OpenAI 与 SpaceX 提交 S1｜Cloudflare 与 ClickUp 裁员｜OpenRouter 与 Polsia 完成超大额融资

I can’t remember a time when corporate America was as convinced of the ROI of something as it is right now of AI.

Rory O’Driscoll

Anthropic has done as much in Q1 as all of last year, and OpenAI has done 30% of what they did last year in Q1. Within a couple of quarters, Anthropic will be visibly and obviously ahead—profitable, growing more quickly, and bigger.

That tornado dominates all 3 vectors this week. Starting off, OpenAI confidentially files its S-1. What does this mean for Anthropic’s plans to go public? SpaceX then drops its S-1—the largest IPO in history. Anthropic then hits $44 billion in ARR and laps OpenAI in revenue. Then NVIDIA prints $81.6 billion in revenue, and the market yawns. Then layoffs at ClickUp and more.

Guest 2

It could be the GeoCities deal of the AI era. We could look back on this and—listen, I’m on Team Elon, okay?—but 100 times trailing sales. We may look back and say these were some good companies, but at 100 times—my God.

Rory O’Driscoll

I love the S-1, but I think it’s all madness. I wouldn’t buy a share. I just love the optimism.

Guest 2

It’s SolarCity on steroids, and we’re all here for it because we love AI, but it makes no sense.

Rory O’Driscoll

The cliché—the picks and shovels of the agentic revolution—is just a good place to be investing.

1. Nvidia Blowout Quarter: $81BN Revenues and Stock… Flat!

Ready to go, boys. I am so excited for this. I have to admit, I was wondering where we start, given the sheer amount of news that came out in the last few days. Jason, you lied to me on email: “Oh, not that much this week.” I’m like, “If this isn’t much, I’m glad we do the show today and not 3 or 4 years ago,” because, Jesus, NVIDIA—we’ve got to start with NVIDIA.

An $81.6 billion revenue quarter, $91 billion in Q2 guidance, and an $80 billion buyback. After all of that, the stock barely moved. How did you guys analyze NVIDIA’s quarter and the stock not moving? Help me understand.

Guest 2

Let’s do the trivial bit. First of all, I think you forgot the most compelling number. It’s not the $81 billion in revenue, which is pretty damn compelling; it’s the $50-some billion in profits, which makes it the most profitable company on the planet. Google typically clocks in at around $100 billion to $120 billion a year. These guys, in a quarter, made $50 billion-plus—$56 billion—of profits. If that were continued all year, and there were some nonrecurring items, that’s $200 billion of profit a year. That, to me, was the big-ass compelling number. It’s not just a great revenue business growing at 80%; it’s a wildly profitable operating-margin business.

The second thing, to your point, is what do you make of it? You were kind of whining about the stock on the day, and honestly, don’t look at the stock on the day. The big picture is that over the last 6 to 9 months, the stock is up 20%. Basically, it had a one-off jump when everyone, from 2023 onward, totally internalized in the space of a year the amount of spending that would take place in AI capex. It gapped up and got to $140 or $150 about a year or a year and a half ago.

Since then, it’s just grown almost steadily, because even though the news is amazing, you have to ask yourself: the stock moves not on the total news, but on the delta news. They beat slightly, and it’s gone up slightly over 6 months, but I wouldn’t look at the movement. I’d look at the overall big picture.

A year or a year and a half ago, it was a credible question to ask: Is this capex thing sustainable? I remember asking it myself. I even priced NVIDIA puts to say, at $140 or $150, “Is this thing done?” Thank God I didn’t, because here we are a year later, and the company continues to grow at 80%.

The stock doesn’t go up 80% because stock markets anticipate. About a year and a half ago, the stock grew 3 or 4 times, and now it’s growing at 20%. What it’s saying here is, “This is great. This is amazing. We trade at a mid-20s P/E. We think this is really cool, but we’re not sure it’s going to 5 times in terms of growth from here.” It all makes sense. The market is pretty happy with NVIDIA right now.

Rory O’Driscoll

Maybe 2 thoughts. For sure, that’s a good summary. One is that, in today’s market, not falling after your quarter is a strong sign because the markets are expecting insane growth. This growth was just what we expected; it’s just good enough. I’m not being facetious, especially when NVIDIA is 7% of all of our 401(k)s. It’s not just a niche stock. It’s not even just us. It’s 7% of all Americans’ life savings, essentially, in NVIDIA this morning, with that concentration.

We’re all in the S&P 500 and VTI index funds, so everyone’s on board with NVIDIA, whether they realize it or not. When NVIDIA falls, we all fall with 7% of our life savings. We’re all AI investors, whether we realize it or not.

But it’s so hard. You’ve got to beat, raise, and accelerate. That’s the mantra today: beat, raise, and accelerate. Even just trading flat—I’m all for it. That’s a good quarter.

2. Polsia Raises $30M at $250M Price: Is this the Peak?

AI capex is going to be about $800 billion this year. You look at it and go, most of the time GPUs are 50% of total capex—$400 billion. NVIDIA has pretty commanding market share, though it’s not alone. That probably says that the NVIDIA business, no surprise, is a $300 billion-a-year run rate. And there it is at $320 billion.

When you look at the announcements from the hyperscalers last week, you kind of know what they’re going to do this week. With NVIDIA, it’s plus or minus 2%. We’re really dealing with small numbers, and at that point, I think the market reaction is in the noise.

likely Jensen Huang said this week that, on AI capex infrastructure spending, we’d reach $3 trillion to $4 trillion by 2030. Do you think that is an extrapolation or an exaggeration, or can you feasibly see that happening?

Guest 2

2030. Let’s just do the quick math here. What does it mean for us, as Harry always says? If we use the same idiotic guideline of 50% of $2 trillion to $3 trillion, that would be $1.5 trillion of total semiconductor capex. Let’s just say that, at that point, there were only 70% market share because of Trainium and all the others. That’s roughly $1 trillion in revenue from $300 billion now.

I don’t think the market is quite anticipating that, to say the least. Extrapolating the past gets you there. Extrapolating the growth rate of the last 3 or 4 years gets you there. But I think the $64 trillion question is whether that growth continues or whether you start to hit—not technical constraints, obviously, but economic constraints.

Is the ROI there economically on the next $2 trillion, to go from roughly $1 trillion of capex right now to $3 trillion, which is what he’s saying will happen in 2030? It’s now, shockingly, only 4 more years. There has to be an ROI on that next $2 trillion, and I think that’s the question.

3. Uber and Microsoft Declare Productivity Gains Questionable from AI

Well, the Uber COO says no, right? The Uber COO says we’ve already reached it. And even Microsoft is allegedly moving off Anthropic, saying it’s too expensive to use Opus—which I think is talking your own game. But at least, to the markets, they’re saying the ROI, the incremental ROI, isn’t there. To Rory’s point, it’s not there.

Jason, how did you read that? The Uber COO said they spent the whole year’s Anthropic credits in 4 months, but they weren’t seeing the gains or the productivity or efficiency gains that they—

Guest 2

Did he say the second part of that? I just want to be clear.

He said it wasn’t measurable. He said he thought it was probably there, but it wasn’t measurable—the COO.

Guest 2

Yeah. Right. Oh, got it. What do I think? Well, first of all, a COO is often a bean counter with a different name.

I think we all need COOs to scale, but sometimes they’re not the most creative person in the organization. Sometimes they think they’re product-centric, but they really just sit in meetings. I always take it with a grain of salt when a COO says, “We don’t need more engineers. We don’t need more product. We don’t need any of this,” because they’re not usually from a deep engineering or product background.

They often live in a pre-AI world where products are static. I would argue Uber, in many ways, is a static product. Uber, Uber Eats, and everything else—these are not radically different products from 4 or 5 years ago. These are clever marketplaces that scale like nobody’s business.

The second point I’ll make is that I hate to say everything’s bifurcated, but I do think we’re going to enter a world where things are more and more bifurcated. We will see folks who get more and more gains from AI. They will token-max forever.

Often, folks well north of $1 million in revenue per employee—$2 million in revenue per employee—will token-max until there’s no tomorrow. If you’re already hyper-efficient, you will find more ways to use AI. The folks that are less efficient, that are larger organizations and more traditional, I think will become more skeptical as the year goes on, especially if prices go up.

Maybe prices are—if you look at Anthropic raising prices, if you look at Nebius raising prices—as prices go up, the era of experimentation is ending, and people will see very different results from that.

I think it’s going to be the core question of the year. It’s going to be the question of the year. I think we said that last week, right? And, as I say, one of the reasons I raise it—it almost sounds like heresy to raise it—is...

Rory O’Driscoll

It's like maybe you won't get a return on your marginal dollar of AI. As I say, it feels like sometimes when I say it in San Francisco, it's literally like wandering around the Vatican saying, “Is this pope guy the right guy?” Right? It's criticism of the core thing.

But I think you've got to—when it's the only question that determines whether you get to $3 trillion in revenue and capex or not—you've got to ask it. I was reading—it's funny, over the weekend, our data scientist actually sent me an arXiv paper, “The Price of Progress: Price Performance and the Future of AI.” It's just a recent paper, basically benchmarking exactly this, right? The summary, no surprise, is basically, yeah. The comment they're making is, don't just look at benchmarks in the abstract; look at benchmarks and pricing at the same time, because implicitly, that's where you have to start thinking about ROI.

