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

And I think we're all pretty AI-pilled. If you're AI-pilled, that means we've got to build a lot more compute than the world thinks and that these models are going to be a lot more valuable than people think. You combine that with their core business, and I don't know another entrepreneur or another business that's a better bet on the future than SpaceX. I think for most institutional investors, it's a must-buy, must-own, set-it-and-forget-it in order to have a real bet on both the space and AI future.

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 Clark Tang into the mix, my partner, to talk about some of the big questions of the day.

How should we be thinking about the SpaceX IPO? What are the big levers? There are big numbers out there for what's going to happen over the course of the next few years, so let's break that down a bit and help simplify it for folks. Mythos launched yesterday, and I want to talk a little bit about who's up and who's down in the race for superintelligence. Where are we, and what did we learn with the Mythos launch?

Clark was in Taiwan last week with Jensen at Computex and GTC. What was our takeaway there? What's going on with GPUs and memory? Where are the bottlenecks, and where do we go from here? To start everything off, maybe just kick it over to you, Gavin, to talk about the SpaceX IPO.

The IPO is in 2 days. You're a big shareholder, so congratulations. We're also a shareholder, and we expect to be buying in the IPO. The Wall Street Journal is reporting, and Goldman Sachs is saying, $160 billion in revenue in 2028. We know that the IPO is $135 a share, or $1.77 trillion.

When we think about what the big levers are, there are so many moving parts in this IPO. Nobody's better than you at breaking it down and simplifying it. What are the key levers that we ought to be thinking about, and that you're thinking about, over the course of the next few years?

Sure. Great to be here. Thank you for having me. I thought we were going to call it BGGB, but we can stick with BG2. I've been to your house.

I think there are 2 big levers or variables that people should focus on. I'm not going to comment on where I think those variables go, but one is—you guys have this chart. Did you post this on X?

I did. I did. We also included a new addition with xAI's new deals as well.

1. xAI's Google & Anthropic Deals: Highest Operating Profit Per Gigawatt

Yeah. Clark, who 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. xAI's deal also generates probably more operating profit than anyone but Anthropic.

Your colleague at Altimeter, Freida, also calculated a 55% IRR on Claude's one.

Mhm.

If you can borrow money at 6%, 7%, or 8% and invest in something with a 55% IRR, I'm not the most sophisticated thinker, but the math maths.

Right.

I think the most important variable—one of the 2 most important—is how quickly they bring on terrestrial data centers.

Mhm.

We do know from Jensen that Elon brings data centers up faster than anyone: 122 days. Speed is literally cost because every day you're paying electricians and plumbers. That's cost. And they're now monetizing them at arguably the highest rate.

Everybody should run their own math on that, but that is a massive variable. Truly massive.

The second thing is, we have a chart, and it's wildly out of date now. It's kind of freaking amazing. Is this chart from 10 days ago? In the 10 or 12 days since we made this chart, which shows the Pareto curves for Opus 4.7, for coding, for Codex from OpenAI, we've had Opus 4.8. It was already out of date. And now we have Fable 5 and Mythos, which is freaking wild. In 10 days, we would have had to update the chart twice.

Totally.

What the Pareto curve shows is how much intelligence you can get for a given amount of cost. I do think that all revenue will accrue to the Pareto curve—at least, all frontier-model revenue will accrue to the Pareto curve. This is the Pareto curve for coding.

What's so impressive is that you can see in the chart that Composer 2 was Pareto-dominant at the lowest level of intelligence, with very little training. This just reflects—and I know you know Cursor well. I think you know Cursor a lot better than I do, by a vast amount.

My understanding is that Cursor and Anthropic have more tokens of proprietary coding data than anyone else. They have more tokens of proprietary coding data than exist on the public internet. So Cursor used Kimi K2.5, used its own private data, did some RL and some supervised fine-tuning, and got a really good model.

Then they spent 3 weeks in the Colossus 2 cluster, and they got a model that 12 days ago was Pareto-dominant with Composer 2.5. That's on their own benchmark, CursorBench, so maybe take it with a grain of salt. But I think this suggests that Cursor's data is very valuable for coding, and when it is trained to Chinchilla-optimal or beyond Chinchilla-optimal with reinforcement learning, I think it suggests that xAI and SpaceX AI have a shot at being real players in coding.

I mean, I think one of the interesting things is the way he answered the question. We didn't talk about launch. We didn't talk about Starlink or communications. Those, up until really 6 months ago, were the business. Then we merged in xAI and we merged in Cursor, and then we announced these deals where it was very clear he was kind of building AWS right under our nose in terms of this.

What I want to do is go to Fox. Give us the breakdown. We've got 3 big lines of business: the communications, Starlink, and launch business; the AI compute business; and then I want to come back to xAI, which you were just clicking on.

If we just go to the core business, what do we have to assume goes right in the core business, both with launch and with Starlink, in order to achieve the numbers that are out there?

Yeah, sure. So, look, I think the thing that's foundational to everything is the launch business. This is the kind of crown jewel of SpaceX. It's something that no one else really has, most notably reusability, and soon rapid reusability.

This is, I think, what you need to believe in to get to the economics in AI that make orbital compute something that's very economically attractive, alongside the idea that we're in a shortage of power and a shortage of chips. So I think rapid reusability is the main thing that we're watching for, and I think most people should watch for.

Elon talks about it a lot, but getting these rockets to fly at a cadence that's comparable to an airline is the goal. Gavin has used this analogy before, but the old rocket industry was kind of like imagining boarding a plane, flying to California, getting off the plane, and then the plane explodes.

What SpaceX is ultimately trying to achieve is to have a Starship fly both stages, not just the booster, 30, 40, or 50 times before you have to retrofit that ship. When you do that, you're amortizing the cost of the vehicle over many flights, and that's what brings the cost down significantly.

But that's a really hard problem to solve.

Extremely difficult. And look, I think the company has been loud and clear: they're going to attempt to bring back the second stage of Starship later this year and then make it reusable—re-fly the second stage next year. From there, ramp up the cadence.

But at the end of the day, driving down the cost of launch is what enables all of these other businesses and is what makes them so attractive relative to incumbents.

How many launches do you think the consensus out there is assuming 2 or 3 years from now? Starship 3 just launched, you know. Are we launching one of these every day, every week, or every month? Where are we in terms of expectations?

Yeah, so look, I think expectations for now are that we're going from, call it, 160 or 165 launches last year up into the high hundreds of launches in several years, and getting into the thousands of launches probably in the next 3 years thereafter. I think the company has aspirations.

Thousands of launches? You're doing 2 or 3 launches a day.

Right.

Right. And then talk to us a little bit: What does this enable? Obviously, I'm here in Silicon Valley. I can't even keep a call on Sand Hill Road, 2 decades into the mobile revolution.

I mean, it’s the craziest thing. It’s like a third world.

Bill Gurley

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

It’s crazy. It’s crazy, right by the Starwood dead zone. I’m like, 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.

Okay, 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, just again at a high level. It’s a big portion of the revenue growth expected in the business over the course of the next 2–3 years. My hunch is a lot of this is driven by direct-to-cell connectivity. Walk me through those economics.

Yeah, so look, it’s actually interesting. The broadband business is still very early stage when you think about the percentage of households that have actually been penetrated to date. You look at the percentage of global households with Starlink, and it’s less than 1%.

