if you telescope out, is there any company better positioned as a bet on the future? Based on what they’ve shown over the last five weeks, they are probably number one. But let’s move on.

Gavin Baker: Can I just say one thing about the employees? I think another unprecedented factor here is that the employees, and to a large degree the investors, have had liquidity every six months.

Brad Gerstner: Exactly.

Gavin Baker: This has been the case for about the last ten years. So, if you’re a SpaceX employee or a former employee and you wanted to sell, you’ve had nearly 20 chances. It’s a matter of historical record that large investors have been able to sell as well.

Brad Gerstner: Great point.

Gavin Baker: The people holding today have chosen to own it. Now, there’s a new valuation, and we’ll see what they do, but this level of pre-IPO liquidity is utterly unprecedented.

Brad Gerstner: Yeah, that’s a great point. We’ve actually called these companies “quasi-public.” You and I both know that SpaceX—and I’d put Anthropic and Databricks in this category as well—has, in many ways, been more liquid over the past three years than some public biotech companies we know. You’re absolutely right. We often treat the distinction between private and public as a binary, but it’s really a continuum of liquidity. Let’s keep moving on to the models.

42:46-45:08

Brad Gerstner: Let’s keep moving on models. Anthropic launched Fable 5 yesterday, which you referenced. It is essentially Mythos, but with added classifiers and safeguards around cyber, biology, chemistry, and distillation. When those safeguards are triggered, it fails back to Opus 4.8.

Andrej Karpathy tweeted about this yesterday, noting that while it is state-of-the-art on all benchmarks, what really makes it special is its performance on long-running tasks. You also retweeted our friend Noam Brown, noting that ChatGPT 5.5 exhibited similar capabilities. This led Noam to suggest that snapshot benchmarks aren’t really relevant anymore. The X-axis now has to be time, tokens, or compute, because we can solve most problems today if we just let these frontier models run for an extended period.

So Gavin, what is this new class of model—Fable 5, ChatGPT 5.5—and what does it mean for the race toward superintelligence? Who’s up, who’s down, and who is still on the frontier? Give us your thoughts.

Gavin Baker: It’s hard to say that Anthropic isn’t “up” right now, especially after the revenue numbers they put up and the Fable 5 release—and Mythos is evidently even better. But I think that post from Noam Brown yesterday is just so profound.

Brad Gerstner: It is profound. Say more about that—why don’t we know how smart they are?

Gavin Baker: We don’t know because nobody has run Mythos continuously for a year. We may never truly know how smart each generation of models actually is or was, because we don’t have the time to appropriately evaluate their intelligence before the next model is released.

It’s a profound statement. Just imagine—I always use this analogy for FSD—imagine a human being who never gets distracted, never gets tired, never talks on the phone, and never loses focus.

Gavin Baker: It never talks on the phone in the car, never drinks and drives, never yells at its kids, and never has to reach into the back seat to give a baby a bottle. Naturally, you’d think that over time, that is superior to a distracted human. I don’t know—how long can you think deeply about a single topic, Brad?

Brad Gerstner: Well, maybe an hour.

45:10-46:34

Gavin Baker: An hour? Oh, man. That makes me feel terrible because I think I can only think deeply about one topic continuously for maybe five minutes before a stray thought enters my mind. Now, I can always come back to it, but still.

Andrew Fox: Imagine if Albert Einstein—who was clearly an exceptional intellect—had been able to think for three hours at a time instead of whatever his limit was. Now, imagine if Einstein could have thought about fundamental physics 24 hours a day.

Gavin Baker: Exactly. He doesn’t have to eat, sleep, or relax. He doesn’t drink, he never gets old, and he never suffers a diminution of intelligence. If he thought like that for a single year, we might have already solved many of these intractable problems.

Andrew Fox: It’s an extraordinary thought. My takeaway was that however bullish I was on compute before, I’m significantly more bullish now.

Brad Gerstner: Right. We saw that this was likely what really unlocked Opus 4.6. It was the first truly long-running model that could maintain context and memory to solve these more complex, enduring problems. For us, the signal came in January. We felt that was a big moment, but once you saw the revenue start to climb, it was clear that people were voting with their wallets.