Obviously, benchmarks get faster. Price per token and all those metrics are going down amazingly every year, but token consumption is going up every year because, as these models do more agentic reasoning, the total compute cost of that goes up. So the net cost of serving the customer goes up in aggregate, right? It's a reasonably clear concept, and they don't actually come to a definitive conclusion, but it at least starts to say the trend here is exponentially increasing costs as you move from simple chat interaction all the way to full-on agents, right?

So even though all the cost-per-token stories are, “Yay, amazing,” going down, the actual cost to do something goes up a lot. So as you look at those SWE benchmarks, right, the thing that's here is not just one-tenth. It might be 100 times the thing that's next level out to the right. So you can consume vast amounts of token dollars.

So you're right, Jason. At some point—I mean, and maybe it is what you said—maybe there are businesses that have a lot of white space ahead of them where they can write a lot of software and get a lot of value, and maybe there are other businesses where you hit a marginal return much quicker. I don't know. I don't have an answer here, but I do know that it's probably the most important question.

Because when you're spending $3 million, you can be laissez-faire and say, “It's all fine. We're probably getting a return.” When you're spending $300 million, someone probably needs to know.

Guest 3

The other thing is, I think that Uber's gross profit margins are 39.75%. Okay, this isn't NVIDIA, but it's an awfully good business.

Yeah.

Guest 3

Right. It's an awfully good business, and I think folks like that are going to have different perspectives on the ROI of AI. When you're at those 40% margins, it sounds great, but you're very focused on protecting them in most cases, right? We're not all Zuck, right?

And so if job number one is protecting the profit margin, you're going to be really skeptical of spending too much on AI, right? It's a trade-off. In some ways, folks with margins like this, we just have to take what they say with a grain of salt. Not because it's not important—it is extremely important to understand that we can't all spend like a drunken sailor with $2 million of free credits out of YC, right? We can't spend that way.

But it also means there's a backward-looking bias in here. There's a skepticism in the organization because I think what the CEO also said is, “It's there, but we can't measure it yet.” So if you're Zuck—not literally Zuck, but if you're Zuck—you’re going to lean the hell in, right?

When the team is telling you it works, you're going to say, “Let's give it the rest of the year.” If you're a skeptic maintaining those profit margins, you're going to be like, “Okay, this is performative token maxing. Get me out of here.” Right?

And I think maybe there's going to be a case study of DoorDash versus Uber. DoorDash is still founder-led, one of the most aggressive companies out there, right? I mean, GJFC, I haven't seen DoorDash say, “Let's use less token, guys.” I haven't seen it.

If I see it from both—now, interesting—if I see it from both, from one of the most aggressive founder companies out there, Uber's one of DoorDash's biggest competitors, then I'll consider myself chastised, right? If Tony says the same, then I'll be like, “Okay, we got ahead of ourselves.”

For real companies, we're not all— But I haven't heard it from DoorDash. Is it not a very simple realization of what your core business is? Uber's core business is moving people—hard logistics, the real world—and Facebook's is precision advertising between consumer and provider, or advertiser and consumer. So, of course, you're willing to spend on one: it is your core business, and then there's the physical world.

Guest 4

No, that sounds credible on some level, but I mean, if you think of it really in terms of both of them having lots of software developers, the big-picture question is, how much leverage does a software developer get using AI? I would argue that should be much more similar across different entities.

It's not obvious to me that a software developer working at Uber still has a software problem to work on. If AI gives a lift across all industries in terms of software development, then he or she should experience the same lift.

So no, what might be true, Harry, is, to Jason's point, if you start with 90% gross margins and you're ideologically committed to the program anyway, then you might close your eyes. And I think it's exactly what Jason said: you don't impose the burden of proof.

There's no one at Facebook yet—though that may change—saying, “Are we getting value from our billions of dollars in AI?” We just assume we are and keep going, right? I do think in companies where they have the same idea and the same software engineers, but where the way they run their business, because of the overall margin structure, is more careful, you might have people asking the same question.

I mean, to be clear, just to nail your point: if what you're saying is that you only get an AI software-productivity lift in industries that themselves are digital, that would be really bad news, because that's a significant subsegment of the economy. I'm saying your willingness to spend—

Is it willingness to spend in a digital business versus a normal business?

Guest 4

Agreed, then we're saying the same thing. The more you're ideologically committed to the belief, the more you spend. But just to put it out there, I can't remember—except maybe for a year on the internet in the mid-'90s—where corporate America was so convinced of the ROI of something as it is right now of AI.

The floodgates are open; the money is being authorized. I mean, the proof that the floodgates are open is you just look at Anthropic going from $5 billion in run-rate revenue to $10 billion in run-rate revenue. No one is stopping. There's not a ton of questioning going on right now, right?

So there is a willingness to spend, and I therefore do think the next shoe will be, “Okay, that was fun. We spent $10 billion. What did we get?” It's just the way the narrative has to play out.

Well, Roy, I think that was a very prescient recognition of Anthropic's revenue growth, but you missed the core number there for me. I think you missed it personally. Gross margins expanded from 38% to 70%, and they're going to have a $559 million operating profit in Q2, projected.

Guest 4

Yeah.

Guest 3

Wow.

Guest 4

Predictable. This business is getting better.

Yes. And if you remember, we talked about last year's margins, which I think were 34%. Don't quote me, right? The prior year they were negative 60%. So the trajectory was really strong, right?

In other words, the margin profile went from negative at the gross-margin level in 2024 to positive 34%. Right? That's 2 data points. That's a trend. So no surprise when, on top of that, your revenue keeps going. You have more fixed-cost recovery and you've got a little bit of pricing power. Not surprised our margins went up.

And as long as your margins continue to improve, if you're doing $5 billion in revenue and losing a little money, and then the next quarter you add another $5 billion in revenue at any kind of margin—let's just say, even if it was not 70%, let's say it was even 50%—right? You have $2.5 billion of cash coming to cover opex. It's just hard to spend that much money.

High growth and improving margins meant profits were inevitable. So, yeah. Now, will they keep them? Are they about to spend a whole ton with SpaceX? Yes. But I wasn't surprised.

Funny, we run an internal model just to understand the ecosystem, and it's funny this happened because Anthropic had only talked about getting to profits in 2027. I kept running this idiot internal model, which is just growth, gross margins, and training costs. And I'm like, it's hard to lose money with this level of growth and any kind of decent margin structure.

So I wasn't surprised by it, but obviously, damn impressed.

Guest 3

One caveat: some folks say some of it may be partially due to a discount with SpaceX in the first couple of months, right? So I just put an asterisk and a dagger. It doesn't have anything to do with Roy's trend; it just may mean that the headline metric is not as impressive as it looks.

It is great, but the flip side of it is there's a bit of a bear case on Anthropic out of this, which is that Anthropic has benefited from 2 things versus OpenAI, at a minimum, right? It has become a premium product since the end of last year.

It is twice the cost of its competitor, right? Especially with Opus 4.6, it started to charge enterprises clearly per token, per use case. They charge them directly. The price has doubled. It is a premium product. It is twice the price of its competitor.

It lucked out, in a sense, by not going into video, which ended up being the biggest cash sink that OpenAI ever did. Probably Grok abandons it too. Everything is go, go, go at Anthropic; every single light across the room is green. But there is a bit of a bear case on Anthropic out of this, which is that Anthropic has benefited from 2 things versus OpenAI at a minimum.

It has become a premium product since the end of last year. But if the market—if even half the world says the ROI is not there for a premium product—Claude's growth is impacted. Either it has to cut its prices to match the competition, or it will maintain being a premium product like Apple but won't be able to maintain its market share. At the end of the day, it's twice as expensive as its competitor. Given the pace of this world, how long can it maintain that?

I'm not sure it's clear that it can maintain it indefinitely. It doesn't feel like it today, but that might be a hint of the bear case: the premium cannot be maintained. The demand is there, but the premium may not be able to be maintained.

Right now, even though they are the premium product, the proof is the revenue traction. The revenue traction from Q1 to Q2 is so strong. What it means is, as I say, it requires a change in behavior from where the default switches are now in corporate America for that number to change.

Rory O’Driscoll

Now, you're exactly right, Jason. It could happen, and it will all be about probably 2 things. One, is the ROI there? We just talked about that. Second, are the competing products in the marketplace competing strongly? Obviously, we're talking about OpenAI and, to some extent, Gemini.

The truth is, right now, Anthropic stole the show in enterprise, and 1 of the other 2 was distracted. I think that's OpenAI with consumer, and then Google, despite having amazing raw tech, just didn't do a great job productizing. So right now, the inertia bet is more of the same, and something would actually have to change in terms of competitive dynamics for that to be true.

As I say, I think there can only be 1 of 2 things. Either the green eyeshades come out and start talking about ROI, or the competitive dynamics change. Absent 1 of those 2 things, the trend is your friend, as they say in trading.