That’s the broadband business. You kind of have a base terminal at your house, on your car, on your boat, and now on airlines as well. I actually think broadband can scale to hundreds of millions of terminals and hundreds of millions of users. Today, the subscriber base—

Hundreds of millions if they get rapid reusability of Starship, which is really hard. If there’s no competition, hundreds of millions is possible. But maybe—

I always say around here, it’s funny: I love seeing PMs and analysts in this situation. It’s exactly what I do with Clark. Clark will say something, and I’ll say, “The future is a distribution of unknown probabilities. It’s either more likely or less likely, so give me the distribution. Are we talking 20%, 30%?” It’s hilarious. It’s the same—

Well, no, 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 think that they’re going to succeed with rapid reusability, but I just think it’s important to acknowledge that orbital compute, Starlink V3, and Starlink direct-to-cell all require first-stage reusability for Starship V3. Then rapid reusability unlocks a lot of this.

Right. When I see the models that the banks are putting out there—and The Wall Street Journal and everybody else has reported on these—these same things have been widely leaked. They largely have the revenue from connectivity, so let’s call it Starlink direct-to-cell, going from $10 billion to $50 billion by 2028.

I’m not asking you guys to react or tell me your specific numbers. 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 course of the next 3 years? Is there enough TAM, both in terms of broadband and direct-to-cell? I think the answer to that is yes.

Yeah. Here’s what I’d say very simply: I travel with Starlink. I’m a big video gamer, and very consistently, wherever I am in the world, Starlink is the best connection.

Yes.

It’s the fastest, it has the lowest latency, and I do think once they get to rapid reusability, they’re also going to have the cheapest cost per gigabyte or megabyte delivered. Better, faster, cheaper has been a winning formula.

So, $50 billion is 0.3% penetration of the global telecom market. Maybe there’s some deflation with Starlink pricing, but that’s the way I’d frame it up.

2. "Elon Web Services" — Nobody Had AI Compute in the SpaceX Model

Yeah, I like betting on better, faster, cheaper.

What they achieved is singular. It’s never been done before. Just to put it in perspective, 100,000 GPUs is easily the fastest supercomputer on the planet as 1 cluster. A supercomputer that you would build would normally take 3 years to plan.

Right.

Then they deliver the equipment, and it takes 1 year to get it all working. We’re talking about 19 days.

Wow.

N of 1 is right. Elon is an N of 1.

And his ability to secure supply, stand up the supply, and deploy it in a way that’s coherent and effective for both himself and, I guess, now for others. Walk us through that. It looks to me, again, like this is a major component of the revenue story.

Totally. We were all at the Macrohard data center, and it was just very evident how much engineering had gone into building these sites. People always talk about Google and its ability to build a TPU and sell the TPU to Anthropic to generate revenues for AI. I think it’s a pretty similar dynamic here, with Elon able to secure power, build these sites faster than anyone else, and now monetize them in the massive AI market that’s ahead of us.

If you look at the relationships that he’s forged with a lot of his suppliers—be it Jensen Huang, or all of these different sites that actually want xAI as a tenant—his ability to finance these deals at very attractive financing rates relative to a lot of the other players in the space, these are advantages that compound over time.

When you’ve built the credibility to stand up these sites and monetize at these levels, it’s actually a very attractive proposition for a lot of folks involved. If you look at these deals in particular, Gavin, you pointed out that they’re actually monetizing perhaps better than other players in the space by selling this infrastructure—

A lot higher.

Google is obviously paying SpaceX a huge premium for this compute. Fox, you said something that I thought was really important: It may very well be that in order to get first in line for space compute, which Google certainly wants to do, they’re willing to pay a premium for their terrestrial compute. To me, that’s how you square the circle as to why there’s a premium. Any thoughts?

Yeah, look, I think there’s some of that embedded there. At the end of the day, SpaceX can stand up compute quickly, stand it up coherently, and stand up a lot of it in 1 place and have it readily available. I think that’s most of the premium, but outside of that, certainly people are going to space over time.

I have to pay a little call option to get first in line for space.

There you go. Good one.

We’ve all been investing in the neocloud space. There’s a fundamental belief around this table that we lack the compute needed to continue to push the frontier on intelligence, so we have to build a lot of compute.

There’s competition going on. On 1 end, you have the hyperscalers, who are building out that capability. Then we have AI-dedicated clouds that are building out that capability. Now, literally in a matter of weeks, we have a giant that’s emerged in this category, which is SpaceX.

The question to you, Gavin, is: Can they consolidate this market? If I think about it as a marketplace, Elon has a unique ability to get the supply. He has a unique ability to cut deals on the other side, and nobody can stand it up like he can stand it up. I think there might be real consolidation in the AI compute market, where you have the hyperscalers on the 1 hand and, on the other hand, he may emerge as the largest, strongest player in the AI compute market.

Yeah, so I think they’re the number 4 or number 5 hyperscaler today after the Google deal. It will be number 4.

Kind of wild.

Yeah. In 30 days, we went from not being an AI hyperscaler to being number 4. We passed a lot of companies, including Oracle.

CoreWeave is a huge business. We’re investors in it and have been investors in it. But there are a lot of other players: the Nebiuses of the world, the IRENs of the world. I would say there are probably 50 neocloud labs being funded in Silicon Valley right now as we speak because of the shortage in compute.

3. Data Centers Are Not Commodities: First-Principles Design

Absolutely. That’s kind of crazy in 30 days. That’s just extraordinary. What I would say is that there’s a belief that these data centers are commodities.

I do not share that belief. I don’t think anybody around this table shares that belief. In the same way that Elon was able to reengineer a rocket from first principles and make it reusable, he engineered an electric car from first principles. Everyone else was trying to make an electric car like an internal-combustion-engine car, and he thought about it differently.

I think he looked at data center design from first principles and designed something fundamentally different. I did actually ask the team—

I said, “Hey, guys, maybe I’d be a little less public about things that are very obvious to you [laughter]—about how to design a data center, but are revelations to other people, because I think what you’re doing is maybe more differentiated than you perhaps realize, because what you’re doing is so logical to you, but maybe not logical to everyone else. And that’s how he was able to do it in 122 days.”

Right. Yeah, to that point, Brad, yesterday we were meeting one of our portfolio companies, and we were talking about behind-the-meter. We were really thinking about it. There are only maybe 2 or 3 players now that can actually reliably engineer behind-the-meter data centers, and there’s real engineering work that goes into all of this.

So if you think about this, if you’re Vernova and you say, “We only have a certain number of gas turbines. Now, we can sell them to xAI, or we can sell them to one of these startup neoclouds. Who are you going to sell them to?”

Well, there’s another dynamic: everyone starts making more money when the GPUs get energized and sold faster. So literally, speed is money for all of the suppliers—power, land, turbines. So I think we’ll see.

Bill Gurley

Right. Hey, Brad, man.

But this is just—we’re just talking terrestrial. I do want to hit on that, and then you can flip it back on me. Talk to me, okay? So let’s assume that they continue to build out the terrestrial landscape. They continue to find buyers for that. Walk us through what this unlocks and how this is related to space data centers, because I think once you start talking terawatt capacity and beyond, we’re talking 1,000 gigs, right?

And this year, what are we doing—25 or 30 gigs, just to put it all in perspective?

20, yeah.

Right?

20, 25 gigs.

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? And then there’s obviously this debate in the world. I heard Jeff Bezos say, “I think it’s more like 6 years,” but Elon’s going to say 3, because if he says 6, then it will take even longer. So say 3, and we may get it in 4 or 5.

But are space data centers integral and essential to the IPO? And what do you think the timeline is, Andrew?

[Speaker?]