46:35-48:43

Brad Gerstner: When people began voting independently, it was a profound moment because the models became significantly more useful. One of the big questions going into this year was whether AI revenue would actually show up. Would we reach the thresholds of intelligence that would cause enterprises and consumers to use them more?

The consensus at the time—at least on this podcast and in my debates with Bill Gurley—was that open-source models and cheap tokens were catching up to the frontier. There was a sense that perhaps these models were beginning to asymptote and that people wouldn’t really pay for premium tokens.

However, six months into the year, the evidence on the field suggests the exact opposite. Frontier tokens are capturing the vast majority of all revenue. In fact, if you believe in long-running capabilities and that more compute enables greater performance, the frontier models may actually be extending their lead over models built on distillation. I’ll open it up to the table: Have we challenged the thesis that cheap open-source tokens will always close the gap, or are the frontier models pulling away?

Clark Tang: I think this debate has existed since we first started training these models. The argument is always that open source is only three to six months behind the frontier. But empirically, you can see that almost all the revenue has occurred at the frontier. I think that’s because every time a new frontier model is released, it unlocks a whole new slew of use cases that we could never tackle before—like advanced coding. We’ve been locked at our desks for the last day just hammering Claude because it’s fascinating. There are things we can do now with Fable 5 that were simply impossible with Opus just a day ago.

Brad Gerstner: So, what are some of those things? I’m curious.

48:43-51:04

Clark Tang: I think it’s gotten really good at multi-agent orchestration now. Anthropic released a blog post about six different agent orchestration patterns they’ve discussed. Once you can manage all these agents, the harness and the model itself are being RL’d (Reinforcement Learned) with one another. They are being fused closer and closer together, allowing the model to understand the full extent of your work.

For instance, I just threw in seven of our models and said, “I want to create a master view of my beliefs given all these assumptions across these companies and TSMC’s capacity. Produce a report on all of this.” The model was able to reason through all of our assumptions and identify contradictions—like, “If you believe this, then this other thing must be true.” It was fascinating. We could never do that before.

I think we’re only at step one of multi-agent orchestration; we’re going to take this even further. As another example, I dumped all my notes from the last three years into it. It reasoned across them and pointed out which ideas were consistent and which sources provided the highest signal for what actually played out. It was just super fascinating to see. We’ve already blown through our usage limits.

Brad Gerstner: It’s unlocking so much. Anthropic gave examples yesterday in their release, like a 50-million-line Ruby codebase at Stripe that was refactored in a single day, whereas it would have previously taken many people many weeks. Think about the impact this will have on biology and life sciences across the entire spectrum. To me, it really gets back to a fundamental point. Number one, if you believe this to be—

Brad Gerstner: First, if you believe this to be true about long-running agents, then we are going to produce and consume more tokens in the future as far as the eye can see. This brings me back to Terawatt, orbital space, and all of that. We may, in fact, unlock real thresholds of intelligence, but we’re going to have to let these horses run for a long time to get there.

Gavin Baker: Yeah. I would just say that two things can be true at once.

Brad Gerstner: Mhm.

51:04-53:32

Gavin Baker: The majority of economic value may continue to accrue to the frontier—and man, has it ever accrued to the frontier thus far, especially in the first six months of this year—but the majority of tokens consumed globally may be open source.

Brad Gerstner: And they are today.

Gavin Baker: Yes, and I think that current state is likely to persist.

Andrew Fox: Harvey had a great blog post that they put out on X. It’s amazing how everything becomes outdated in five days, but they used their own proprietary legal data to perform reinforcement learning and supervised fine-tuning with Fireworks on an open-source model. Then they used a router—a tool that determines which model to send a specific query to and which model to use for verification—and they achieved better outcomes than Claude 3 Opus or GPT-4.

Gavin Baker: Either 4o or o1, and at a lower cost.

Andrew Fox: Exactly. I think that is the future. The reality is they were still consuming a lot of Opus, but the majority of the tokens they were processing were likely through their own open-source model.

