OpenAI收购TBPN及管理团队重启｜Mercor遭黑客攻击，以及为何此刻正是网络安全的时点

I'm going to call it start to finish on this whole discussion. So, what do we have on the agenda this week? OpenAI reboots its management team. OpenAI buys TBPN.

Jason Lemkin

I thought the acquisition was just insane. Owning a media asset invariably takes way more time than you think for way less money than you expect. See Jeff Bezos for details.

Rory O'Driscoll

There's no way that deal's going to happen today. It's dead because of the management change.

Anthropic hits a whopping $30 billion in revenue, surpassing OpenAI.

Jason Lemkin

Their training costs are a quarter of OpenAI's.

Rory O'Driscoll

The big 3—SpaceX, OpenAI, and Anthropic—their value at IPO will exceed every other IPO for the last 20 years combined.

1. Anthropic Surpasses OpenAI in Revenue

Ready to go? Boys, welcome back. I've been looking forward to this one. I was doing the schedule over the weekend and last night, and I was like, “Wow, this week we really have a lot of meat to get into.”

I want to start with OpenAI and Anthropic. Anthropic now has $30 billion in revenue, obviously surpassing OpenAI. It's all intertwined with the subsequent things that we will discuss with OpenAI, but as Jason put in an email to us all, “Holy cow.” Jason?

Jason Lemkin

Holy cow indeed. What did you think? Even in an era where we're getting ignored and anesthetized to crazy numbers, this one I did fall out of my chair, right? Getting to $30 billion, up from $9 billion at the start of the year.

I think Salesforce is the largest cloud software company, right? It took them 25 years to get there. Anthropic got there in 5, but maybe they really got there in 3, depending on how you count. We couldn't believe it when we saw where they were in February, and then they essentially added $10 billion of net ARR.

Let's not debate how many Rs there are and whether it's recurring. At this level of growth, it really doesn't matter. They're still capacity-constrained, Claude still shows us when we're in there that it can't finish chats, and every engineer and tech has been told to consume more tokens and move faster, right?

The crazy thing is, what will it be at this rate at the end of next year? If it grew 3.3x in 4 months, we need Rory's math help to figure out what Anthropic's run rate will be at the end of 2027. The estimates that we were just looking at 2 months ago just look incredibly wrong at this stage.

Yeah. These are amazing numbers, right?

Rory O'Driscoll

Yeah, and I think a bunch of other interesting things start to happen here. One is, you're kind of mixing in some stuff. One is their announcement on OpenClaw and not allowing that to be in the base plan.

I think it gets back to the fact that they're in a massively interesting situation: their revenue is exploding, but despite the revenue explosion, they're still compute-constrained. In other words, they could sell more if they had more, right? What do you do when you can sell more if you had more, but you can't make more? You can't magically make data centers, though obviously they have that big announcement to do that.

What you start doing is allocating capacity based on money, right? One of the first things they figured out is that people using these OpenClaw-type agents are consuming vast amounts of tokens on fixed-price plans, and they probably want to stop that, which is what they've done.

You're going to see them do exactly what anyone in economics would say: try to find a way to maximize and extract even more revenue. We saw it even with OpenAI last week, where you deemphasize things like video, which consumes huge amounts of compute for small amounts of revenue.

In Anthropic's case, obviously they have much less of that pure slop, but you deemphasize things like OpenClaw access, where it consumes a lot of your compute and doesn't make you a ton of money. I think you're just going to see a continued trend toward pricing tokens closer to the value.

It's not a huge trend, because you want to get people addicted to the product. The truth is, the thing you have in your favor with any digital good is the complete certainty that prices per token go down over time. But you do at least want to start allocating it a little more sensibly while you're constrained. That's the trend here. We could talk a little bit about the OpenClaw stuff.

Just on that, before we get there, the other interesting thing was that The Wall Street Journal today had a bunch of leaks on the financials for Anthropic and OpenAI. The one that jumped out at me, when I contrast it with the fact that Anthropic has caught OpenAI in half the time, is that their training costs are a quarter of OpenAI's. Their training costs for models are a quarter of OpenAI's.