So I think if you think about those variables around what Crusher could mean for xAI, we do have an existence proof that once you really get on that Pareto frontier, revenue can scale rapidly, and it’s called Anthropic. There also seems to be a lot of demand for coding. And I do think John Massad posted something very interesting.

The founder of Replit.

The founder of Replit. He called it “Bitter Lesson-adjacent”: that coding may be the fastest path to AGI and ASI, because if you really go to coding, you can write code—if a model’s good at coding—to do anything. So I think that’s a profound point, and I think coding is going to continue to be very important.

So I think if you think about that variable, if you think about Starlink direct-to-cell enabled by Starlink V3, and you think about how quickly they can or cannot bring on terrestrial compute, I think orbital compute is necessary for the IPO valuation, but it’s certainly important, and it’s—

Well, maybe another way to say it is you may think we’re going to get to ASI faster than we’re going to get to orbital compute. That may take us from 300 IQ to 400 IQ, 500 IQ, and beyond, and the ability to scale it up to consume 10% of global GDP. But maybe that’s where we should move next.

Bill Gurley

No, no, I think on orbital compute, I think Foxy would be great, or Clark, to lay out the math from first principles. Clark has this great chart on the gigawatts it costs—the dollars per gigawatt.

Right. Walk us through the economic case.

Yeah. Yeah, so on this point of whether orbital is key to investing here, I don’t think it is. The first point I’ll make is: what are the implied monetization rates based on expectations today for the AI business?

I think you threw out the $160 billion number that’s been leaked out there, that people are talking about. The implied monetization rate on that number is something like $14 billion per gigawatt per year for the AI business. They just signed Anthropic at 22 to 23. They just signed Google at 50.

Right.

Right. So I think you can invest behind the AI business terrestrially and still be excited about it.

But with orbital—an important point—you’re excited about it if they can get the land and the power.

Right. But I think for most investors, they have an easier time getting their head around how SpaceX wins terrestrially. Can they go get land, power, and chips? The answer to that is high probability, yes, okay?

And what we’re saying is, at the rate they’re monetizing that, that gets you to the numbers that are being leaked out there before you even have to take the leap of faith that they’re going to extend the lead with orbital data centers. But take us there on that, too.

Sure. Yeah, so look, with orbital, I think the key thing is 2-stage reusability.

Yeah.

And beyond that, rapid 2-stage reusability. So today with Starship, they’ve shown that they can successfully reland the booster. The second stage, we’ll see what happens later this year. I think they’re attempting to bring that back and then make it reusable by next year.

But the thing that’s important about 2-stage reusability when it comes to the economics for orbital compute is the cost per kilogram comes down significantly. We’re talking about going from $1,500 per kilogram on Falcon, somewhere in that range, to $250 per kilogram, something lower. And the more that you can reuse the rocket, the more that price comes down.

Right, because you’re just depreciating the cost of the launch. And eventually, you asymptote to the cost of the fuel.

Right. Assuming you can use a rocket forever.

Yes. Right, which will take a very long time for us to really achieve that.

4. Orbital Compute Economics: $5B Per Gigawatt in Space vs. $25B on the Ground

But at that point, we’re talking about something well south of $250 per kilogram. So then you look at the specs of these AI satellites.

Yeah, that post was incredible—the one Elon laid out the other day, the specs on the satellites.

It was really great, because I think they are finally showing people, “Here’s how you could viably design one of these satellites.” How heavy is the satellite? How many could you fit into a Starship launch? And when you back into the numbers, you get to something like 5 MW of capacity per Starship launch.

Right.

There’s 100 metric tons in one of those Starships. So you can back into the math of how much it will cost per gigawatt to launch this compute into space. And the math that you get to, before you account for things like bad GPUs and bad satellites—which will all be things that happen—is about $5 billion per gigawatt of CapEx to put these in space.

Right.

For comparison, terrestrially, when you talk about the switchgears, the generators, the transformers, the shell, getting the power—that today is about $20 billion to $25 billion per gigawatt. So we’re talking about a 5× reduction in cost on half of your bill of materials—

Right.

—for the data center.

Right. Yeah, just very simply, to say that it costs $60 billion to put a gigawatt on the ground today. We’ll call it $35 billion of that for the GPUs and the silicon that’s doing the training and the inference, and $25 billion is the land, the shell, the power, and the cooling.

I would hypothesize that those elements are probably going to be inflationary, so that $25 billion may not go down. Because space, power, and cooling are effectively free in space—and when I say space, I mean land. There’s no land in space, but there is space—you’re talking about putting a gigawatt into space for $30 billion and having lower operating costs.

Now, the dynamic versus $60 billion is that the $60 billion is inflationary, and that $30 billion, that $5 billion, may be deflationary over time. But what we need to consider is the reliability and the maintenance.

So as long as these satellites in space aren’t failing at an astronomical rate, the math maths. As you can see, we know GPUs melt and lasers fail. We know this happens in data centers, particularly during big training runs. So as long as the reliability and maintenance are not dramatically lower, the math is there once we have reusability, and then rapid reusability, for Starship V3.

When we look at this, okay, we went through Starlink and said, “Okay, it just stands to reason we’re going to have direct-to-cell on Starlink.” The assumptions there, again, seem like you can get your head around them.

5. The Most Underrated Variable: What Cursor Does for xAI's Model

Then, when it comes to building terrestrial data centers, again, it’s not a hard one to think that, based on these couple of deals, SpaceX is going to build a much bigger business there. And then you have this call option on space that would drop the price even further.

The one thing we haven’t talked about is their model, right? And I find this surprising. 6 months ago, xAI was competing—they were doing pretty well—but they’ve done something dramatic over the course of the past couple of months, which is they bought Cursor, right? Cursor is 700 or 800 people and was already doing incredibly well from a revenue perspective.

Our own projections were that they could exit this year at up to $10 billion of revenue, so they were growing very fast—one of the leading coding agents—but they also had this incredible team with the potential to really build a frontier-level model. But they were compute-constrained. So, all of a sudden, they get bought by SpaceX. SpaceX has massive compute that they can now train on.

When I think about the revenue in AI, if I look at that line item in the models and see it going from $10 billion to $150 billion, yes, a lot of that will be the CoreWeave-type business that they have. But the question is, how much of that is going to be the core xAI business that's really powered by the new team from Cursor? So, any thoughts on that, Gavin?

Right now, Composer 2.5 was Pareto-dominant 12 days ago. It was trained on the Kimi K2.5 base model.

Now, what's happening is the Grok 4.3 1.5-trillion-parameter model is training. One would hypothesize, based on scaling laws, that that will be a better base model. And the Cursor data is being injected into the pre-training process, not just reinforcement learning. We'll see, and I think that is going to be a very important data point when that comes out. 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.

You know, that, to me, is—if I had to say what the one piece that's being lost in the story is—it's easy for everybody to get excited about the deals with Anthropic because you can put your hands around that. You know how much revenue it is. I see debate about the 90-day termination, how long they last, and what multiple you put on those revenues. But I think the thing that's 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. This is an extraordinary team that he just downloaded into SpaceX. SpaceX was already building good models. And what they have is this way to monetize compute that gives you this call option that you can pull all that compute in-house to train a model and then to run the model. I suspect, if there's an upside surprise—if we went around the table, I'd say this is the place that's getting the least amount of attention and could have the biggest upside surprise. Any thoughts, Clark, on what you think is being overlooked or areas that you think are misunderstood about the business today?