Clark Tang: We hear the same thing. We conducted an enterprise survey of 300 companies that we’ll be posting soon. We looked at which ones were optimizing—folks who are implementing model routing and deciding to send certain tokens to specific places—and which ones aren’t.

Andrew Fox: We are looking at which companies are optimizing and which aren’t. More importantly, what is their expected consumption of frontier model tokens? Even though they are already in the process of optimizing, they all expect to consume significantly more. Think about it in the context of JPMorgan. If they are handling back-office tasks like customer service, they may very well use an open-source model.

I think they are loath to use Chinese open-source models, so they are waiting for US-based open-source models to deliver the performance they need. My hunch is that for these enterprises, a lot of that back-office work will be routed there, which will likely account for the majority of the tokens. However, for high-value tasks—coding, for example—they don’t want to write second-tier code. I believe the vast majority of that high-value work will continue to reside on the frontier.

Gavin Baker: Exactly. You don’t need Albert Einstein to book a trip, and you don’t need Albert Einstein to handle KYC (Know Your Customer) compliance.

Brad Gerstner: But this is the debate. We had this exact conversation at this table two years ago. When people looked at those use cases, they concluded that frontier models would not accrue most of the revenue. However, if you look at the revenue curves right now, frontier models are capturing 90% of it.

53:33-55:42

Gavin Baker: That conclusion has been decisively wrong. It’s probably more than 90%, and it may continue to be that way. The frontier might represent 90% of the economic value, even if open source accounts for 80% of the tokens.

Andrew Fox: There is something very important to note about open source. There is a belief that it’s bearish for AI or bearish for frontier models—that’s the bear case you mentioned.

Gavin Baker: It’s actually incredibly bullish for compute and hardware. If the frontier models are capturing less of the margin, then you’re going to spend more on compute. So, the better open source performs, the more compute is required.

Andrew Fox: The better open source performs, the better it is for compute providers. I’ve noticed a deep-seated difference in belief between Silicon Valley and Asia. If you spend a lot of time here in the West, the consensus is that everything will be closed-source cloud and all traffic will flow in that direction. However, in Asia, the overwhelming belief is that we will find the right model for the right workload to avoid overspending.

Andrew Fox: I think this coming year will be the most indicative of which way the market falls. Closed-source models have captured so much value because they actually understand user intent and carry out the work. This was the first year we saw agents transition from simply answering chatbot requests to actually producing useful work.

Andrew Fox: Intelligence has scaled so rapidly that we continue to push against the most economically valuable tasks, such as coding, finance, and other knowledge-work sectors. But for the long tail of tasks, if open source continues to maintain only a six-month lag, we might see it used for many of our everyday requirements.

55:42-58:04

Brad Gerstner: That is essentially Jensen Huang’s argument, right? His view is that we will eventually have model routing. We are just at a moment in time where frontier models have the advantage and can handle long-running tasks. Open-source models couldn’t do that very well, so the frontier models are accruing all the value. But as soon as open-source models can handle those long-running tasks—which isn’t far off—they will capture a significant portion of that revenue as well. Are you investors in Reflection?

Andrew Fox: I am not.

Brad Gerstner: Okay, neither are we. But I...

Brad Gerstner: I’m not either, nor are we, but I am very impressed by Misha and the team and what they’re doing. I very much want a frontier open-source US lab to win. I heard you say recently—and I believe it to be true—that Nvidia could do this any day they really wanted to. They already have some great open-source models, and they could absolutely build a frontier open-source model whenever they chose to. So, it’s not a question in my mind as to whether the US will have a frontier open-source model; it’s just a question of timing. At that point, let’s assume they achieve these long-reasoning capabilities—will the frontier labs have achieved something else by then that allows them to maintain their stranglehold on the revenue?

Gavin Baker: Yeah. And I just think it’s a case of, “Wow, that’s a cute ASIC you’ve built there. That is so cute. How would you like open source to join the frontier? How do you like them apples?” I’m not sure that’s the explicit calculation, but I do think Jensen—

Brad Gerstner: Just to double-click on that for everyone at home: if they were to put a frontier open-source model out there, how does that impact the ASIC landscape?