Maybe that's because they're focused. They don't have to do video, they don't have to do images, and they don't have to do a lot of consumer stuff. But if you just think about it for a moment, the compounding effects of catching OpenAI in half the time, at roughly the same revenue or more—$30 billion in 5 years—and having training costs that, for now, are a quarter of OpenAI's, that's a double code red.

It's one thing if you have 2 classic startups where one has bled money and it's artificial—there are other things—but if you have a dramatic cost benefit, you're out-accelerating your competitors, and there's management-team turmoil at your competitor, it really feels like the investors in OpenAI got a much worse deal in the last round than the Anthropic ones did.

It's just crazy to have both. Usually you don't have both together with your competitor: you're out-accelerating your competitor and your training costs are a fraction of your competitor's. Good God, that just compounds.

Jason Lemkin

That's actually a good point, because you take the Uber–Lyft struggle, right? Uber had the, “Oh my God, we're out-accelerating. Oh, but by God, we're spending every dollar we have to do it, and we show no fear,” right?

In this case, you're out-accelerating the opposition while being more efficient on a bunch of interesting measures. That's a bad fact pattern.

Rory O'Driscoll

No, you're right, Jason. That's a scary fact pattern if you're running the game theory and you're the other guy. It's like, “Hmm. That's not good, right? They're going faster than us. The gross-margin economics are roughly the same, or slightly better, and their costs below the line—to rounding, our costs are compute and scientists to run the compute for training—are better.”

Have we ever seen a bigger seeming chasm between where they're at? With the greatest of respect, it seems like Anthropic is accelerating faster than ever, and OpenAI is having more challenges than ever, all at once.

Jason Lemkin

I'll tell you the one I think about, and where I didn't realize it when we did this show last time, because the press is always focused on the headline stuff, right? The OpenAI round was barely real.

Yeah, you said Anthropic.

Jason Lemkin

Barely real, and Andreessen's money appears to be real. It came in out front, right? The $13–14 billion, whatever they put in, that's real.

$11 billion.

Jason Calacanis

All I have to do is get 20% carry and double that, and it's a nice side bet in an SPV, but that was real. The SoftBank money comes in tranches. They have to borrow money to pay it. The Amazon money is tranched in part on an IPO or AGI, right? And the NVIDIA money is almost all not money. It's almost all offsets and compute.

I thought about it, but then, in the context of Anthropic's growth, OpenAI would have rather had all the cash. It's not a sign of strength where the majority of the round is not cash up front. That's not, I don't think, a sign of strength. That's classically, at least, barely getting the round done.

Barely getting the round done. Versus getting it—why wouldn't you want all cash up front? Why wouldn't you want $140 billion up front?

Rory O'Driscoll

A bit harsh on the “barely,” because I thought they tacked on another $10 billion. I can't believe I said that sentence. Think about that sentence sometime. That, I think, was cold, hard cash.

Your comment is correct, Jason. The vast bulk of the dollars weren't cash, but enough money changed hands that it represented a bona fide price at the time.

Anthropic does some of the same stuff, in the sense that, given your biggest expenses are compute and then distribution, from Microsoft onward, OpenAI to all the recent Anthropic deals, there's a lot of this round-tripping business.

In both cases, there were enough hard dollars changed hands for them to represent price estimates. But, to your point, based on what you know now, given the roughly equivalent revenue—you shouldn't assume until OpenAI releases its numbers; maybe they've exploded too—but definitely Anthropic at $370 billion feels a little more comfortable, let's just say, than OpenAI at $820 billion or $840 billion, or whatever the final closing was, right?

Jason Lemkin

Well, OpenAI did say $2 billion last month. I think that's why Anthropic rushed out the $30 billion, right? They're at a $20 billion run rate.

And look, we have the whole gross-and-net thing, but the bottom line is this: when you look at those 2 graphs, you definitely don't say to yourself—I mean, I think what you would say, if this was a public stock, let me put it this way.