I would say what the last few weeks have proven is that Elon and their team can stand up all this compute. Actually, if you just went back 1½ years, they were behind in the race to stand up compute. They didn't have that many H100s. They brought in Colossus, then they brought in Colossus 2 at a scale much larger than anyone else.

And now, as we gear for Vera Rubin, from a lot of my conversations, it looks like they've secured maybe up to 20% of Vera Rubin capacity, especially in the early days when these chips are very scarce. They're going to have a lead on all of this because people think that they can stand up this compute better. So, I think what the last few weeks have actually shown is that Elon will take a shot at hitting the frontier.

But if, for whatever reason, they have overprocured some capacity, this is a very scarce asset that they've shown they can monetize at best-in-class margins and payback periods.

The irony is, you and I've been doing this long enough to know—I mean, that's why Bezos built AWS, right? He had to build capacity for Black Friday. But then the rest of the year, he sat on all this capacity they had to build, and he figured out a really incredible way to monetize it. And, by the way, investors at the time—2009, 2010—when he was building out the capability around AWS hated it.

Bill Gurley

Yeah.

Because he was consuming all that free cash flow. Meanwhile, he was digging the biggest gold mine in the history of the world. One of the biggest.

Bill Gurley

One of the biggest.

Among them, at the time, it was probably the biggest.

Bill Gurley

Yeah, Google Search might want to have a discussion.

By the way, I do think it is important. Grok 4.3—I think the Cursor acquisition, if they acquire it, may end up being very important. But Grok 4.3 was on the Pareto frontier as of 10 or 12 days ago, and these things move fast. It was the most intelligent 500-billion-parameter model in the world.

There are 4 companies on the frontier: xAI, SpaceX AI, and Google with Gemini 3.1 Pro. The rest of it was dominated by Anthropic and OpenAI. But they were on the Pareto frontier, and now we'll see what they do with Cursor.

Yeah. I want to come back to that in a second.

Bill Gurley

By the way, man, I want to ask you some questions.

Go, go, go.

Bill Gurley

What do you think? So, you think the biggest source of potential upside is the model?

Yes.

Bill Gurley

What do you think?

I think that's the thing that's least talked about.

Bill Gurley

Least talked about.

6. Bull & Bear Case — Can SpaceX Really 8X Revenue in 4 Years?

Right? And so, listen. When I look at the bull-bear case on the IPO, the bears are looking at last year's revenue, say it was $18 billion, and they're looking at the forecast from the banks of $160 billion 3 years from now. They're saying, “Listen, not many companies in the history of the world have basically 8x'd their revenue over 3 to 4 years.” That's where people get nervous about the valuation.

When I look at this, again, when you break it down as an analyst, first principles, part by part—which is what I tried to do here—when you look at Starlink, it looks totally doable. When I look at what they're building in AI compute terrestrially, it looks totally doable over the course of the next 3 years. When I look at the model itself after the acquisition of Cursor, combining those things around the compute they have, that looks to me like it could be an upside surprise.

So, I would say that, in the IPO, I think when you look back 3 years from now, there's a decent chance that everybody's like, “Oh my God, that was super obvious.” Even though today, all of these things have risk associated. And, back to where we started, none of us are here to pump the IPO at $1.77 trillion. It's really to just break it down as we do inside our shop and to say, “What is that distribution of future probabilities? What's the probability that it's higher from here?”

I think we're all pretty AI-pilled. And if you're AI-pilled, that means we've got to build a lot more compute than the world thinks and that these models are going to be a lot more valuable than people think. You combine that with their core business, and I don't know another entrepreneur or another business that's a better bet on the future than SpaceX. So, I think for most institutional investors, it's a must-buy, a must-own, set-it-and-forget-it position in order to have a real bet on both the space and the AI future.

Bill Gurley

From your lips to God's ears.

7. Post-IPO Drawdowns, Lock-Ups & How to Size the Position

I mean, listen, I again think that you're going to have to wait. But we had this chart last week that came out. Everybody was sending it around Twitter, conveniently timed, and it showed the average max drawdown post-IPO for about 20 companies—from Facebook, Twitter, Alibaba, and Shopify—is over 50%. So, maybe that will end this section here.

Gavin, you and I've been doing this a long time. We know it's going to be bouncy around the IPO. How do you, as a manager, try to manage that? Do you try to trade around the IPO? Do you set it and forget it? I would say, from an Altimeter perspective, what we tend to do is take a base position that we set and forget, and then we may size up or size down depending upon how the market reacts in a particular moment. But any thoughts on this chart, or how you guys are thinking about it in particular? You obviously own a lot going into it.

First, I agree with absolutely everything you said, and I actually think about it the same way: set it and forget it. You've talked about how you have ballast. You move it around, and you move the ballast to one side of the ship when you want the ship to lean into the wind to go faster, and you move it to the other side when you don't want the ship to tip over. I think that's a great analogy. I think about all important companies in the portfolio the same way, so 100% agree.

I mean, this chart is a bummer. What I would say is, this data on IPOs—but what I would just say is, this is a really unprecedented situation.

Yes. We've never had an IPO this big. We've never had an IPO that's going to go into an index this quickly. We simply do not know how much selling there will be from investors. I would hazard a guess—I mean, I don't know—but Elon, I don't think he needs liquidity, and I think he owns—what does he own, Foxy?

It's 50%.

50% of the company. By the way, he's locked up for 365 days or 366 days. So, we know he's not selling, right?

I just think it's an unprecedented situation, and the right answer is, I don't know what's going to happen in the short term. The right answer that I would encourage every investor making their own decision is to just think exactly the way you articulated it.

We have these different levers. We have these different variables. Think about each one of them from first principles. Make your own decision. Do your own due diligence. Be thoughtful. But there are a lot of variables here, and it is a little funny to me that it was 100 times trailing TTM revenue. Well, after the deals they signed, I think it’s at 39 times.

That can change fast.

So, they added $29 billion in a month.

Yes. [laughter]

By the way, have you ever seen that happen?

Never. Never. And it just goes to show, first, Elon is not only a great engineer. He and Gwynne and the team are great at business. They understand what needs to be done to raise the capital to get to the next phase. They have a long-term mission in the business.

And so, to me, again, what we saw in the course of the last few weeks with Cursor, what we saw with these deals that they cut, I don’t know that any of the Mag 7 could have moved that quickly to adjust the business as they did. It’s exceptionally entrepreneurial at scale, which we very rarely see in businesses.

Two other things I would just say. Number one is people talk a lot about the total amount of capital being raised. If you add up the capital here—for Anthropic, what they may raise; what OpenAI may raise; what SpaceX may raise—let’s call it $250 billion. That’s 1% of the Mag 7. Okay? It’s 1% of the Mag 7.

Bill Gurley

Can I give you a hug, Brad?

Two other things I would just say. Number one is people talk a lot about the total amount of capital being raised. If you add up the capital here, right, for Anthropic what they may raise, what OpenAI may raise, what SpaceX may raise, let’s call it $250 billion. That’s 1% of the Mag 7. Okay, it’s 1% of the Mag 7.

Yeah. And we will as well. You know, to me, that is a bet on the future that we all believe in. And so, if I said, “Where are we out of consensus? What is our variant perception?” we actually think it’s going to be bigger, faster, and we’ve thought that for a couple of years.

First, it’s only 1% of the Mag 7 market cap. And then you referenced it, the amount of selling. I’ve got a chart we’ll post here. This is the dribble-share release for SpaceX shareholders. There’s not a lot that can be released up until after the first earnings.