Listen, the Uber complaining and the Microsoft stuff—we can take some potshots at it, but I also just think it's showing you the future. More folks are going to come out and say the same, by definition. It's going to happen. The bear case is just that, as that happens, it might seem like the reaction is, “No more AI, guys. Go back to pulling up your old IDE and coding by hand.”

More likely, we will just get better at token budgets. As folks struggle to manage it themselves, that may degrade pricing power for Anthropic, as people say, “You know what? I'm just going to use a combination of older GPT-4, some Sonnet, and some DeepSeek, because that's the way I'm going to spend my budget.” It just benefits them. It's just a bear case, that's all.

Guest 3

I think you're exactly right. The reason I think this is going to happen—that's why I was reading that paper—is this: at some point, when you go from experimental to, “We spent $300 million on tokens. We don't want to change our guidance. We got some lift in efficiency, so we can tout that we've done $100 million of better operating income. So, yay us.”

But that means we have $200 million of opex we didn't expect to have, and we have to terminate a whole bunch of people. At that point, HR comes in, right? If you think you can do that just on vibes or no ROI, you're fooling yourself, right?

I actually agree with you. Once AI spend starts to eat big chunks of wages in large companies, you're just going to have to have that discussion, and it's going to become more quantitative. It's going to have to be more provable, right?

And not just for human and political reasons, but just because economically it makes sense. Someone should stand there and say, “Okay, we've never done a layoff in 5 years. We're about to terminate 10% of our people. Are we sure that we're getting the ROI from this investment?”

If the person leading the investment says, “I don't know. I think so, but we haven't really checked,” hopefully a CEO will say, “Well, before we terminate 1,000 people, shouldn't we check? Perhaps take another week here, guys?”

4. The Layoffs Continue: ClickUp and Cloudflare

The ultimate symbol, though, is that the layoffs we're seeing are ultimately a case of overhiring during COVID, not AI efficiency.

Guest 3

That's the dumbest thing I've ever seen on Twitter. This is the dumbest take from the smartest people, including Marc Andreessen, who's got 40 IQ points on me. It's the dumbest take I've seen since our last show. The dumbest take.

Whoa, whoa, whoa. Why?

Guest 3

Harry, I'm starting to forget COVID. I mean, what is natural attrition across your portfolio? 15% a year, 20% a year, 25% a year, right? This is not overhiring. It might be that your worst employees didn't leave. That's a related but different issue, right?

That's that. Actually, I think that's kind of part of it: the unreskillable need to go. But this is not overhiring with 20% attrition. I can't—what's 20% times 6? How does that compound?

Yeah, exactly. You get to the same place. So you're saying the 16,000 cut from likely Intuit, the thousands cut from Coinbase, and the 800 cut from LinkedIn—you're saying that is not overhiring. That is AI efficiency?

Guest 3

I'm saying there was already more natural attrition since 2020 than these numbers. So blaming it directly on overhiring is clickbait. It's clickbait because it's not mathematically true.

I saw a really cool tweet that basically said, “It's 5 years since COVID, people. Haven't you heard of performance reviews?”

Guest 3

That was me.

That was you.

Guest 3

That was my tweet. That was my point. You have no excuse. You could manage—you had 5 years to manage your low performers out, and now you're blaming it on overhiring.

Hang on. The only thing that could be—what? Just articulate the case. The only thing that could be true is that you hired to this level and you have this kind of human anchoring whereby you perceive you need all these people.

Every time they churn, you just rehired because you had myopia and you didn't realize you could get by on much less. Now suddenly terror has opened your eyes. It is possible in mature organizations with monopoly-type profits to see that behavior.

But I'm with you, Jason. It's weird to see it across the board, right?

Guest 3

Yeah. It just can't be true. It's an excuse, and I don't totally get this mea culpa, this ClickUp-Cloudflare mea culpa on Twitter. I do think they're 80% honest and direct, but this bloat thing seems like the excuse.

Jason, can you just explain the ClickUp and the Cloudflare situation, just so people have some context? They're publicly explaining why they're doing 20% layoffs in 1,000-line, 2,000-word tweets, and I just don't understand the PR benefit. I'm literally saying I don't understand the PR benefit.

The meta problem I find is that, even if it's true, they're sort of blaming the employees—or, even if they're not, it's perceived to be blaming the employees—for not understanding AI, not being able to change, and not being able to evolve. So we've got to let you go.

Before the Cloudflare and ClickUp layoffs, they were always like, “Our very best people have to go. Oh, please hire them. Here's a Google Sheet. There's no one better than Harry, Rory, and Jason we've ever hired.” It had nothing to do with organizational change or the world, or that there were low performers. It was just a flip of the coin.

We had to let somebody go. We flipped the 3-headed coin, a 4-headed coin, and Harry, Rory, and Jason came up, but they're just as good as the folks we kept. That was the vibe, even through last year. Now it's ruthless.

With Cloudflare, Matthew Prince is like, “Cloudflare is doing great, and we're still walking 20% of the people out the door.”

It's funny because you're damned if you do, damned if you don't. The Intuit CEO did say it's not about AI and also got slammed for that. People said, “Of course it's about AI, because look at these deals you did with Anthropic. You're being competed with.”

It's always about AI one way or the other, isn't it?

Guest 3

Well, directly, or our growth is slowed because we don't have an agentic product, or we're not being renewed by CEOs because they need budget for Anthropic tokens.

The world isn't static, right? That's the honest message. That would be my tweet, guys. Sorry. The world isn't static. I partially screwed up. This is what I'm doing to try to fix it. Blame me. There's no easy way to give bad news, right? There's no easy way to lay off 1,000 people. So no matter what you say, you're going to get dinged on it. And I think it's been evolving over the last 6 months how people have approached it.

I liked what Zeb said, though, from ClickUp, even though he got hazed more than most. One of the things I liked that he said was very transparent. It's like, listen, I'm laying off 22% of my company so I can pay $1 million to my high performers in the age of AI. He didn't really say, “It's the market,” but that's one reason you have to do it.

The other point is, if a 10x employee—if a 10x engineer is a 100x engineer, or a sales rep bringing in $1 million in bookings a year is now bringing in $2.5 million—I got to pay them. I got to pay them. I got to pay what was a $300,000-, $400,000-, or $500,000-a-year employee $1 million because they're delivering that much value. And I've lived it. We've gone—not to be a broken record—we're doing with 2.5 people what we did with 20.

Everyone's getting paid more on my little team, right? Everyone should make $1 million because we don't have the 18 other people's salaries to pay and we're more productive, right? And I don't know if it's going to happen at Uber's scale, but it's going to happen all across startups, right? And with $2 million or more per employee in startups, you got to pay, man. You got to pay your high performers.

And to be fair, you're consistent, because when I did mine 3 months ago, I was doing my 4 different ways, 4 different reasons for AI layoffs, only some of which are AI. You added that fifth, which is, “I'm terminating people to get different people,” which effectively is what this guy's doing. He's saying, “I'm terminating people to make room for people who might cost more per head, who have just different skills.”

Guest 3

Well, that might be what he's doing, but what Zeb really said is, “I just want to pay my current high performers.”

It's to say, “I'm making space for people to whom I can allocate more salary dollars.” And the funny thing is, guys, this is the counterargument to the last discussion, right? We're actually repeating ourselves in a weird way. We just had a discussion on whether 20% of R&D spend on tokens makes sense. And if it does, that's—you know, first, because the first section of the podcast was: Does token spend make sense? Does it have a positive ROI? Is it good for Anthropic? And thus, it's good for Anthropic.

Then the second section is, “I'm shocked to discover there's layoffs.” And if you zoom out a million miles, that's the consequence of the first, right? These are actually the same discussions. If there are efficiency gains from AI across R&D and sales and marketing in particular, and then on top of that, if there's also product impact—which I always like to separate, internal impact versus product impact—if those things are true, then everyone's going to be doing 20% layoffs just because it's 20% more efficient.

My aha is this discussion and the last discussion are, in fact, the same discussion. In one case, we're focusing on the winner, which is Anthropic, and in one case, we're focusing on the loser, which is employees.

Well, I think—yeah, but the top employees. I think what ClickUp shared—I mean, we read that, you see them on social media—Anthropic's hiring a social media manager in King's Cross for $450,000 a year. And you're like, how could that be? But their revenue per employee is so astronomically high that it all ties, right?

5. Data Centers In Space is BS and Will Not Be Core to SpaceX

And I think the new paradigm—when the 3 of us met, $2 million per employee was what you reached someday, right? That was Apple and others. You didn't even have to think startups never got there. I think $2 million per employee is going to become the new normal as you scale, and that will mean that your high performers can make 2–3 times what they used to make, just like at Anthropic, because you just don't need as many people. And you're going to do what ClickUp did: funnel your compensation to your high performers instead of the high performers making 40% or 50% more than your mid-pack. It should be 5x.

What you're basically saying is a new set of tools have come to the fore in the last 2 or 3 years. And if you know how to use those tools, your productivity relative to someone who doesn't have those tools is so great that we can afford to pay you half a million bucks because you're obviating the need for 3 people.

Guest 3

Being an agentic expert should be like Excel in 12 or 24 months. Everybody should be able to do it, because prompting's gotten so much better you don't even need to know how to prompt anymore, right?