Rory O’Driscoll

If both of these companies were public, there would be a bunch of New York hedge funds shorting OpenAI, longing Anthropic, and saying they have the perfect AI bet, right? Would you short OpenAI at $870 billion and go long Anthropic at $372 billion? I’m not a risky guy, but even I would contemplate doing that. It feels like a no-brainer bet. You have roughly the same revenue, a better trajectory, and a management team for half the price.

If you short the one and long the other, you’re diversifying away the overall AI risk, and you’re just making a relative-performance bet.

Hmm. Yeah, that would be an interesting one. What would you say to an OpenAI employee who is now looking at that incredible stock-price appreciation, with tens of millions of dollars in equity that they now have at the $820 billion price? Sell it all at $820 billion the minute a tender comes. What would you say to them?

Rory O’Driscoll

I think I wouldn’t pile on. In general, look, I have some things on OpenAI I want to pile on this time. And if you recollect, in the last couple of weeks, I’ve tried to avoid the pile-on when someone’s down. I think you always want to be more tempered.

But I always say to everyone in any private company, when the liquidity window opens, take it seriously, because it might not open again for a while, right? So, yeah, when your liquidity window opens at $0.8 trillion, the alert reader should say, “If you’re planning to buy that house in San Francisco, you might need an extra few million,” just based on what I’m seeing in the market now in terms of house prices. So take advantage of this thing, because all your brethren have, right?

I wouldn’t—I just say, I think the people that I don’t want to pile on are the individual employees. Remember, the company’s still doing a lot of great stuff. You’ve got a lot of turmoil. We’ve got a lot of drama at the top. We’ll talk about that, but I think you take advantage of liquidity just because you should always take some advantage of liquidity.

Let’s knock it on the head. Let’s talk about the drama at the top. Talk about a management-team turnover. You have Brad Lightcap, the COO, who’s been moved to special projects. My dream job is being moved to special projects—the SVP of special projects for 20VC in my next phase of life. Jason, I would love you to be the SVP of special projects.

Jason Lemkin

Just special projects.

2. OpenAI Management Reboot

Just special projects. We have the CMO stepping down due to health reasons. We have the CRO out. We have Fiji, who was head of apps, taking a short leave of absence with health problems. How do we read this very significant multitude of changes at the management layer?

Rory O'Driscoll

If we step back a minute, it ties to Anthropic passing them. You don’t just sit there and make no changes on the team when your competitor, over the last 6 months, has radically changed the competitive posture.

So, look, I don’t think any of us like the amount of change in any management team, right? It feels almost like a wholesale change at some level. It’s risky, but calling code red 3 months ago didn’t magically change the trajectory here. So it ties. You’ve got to try to mix things up in some fashion.

Hopefully, you can do it with the team you have, but in the context of Anthropic now out-accelerating OpenAI, it just makes sense to reboot the team. It just makes sense.

Yeah, there’s some rebooting. I mean, to use your phrase, who’s been hired? What’s the additive reboot?

Rory O'Driscoll

Well, the dramatic one, which is always risky for any startup, is you take Denise Dresser, who was CEO of Slack and came from Salesforce just a couple of months ago, and you put her in charge of basically everything go-to-market and related, right? That’s a good bet on a seasoned executive, but that’s the type of change that we’ve all seen as investors that’s super risky, right?

You bring in the person with the perfect LinkedIn and the perfect background. They’re still getting to know the product, they’re still on a get-to-know-you tour, they haven’t quite been to the New York office yet, they’re getting to know the product, and all of a sudden you give them this massive portfolio because they’re a proven executive.

In my experience—I don’t know what you guys think—in my experience, that has about a 30% chance of success, just roughly. Bringing in the perfect LinkedIn, giving them a massive portfolio, and attaching them to something in tumult: if it’s executing to perfection, it always seems to work. Bringing in Mr. LinkedIn—it always—but when you’re in tumult, there’s not a lot of time to learn everything, right? There’s not a lot of time for the get-to-know-you tour.

So it’s just risky, but it’s a play. It is a play.