Gavin Baker: Well, you might not have the revenue or the margins to fund those ASICs anymore. I think Nvidia is highly likely to become the world’s dominant provider of open-source AI. I believe Jensen will bring open source—which is currently maybe six months behind the frontier—and we might see it creep closer and closer.

Brad Gerstner: And I think Jensen has a major business decision to make. I see the chart here, so let’s discuss Nvidia. As you say, if all of his customers are going to compete with him—

Andrew Fox: Yes. Then why not compete with his customers? We already have all these “Neoclouds.”

58:04-01:00:00

Gavin Baker: So, that is a cloud computing business that can compete with all the existing cloud computing businesses. Jensen has his own models that are actually very good. Nemotron 3 or 3.1 was really cool from a compute efficiency perspective. He is always careful to release small models so as not to tread on the toes of Anthropic, OpenAI, or Google, but I do think that is a conscious choice he is making.

Andrew Fox: And if the economics change, I think Nvidia could join the frontier and become one of the world’s largest cloud computing companies much faster than people think.

Brad Gerstner: Interesting. Clark, walk us through this chart.

Clark Tang: Yeah, so I think one of the takeaways from spending time in Taiwan was that there is certainly a lot of excitement around the next wave of ASICs. It used to be an argument of Nvidia versus ASICs—one or the other, total domination by one side. Now, I think it’s becoming clearer every year. Everyone assumed Nvidia was going to lose share dramatically on a revenue scale, a gigawatt scale, and a unit scale. But if you look at the last few years, they have actually maintained their share very handsomely. In fact, if you account for the fact that Anthropic wasn’t really using Nvidia, they probably actually gained share heading into ‘25 and ‘26. What was very interesting, though, was a new class of accelerators or ASICs. MediaTek, with their new V8T versus Broadcom’s V8i for TPUs, was a big topic of discussion. I think the argument for ASICs now is that more and more of them will be custom-tailored to the actual workload.

1:00:01-01:02:30

Clark Tang: We are seeing a shift toward hardware that is custom-tailored to specific workloads. That is one vector people are moving in, whereas NVIDIA has established itself as the predominant provider of compute for much of the world. For internal workloads, companies may move further down the stack toward custom solutions. I remember just a year ago, it felt like a binary battle between Broadcom and NVIDIA. Now, there is much more nuance regarding which accelerators fit specific workloads, customers, and business models. I thought that was a significant new realization.

Brad Gerstner: Actually, I think we’ve all shared that view for a while.

Andrew Fox: I was honestly shocked. I’m out here for a board meeting with one of our companies, and the one thing they emphasized was that they expected the world to be consuming less NVIDIA by now. If anything, NVIDIA’s lead is accelerating because they simply continue to out-execute their competitors.

Gavin Baker: I think a lot of people are indexing to this OpenAI gigawatt-scale project. You have to remember, NVIDIA has ten of those.

Brad Gerstner: Broadcom has ten as well.

Gavin Baker: And who has six?

Andrew Fox: AMD has six, and they have warrants. And then there is Cerebras, our shared portfolio company.

Gavin Baker: One portfolio company has a gigawatt.

Andrew Fox: They have a gigawatt?

Gavin Baker: Well, that is what is on paper. Let’s see what actually gets deployed.

Brad Gerstner: I would be very surprised if—what’s the math on 10 out of 27? Let’s see who is best at math here. What percentage market share is that?

Gavin Baker: It’s about 30%.

Brad Gerstner: Yeah.

Gavin Baker: I’ll be very surprised if that is where they land. I think that is an extremely unlikely outcome, especially as long as we are in a watt-constrained world. If you can get more tokens per watt—which literally translates to revenue with Nvidia—than you can with alternatives, it changes the math. If you build your factory with another chip, you might save some money upfront, but you’re going to have less revenue and your margins may be lower. That is a point Jensen Huang keeps hammering, and I think it’s really important.