It should be true, but I have some lingering concerns that instead what happens is the agentic experts get another 10x better, and everyone else falls further and further and further behind and becomes unemployable. We go to $4 million in revenue per employee and $5 million in revenue per employee, and it's exhausting. I'm the most tired I've ever been, in a sense, but my productivity rate is off the charts.

Rory O’Driscoll

You're right, Jason. What you're saying is, if you're on this exponential curve that Dario and Sam keep talking about, then the knowledge to be at the frontier keeps growing every year, and it's just super hard to keep up, right? In which case, it does get pretty brutal.

I mean, one of my favorite Max Planck quotes is, “Science advances funeral by funeral.” In other words, people don't learn and change their views; it's just the people with the wrong views die, right? And it's pretty brutal, but it's true, right? Gradually, all the people who believe the Earth is flat died, right?

And retraining is hard. It is, which explains the unpopularity. It's uncomfortable to have to learn new stuff beyond a certain age, and it must be pretty dystopian to have to do that or die.

Don't worry, Rory. As a podcaster, now you're safe. The venture investors are screwed, but now you're a podcaster.

Rory O’Driscoll

Good to know, Harry.

I got you. Don't worry.

Rory O’Driscoll

I look forward to seeing the passive income from your advertising dollars coming into my account.

Guest 3

I've looked, and they haven't hit, Harry.

6. OpenAI S-1: Is it a Race? How Will it be Received?

Oh, don't worry. Don't worry. They're coming soon. They're coming soon, just like productivity gains. Okay.

Listen, S1, let's start with OpenAI. Rory, you teased it out at the end of last episode brilliantly, as it kind of came breaking through. OpenAI confidentially files an S-1 at an $852 billion to $1 trillion valuation for a Q4 listing. The main question that I had, I think, probably is on the minds of everyone: Is OpenAI forcing its way out before Anthropic? Let's start with that.

Rory O’Driscoll

Yes. They have to, they should, they must, and they've got to go now because they've been lapped. Last year, Anthropic GAAP revenue was $4.5 billion; OpenAI GAAP revenue was $13 billion. Q1 Anthropic GAAP revenue was $5 billion, more than the entire last year.

Q1 OpenAI revenue was $5.4 billion, so still bigger than Anthropic, which is the point they were making by leaking that number. But if you look at it, it's only, what is it, 35% of last year's revenue. So Anthropic has done as much in Q1 as all of last year, and OpenAI has done 30% of what they did last year in Q1, and they're only slightly bigger than Anthropic.

You play that out for a couple more quarters. If Anthropic does anything close to $10 billion in Q2, Q3, and Q4, that's $35 billion in GAAP for the year, right? One company is growing 10x year on year and the other company's growing 2–3x year on year, and they're pretty close now.

Within a couple of quarters, Anthropic, on the current trajectory—and it could change—will be visibly and obviously ahead. It could be something like high-$20-billion GAAP revenue for one and high-$30-billion GAAP revenue for Anthropic. That's not a good look, right? Especially when the one with the high-$30-billion revenue is profitable, growing more quickly, and bigger—dominant on all 3 vectors—and OpenAI could be smaller, growing less quickly, and still unprofitable. That's a horrible strategic place to be, right?

That's a tough message to come into the market with.

Rory O’Driscoll

Yes, it is. But I'll tell you this: you've got 2 choices. You go first, and at least you can say, “Here's a chance to buy the first foundation model. Go now.” Right? ChatGPT—everyone knows who you are.

Or are you waiting until the other guys go, and then you go right after them as, “Hey, we're like Anthropic but smaller and not as good”? Or are you saying you wait 2 years and grow your way out of it because nothing's forever? You reaccelerate growth and go public in 2029. No, that's craziness, right?

Unfortunately, the rule of life in all these markets is the number 1 gets to choose what happens and the number 2 has to respond. OpenAI was number 1 and had those degrees of freedom. Now they could be number 2. When you're number 2, what you do not do is wait until the number 1's crisis goes public and then come out as a pale version of that.

So I think they're smart to go, and they've got to go.

And this might be one, actually, where the risk is you exhaust some of the capital pool available, because these IPOs—this is always something people would say, but it was never really true, right? Traditionally, folks that bought into the IPOs—there's a fairly slim segment of the market, because you can always buy the next day.

Now we have Cerebras, which really was maximally oversubscribed, right? As Rory made the point, it priced to the exact highest thing you could do without refiling, without taking the risk of refiling with the SEC.

But it’s still a relatively thin investor pool, right? Even with Schwab and everyone coming in for SpaceX. So this might be the one time where it really does help to go public, just to not exhaust the capital pool.

Rory O’Driscoll

And tell the story. Given the trajectory of the last 5 years, they’re entitled to tell the story. They did build the category, right?

7. Do Anthropic Rush Out Their IPO Also?

And I would much prefer to go public as the first foundation model in the category. If, in fact, the numbers that we’ve had leaked—and I always caveat this because no one sees the numbers—are correct, such that by the middle to latter part of this year you’re visibly number two, losing money and growing more slowly, then you do not want to wait until that becomes painfully obvious.

If you’re Anthropic, do you change anything on seeing OpenAI now force this out?

Rory O’Driscoll

Again, I’m going to repeat myself: because you’re number one, you don’t have to change based on what the other guy does. This is a totally historical analogy. I remember when Britain ruled the waves. I like to do my little Britain digs here. When Britain ruled the waves and the British Navy ruled the world, Jackie Fisher, the First Sea Lord, was asked, “Why do you have the British Navy?” He said, “To go anywhere you damn well please with whatever force you want.”

It was, you can do whatever you want. That’s the point of being number one. You don’t want to be sitting there agonizing about what the other guy does. It doesn’t matter. Anthropic could decide, “Oh, my God, OpenAI is going to go public in September. Maybe we go in November.” Maybe they do bring it up a little.

Well, to your point on exhausting capital supplies, we’ve got SpaceX, so I’m less worried about that. To some extent, it’s a thing, but remember: if you’re profitable and growing nicely—if you are the better company and you have the more attractive metrics—the way this market works is that if the other guy goes out 3 months ahead of you, everyone who’s investing will think, “I’m going to put some money in this, but I’m going to keep some money for the good one,” especially if they’ve telegraphed when they’re going out.

The truth is, this operational excellence gives you a fair amount of strategic independence. You have way more degrees of freedom.

Guest

Or maybe if Anthropic truly is profitable and truly has infinite demand for its shares, it doesn’t need to IPO at all. Why would I? If I’m Dario, I’ve already pledged 90% of my shares to charity. I still have an altruistic mission at my heart. If I’m profitable and I have infinite demand for my shares, maybe, ironically, they stay like Stripe. They’re like, “Listen, we crushed it. Why would I want to float a small amount of my shares and deal with the headaches? I am going to stay true to my mission.”

I’ll tell you why. I actually think that the dynamics of staying private longer are super different for people like Stripe than for Anthropic. At Stripe, you’re in a fair degree of predictability. The business model is understood, the capital needs are light, you’re kicking off cash and you’re buying back shares. Those guys can actually stay private as long as they want. It’s a payment-processing company at scale with profits. It’s a glorious thing.

Who knows how it’ll be valued? That’s still to be determined in the public markets, but you can chug along forever. I just think the capital needs for Anthropic are such that you would be wise to access the public markets when you can. Unlike Stripe, where you know your capital needs are light, in the case of Anthropic, even though you’re not bearing those capital needs today because you’re lucky enough to have hyperscalers foolish enough to build on your behalf, for every dollar of revenue that you add, someone has to invest $4 or $5 in CapEx up front.

So if one day you want to do $100 billion in revenue, someone’s got to invest $500 billion. You probably want the operational freedom to get some of that money for yourself, and there’s a dollar sum above which only the public markets can do it. Again, people talk about, “Oh, my God, the private markets.” SpaceX, as its prospectus points out so wisely, was pretty capital-efficient: $23 billion raised. It’s not $100 billion.

It’s entirely credible—entirely credible—that if Jensen’s correct, going back to the thing, and CapEx is $3 trillion a year, and you’re the leader in that industry, you might need to raise $300 billion. Maybe 2 years from now Microsoft doesn’t want to, or Amazon doesn’t want to, be your hyperscaler provider. So I think they go public because the capital needs here dwarf anything we’ve seen.

Or at least it may provide some independence, right? NVIDIA’s got $200 billion of cash a year it needs to do something with. The other thing, going back to our conversation, is that’s why I don’t think this circular revenue is as negative as you think. NVIDIA has so much cash it has to recirculate this capital. It has no other use for it.

Going back to this, before we do SpaceX, if OpenAI were to go out, what do you think the reception would be? Do you think it will be overwhelmingly positive?

Guest

In today’s market, yes, because we’re risk-on and people want to make an AI bet, and they’re the leader. They are the leader, and they built the category. You go outside the Valley and utter the word Claude, and people look at you. ChatGPT is the thing, right?

They’re going to have stellar revenue, albeit clearly losses aren’t a problem at the moment. When you look at where SpaceX is going to come out, I think that would be really good. Really good might mean not quite as good as Anthropic, which is all that plus profits, too.