I get where you’re going there, Rory, which is that for the replacements to be added, there needs to be great talent added, and there seems to be a lack of people coming onto the field when they’re coming off.

All I have, as I say, is a couple of comments. One of them I’m always loath to comment on, because the illness-related changes—you just don’t know what’s going on in people’s lives, and that’s tough. People have challenges, and you wish people all the best, especially with these kinds of chronic diseases, and hope they can get back to full health. Let’s just start with that, because that sucks, right?

At the same time, you have a lot going on here. To me, even with all these people changing, I’m tempted to make the famous Oscar Wilde quote in The Importance of Being Earnest: “To lose one parent may be regarded as a misfortune; to lose two looks like carelessness,” right? Well, you are getting to the stage of carelessness here, right? But I actually don’t think that’s the real issue. It’s fun to say.

I’ll tell you, we haven’t mentioned the 2 most surprising things in the last week on OpenAI. One is—I’m just going to say it—I thought the acquisition of TBPN was just insane. Not on its own, in particular—it doesn’t matter—but you don’t launch a code-red deck and a focus deck and, you know, a no-more-side-projects deck, and then within the space of a week do something that’s so obviously a side project.

To me, no matter what you get—I mean, we can discuss whether it’s stupid on its face and whether buying media assets is the way to go. And I acknowledge Andreessen’s articulated thesis that you have to control the media story, though it doesn’t seem that Anthropic has any need to do that.

But stepping back one level, you’re running a $25 billion company, the most exciting company on the planet, and you just told your entire internal team that you need to focus. Buying a media company is nothing more than a vanity project. Right? I mean, look at—

Rory O'Driscoll

Just one thing—we could talk about it more or less. The one thing to add, when you look at the press, is interesting: the outreach for that deal was in January. That’s a lot of time in OpenAI and AI time, right? Fidji was new and thought this would be a great thing to elevate OpenAI in January. Now it’s April, and maybe the deal seems a lot different, but in January it was a different world, right?

At some point, we remember—yes. I think it didn’t happen last week; that’s my only point. It happened in January and took some time to close, right? And I will say one thing: I’m 90% sure it wouldn’t happen today, to your point. Priorities—if nothing else, priorities change, right? It probably wouldn’t happen today.

If you only noticed in the last week that you need to focus, then yes, I’ll give you that, right? But you didn’t just notice in the last week that you need to focus, right? And if you did, maybe you need to focus. If you haven’t realized you’re in code red for the last 2 or 3 months, and if you haven’t realized this is the kind of thing you don’t do when you’re in code red, then you’re just not paying attention. So I challenge that. I think it’s a vanity project and absurd.

Jason Lemkin

Can you just pause? I know you want to—because you want to say, like, “Where’s your $200 million, Harry?” But, yeah, let’s pause on your lack of $200 million.

3. OpenAI Buys TBPN

No, no, I know. I just want to articulate a bull and a bear case rationally for an audience for how this acquisition could be seen from both sides, because it is very confusing. So if we were to start with a bull case, Rory—and Jason, please chime in too, because you’re the master of media and venture as well—what is the bull case for it?

Jason Calacanis

I’ll give you the bull case. There are 2. The strategic one’s more interesting, but let me hit the tactical one, because Rory made a good point.

Look, this is not going to make or break the company, right? There are certain acquisitions that can. Let’s stipulate that. But there are some things you acquire where they run almost on autopilot. They are not massive distractions. If the price is small relative to what you hope to get out of it, that does factor into the equation.

If you have to rebuild your whole team, it’s a total distraction. You’re going to rip out your guts. That’s a big deal. Once in a while—it’s pretty rare—you can acquire something that isn’t massively distracting at some management-team level.

So, even if it’s not the perfect acquisition, I don’t think it’s huge. It’s not going to require a huge amount of senior executive time, so it’s just important in general to the calculation. The one point I’ll make—and I wrote a post—is that every profitable public company should do a deal like this, of which OpenAI is neither, right? It is clearly not profitable. It is clearly not public. But other than that, let me tell you why Rory and you might end up agreeing with me on this.