Gavin Baker: By the way, credit where credit is due. One of the most surprising things to me in this ASIC landscape is that Meta and Microsoft have probably been disappointing.

Brad Gerstner: Yes.

Gavin Baker: Do you know who actually made a good ASIC?

Brad Gerstner: Yes. Well, I know you know.

Andrew Fox: Jalapeño.

Gavin Baker: Exactly. From OpenAI. They made a great chip.

Brad Gerstner: Yes.

01:02:29-01:05:08

Gavin Baker: Yes. Now, unfortunately, it needs to run at a much lower temperature than the Nvidia GPUs, which means you have to spend more money on cooling, and that consumes more power. They made a great chip, but I think the real question for them—and for everyone—is whether that is the highest and best use of their time.

Brad Gerstner: I tend to think that while there is a belief that frontier companies need to be vertically integrated, if you believe—as I do—that the race to superintelligence may be over in the next two to three years (especially as these recursive loops start working), then the priority has to be focus, focus, focus.

Brad Gerstner: Your purpose is to build and deliver the best intelligence in the world. To do that, you need to capture all the revenue possible, because building out the compute required to continue pushing the frontier requires massive revenue to support it. So, subject to that question of focus, I think they’ve certainly made progress.

Brad Gerstner: This all brings me back to a reality check, though. We just finished talking about test-time compute, inference-time compute, and long-running agents. These are the things that have truly unlocked revenue this year, but they all push us toward higher CapEx. Google just raised their guidance to $80 billion, right? We’ve now seen the free cash flow of the Mag Seven move down dramatically.

Brad Gerstner: Free cash flow is down dramatically—about 80% from just a few years ago. Looking at this Morgan Stanley chart, they’ve upped their 2027 CapEx forecast from $950 billion to $1.1 trillion. We were discussing this with Jensen Huang, and that was his forecast two years ago. Obviously, this doesn’t even include companies like SpaceX or CoreWeave. I suspect the actual number for 2027 is likely closer to $1.5 trillion.

If we compare this to total incremental inference revenue—which is what the market is worried about—it brings us back to the conversation I had on the Sam Altman podcast last October. Can we really afford to spend $1.5 trillion in CapEx per year if we’re only generating a certain amount in inference revenue?

I think what lit the fuse this year was Anthropic showing up with significant revenue. If you combine the revenue of all the AI labs, it’s projected to be around $300 billion by 2027. So, we’re looking at spending $1.5 trillion in CapEx for $300 billion in inference revenue. Does that math work for you? What would make you nervous about our ability to sustain these investments? Because the moment the market gets nervous, the entire semiconductor complex is going to take a massive hit.

Gavin Baker: Well, what do you think the gross—

1:05:09-1:06:56

Brad Gerstner: What do you think the gross margins are on that $300 billion?

Gavin Baker: Let’s call it 50%.

Andrew Fox: I would guess they’re probably a little bit higher than that—maybe 60% or 70%. But even then, the math starts to make sense.

Brad Gerstner: What I would say is that I think $300 billion is actually a low estimate. I really think it’s low.

Gavin Baker: From your mouth to God’s ears. I think we’ll end this year well over $200 billion in inference revenue—well over. So, the math really does work out.

Andrew Fox: We have to give our friend Jensen Huang some credit here. He said some things that seemed outlandish a couple of years ago, right? He predicted a trillion, and it turns out he was actually being conservative. He was low. We should give the guy some credit and really consider what he’s saying right now.

Brad Gerstner: For sure. And listen, I would say consistently, Elon Musk has been taking the “over.” Sundar Pichai has been taking the “over.” Sam Altman and Dario Amodei too. You know, Dario did that podcast with Dwarkesh Patel where he talked about having “country-sized geniuses” in a data center. He said that will be here by 2028. He predicted revenues would reach the low hundreds of billions by 2028—let’s call it $300 to $400 billion. He said that a while ago, so he may even be revising those numbers upward now.

Brad Gerstner: He may even be revising his numbers upward. He mentioned that it’s hard to imagine we won’t see trillions of dollars in revenue before 2030. If we are on that trajectory—hitting $200 billion by the end of this year, maybe $400 or $500 billion next year, and on a path to over a trillion by 2029—then the math actually works out.