But right now, the public doesn’t have any pure-play AIs other than something like CoreWeave, now something like Cerebras. It’s obvious that the top of the heap are the foundation-model companies, and I think it’ll get a great reception. That’s why they’d be crazy not to do it.

You’ve got to remember, all this leaked “we’re not going to be ready yet” from the CFO—2027, 2028. I would imagine we had a good internal session. We were like, “We’re doing this, guys. Thank you for your input. But the day after the litigation is settled, we are going for this thing.” I think they’re right.

I just think 2 things. One, we’re going to get a little preview, which is: if this retail demand for SpaceX is super high, then OpenAI can copy it, right? Elon’s putting 30% of this massive IPO—the largest in history so far—into retail.

It sounds like he’s being democratic and meritocratic and Robin Hood, and maybe he is, but I think it will benefit the stock price because they want to own Elon—not the jury in Oakland, but a lot of the country wants to own it. If OpenAI clones that for its fan base and it drives the stock up, they’ll get a little preview of how well that strategy works.

The other thing is, it’s fun to talk about this 2-horse race on the 20VC, but again, I don’t think there’s anything wrong with owning both as a public-market equities investor. I think owning both is a good strategy. Why? There’s so much change. Even if Anthropic is better, why not buy into the IPOs of both? Put in $50 million into each, $100 million, $250 million into each, and see how it plays out over the next few years.

The overall trends are strong. Not everyone needs to be a savant. One of the ways I think about this is that these are 3 CEOs trying to do something incredibly hard, which is stay on top of the most dynamic, evolving industry and tech trend we’ve ever seen.

And no one—none of the 3 of them, and by that I mean SpaceX, Anthropic and OpenAI—is pitching a perfect game, right? You can look at aspects of what any of the 3 of them have done and go, “Oh, that’s not as good.” But the truth is, all 3 of them, from a standing start, have built companies worth plus or minus $1 trillion and are kind of riding this trend, and no one else is even close.

I mean, if you look at the report card, Anthropic did an amazing job of building the model, which ironically meant that its CapEx forecasting was wrong because it underestimated the success of its own model. To some extent, OpenAI has been the other way around: they really nailed it on CapEx, they nailed it on compute, but maybe should have spent a little more time on the enterprise features of their model. So they’re behind there.

Obviously, if you look at the xAI part of SpaceX—leaving aside the other thing, just focusing on their AI business—they’re the world’s best builders of fast data centers. So they get points for that. On the other hand, they lose points because they weren’t able to fill them with their own stuff. You’re right, it’s perhaps not quite as modern or as SOC 2-compliant as some of the stuff OpenAI got.

So, yes, you can make knocks on all 3. But zooming out, from a standing start, they’ve each created a trillion dollars in AI in the last 5 years. What did you do on your summer vacation?

Well, Elon founded both.

Guest

That you have to remember. I mean, yeah, the entire pivot—

Plus Neuralink, plus he’s controlling people’s bodies with their brains, plus the Hyperloop, plus Tesla.

Guest

Yeah, because we can talk about that. The entire—when we come to talk about SpaceX, the proportion of their future value that they attribute to AI—this is Elon effectively losing OpenAI and saying, “God damn it, I’m going to have one of my own,” and made it happen, realized it from whole cloth.

8. SpaceX S-1: "Why I Would Never Invest"

Well, let's discuss SpaceX directly, then, with the S-1 dropping. Rory, given the fact that I didn't correctly identify the most important numbers, what did you think were the most important numbers when you looked at SpaceX's S-1 dropping?

Guest

Cynical answer: I don't think the numbers mattered at all. The point is this: everyone said the same thing. You run the sum of the parts on their 3 businesses. There's a low-ish-growth but tech-enabling, amazing space business; there's a good-growth, profitable Starlink business that has a much bigger TAM than the simple rocket and launch business, right? But it's growing 30% to 40%.

And then there's xAI, which at the time of filing was a great big, gaping hole of $15 billion in CapEx and no obvious revenue. Since then, obviously, he's done 2 amazing deals to find the revenue, and now he's got a $15 billion run-rate business. But you add it all up, right? The truth is, you get to a launch business, you have a Starlink business, which you can value, and you have a CoreWeave-type business, because that's what the AI business has now become.

The sum of those parts is so far lower than the proposed valuation, and the only difference between the 2 is the Elon premium. That's not a number that comes out of the prospectus. So I looked at the numbers, I read them, I understood them, but they offer me no value in trying to value this company, right? Other than saying utterly stupid things like, the fundamental value of Tesla is $200 billion and the trading price is 6 times that, so maybe the Elon premium is 6x value. I don't know.

It could be the GeoCities deal of the AI era. We could look back on this and say, listen, I'm on Team Elon, okay? But 100 times trailing sales—we may look back and say these were some good companies, but 100 times? My God. No, no, no, no, no, no. I'm going to defend them. No matter what happens, these are 3 amazing companies.

Guest

Twitter's growth has fallen off a cliff. It's shrunk 50% since he bought it. Elon's the magician of revenue compression in Twitter: revenue has fallen 50%.

Yes, Twitter agreed.

Guest

Twitter is—I mean, again, if you disaggregate the 3 businesses, the launch business is a good, stable, boring business with amazing technology. It probably would get capped out at a 10% or 20% growth rate. The Starlink business is a great business with a much bigger TAM and a 30% to 40% growth rate, with, I think, $14 billion in revenue and EBITDA-positive, right?

I'll give you the most interesting chart, Harry, and the number in the thing with the TAM analysis. I want to say, if I remember, something like $28 trillion—the largest TAM in history. That's actually not as interesting. The interesting part is, even though 10 years ago this was a launch company, and 2 years ago this was a launch company with a communications business, 90% of the TAM that they identified is not in those 2 sectors. It's all about AI, right?

9. Why Colossus is a Stroke of Genius By Elon

The interesting thing is, the story they're trying to tell is a story that's 90% not focused on the 2 things where they're unique and differentiated. I think they're not unique and differentiated in the AI story, because Grok as a foundation model has gone nowhere. They did make an enormously clever deal in the sense that they had Colossus, they built it quickly, and again, the S-1's very clear on that: they think they build faster than anyone else.

They've sold that to Anthropic. They've rented that to Anthropic for $1.25 billion a month, with, I think, a 90-day cancellation clause on either side. So basically, it's $15 billion a year, right? In 1 sense, that's amazing, because it's amazing to think that you had an $18 billion business built over 20 years, and then you just added $15 billion to that in 1 transaction, on a product you only built a year and a half ago.

In 1 sense, it validates the comment that the AI business is so much more dynamic even than the launch and communications business. If you're willing to spend the capital up front and invest $10 billion in a hole in the ground, and the compute demand is there, you can get to $15 billion in revenues pretty quickly. But what you are is a more efficient CoreWeave. Now they will say that—

That's worth $2 trillion.

Guest

Renting out chips because Jensen's your buddy and you're a good customer? That's worth $2 trillion, $4 trillion, $3 trillion—

Or the rockets growing 10%, or Twitter declining 50%. Which of them exactly is worth $2 trillion to $3 trillion?

Guest

And that's why I jumped on the GeoCities thing. I love the S-1, right? I think it's all madness. I wouldn't buy a share; I just love the optimism. I just love the fact that they've done this thing for 20 years. Some of the assertions of why they're great were totally true: they've been fairly capital-efficient for people who put rockets in space; they build faster than anyone else; and they have organizing principles around engineering—the algorithm to make things efficient.

There's a lot of balls-out claims here that are justified on the basis of 20 years of achievement from probably one of the most talented entrepreneurs ever. It was a great read. Go America, that we can have this, right? So after I say all that, you're right, Harry: I still don't get it. I'm so far from a $2 trillion DCF that my head hurts.

Well, you know, it's almost worse. The 100x trailing revenue is hard to—it's hard to tie, right? But listen, I'm on Team Elon. I've bought 5 Teslas. I've got 3 Starlink subscriptions. I am from the very early days.

In some ways, this is just very cynical. This is financial engineering, which I have mad respect for, but my God, you're taking a bunch of disparate assets that most of them have no connection. You're throwing an S-1 talking about the history of AI, which wasn't even your business a year ago. A little bit of financial engineering—I'm all for it. But, man, this is so much fun.

This is bailing out my failed Twitter acquisition, my Colossus idea that I'll compete with OpenAI just by buying chips. That didn't work. I'm bailing both of them out, right? I had a quiet business, which I should have kept private, with SpaceX, right? It was a great private business. Starlink was doing great things. Now I'm bailing everybody out in this massive confab.

It's SolarCity on steroids, and we're all here for it because we love AI, but it makes no sense. This conglomeration of friends of Elon makes no sense to anyone but the folks getting bailed out on Twitter. It just doesn't make any sense.

The answer is, the launch business enables Starlink, right? The launch business will enable data centers in space, which is how—and we've proven that we can build capacity on the ground. Therefore, we'll be able to build capacity in space, and we'll be able to get 100 gigawatts of capacity a year 5 years from now.