We saw this in the Cerebras IPO. There’s a version of it here in this IPO. And so, again, I think the banks have been thoughtful here, knowing that this is a very large IPO. And I’m not saying that it won’t trade down. There’s a possibility these things trade down.

But again, for me, telescope out: is there any company better positioned as a bet on the future? I think what they’ve shown over the course of the last 5 weeks, they’re probably number one. But let’s move on.

Bill Gurley

No, no. Can I just say one thing about the employees? I think another thing that’s unprecedented here is the employees—

Yeah.

Bill Gurley

—and, to a large degree, the investors here have had liquidity every 6 months—

Exactly.

Bill Gurley

—the last 10 years.

Yes.

Bill Gurley

So, if you’re a SpaceX employee or former employee and you wanted to sell, you’ve had whatever that is, close to 20 chances. And it is a matter of historical record that large investors have been able to sell. So, I would think a lot of the people—they’ve chosen to own it.

Now, there’s a new valuation, and we’ll see what they do, but this is utterly unprecedented, and we’ll see.

Yeah, I know. It’s a great point. We have, in fact, called these companies quasi-public. You and I both know that SpaceX—and I’d put Anthropic in this category as well, Databricks in this category—these things, in many ways, have been more liquid over the course of the past 3 years than some public biotech companies we know, right?

And so, there’s a continuum of liquidity here. We treat it as a binary, private versus public, but it’s really about this continuum.

Let’s keep going on models. Anthropic launched Fable 5, which you referenced yesterday. It’s basically Mythos with some classifiers and safeguards around cyber, biology, chemistry, and distillation. When those things get triggered, it fails back—[snorts]—to Opus 4.8.

There was a Copart tweet about this yesterday. He said it scored on all the benchmarks, but what really makes it special is long-running tasks. You retweeted our good friend Noam Brown. ChatGPT 5.5 also exhibited these capabilities. It led Noam, right, to suggest that it’s not very relevant to do these snapshot benchmarks anymore.

8. Fable 5, Mythos & Why Snapshot Benchmarks Are Broken

The x-axis has to be time or tokens or compute, because we can solve most problems now if we just let these frontier models run for a very long period of time. So, Gavin, what is this new class of model—Fable 5, ChatGPT 5.5? What does it mean for the race in superintelligence? Who’s up? Who’s down? Who’s still on the frontier? Give us your thoughts.

I mean, it’s hard to say that Anthropic’s not up.

Yeah.

After the revenue numbers they’ve put up, after the Fable 5 release, and Mythos is evidently even better. But I just think that Noam Brown’s post from yesterday about polynomial time is so profound. And just the idea that we do not know how smart these models are.

Say more about that. Why don’t we know how smart they are?

Because nobody has run Mythos for a year continuously. And we may never know how smart each generation of models actually is or was, because we don’t have time to appropriately evaluate their intelligence before the next model comes out. I mean, this is a profound statement.

And just imagine. I always say, when you think about FSD, just imagine a human being who never gets distracted, never gets tired, never talks on the phone in the car, never drinks and drives, never yells at their kids, never has to go to the backseat to give their baby a bottle. Of course you would think that, over time, that is superior to humans who are distracted.

I don’t know how long. How long can you think deeply about one topic, Brad?

What do you mean? Give me an hour. Give me an hour. [laughter] Give me an hour.

A bit. That makes me feel terrible because I think I can think deeply about one topic continuously before having a stray thought enter my mind for maybe 5 minutes. Then I can come back to that.

Imagine if Albert Einstein. Maybe he could think for 3 hours at a time—clearly, an exceptional intellect. But imagine Albert Einstein had just thought about fundamental physics 24 hours a day. He doesn’t have to eat, he doesn’t have to sleep, he doesn’t have to relax, he doesn’t drink—

Never gets old.

Never has diminished intelligence, and he thought for 1 year. I mean, we might already—

Have solved a lot of these intractable problems.

So, I just think that’s an extraordinary thought. And my takeaway was, however bullish I was on compute before then, I’m just a lot more bullish.

Right. Right. Right. So, that is what we saw. That was probably what really unlocked Opus 4.6. It was the first really long-running model that could maintain that context, maintain that memory, solve some of these longer-running problems, right?

For us, the signal was in January. We felt like that was a big moment, but then when you started to see the revenue go up, we knew that lots of people were voting independently, that that was a profound moment, that they became much, much more useful.

One of the things that the consensus going into this year—the big question going into this year—was, was the AI revenue going to show up? Were we going to get to these thresholds of intelligence that caused enterprises and consumers to use them more?

I think the consensus at the time, at least on this podcast, was the debate with Bill: the open-source models and cheap tokens were catching up to the frontier, that perhaps these models were beginning to asymptote, that people wouldn’t really pay for premium tokens.

And it seems to me that the evidence in the field, 6 months into the year, is just the opposite: that frontier tokens are capturing the vast majority of all the revenues, and that, in fact, if you believe in the long-running capabilities and more compute allows you to do that, they may actually be extending their lead on some of these models that were built on distillation.

So, I’ll just open it up to anyone around the table. What are your thoughts on whether or not we’ve challenged this thesis that cheap, open-source tokens are going to always close the gap on these frontier models, or are they extending their lead?

I think this debate—this same debate—has existed since the beginning, since we started training these models to begin with, which was, “Hey, we’re always kind of 3 to 6 months behind the frontier.”

But empirically, you can just see all of the revenue has actually just accrued at the frontier. And I think that’s because every time we release the frontier, a whole new slew of use cases—

Right.

—that previously we could never have tackled before, like coding. But also, we’ve just been locked at our desks for the last day, hammering Claude, because it’s fascinating, the things that now we can do with Fable 5 that we just couldn’t do with Opus 4.8 just a day before.

So, what are some of those things, man? I’m curious.

I think it’s really, really good at multi-agent orchestration. Anthropic released a blog post about 6 different agent orchestration patterns that they’ve talked about.

But really, once you start being able to manage all these agents, the harness and the model itself are blending with one another. They’re actually being fused closer and closer together, but the model can understand the context of your work.

So, one of the things, for instance, is I just threw in 7 of our models and said, “Okay, I want to create a master view of my beliefs, given all of these assumptions of all these companies, TSMC capacity, and then produce me a report on all this stuff.” The model is able to reason through all of our assumptions. Like, actually, if you believe this—

Right. What are the contradictions exactly?

Yeah, it was fascinating. Before, we’d never do that, but now I think we’re just at step 1 of multi-agent orchestration. We’re going to do this even further, and that’s one example. I’ve also dumped all my notes into it and had it reason across all my notes from the last 3 years and say, “Here are some of your ideas that were consistent. Here are the sources that were actually the highest signal to what actually played out.”

It’s super fascinating what you could do, and we’ve just blown through our limits.

I mean, it’s unlocking all this. They gave examples yesterday in the release Anthropic did: a 50-million-line Ruby codebase at Stripe that was refactored in a day versus many weeks with many people. You think about where this is impacting biology and life sciences, just across the spectrum.

9. Frontier vs. Open Source: 90% of Revenue Accrues at the Frontier

To me, it gets back to this fundamental point: number 1, if you believe this to be true about long-running agents, then we’re going to produce and consume more tokens in the future as far as the eye can see. So the world—this gets me back to terafab and space orbital and all this—because we may in fact unlock real thresholds of intelligence, but we’re going to have to let these horses run for a long time in order to get there.

Yeah, I would just say 2 things. 2 things can be true.

Mhm.