Clark Tang: We also have to keep in mind that a significant portion of that spending is for training. It’s maybe a little less than half? What do you think, Foxy?

Andrew Fox: It probably depends on the specific lab, but I would say it is increasingly less than half.

1:06:56-01:09:18

Clark Tang: Right. So let’s say 35% of the spending is not yet revenue-generating because it’s being used to develop the next model. I think the math holds up, and we are still in a prisoner’s dilemma where opting out could be an existential mistake.

Coming into this year, several narratives were challenged. Everyone expected token pricing and the cost of compute to be purely deflationary—a smooth downward line over time. However, what we’ve seen this year is the opposite. It all comes back to supply and demand; the demand side of the equation is far outstripping supply. If you look at the deals signed by SpaceX and others, the monetization rates per watt are actually increasing.

Andrew Fox: Look, that is on a pretty nascent, small base of users. Alex at Whale Rock has a great way to frame it: less than 0.2% of people on Earth are actually using AI in an agentic way.

Brad Gerstner: Right.

Andrew Fox: Exactly. I’m not a technical person, but I’m consuming 500 CPU cores in a VM instance and five GPUs 24/7.

Brad Gerstner: If you extrapolate that to any meaningful percentage of the population, we’re going to be in a shortage environment for quite some time.

Andrew Fox: I think that is all positive for the ROI question.

Brad Gerstner: Man, Foxy, a 100-to-1 CPU-to-GPU ratio? That’s quite a workflow. Of course, it’s fine.

Andrew Fox: I’m being smart with my spend.

Brad Gerstner: Good, good. Excellent.

Clark Tang: I will also say that regarding the ratio of $300 billion in revenue to $1.2 or $1.5 trillion in CapEx, there is a physical limit to how quickly we can expand production and increase that spend. Conversely, we are seeing the opposite regarding the willingness to pay for these tokens. The monetization per gigawatt is actually increasing—from maybe $20 billion in the best-case scenario at the beginning of the year to $30 billion or even pushing $40 billion per gigawatt now.

1:09:18-01:11:41

Clark Tang: It’s all about the operating profit per gigawatt. While it’s a very heavy fixed-cost base, everything coming in now is essentially pure margin flow-through. As we scale, the willingness to pay for all of this is increasing. This is, of course, subject to the dynamics we discussed regarding open source versus proprietary models. But as we climb this curve, revenue might actually outstrip our fixed-cost base by a significant amount. I think that’s why all the labs are stepping on the gas; they realize that if we stay on this trajectory, we are going to be incredibly short on compute within the next three years.

Brad Gerstner: That’s a great point. If you look back at when these investment decisions were being made—

Gavin Baker: In November of 2022.

Brad Gerstner: Exactly. You expected a certain return back then, but you might be seeing triple that return today.

Gavin Baker: There’s no way. There is no way they thought they would be anywhere close to breaking even at this stage of the curve. The reason I called it “accidental”—

Brad Gerstner: I’ve called it “accidental profitability.” People have been talking about it because these companies actually want to spend a lot more money on compute, but they’ve just had a hard time doing it. Now, perhaps with SpaceX, they can take some of those dollars and deploy them elsewhere. To me, that represents a fundamental change.

Brad Gerstner: The first argument against frontier labs was that they would never generate revenue. We saw that get blown up. Then the argument shifted: even if they generate revenue, the gross margins will be terrible and they’ll never make money. That argument has also been debunked.

Brad Gerstner: Now, people are falling back on the claim that they are overcharging—calling it “token maxing.” My good friend Chamath says there’s no ROI on any of this spend and that it’s all just token maxing. Of course, when anyone spends this much, it’s not always optimal. At Altimeter, we aren’t optimally spending every single dollar either.

Brad Gerstner: But the real question is: why are millions of independent businesses—small, medium, and large—and millions of consumers all choosing to do the same thing? These are rational economic actors. They aren’t dumb. They are all simultaneously saying, “I want to do this because it makes my life better and my business better.” To me, that is the best evidence that this revenue growth is sustainable.