Data centers in space is the thing that joins all the dots together, which is different from me saying, even for a second, that I ascribe a high probability to that. There is a coherence to the narrative, but only if the next thing happens. That's always the way: if you buy into data centers in space working, then it makes sense to have a launch business in the same company that has a data center business.

And the tweets—where do the tweets fit into all this?

Guest

The tweets don't fit in at all. [laughter] Let's get real here: the tweets don't fit in at all. The truth is that the $44 billion on Twitter was a miserable acquisition, and the value of that company is less today than when he bought it. That was pretty obvious 2 years ago, right?

He's chosen to bail everyone out by rolling it into xAI. He's chosen to bail xAI out by rolling it into SpaceX. And, to be fair to xAI, even though we just used the words “bail it out,” if they can keep the $15 billion a year they're going to cut, that's a very high—I mean, the total CapEx in xAI over the last 2 years, I think, was $12 billion and $7–$19 billion, right? One year and a bit of Anthropic revenue at $1.25 billion a month, and you covered your nut, which is a pretty high cash-on-cash return for a data-center project.

So I sat there thinking it's all crazy too, and then I'm like, he built something at a time when capacity is at a premium, and he's found a way to monetize it. I mean, again, Harry, I repeat: I don't get it. Even so, you apply the CoreWeave multiple and you get to under $100 billion. I'm not sure what to make of that information, but it was definitely a get-out-of-jail-free card when he got the Anthropic business.

By 2030, what will that core business be? Will it be Starlink, will it be data centers in space, or will it be a neocloud?

Guest

Good question. I think Starlink will be the vast portion of the value. I don't think data centers in space will be a meaningful percentage of revenue, and I think their existing data-center business will be a high-revenue, relatively low-return-on-equity business. How that compares to Starlink, TBD, but I think we'll discover that.

As I say, it's a very—it's CoreWeave with the world's best engineering and construction team. So implicitly, I'm saying not a whole ton of data-center-in-space revenue on top, and that will only be worth anything like the current value if the narrative is tracking toward data centers.

Again, it goes back to the first sentence of the show, which means if Jensen is right that you're spending $3 trillion a year and the ROE is there for the full $3 trillion, as Harry has pointed out, and if we can't build anywhere else, then at some point you'll want to stuff it into space, regardless of all the pain and suffering involved in doing that. And then Elon will be—

If any of those if-clauses turns out not to be true, then it'll be a bridge too far, or, more correctly, a rocket too far.

Guest 3

(Laughter.) But who knows? If the one human being on the face of the planet who's earned the right to say, “Give me a trillion dollars. I'm going to make the bet,” it's Elon. And if you want that bet, I'm glad it's going to be available for people. If you want to scratch that itch, go scratch the itch.

Look, I think he made a compromise to get it all done. I think he's going to have to build a $50 billion CoreWeave business to make the math tie. And I think he had to give up on going to Mars and go to the moon. We reach a certain point in life where, to achieve our goals, we have to be practical again. He's got into a practical phase. He's got to take this thing public. He's got to mash it all together. And this Anthropic thing either has to die or keep going to make this work.

So he's got to commit to building a $50-plus-billion CoreWeave and going to the moon, which is pointless, but it doesn't mean it won't get him there in the end. He could also, if the world changes—I mean, we're on such an intense trajectory—but if the world changes, he could also dump this whole CoreWeave business and just move on. You can write off the chips; you can write it all off if the world goes a different way than we expect. And proving how little we all know at this point, how little you can know, let me give the other option: if everyone else proves incredibly mediocre at building new data center capacity and Anthropic keeps continuing to grow.

I was thinking they did this at $1.25 billion a month. It's kind of like we all buy storage capacity. When you buy storage for your stuff, when you send it away to storage—extra furniture—and you pay $100 a month, and you think, “I'll be out of this thing in 6 months,” and then 5 years later you're still paying, and they're now charging you $400 a month. Basically, this is Public Storage but for compute, and he could well be there for the next 5 years with Anthropic just needing the capacity and making a ton of money. Consumer storage solutions are one of the best businesses. Yeah, agree.

Guest 3

Exactly. This is like that. You go in—I'm sure Anthropic went in thinking, “We're only going to pay him $1.25 billion for 4 months,” but they're relying on other people building data centers, and life is tricky, and not everyone will bulldoze through like Elon did in a year or 2 from now. And because, remember, this is very profitable revenue. You cannot take it away in his CapEx. In a shitty CapEx business where CoreWeave is a good company but makes only okay profits, at $1.25 billion it's a great deal, right, relative to his CapEx. So his ROE on that is pretty damn high in a low-ROI, return-on-equity business.

And don't forget, it's a poor analogy, but it's the same guy. There's no one in the US or Europe that can produce EVs remotely as efficiently as Elon, like the investments he's made. And I'm not saying it's the same; he can't control the whole supply chain. But the man has a history of doing things that are both at the cutting edge and radically more efficiently by investing hard, like Colossus and all of this.

Maybe he does build his own massive fabs in the US and they work. We can take shots at it like we did when the Tesla Roadster came out, but no one can approach what he's done in EVs outside of China. Maybe it happens in data centers. I mean, CoreWeave has nothing. What do [likely Nebius] and CoreWeave have? Nothing. No special technology, right? He could accelerate past everybody. He has the capabilities, right? And we could turn around in 5 years and say, “My God, he owns every data center in the galaxy.”

Because, I mean, if nothing else, who else? You can't compete with a Model 3. You can lease one, Harry, in the US for $299 a month. It's self-driving, it needs no fuel, and it's better than all its competition. And the only reason we don't all drive them in the US is that 60% of the country hates them. Otherwise, there would be no other car in this country, right? What if he does the same for data centers? It might be with Colossus and friends and everything. He's actually doing this for real.

Because, remember, the one thing you need to be successful in a capex-intensive data center business is a low cost of capital. And I think we can agree that a $2 trillion pre-money is a pretty damn low cost of capital.

Guest 3

Yeah, no, he can sell. I mean, if he gets $75 billion, he can put $25 billion into another Colossus, sell that for another $15 billion, and away you go.

There are 2 private-market stories that I want to touch on that we haven't touched on. One, AI slop spelled backwards is likely Polsia.

Guest 3

I feel like we're being punked.

Listen, it is a company that enables a single person to build a business, and that business is then run by AI. It is, bluntly, not very popular. It sent out a huge number of unwanted emails, which is why a lot of people don't like it. The founder of OpenClaw, Peter Steinberger, even responded to the funding announcement, saying how much he disliked these unsolicited emails.

But they raised $30 million to $40 million—I can't remember the exact amount—at a $250 million valuation. Some pretty well-known firms are in there. Jason, I'm intrigued. I think you'll have a thought on this one.

Guest 3

I feel like I could do it. I feel like I could just call it AI slop instead of Polsia. If they can raise it at $250 million, the only question is, if I did this, could I take the $40 million out in secondary? Is that okay? I would do this if I could do it all. It would be nice. Unfortunately, I wouldn't get QSBS on it, right? That's the negative, but I wouldn't mind taking $40 million. I think I could do $10 million like this if I knew I could.

I don't want to raise from VCs and be stuck with the moral obligation of turning it into $2.5 billion. But I don't know. Listen, I don't want to just—the fact that you have a startup named AI slop raising at $250 million, if this is the peak, we're going to look back and make fun of this one. But I don't know enough about it, in all fairness, to be sure I'm wrong on this.

It's just, I feel so punked by a startup called AI slop raising at $250 million. I just feel like this is the bigger FU than Moltbook. Moltbook pretends these agents are talking to each other in a social network, when it's all humans telling the agents to go onto Moltbook and talk to each other. I just don't know—is this the next Moltbook? I literally don't know. But I would at least have asked them to please change the name.

I did admire the tweak, though. It is such a two-finger salute to the great AI marketplace.

Guest 3

I'll tell you what I don't like about it. And listen, we can talk about what it's actually doing. I've just begun to build an app while we're on this Zoom, or this whatever. From a marketing perspective, it's pretty good.

I asked it to build an inbound AI SDR for me. I've already done this and bought and used a bunch of these tools. The way it qualified and specced out this product—pretty good, pretty good, right? Certainly better than Replit or Lovable, which are horizontal tools. The journey it guided me through—it did what Manus couldn't do while we're here. It researched SaaStr.com. It figured out what I did. It analyzed what tool I would likely want to build. It specced out the tool. It did a great job.

I give it a 10 out of 10 in terms of the journey so far, and then it immediately asked me to give a credit card and pay $49 a month before I got any value. My general experience in AI tools like this is that when they ask for money before they deliver any value, it's not that good. The marketing exceeds the value of the tool.

I can't think of an exception. I'm not saying they're out there, but every time I go through one of these, I immediately try to figure out where the cancel link is, and it doesn't exist. But I give them a 10 out of 10 for the before-I-put-in-my-credit-card journey. It's pretty good. They've definitely made you feel like they can build something pretty badass for you.

I have to say one thing that is impressive. I met them as part of the fundraiser, to be very candid. It is 1 founder.

Guest 3

There's no team. There's no them.