If you are a profitable B2B company, especially, you are under insane pressure to get more profitable. I actually can’t overstate how intense the pressure is. They’re looking at every headcount, every sales efficiency, everything. It is brutal, right? Your cash is trapped on your balance sheet, and so it is very difficult to increase marketing spend. It is very difficult to spend another $100 million this year on marketing.

But at least in the short term, if you can buy a marketing asset that is at scale, you can turn your balance sheet into marketing, which is hard to do. I think maybe TBPN is not the most successful way to get OpenAI’s brand out there. We could debate that, but it is a way to turn a balance sheet into a marketing asset.

I’m going to call you out from start to finish on this whole discussion. OpenAI is the best-known company on the planet, perhaps other than Apple, right? Within the last 2 years, the CEO of OpenAI has been able to meet every world leader he wants. He’s gone on world tours. He’s met Macron, the president, every single prime minister of India—whatever, right? They get constant attention.

The AI story has been the entire zeitgeist for the last 3 years, and they’re the leader of the AI story, right? In terms of media minutes, there’s nothing left to get. If what you’re saying is, “I don’t like what they’re saying about me. Oh, they were mean to me,” then, yeah, maybe you can pay these guys to say nicer things about you than average. But you don’t need more. It’s not like you’re making widgets in the heartland here, right? You are the most exciting tech story on the planet. You don’t need a little bit of help just to get out and get covered. Literally everything Sam does gets covered.

So, I hear you, Jason, most of the time, but not for these guys. If you were to pick the 1 company that doesn’t need media attention and does need to focus, it would be OpenAI. This is unfocused and getting media attention. From a signaling perspective, we’re 100% aligned.

The only thing I would come back to you with is that they have consistently shown an inability to respond to negative moments on social. Whether it’s the lemonade stand or Anthropic adverts, they’ve consistently messed up crisis PR and crisis communications and made themselves not look great. The only way I could justify this is by saying they are vibe maintenance for those times, to make us better, cooler, better responders to bad things, because they have no editorial control.

This is the most important thing. Andreessen is right about the importance of owning media, but they have no ability to own the content, to influence it, or to impact it in any way. It is editorially completely impartial, so they have zero benefits. This is the only reason this does not make any sense. If they had the ability to own the media properly, it would make sense, but they have zero impact on it.

History is riddled with people who buy media assets to try and change outcomes. My observation is that owning a media asset in a very big way takes way more time than you think for way less money than you expect. See Jeff Bezos for details. You end up getting a sinkhole.

If you can’t control the story, hire a better storyteller. Hire a better comms person. Hire a better marketing person. Think before you speak and before you hit send about lemonade stands.

Look, in the short term, it’s in the noise. And, Jason, you are right: they’re not going to spend a lot of time managing it, at least in the short term. My comment is more that it’s really silly when you say, “We’ve really got to focus. Nothing else matters but these 2 or 3 big things. Oh, but by the way, here’s 1 last plaything project.”

The 1 thing I will say at a meta level, especially to founders who listen to this, is that this is why you should default to yes on a good deal. Let me be clear: I’m pretty sure this is the deal—I just read the press. The thing is, at OpenAI, there was stress in January, but it’s not like today. It was not like today.

Fidji Simo comes in and she has an idea. This is not the biggest bet the company’s going to make, but they have a team meeting. She’s like, “I love TBPN. What if we brought them in for a little bit of good promotion?” Everyone around the corner is like, “Whatever. Yeah, let’s go talk about buying some OpenClaw or something.” But they say, “Fine.” Things are good, and they kind of shake hands on a deal. It takes a little while to happen, and it closes last week.

There’s no way that deal’s going to happen today. It’s dead because of management change. I can’t tell you how many times I’ve seen this for portfolio companies, and it’s happened to me twice. It’s not just time that is the enemy of deals; it’s management turnover and priority turnover.

So, the meta lesson is that I just don’t think this deal would have happened today. It has nothing to do with the team at TBPN. When you say no to an attractive deal, just be sure you’re okay if it’s no forever, because the odds that the VP who wants to do the deal is still there in 12 months and that their priorities have not changed approach single digits. It’s back to the liquidity-window comment. You’re exactly right, Jason.