1:11:40-01:14:07

Andrew Fox: Yeah, and Clark, I think the point you made is dead on. You want to own asset-heavy businesses in inflationary environments, especially when token pricing is going up and the supply-demand balance is tightening. I totally agree.

Brad Gerstner: As we begin to find our way to the exit ramp and wrap things up here—Gavin, you and I have been doing this for a long time, a couple of decades now. You might have even been at it longer than me, even though I’m a little bit older than you. I always like to do a market check because I find that a lot of analysts come on these shows and just talk their book. There are a lot of retail investors and others listening who just want to know what we really think. I always characterize my positioning as small, medium, or large. Am I carrying small exposure, medium exposure, or large exposure?

Brad Gerstner: If you look at what’s happened in the markets, semiconductors have absolutely ripped this year. I’ve been doing this a long time, and I’ve never seen anything like it. I’ve never seen doubles and triples across the board like we’ve seen recently. However, there’s been huge dispersion in the market. Internet is down 16%, and software is down 8% on the year. The SPY and the NASDAQ are up, but they are really only up because of the specific components related to AI and compute.

Brad Gerstner: The market itself has struggled with anything not related to AI and compute. Meanwhile, if you were in the sectors we’ve been invested in, we’ve all done pretty well. I’ve said it a couple of times: if Anthropic’s revenue hadn’t shown up this year—which was the major overhang on the market—I think the entire market could have been down. But that revenue did show up, and we saw huge months in April and May.

For us, because prices rose so much, I have some concerns about geopolitics and the short-term macro backdrop regarding inflation. I think we need a little consolidation in this market to answer some of these questions, especially now that expectations are higher. Consequently, we dialed back from what I would call a “large” position for Altimeter to something more “medium-small.” Again, it’s never all-or-nothing for us; it’s always about the risk-reward at a given price. We think this might be a period of consolidation on the way to much higher highs. I’m curious how you run your book and how you think about this as a portfolio manager.

Andrew Fox: Very similarly, man. I always imagine stocks and the markets as runners.

01:14:07-01:16:30

Brad Gerstner: And in 2022, that runner was going downhill. It had a lot of energy, but it was painful and not very fun. However, coming out of that, there was a lot of pent-up upside in the market.

Gavin Baker: There is a lot of pent-up upside in the market, but the market—particularly over the last two months—has run up a very steep hill. Ironically, many semiconductor companies, specifically Nvidia and Broadcom, have actually been laggards.

Andrew Fox: I see a lot of talk on X about trying to find “the next bottleneck.” I think that was the last game, and that game is over. You’ve had a lot of stocks that didn’t just climb a mountain or a hill—they went straight up a cliff.

Gavin Baker: Exactly. They’re tired and they need to rest. We’ll see if they just hang out in their harnesses at the top of the cliff they just climbed, or if they need to head back downhill for a bit. We’ve already seen some retracement over the last week.

Brad Gerstner: I’m thinking very similarly to you. The market is seasonal, but I think there are also very real concerns regarding inflation and interest rates. What was the CPI reading this morning?

Clark Tang: It was 4.2%. I believe core CPI came in at 0.2% versus the 0.3% estimate. So, a...

Brad Gerstner: The numbers came in at about 0.2 versus 0.3, so they were a little bit better. But clearly, we are above 4% again.

Gavin Baker: There is short-term pressure on core PCE and other metrics. We have some “unknown unknowns,” but look at the market. If I had told you the fact pattern for this year—that we’d be in a war with Iran, oil would be at $100, CPI would be creeping back up, internet stocks would be down 15%, and software would be down 8%—you would have said, “I want nothing to do with that market.” Yet here we are. The market has done quite well in the sectors we traffic in because the world underestimated both AI revenues and the sheer amount of compute that would be required.

Andrew Fox: I would just add that we’re heading into a seasonally weak period amidst all these fears. Interestingly, AI has actually shown seasonality over the last three summers.