The majority of economic value may continue to accrue to the frontier, and, man, has it ever accrued to the frontier thus far—and for sure in the first 6 months of this year. But the majority of tokens consumed in the world may be open source.

And they are today.

Yes. I think that this current state is likely to persist. Harvey had a great blog post that they put out on X, and it’s just amazing how everything gets out of date in 5 days. They used their own proprietary legal data to do reinforcement learning and supervised fine-tuning with Fireworks on an open-source model, and then used a router—a router being something that picks which model you send each query to, and which model you use to check which model.

They got better outcomes than Opus 4, either 4.7 or 4.8, at a lower cost.

Yes.

And I think that is the future. The reality is, they were still consuming a lot of Opus, but a majority of the tokens they were processing probably were in their own open-source models.

We heard the same thing. We did an enterprise survey that we’ll post of 300 companies: which ones were optimizing? These are folks who are looking at model routing and saying, “We’re going to send certain tokens over here.” Which ones are thinking about optimizing, which ones aren’t optimizing yet, and then what is their expected use of frontier-model tokens?

They’re all expecting to consume a lot more, even though they’re already in the process of optimizing. Think of it in the context of JP Morgan. If they’re doing some back-of-the-house stuff, on customer service or whatever, they may very well use an open-source model. Now, I think they’re loath to use Chinese open-source models, so they’re waiting on U.S. open-source models to be able to really deliver the bang that they need.

My hunch is, for these enterprises, a lot of that back-of-the-house stuff will get routed there. That will probably be a majority of the tokens, but I think the really high-value stuff—coding, as an example—they don’t want to write second-tier code. I think the vast majority of that will continue to be on the frontier.

Bill Gurley

You don’t need Albert Einstein to book you a trip. You don’t need Albert Einstein to do KYC.

But this is the debate we had at literally at this table 2 years ago. However, if you just look at the revenue curves, what folks concluded when they said that, they said, “Therefore, the frontier models will not accrue most of the revenue.”

Bill Gurley

That has been decisively wrong. Probably more than 90%, and it may continue to be decisively wrong. Frontier might be 90% of the economic value. Open source—

Might be 80% of tokens. Something that I think is very important about open source is that I think there’s this belief that it’s bearish for AI. It may be very bearish for the frontier models. There’s that bear case you talked about.

It’s actually really bullish for compute and hardware, because if the frontier models are capturing less of the margin, then you’re going to spend more on compute. So, the better open source does, the better it is for compute providers.

10. Intro — SpaceX IPO in Two Days, Mythos Launches, Taiwan Takeaways

And I will say this: between spending time in the heart of the West—Silicon Valley—and also spending time in Asia, there is a very big, deep-seated belief in one versus the other. If you spend a lot of time here, it’s all closed-source cloud; all traffic is going to go by way of this direction. Then you spend time in Asia, and the overwhelming belief is that we’re going to find the right model for the right workload, and we’re not going to overspend.

Right.

And I think the next year is probably going to be the most indicative of which way this falls, because the reason why closed-source models have captured so much of the value is that the models actually get the intention and carry through the work. This is the first year where we actually had agents that carried out user intention, from just answering a chatbot request to actually producing useful work.

Right.

Now, the level of this intelligence has scaled so rapidly, and we continue to push against the most economically valuable tasks, which are coding and finance and all these knowledge-work tasks. But for the long tail of tasks, if open source continues to maintain a 6-month lag, we might actually see a lot more open source used for our everyday tasks.

Bill Gurley

Basically, Jensen’s argument, right? Jensen’s argument is that you’re going to have model routing, and we’re just in a moment in time where the frontier models gain the advantage, can do long-running tasks that open-source models couldn’t do very well, and so they’re accruing all of the value. But as soon as the open-source models can do the long-running tasks as well—which is not far away—they, too, will grab a bunch of this revenue.

Are you about to burst into reflection?

Bill Gurley

I’m not.

Okay. No, no, no, no—are we? But I’m very impressed by Mistral and the team and what they’re doing. I very much want a frontier open-source U.S. lab to win. We know that. I heard you say recently, and I believe it to be true, that NVIDIA could, any day that they really wanted to. They already have some great open-source models. They could absolutely build a frontier open-source model whenever they chose to do it.

So it’s not a question in my mind as to whether or not the U.S. is going to have a frontier open-source model. It’s just a question about timing and then, at that point in time, let’s assume they get these long-running capabilities: have the frontier labs now achieved something yet again that allows them to keep the stranglehold on the revenues?

Yeah, and I just think it’s—wow, that’s a cute ASIC you’ve built there. That is so cute. How would you like open source to join the frontier?

Right.

How would you like that? How do you like them apples? I’m not sure that’s the explicit calculation, but I do think Jensen—

Bill Gurley

Say more. Just double-click on that for everybody at home.

Yeah.

Bill Gurley

If they were to put an open-source model out there, how does that impact the ASIC landscape?

Well, you might not have the revenue or the margins to fund that ASIC. And I do think NVIDIA is highly likely to be the world’s dominant provider of open-source AI. I do think Jensen will bring open source—right now, it’s whatever, 6 months behind the frontier.

Yeah.

We might see it creep closer and closer and closer. And I do think Jensen has a big business decision. I see this chart here, so let’s chop it up about NVIDIA, as you say. But if all of his customers are going to compete with him—

Yes.

Then why not compete with his customers? We have all these neoclouds. So that’s a cloud-computing business that can compete with all these cloud-computing businesses. He has his own models that are really, really good. Nematron 3 or 3.1 was actually really, really cool from a compute-efficiency perspective.

And he’s always careful to release small models so as not to tread on Anthropic, OpenAI, and Google’s toes. But I do think that is a choice he is making. At some point, if the economics change, I think NVIDIA can join the frontier and become one of the world’s largest cloud-computing companies much faster than people think.

Interesting. Interesting. Clark, walk us through this chart.

Yeah, so I think one of the takeaways from spending time in Taiwan was there is certainly a lot of excitement around the next wave of ASICs. But I think it's a very clear moment now where Nvidia—it used to be an argument of Nvidia versus ASICs, one or the other, and total domination, one or the other. Now, increasingly, every year, everyone assumed that Nvidia was going to lose share dramatically on a revenue scale, on a gigawatt scale, and on a unit scale. Actually, if you look at the last few years, they've maintained their share very, very handsomely. If you accounted for the fact that Anthropic was not really using Nvidia, they probably actually gained share in 2025 and 2026.

I think 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 for ASICs, the argument now is that more and more will look custom to the actual workload, and that is one vector in which people are moving against Nvidia.

Nvidia has now shown itself as the predominant provider of compute to a lot of the world. For internal workloads, perhaps companies will go more and more custom and more and more down the stack. I remember just 1 year ago, when it was kind of a Broadcom-or-Nvidia battle. It seems there's a lot more nuance now to what type of accelerators will fit which workloads, which customers, and which business models. I thought that was a new topic.

Bill Gurley

It's actually a new realization, though. I think we all kind of shared this view for a long time.

Yeah, I was just shocked. I'm out here. I did a board meeting with one of our companies, and the biggest thing they emphasized is, “We thought the world would be consuming less Nvidia than it is.” If anything, Nvidia is accelerating, and they just continue to out-execute their competitors.

I think a lot of people are indexing to this OpenAI gigawatt. Nvidia has 10. Broadcom has 10. Who has 6? AMD has 6, and they have warrants. Then Cerebras—our shared portfolio company—has a gigawatt. That is what's on paper.