Jason Lemkin

Yeah, but my God, I think it’s worse for M&A, because so many times that guy just isn’t there next year. I’ll tell you, 1 of the things about being the big boss is that even when you’re a long way down, if you don’t think it suits what you’re doing now, you should stop it.

I remember—fun story—25 years ago, we were selling a company to GE. I’m not going to name the company, right? It was a mediocre company, and we were darn lucky to get the bid. It was going all the way through, and it went through every level at GE. Then it came to the CEO. Jack Welch wasn’t perfect, but he was willing to take a tough decision.

We were a long way down, everyone was about to sign and be happy-happy, and he looked at the numbers and said, “No.” I remember thinking, “Damn, I thought we’d get away with it, but he’s right.” At some point, I remember thinking, “That’s impressive.” All these people were in, it was a long way down in the process, and he just said, “I’m thinking no.”

I’m going to give you a hard one before we move to SpaceX. You have the chance to buy Anthropic at $850 billion or OpenAI at $380 billion. Which would you rather buy? You may have that opportunity in the secondary market as we speak.

Jason Lemkin

I wouldn’t be surprised. I think I’d buy—I mean, having said I’d do Anthropic last time, 6 months ago, at the $300-something-billion valuation, I think I’d go the other way this time. If the choices were Anthropic at the last OpenAI price of $850 billion post-money, or whatever it is, $820-something billion post-money, or OpenAI at the last Anthropic price of $370 billion post-money or $380 billion post-money, I would argue that you would buy OpenAI on 1 proviso.

You could sit down with the board and say, “What are you going to do about this?” Because it doesn’t take a lot to fix this thing, right? Just stop screwing around and focus.

No, no, no, because you’ve got to stop that—a machine that is Anthropic, which is now picking up more and more pace with every day that goes by and becoming more—

Jason Lemkin

Google. First of all, you still have the consumer asset, where you are by far the dominant thing. Again, this goes back to what we said last week. You have to do 2 things: figure out a consumer monetization model, and get a Codex—the Codex competitor to Claude—out there. The mission at OpenAI is pretty simple.

You do have 1 big advantage we didn’t talk about, though, and it’s changing a little bit. Give Sam Altman some credit: he was more aggressive on compute purchases. I’ll admit I was someone thinking from the peanut gallery, “Hmm, is that a bit aggressive?” But now it looks like compute constraint is a real thing in 2026 and early 2027.

You have that asset. Maybe you figure out how to deploy it aggressively with Codex. There are buttons you can press and things you can do if you focus and do them. Jason, you’ve got that same choice. I say buy both if you can invert the valuations.

Yeah, that’s actually a good idea. That’s what all the growth VCs are doing, if they can get away with it anyway. So, let’s invert: that’s amazing, but you can only buy 1.

Jason Lemkin

Yeah, but conflicts aren’t important in our firm anymore. They don’t matter at Pace, and they don’t matter at Growth.

You only have 1 check left. Well, look, I’ve said the same thing on the show. I’m just not into the tumult at OpenAI. I’m not into the drama. I’m not into non-tech, non-deeply technical founder leadership. It’s just not my vibe. I wouldn’t invest in anything like OpenAI at a high price.

Rory O'Driscoll

It doesn't matter what it is, because I just find it so risky. The turnover and not being led by a deeply technical CEO—that's just... In my life of investing, I ain't doing those risks anymore, right? Maybe I'll miss a lot of opportunities. I want someone like Dario, or someone technically smarter, running these companies, or it's just too much change. You get too lost on the PAN and the TBPNs.

Although, I don't think Sam had anything to do with TBPN, in all fairness. I think he said, “Fine,” in a meeting and moved on.

Related to that, for folks, I don't know how M&A works at OpenAI. It's not that sophisticated, okay? But I will tell you, when I was at Adobe a long time ago for M&A, basically every senior executive got a big chip and a small chip.