Brad Gerstner: Yeah, that is interesting.

Andrew Fox: Token consumption has plateaued and slowed down a bit. That’s because college students are major AI consumers, and they don’t use it as much during the summer. Hopefully, they’re using it to learn rather than cheat. But that slowdown might not happen this time because the fifteen-year-olds...

Brad Gerstner: ...are building swarms of agents and SpaceX models. My son is coming to the SpaceX IPO with me at the exchange on Friday, but he had to build an AI model using AI first.

01:16:33-1:18:46

Brad Gerstner: He had to build an AI model using AI agents. Specifically, he had to build a DCF model before we went to the exchange. He is absolutely mesmerized by it, and what he’s building is truly extraordinary.

Gavin Baker: So, he’s one kid who is definitely not using less compute.

Brad Gerstner: He’s burning through it. He’s absolutely burning it.

Andrew Fox: Yeah. But you know, if token consumption plateaus or if open-source takes some share—there’s a silicon data index that tracks consumption and pricing. I think there may have been a slight shift over the last two weeks toward cheaper open-source tokens. People looking at that data might see it as bearish, or perhaps they just don’t fully understand it.

Nonetheless, I think there are reasons to look around, be careful, and be thoughtful. I always assume a bullet is coming for me, so I keep my head on a swivel. It’s the bullet you don’t see that gets you, so I try to spin as fast as I can. The market may need to take a breather, but man, when I think about what Noam Brown said...

Gavin Baker: I know.

Andrew Fox: ...and when I see the capabilities of Fable, it’s just hard for me to get too bearish.

Brad Gerstner: I mean, to me—and we have two of the most extraordinary guys of the next generation sitting in this room—we at Altimeter have deep admiration for the work you guys do. I always appreciate it when you send me a note about...

Brad Gerstner: I appreciate it when you send me notes about the work we publish. For those who are newer to this business, they might think this is how it has always been. But the steepening of the line of creative destruction and the acceleration of scale advantages—while I always believed it would happen, I never imagined it would occur at this rate. I looked back last night: in the last seven years, we’ve added $1 trillion in revenue to the “Magnificent Seven.” To put that in perspective, it took over 20 years for them to reach their first trillion.

Brad Gerstner: In just the last seven years, we added another trillion in revenue, which created $17 trillion in market cap. Now, the forecast suggests we are going to add another trillion dollars of revenue from just three companies—SpaceX, Anthropic, and OpenAI—over the next four to five years.

1:18:47-01:20:22

Brad Gerstner: Think about that: not seven companies, but three, and in half the time. We are certainly going to hit some bumps in the road—I know it won’t be a straight line—but we are headed toward higher highs because of the sheer size of the prize. This technology is going to transform 5%, 10%, or even 15% of global GDP. There is no doubt in my mind. Ten percent of global GDP is $10 trillion. It’s an exciting future to be a part of, and it’s fun to navigate it with you guys. We have to do the work to ensure America wins, evolve the social contract, and keep moving forward.

Brad Gerstner: We need to evolve the social contract to ensure we lift the floor and take everyone with us on this ride. It is a incredibly exciting time to be doing what we’re doing, and it’s a privilege to do it alongside you guys.

Gavin Baker: Yeah, I just want to say, Brad, thanks for having us. And thank you for what you’ve done with the Invest America accounts. I actually think it’s super important for America and the world to give people an equity stake at a very young age so they can see it compound over their lifetimes.

Andrew Fox: This is a great thing you’ve done for the world, so thank you. I’d echo all those comments—I have deep admiration for you and your team, and I’m grateful for the collegiality and friendship between our firms. I know Clark and Foxy hang out all the time.

Brad Gerstner: People often think that in our business, you shouldn’t share anything. Our view is that we “open source” it. However, there are very few people we actually call to ask for an opinion, because very few people put in the thousands of hours of work that you do to truly add value. We appreciate that, Gavin, and we appreciate you as well. So, with that love fest concluded, let’s call it a wrap. Thanks for being here.

Gavin Baker: Thank you.

Made with: The Transcript Desk Chrome Extension