Bill Gurley

Right.

What actually gets deployed, let's see. I will be very surprised if 10 out of 27—what's that math? Let's see who's best at math. What percentage market share is that?

30%, yeah.

Yeah. I'll be very surprised if that is where they land. I think that is an extremely unlikely outcome. Especially as long as we're in a watt-constrained world, if you can get more tokens per watt—which is literally revenue—with Nvidia than with a lot of alternatives, if you build your factory with another chip, you may save some money, but you're going to have less revenue and the margins may be lower. That's a point that Jensen keeps hammering, and I think it's really important.

Bill Gurley

And by the way, credit where credit is due: one of the most surprising things to me in this ASIC landscape, I'd say Meta and Microsoft have been probably disappointing.

Yes.

Bill Gurley

You know who made a good ASIC?

Yes.

Bill Gurley

Well, I know you know.

Yes.

Bill Gurley

Jalapeno—

Yeah, exactly.

Bill Gurley

—from OpenAI. They made a great chip.

Yes.

Bill Gurley

Now, unfortunately, it needs to run at a much lower temperature than the Nvidia GPUs, which means you need to spend more money on cooling, and that consumes more power. They made a great chip.

Well, I think the question there—and the question for everybody—is going to be: Is that the highest and best use of your time? I tend to think that the frontier companies—there's this belief that they've got to be vertically integrated. But if you believe, like I do, that the race to superintelligence, particularly as we get these recursive loops working, may be over in the next 2 to 3 years, then I think: focus, focus, focus, focus.

You exist to build the best intelligence in the world and to deliver the best intelligence in the world, and that means you have to have all the revenue. Because if you want to build out the compute that's going to be required to continue to push the frontier, you have to have the revenue in order to support it. So, subject to the focus question, I think they certainly did.

This all brings me back to kind of a reality check, though. We just got done talking about test-time compute, inference-time compute, and long-running agents. This is really the thing that's unlocked the revenue this year. It all pushes us in the direction of more CapEx. Google just raised $80 billion. We've now taken the Mag 5 or Mag 7 free cash flow down dramatically—80% from just a few years ago.

Morgan Stanley, you've got this chart in front of you, up to their 2027 CapEx forecast from $950 billion to $1.1 trillion. We were talking about this with Jensen. That was his forecast 2 years ago. Obviously, this doesn't even include SpaceX, CoreWeave, and so on. I think the number in 2027 is likely closer to $1.5 trillion.

11. 1.5T in CapEx vs. $300B in AI Revenue — Does the Math Math?

If we compare this to the total incremental inference revenue—the thing that the market gets worried about, back to my Sam Altman podcast in October of last year—can we really afford to spend $1.5 trillion of CapEx a year if we're only generating X amount in inference revenue? The thing I think that lit the fuse this year was Anthropic showing up in a major way with revenue.

We have the AI lab revenue, everybody combined, at around $300 billion next year, right? So, that's 2027: $300 billion. We're spending $1.5 trillion of CapEx on $300 billion of inference revenue. Does that math math for you? What would cause you to get more nervous again about our ability to continue to make these investments? Because the second we get nervous about it, the entire semiconductor complex is going to come down a lot.

What do you think the gross margins are on that $300 billion? Let's call it 50%.

I would guess they're probably a little bit higher than that. I might say 60% or 70%. But that math starts to math, and what I would just say is I think that $300 billion is low, man.

Bill Gurley

Yeah. Yeah.

I just think it's low.

Bill Gurley

From your mouth to God's ears.

Yeah, yeah, exactly. I think we end this year well over $200 billion in inference revenue—well over. So, I think the math really maths, and I do think we have to give Jensen—

Bill Gurley

Yeah, our friend.

—some credit because he said some things that seemed outlandish, and he was conservative. He was low. He said $1 trillion 2 years ago, and he was really low. So, let's give the guy some credit and think about what he is saying right now.

Bill Gurley

For sure, for sure. And listen, I would say consistently, Elon has been taking the over. Sundar has been taking the over. Sam and Dario—Dario did the podcast with Dwarkesh when he was talking about a country of geniuses in the data center. He said that will be here by 2028. He said revenues will go into the low hundreds of billions by 2028.

Let's call that $300 billion or $400 billion of revenue by 2028. He said that a while ago now, so he may even be revising up his number. He said it's hard for me to see that there won't be trillions of dollars in revenue before 2030. If you're on that revenue trajectory—if we're on a trajectory to $200 billion by the end of this year, let's call it $400 billion or $500 billion by next year, and a path to $1 trillion-plus by 2029—then the math maths.

We've got to keep in mind that half of the spending is there for training, maybe a little less than half. What is it, Foxy?

That depends on the lab, but it's increasingly less than half. I would say it's increasingly less than half.

Yes.

Bill Gurley

Okay. So, we'll call it 35% is spending that's not revenue-generating, but it's going to make the next model. So, I think the math maths.

Right.

Bill Gurley

And there's still this prisoner's dilemma where, if you opted out, that may be an existential decision.

And I think coming into this year, going back to what narratives were violated, everyone expected token pricing—the price of compute—to be deflationary, and it would be a kind of smooth-line deflation over time. But I think this year what we've seen is the opposite. It all comes back to supply and demand: the demand side of the equation seems to be far outstripping the supply.

I think you look at the deals signed by SpaceX and others, the monetization rates per watt are increasing. Look, that is on a pretty nascent, small base of users, right? Alex at WellRoc has this great way to frame it: less than 0.2% of people on Earth are actually using AI in an agentic way.

I'm not a technical person, but I'm consuming 500 CPU cores in a VM instance and 5 GPUs 24/7. If you draw that out to any meaningful percentage of the population, we're going to be in this kind of shortage environment, maybe for some time. I think that is all positive for this ROI question.

Bill Gurley

Man, Foxy, a 100-to-1 CPU-to-GPU ratio. [laughter] Kind of an agentic workflow.

He said, “Of course. Five.” [laughter]

Bill Gurley

Five.

Five, yes.

Bill Gurley

Yeah.

I'm being smart with my phone.

Good, good, good. Excellent.

I will say also that the ratio is 300 to 1. You know, call it 1.2 or 1.5. There is also a rate at which, physically, we can only expand how much we can produce and how much we can actually increase that spend, whereas we're seeing the opposite right now in the willingness to pay for these tokens. Actually, the monetization per gigawatt is increasing from, call it, $20 billion in the best of cases at the beginning of the year to now $30 billion to even pushing $40 billion—

Per gigawatt.

Per gigawatt. All of that is a very heavy fixed-cost base, but all of that is pure margin flow-through now. As we scale, the willingness to pay for all of this—and all of this is stipulated by everything we're talking about, how much is open source versus not, and all of these different flows—the revenue might actually outstrip our fixed-cost base by a significant amount. I think that's why all the labs are pushing the gas to the pedals, because they all see that, within 3 years, if we continue this curve, we're just going to be so short on all the compute.

It's a great point. If you thought you were getting a certain return when you made these decisions in November 2025, you may be getting triple that return today.

Yes.

At Anthropic, no way did they think they were going to be anywhere close to break-even.

Yeah. Right? In this part of the curve, the reason I called it accidental profitability—and people have been talking about that—is that they want to spend a lot more money on compute; they just had a hard time doing it. Now maybe with SpaceX, they could take some of those dollars and go spend them other places. But that, to me, is a fundamental change.

The first argument against the frontier labs was, “They’ll never generate revenue.” Then that got blown up. Then it was, “Even if they generate revenue, they’ll have really shitty gross margins, and they’ll never be able to make money.” And then that got blown up.