You're right. It is 1 guy, right? It's 1 guy, right? I'm impressed with the aesthetics of it. I know a lot of founders disagree with me these days, and a lot of accelerators tell you to charge instantly for these generic products, but I know the ones that I love, the ones that I use every day: they give you value before you have to put in a credit card.

And this is going back to the Y Combinator $2 million of tokens. Invest $5 in me, or $1 of tokens in me, or $2 to earn $49 a month, which is the entry point for the product. You can't invest $2 in me. I'm disappointed, right? But I get why a lot of young founders or others think it's not worth $2, but to me, I'm out.

Boys, what other stories have I missed that you think we should cover?

Guest 3

Exa raised $250 million at a $2.2 billion valuation to build the search engine for AI agents. Today, OpenRouter announced their round, which was, I think, $150 million at a $1.3 billion valuation, led by Capital G. The likely Manus founders are trying to buy back the company in quite a contorted deal.

Yeah. I mean, look, you know what all these have in common—and the first 2 have in common, and the third one's obviously very different—is super-interesting infrastructure companies, all about the buildout in AI, right? And the big picture here is enterprises are adopting AI. They're using LLMs, but there's a series of other tools that developers and people need to build great agents.

Exa caters to that need. It's done really nicely. In parallel, there are 2 companies doing that that have recently raised, and my point is merely there's just a whole bunch of infrastructure going on 1 level below the foundation models that are super interesting. We talked about OpenRouter before, which is a company that allows its customers to switch between various different models, both the foundation models and the open-source models.

They have, I think, 50-plus models hosted on there to allow enterprises to access the lowest-cost model. These are all interesting trends because I think the OpenRouter one is interesting because, going back to what Jason says, if people are spending $300 million on a premium product and having to lay off people as a result of that, there’s going to be some interest in exploring cheaper costs. I think companies like OpenRouter have a value proposition that says, “Hey, not every query you need to make needs to be done on the most expensive foundation model, and there are other places to do that.”

I think, again, the cliché—the pick-and-shovels of the agentic revolution—is just a good place to be investing, and I think it will be going forward. There are strong, interesting teams doing good stuff. You can squint one way and say foundation models hoover it all up, but I don’t think they do. I think there’s a lot to be done around building great next-generation agents. Jason, do you guys use Parallel or Exa, or what, for web search when you’re building your agents?

Guest 3

Yeah, and I’m a small investor in Exa. It’s just a great product, right? It’s another one where I was a user before an investor—an early user. The reason I think it’s super interesting, even more interesting than OpenRouter, which is a great one, is—

OpenRouter is really interesting because, first of all, they nailed the ability to dynamically and easily pick your LLM. They nailed something, and so it’s a proxy for all these discussions. It’s a great proxy insight into that product. Exa is more interesting than some of its peers because—

Guest 3

These are products that have no use without agents, and we throw this term “agents” around, but agents have a lot of meaning. Literally, if you’re not, at least as far as I’m aware, building an agent product, you have no use for a tool like Exa. Agents don’t hop on Google and do search. They don’t create files on Dropbox, and they don’t hop on Zoom.

True agents have a different set of workflows and tools they use, and Exa nailed one of the core ones. It wasn’t even obvious to me until I started using it, but there was an obvious issue: your agents do need to find current information. It’s just that simple. Just like Google was the killer use app for humans, your agents need it, but they don’t need Google.

Putting aside whether Google should have built it rather than promote Exa, it was a ball bouncing the right way to see the type of growth, even at the early stage, for something that agents need and humans directly don’t. It’s a look into the future, and it also says that a lot of this stuff isn’t BS. A lot of this stuff is real, and the idea that we may all manage more agents than humans isn’t just BS. It’s real. It’s real.

They need a whole different set of tools. They don’t need Zoom. They don’t need a traditional CRM. They need tools like Exa. Maybe we should only be investing in those things and leave those human tools behind. Guys, enough investing in human software.

10. Exa Raises at $2.2BN to Build Search for Agents

A very dangerous thing to do. Can you paint for me the upside case here at $2.2 billion? What is that upside case from here? Is this a $100 billion company that powers the future of agents? What is the market composition of this market? We’ve got Parallel doing it as well. Is this an Uber/Lyft, where one of the providers takes 90%? Is this an AWS/Google Cloud? Just help me understand that. I think it’s more the latter than the former.

Guest 3

It’s not a winner-take-all market. It doesn’t have huge network effects because it’s a developer tool. You can easily contemplate 2 companies being pretty successful here. We did some work on this. We looked at Exa back in the last round. I think they’re a super team with a good product. Someone bid higher, and they won the deal. Great name, too: Benchmark.

What can you do? I actually think it’s a super good company. I like the team. I think they’re smart, and I think the market’s there. I don’t think it’s winner-take-all. I think it’ll be a couple of players because it just doesn’t pay. You’re not going to have 10 because it’s not economically efficient to have that.

At the same time, it’s not like Google, where winner-take-all emerged because it was a consumer product, and then the ad network and the structure of the market for ad sales pointed to a winner-take-most. In this case, individual developers will be making decisions. So I can comfortably imagine 2 players here dividing reasonable market share, just like most developer tools.

I think the next level—how big can it be—is the question. I think Jason’s right. It’s all about the more you believe in agents, the more you believe they’re going to want to access structured information, and this thing’s going to co-attach to all that. So that’s the bet here, right? You are betting on lots of agents needing lots of information, doing lots of reasoning, and doing lots of search. It’s a derivative bet on agents, and right now, that’s been a good bet.

How you think about getting to $1 billion in revenue, which is probably what you need, in a normal market, to be worth $10 billion in market cap—I’m just thinking about it on the fly here because I remember we looked at the numbers a year and a half ago. It’s not crazy in a world where, going back to first principles, we’re spending $3 trillion on capex. That’s the Jensen Huang comment.

The second comment is, if you look at the SpaceX S-1—remember, I said that 90% of the market is AI—one of the next interesting things is that within that AI, they said 90% of that is enterprise. They’re basically saying the vast bulk of AI spend is going to be enterprise building.

If that’s the case, and they’re building enterprises to frankly replace human work, then I think there’s a pretty compelling market for the product that allows structured information to be accessible to that agent. I think you’ll expand beyond just web search. I think you’ll do curated lists. I think you’ll do internal information. There’s totally a need for that product.

That’s wild. Another Benchmark company. This fund is going to be one of the best-performing funds in history. Seriously.

Guest 3

Yes. Good for them. That’s the job.

That’s the job. Like every company, Jason, the number of hits is extraordinary.

Guest 3

Yeah. And so they did Exa at $700 million, right? Or something like that. I think you’re right. This is 20VC, so it’ll be an epic fund.

I think also going in early and realizing that investing with a smidge of traction and breaking your rules a little bit was the right bet for today—arguably, I think that was maybe another insight too, right? likely Jack Altman just joined, and he just did Monaco, which I know well. I’m a small shareholder, and they marked up the deal very quickly, almost 2x, from Founders Fund. When Founders Fund invested, it had no revenue.

So he got a much better deal for Benchmark by investing a couple of months before the markup. But going from pre-revenue to blowing up, right? That’s—if you can get in at the moment an AI leader blows up, that hour, that’s when you want to invest, right? And so then they’ve tripled their Exa valuation. Going back to “seeds for suckers,” you want to invest the hour it blows up. The minute it blows up, you want to get the DM and just wire the money in AI. That’s the play.

Pushing on that a little, first of all, you’ve got to give huge credit to likely Lightspeed for doing the seed in Exa. For suckers—too early on.

Guest 3

Well, I mean, if a sucker means he won a lot of percentage at low dollars, yeah. No, but I think what you’re saying is actually correct, which is there’s a quantum reduction in risk when you go from no revenue to revenue, and then there’s a linear reduction in risk thereafter.

So the sweet spot, I always think, on the risk-return continuum is early product-market fit, which is why we focus on it.

Now, you’re right, Jason. The amazing thing about these AI companies is that moment used to be a year, a year and a half, maybe 2 years of early product-market fit. Now what you’re seeing is that you go from early product-market fit to, “Oh my God, it’s incredibly obvious,” given the take rate, literally sometimes in weeks, if not months, and definitely not years.

You know the Lovable and Replit numbers. I think the same thing kind of happens here. You’ve got to find that moment, and yeah, you’ve got to pay. In retrospect, I was wrong. You shouldn’t have bid what I bid. You should have bid $750 million.

Guest 3

Yeah. But also, you have to abstract away not only that it happens so fast, but the competitive risk. You have to not worry that Google could do it. Why didn’t Google do it? Otherwise, you’ll do none of these deals, especially if Google wouldn’t do it. The company didn’t even exist a year ago in its current form. What are the renewals going to look like? We don’t know any of these data points.

I agree. This is actually an interesting discussion about our job, as distinct from being public-market pontificators.

You're right. You have to decide that you're willing to pay more on less information than in the SaaS era, simply because the market and traction that you're going to see if you're right come so much quicker and are so much bigger. That is the core of figuring out what's going on right now.