The big chip was a big deal. Back then, it was maybe a $1 billion deal. That would move the needle. If it doesn't work, you get fired. It's that simple, right? And everyone got a small chip, which could be like a $50 million to $200 million deal, and you had to justify it, and you didn't get 5. As a forcing function, you got 1.

But you really weren't challenged that much to do the smaller chip. You picked your 1 a year, and you didn't get fired if it didn't work out. There was an idea that maybe 20% of them would work out. So, I bet he spent 5 minutes on it: “Is this the one that you really want to do this year, Fidji? Then just do it. Let's move on. We've got bigger fish to fry.” That's why I don't think it's that big of a deal.

4. SpaceX Files for IPO Targeting $2 Trillion Valuation

It was a small-chip deal, right? No one loses their job at Adobe over the small-chip deal. Otherwise, it would never happen. No one would take any risk in buying an emerging company, right? They just wouldn't do it.

Okay, we've got to move on. Other things did happen. SpaceX finally, finally confidentially filed for an IPO, targeting a $2 trillion valuation. It would be the largest IPO in history, surpassing Saudi Aramco. They could raise up to $75 billion.

This obviously includes xAI, otherwise known as Twitter, which was incorporated earlier this year. 2025 revenue: $15 billion to $16 billion, $8 billion of EBIT. At $2 trillion, it's a 125x revenue. So, coming in punchy, to say the least. Feels like a Series A these days, Rory. How do we feel when we hear this?

Well, I'll just tell you 1 insight. I think to say that at least venture is different or remade is an understatement. The big 3—SpaceX plus OpenAI plus Anthropic, assuming they all IPO, and certainly SpaceX will in the next 12 months—their value at IPO will exceed every other IPO for the last 20 years combined. All of them, all of the last 25 years. These big 3, every other little deal here—I mean, Rory's had some great IPOs, and there have been tons of them out there—but this exceeds all of them combined, right?

So, I found it almost depressing in a way when I thought about it this way, because it was like, what's the guy we had earlier in the show from Slow Ventures who kind of bothered me a little bit? Sam Lessin. Yeah, and he kept saying Box doesn't matter, and then he said OpenAI doesn't even matter. It's not that important, and he was very triggering to me. I tried not to get triggered directly, but he kind of rattled around in my head. I was like, maybe the guy's right. Maybe nothing we're doing matters because the big 3 dwarf the last 25 years combined. What are we doing, guys? What are we doing here?

Rory O'Driscoll

First of all, I think that is a real phenomenon. What you're simply saying is, especially in SpaceX's case, the longer the holding period, the more dispersion sets in: the biggest get bigger, and the little ones fade out.

You're exactly right. At the tail end of a power law, it does mess with your head because the combined value of the top 3 privately held companies is larger than everything else. In much the same way, it's even more concentrated than the public markets, which are more concentrated than they've ever been. The market cap of the top 4 or 5—you know, Nvidia, Apple, Microsoft, Alphabet, and I think Meta—is approximately 30% of the total S&P, right? Which is everything for the last several hundred years, right?

Psychologically, the thing about a power law they don't tell you is that you can have the 3rd-best outcome in venture history and be only 1/10 as large as the largest outcome in venture history. If you're going to let that in your head, it's going to be a very tough business psychologically for you, because you can have a life-changing event that's down in the noise of $10 billion or $20 billion outcomes, which can be enormously great for you and your family, for your co-investors, and for everyone involved.

If you're going to let it in your head that it's not $2 trillion, then you're doomed and you're just going to need therapy. I wrestle with these things all the time.

I mean, it is the thing that your mama told you, right? It's like you just have to not let other people define you. You said it really—it is a psychologically weird thing, right? You're going to have these 3 deals go public, and they're going to be worth literally everything else that's happened in the last 20 years if they trade anything like their current prices.