I think now people are falling back and saying, “Well, they’re overcharging. This is token maxing.” My good friend Chamath has said there’s no ROI on any of this spend. It’s all this token maxing.

My best evidence is that, of course, when somebody puts on this much spend, like at Altimeter, we're not optimally spending every single dollar. But the question is: Why are millions of independent businesses—small, medium, and large—and why are millions of consumers all choosing to do the same thing? They're not dumb. These are rational economic actors that are all simultaneously saying, “I want to do this because it makes my life better. It makes my business better,” et cetera. To me, that is the best evidence as to why I think this revenue can continue.

Yeah. And Clark, I think the point you made is dead-on, because you want to own asset-heavy businesses in inflationary environments, and token pricing is going up, and supply and demand is tightening. So, totally agree.

As we begin to find our way to the exit ramp and wrap here, one of the things—you and I've been doing this for a long time, Gavin, a couple decades. You may even have been doing this longer than me, even though I'm a little bit older than you. We have—I always like to do a market check, because I find a lot of the time that analysts come on these things and talk their book, and there are a lot of people who listen to these things, retail investors and others. It's just kind of like, what do we really think?

So I always characterize it as kind of small, medium, and large. What am I doing? Do I have small exposure on? Do I have medium exposure on? Do I have large exposure on?

If you look at what's happened in the markets, semis ripped this year. You've been doing this a long time. I don't think I've ever seen it before, right? I've never seen the doubles and triples across the board like we saw. But there's been huge dispersion in the market, right? Internet's down 16%, software's down 8% on the year. SPY and Nasdaq are up, but really up because of their components that are related to AI and compute. The market itself has kind of struggled.

Meanwhile, if you were in the stuff that we were invested in, we've all done pretty well. I think I've said it a couple times: If the Anthropic revenue had not shown up this year, because that was the overhang on the market, I think the whole market could be down this year, right? We just had these huge months in April and May.

For us, because prices came up so much, because I have some worry about geopolitics, the macro backdrop with what's going on with inflation in the short run, and just needing a little consolidation in this market to answer some of these questions, because expectations are now higher, we dialed back from what I would call large for Altimeter to something kind of like medium-small. Again, it's never all or nothing for us. It's like, what is the risk-reward at a given price?

We think this is maybe going to be a period of consolidation on the way to much higher highs. I'm curious just how you run the book and how you think about it as a portfolio manager.

Very similarly, man. I always think of stocks and the markets—I imagine them as runners. In 2022, that runner had gone downhill. It had a lot of energy, man. It was painful. It wasn't fun.

Coming out of that, there was a lot of pent-up upside in the market. The market, particularly over the last 2 months, has run up a very steep hill. A lot of semiconductor companies in particular—ironically, Nvidia and Broadcom—have been laggards.

Totally.

And so a lot of these—I do see a lot on X about finding the next bottleneck. I think that was the last game. That game is over. A lot of these stocks, forget climbing a mountain or a hill, have gone straight up a cliff, okay?

Yes. They're tired. They need to rest. We'll see: Do they just rest at the top of that cliff they climbed? Do they hang out in their harness for a while? We've seen some. Or do they need to go downhill for a bit? We'll see, but I'm thinking very similarly to you.

The market is seasonal. I think there are real, real concerns around inflation and rates. We have some unknown unknowns, but the market—I mean, if I had told you the fact pattern for this year, that we were going to be in a war with Iran, that oil was going to be at $100, that CPI was going to be creeping back up, that internet was going to be down 15%, and software was going to be down 8%—you would have said, “I want nothing to do with that market,” right? And here we are. The market's done pretty well in the stuff that we traffic in because the world underestimated AI revenues and underestimated the amount of compute that was going to be needed.

What was CPI this morning?

4.2. I think core came in at 0.2 versus 0.3, so a little bit better. Clearly, we're above 4 again, and there's short-term pressure on core PCE, et cetera.

It's odd to say we're heading into a seasonally weak period with all of these fears. AI has actually been seasonal for the last 3 summers. Token consumption has kind of plateaued and slowed down, and that's because college kids are big AI consumers and they don't use as much AI. Hopefully, they're all using it to learn and not cheat, but that may happen. It may not happen because of generative AI.

He's building swarms of agents, building a SpaceX model. He's going to the SpaceX IPO with me at the exchange on Friday, but he had to build an AI model using AI agents. He had to build a model, a DCF, before we go to the exchange. He is mesmerized, and it's extraordinary what he's doing.

So he's one kid who's not using less compute. [Laughter]

Or something. He's burning it. He's burning it.

Yeah, but if token consumption plateaus, if open source takes some share, there's a SemiAnalysis index that has shown—which is an index of consumption and pricing—I think there may have been a little bit of a shift over the last 2 weeks to open-source tokens that are cheaper. People may look at that data as bearish or not understand it.

Nonetheless, I just think there are reasons to look around, be careful, be thoughtful. I always assume a bullet is coming for me, head-on. [Laughter] Head on a swivel. It's the bullet you don't see that gets you, so I'm trying to spin as fast as I can.

But yeah, the market may need to take a breather. Man, when I think about what Noam Brown said and when I see the capabilities of Fable 5, it's just hard for me to get too bearish.

I mean, to me, we got 2, I think, of the most extraordinary guys of the next generation sitting in the room. At Altimeter, we have deep admiration for the work that you guys do. I always appreciate when you send me a note about the work that we do and publish.

For the guys who are newer to the business, they might think this is the way that it kind of always was, right? The steepening of the line of creative destruction, the steepening of the line of scale advantages—I always believed it was going to be true. I never thought it would be true at this rate.

12. The Next $1 Trillion: Three Companies, Half the Time

I went back last night. In the last 7 years, we've added $1 trillion of revenue to the Mag 7. To get to the first trillion took over 20 years. In the last 7, we had another trillion, and that added $17 trillion in market cap. That trillion dollars, okay?

The forecast now is that we're going to add another $1 trillion of revenue in just 3 companies—SpaceX, Anthropic, and OpenAI—over the next 4 to 5 years. Not 7 companies: 3 companies, and in half the time. I would say that we're going to have bumps in the road. I know that it's going to be like this, but we're going to have higher highs because of the size of the prize.

This is going to transform 5%, 10%, 15% of global GDP. There is no doubt in my mind, and 10% of global GDP is $10 trillion. It's an exciting future to be a part of. It's fun to do it with you guys. I think we're going to have to do our work to do the things to make sure America wins and that we evolve the social contract, keep everybody—lift the floor, take everybody with us on this ride. It's a really exciting time to be doing what we're doing. It's fun to be doing it with you guys.

Yeah, I just want to say, Brad, thanks for having us, and thank you for what you've done with the Trump Accounts. I actually think it's super important for America, for the world, to give people an equity stake at a very young age. They will see it compound over their lifetimes. This is a great thing you've done for the world, so thank you.

I'd echo all your comments: deep admiration for you and your team, gratitude for the collegiality and friendship between our firms. I know Clark and Foxy—they hang out all the time.

Bill Gurley

People think that—and there are people in our business who don't want to share anything. Our view is, we open-source it, but there are very few people who we actually call and ask their opinion because there are very few people who do the thousands of hours of work that we do that are adding to that. And you do it, and we appreciate that. And you do as well, Gavin. So, with that love fest, let's call it a wrap. Thanks for being here.

Thank you.

Bill Gurley

Mhm.