That's been the bet that's worked, right? I look back on the stuff we did that I feel right about, and those are the criteria. Then, obviously, you have to get the traction. If you don't, you feel like a right idiot. On the stuff you missed, you look back and go, "You should have leaned in even more. You should have been more willing to give conviction on the traction, even though the data points were sparse," because the stuff that's working, that's in tune with what people are trying to build, is just the universality of the propensity to adopt right now.

There's no 10-year cloud journey where some people decide now and more people decide 2 years from now. It goes back to the first thing we said: corporate America flipped the switch a year ago and said, "Thou shalt do AI in 2026." If you're selling anything in the picks-and-shovels business to help corporate America do AI, you get revenue this year. You don't get the luxury of waiting and saying, "I'll wait to see 1 year's renewals before I bid."

There's a thread, which is: if you have agentic products, can you make it up in volume? There is a software explosion. The number of apps we're building may not be benefiting Dropbox, where Drew just stepped down, but overall, the number of apps is exploding at a rate we've never seen before. Apps are being built everywhere, workflows are being built everywhere, and agents are being built. If you consider an agent a piece of software, it's exponential, right?

At Databricks' Neon, which is their Supabase competitor, over 90% of the databases are built by agents, not by humans. But they don't pay as much as they pay. So the question is—and I don't think Exa makes anywhere near what Google makes on a search—where I think we're going to get caught in a lot of these tools is where there aren't orders of magnitude more volume.

The reason Supabase, Exa, and these groups are neat is because you don't need them for every app you build or every use case, but you sure need them for a lot. We're going to make mistakes as investors. I know I've already made 1 or 2, and they're not fatal mistakes, but where it all is an agentic workflow, agents need it, but they don't need it all the time. I can't make it up in volume. We need vast volume for these plays to work—vast.

Guest

You're right, Jason. Basically, what you're saying is: what is the—call it the harness—what are the core primitives that it takes to build an agent, and what are the other things that pretty much most enterprise developers will figure out? "Oh, I need this, this, and the other to make my agent."

If you pick the thing that only 10% of developers need because it's a corner case, you'll pay up, and you'll have a tiny market, and you'll be sad. But there probably are going to be 5 or 6 things, like a database, like a search engine, like some kind of observability, where, over the next 2 or 3 years, revealed preference will say that the vast majority of developers realize they need this.

It doesn't come in the box in the harness from Anthropic or OpenAI, and therefore standalone companies will be built doing it.

Guest 3

You're right. We actually had that discussion last week: what are the 5 or 6 things that you need to own? I think it's a credible argument that agent search, which is Exa, and—actually, I think I could be wrong on the OpenRouter thing—that kind of model-switching layer could be 2 of those things.

11. Is Replacing Your CRM with Vibe Coding Always Ragebait

When you look back 5 years from now, every app company, whether building a third-party app or building in-house, will have said, "Oh, I need that product." So they end up with mass developer adoption. That's the bet.

Okay, we're going to do a final round, which, as always, is rage bait but real. I just tweeted, and Jason and I were talking about it. I just interviewed a CEO who said 3 things:

"Number 1: We replaced our $600K Salesforce contract with a vibe-coded CRM, which we built in 3 weeks.

"Number 2: We will get rid of 80% of the SaaS we use internally.

"And number 3: If Anthropic doubled its pricing, we would not change our usage."

Which do you think is the most rage-bait statement there?

Guest

Oh, it's tight.

Guest 3

I think the first one is clear rage bait.

Yeah.

Guest 3

And it's unhelpful. If you want to build your own CRM on top of Postgres, go for it. Go for it. We moved to headless Salesforce. We don't log into Salesforce anymore.

We could—we're a small enough organization that we could swap that out for our own database. It's not worth it. It's not worth maintaining that database. It's not worth fixing the connectors to our 10 external agents, which are all natively built into Salesforce. It's not worth giving up Agentforce.

If we needed no third-party apps and had time to burn, sure, because we already have our own layer. We already have our own autonomous agents running on top of Salesforce. But this is such a waste of time, talking about vibe-coding away Salesforce, when we have more important threats to our existing portfolio than this one, or more important things to invest in.

I just think it's a dated 2025 take. Even if there's truth in it, this is, of all the threats there are to Salesforce, HubSpot, Monday, and Atlassian, not in the top 10. That's why I think it's rage bait: it matters, but it's not important in today's world. It doesn't even explain the SaaS apocalypse. Other things explain the SaaS apocalypse.

Guest

I like that answer on the first one, and I agree. Yes, if you can get better functionality, knock yourself out, but vibe coding—that's just silly.

I actually want to take the third one and put it back on the CEO. What he basically said is, "If Anthropic doubled its prices, I wouldn't mind," right? Then you should be ashamed of yourself. You know why? Because what it basically says is, the ROI on this AI is so good that I could pay twice as much.

In that case, let me tell you what you should do right now: you should go in and tell your people, "Use twice as much," because logically there is an ROI on every token you spend. The early ones are really great, and eventually you probably hit the point where the marginal cost is exactly equal to the marginal advantage. If you're token-maxing, you're overspending, right?

If you think you're willing to pay—if the projects that you're doing right now with AI really, not just as rage bait but really, have a 2x return, have a return that's 2x better and could support twice the price—then what you should say to your engineers is, "You're not using enough. Go do more," right? The 2 things have to be true at the same time.

If Anthropic could double your prices and this project were still economically viable, then what's the next project on your to-do list? Go do it.

Guest 3

I'm with you. But the company Harry's talking about is at $2M in revenue per employee, so they can afford it. Let's be clear: they can afford to spend more on tokens. You're right. Of course you're right, Rory. If you're that efficient and you don't care what it costs, spend more, right? You haven't reached that—

Guest

We don't care what it costs.

Guest 3

But I will tell you, because we're there, I don't care what we spend on Anthropic either. Our AI VP of marketing and customer success costs $257 a month. I don't care if it's $500. What's 2x? We're getting toward the end of the pod. It doesn't matter. It doesn't matter.

But here's the interesting thing. This is why I actually think it's interesting: if you're really good at this stuff, the issue isn't more money. The issue is idle—idleness. Our agents are idle. We don't have enough brain cycles.

This is why these—what's the term for folks that can't sleep in San Francisco? They're coding all night. They have a—what's the term? I forget what it is.

I could be cynical and say incels.

Guest 3

No, there's this term where you're addicted to it, like your brain's been rewired. The problem with—listen, there's the Uber issue, where we spent too much on slop, okay? Or we don't know how to measure.

There's a different issue coming, which is coming for more of us in tech: our humans just can't process all the output. It's not even about the money anymore. If you wake up and build me 50 features every single day, how many features can I even qualify overnight?

Because we wake up every day, Rory, and our AI VP of marketing nags the hell out of us with 3 ideas we should be doing every day.

Guest

It’s exhausting. You can’t implement 21 ideas a week, right? Every day at 7:13 a.m., AI comes up with and pushes 3 ideas of the day. You didn’t get to my idea yesterday, did you?

Right, but stop. First of all, you’re right. I agree with what you’re saying. I still go back to my comment, though, which is that—I mean, that paper I cited says you can increase the complexity of the task and spend more money, because it’s implicitly saying he’s getting way more value than he’s paying for, right?

So when you look at it, you’re saying the same thing at $250, which is nothing. You’re saying you get 3 great ideas a day and you can’t process them. Correct.

Guest

Yeah. We get more than that, but we get nagged. We’ve got 3 things we need to do.

We’re going to end on a note of optimism. In that case, you need to hire another human.

Guest

I agree.

Because if the AI is giving more good ideas than the humans can process—and this is kind of the whole notion of complements and substitutes—if the AI is now giving more ideas than 1 human can process and they’re good ideas, hire a 2nd human, who will now be more effective, and the return on human capital will go up.

Guest

But you know, you’re right. But you know what the problem is, to tie it all together? I desperately need a human for those ideas. But a $125K laid-off ClickUpper is worth nothing to me. I need a $1 million person to process these ideas. They need to turn that into $5 million, and I will hire them today for $1 million. No joke.

But you’ve got to be the $1 million person—not think, not want to be that person. You’ve got to be that person.

The odd thing is, every single discussion here has gone back to what the optimal spend is between humans and tokens and what the consequences of that configuration are. That is the same question the whole time.

Guest

Right.

Jason also does not like other humans, and so that would be—

Guest 3

I do. I don’t know why people say this. I’m a people person.

That’s what they all say.

Guest 3

I’m a people person as long as I don’t have to interact with too many on any given day.

Guest

Jason, I have to be real: if you have to say you’re a people person, you’re not a people person.

Guest 3

I know. I know. You’re right. You’re right.

If you have to say you’re right, do you want me to be honest?

Guest

Whatever the 2x2 is, Harry and I are at opposite ends of it in terms of extroversion and being a people person. Everyone’s Harry’s best friend. I never get that. Why don’t I see on social media, “I was just in London meeting with my best friend, Harry Stebbings”? I see 11 of these on Twitter a day.

Guest 3

Because you’re not insincere like Harry.

I see. That could be part of it. That might be part of it. It’s because I’m Big Ben. When people are in town, they come see me. Yeah, but they’re always my best friend. How many best friends can you have? 300.

Guest 3

Oh, many.

Guest

And the Dunbar number is 160, just for the record.

For best friends.