Will SpaceX rip and hit the $2 trillion when it does go out? I try not to spend time talking down an amazing company, right? Not least because I'm going to be a buyer of SpaceX 15 days from the IPO. I surprised you, because 15 days from the IPO it's coming into QQQ. I have a big QQQ holding, right? If you're an index fund, you're getting this thing in 15 days, right? I don't know when it'll be for the S&P, but it'll be fairly soon thereafter, right? So, we're all going to be buyers of this thing, right?

In terms of the valuation, you don't know. Look, you do any kind of meaningful analysis, sum of the parts, and you come up with a lot lower number. As we've discussed before, the gap between what you think the assets are worth on any kind of normal basis and $2 trillion is huge, and it's all Elon premium. What you're really asking, therefore, is, how big is the Elon premium sometime in June? I don't know.

Rory O'Driscoll

Well, obviously, I don't think either of us have worked on an IPO quite of this scale, right?

No, by definition, no one has in the universe, because it's the first time it's ever happened.

Rory O'Driscoll

Right, so the process is going to be different. But here's my point: at some level, there is a tough negotiation sometimes between the company and the underwriters on valuation. Oftentimes, some CEOs are like, “Whatever the Lord brings,” and some are extremely aggressive on the number they want. Depending on the situation, sometimes the CEO wins those debates, gets out with a valuation the underwriters are very uncomfortable with, and sometimes it works and sometimes they stumble because of it.

I think Elon—what Elon said publicly on X—is that it ain't going to be $2 trillion, and maybe he'll change his mind. He said $2 trillion was too high. So, whatever his number is, I think he's going to get it on IPO day. He's going to will it into existence. The underwriters are not going to be able to argue with him for more than 5 minutes, and there'll be enough demand between retail—30% of the IPO, it's a lot, right?—there'll be enough whipped-up demand, I think, to support it for 1 day.

He will will it into existence. Whether that valuation is there in 30 days, or possibly even in 1 day, I don't know. But I do think the sheer force of will, the lack of power of underwriters, and the 30% retail will will his $1.75 trillion into existence for 1 day, at least 1 day.

I think that's quite correct.

Rory O'Driscoll

I think less than 12 months ago, there was a meaningful transaction valuing SpaceX at $400 billion, right? Then there was this much smaller—I don't know if it even happened in the end—secondary at $800 billion. Then, in conjunction with the merger with X, or X/Twitter, SpaceX was valued at $1 trillion to value the other asset, with its negative $12 billion in cash flow, at $250 billion.

So, they added that in to get to $1.25 trillion, and now you're talking about $1.7 trillion to $1.8 trillion. It's all been walked up in a very interesting way. It is worth remembering that the last time the useful asset was valued on a standalone basis, it was worth $400 billion.

If the deal went public at $1.5 trillion or $1.6 trillion, less than the whisper number, I still think they'd have done a magnificent job of walking the value of the asset up. It's not clear to me that the xAI asset has a positive NPV in anything like the near term.

We just had a long conversation on Anthropic versus OpenAI, and they're number 1 and number 2 in this space. Google's Gemini is almost certainly number 3. So, xAI is number 4 in the model and LLM space at best, burning $12 billion a year. I don't know what that's worth, but I would argue that they won't be talking about that on page 1, 2, or 3 of the slide deck at the IPO.

They'll be talking about SpaceX, which means the entire addition of that probably is net negative. So, I go back to my comment: I think you're right, Jason. They'll will something amazing into existence for a short period of time because Elon just has all the leverage and the drive.

Only in the long term are markets weighing machines; in the short term, they're voting machines. We'll see over time how it settles as people look at the dynamics of a $20 billion, plus-or-minus, business. Cash-flow positive, apparently—well, EBITDA-positive; CAPEX is not clear, excluding xAI. Then add xAI, and it'll settle into a long-term value over time.

5. Doug Leone Returns to Sequoia Capital

What happens on the day, I think you're right, will be much more a function of the will. It's a small float, so people will push. I'm switching it up here. We're going to the private market.

We had big news from Sequoia this week. For context, Doug Leone had taken a step back from the firm, back from day-to-day operations and back from investing. Pat and Alfred had recently taken over the leadership from Roelof, and now Doug is back in an investing capacity—not in a leadership capacity; that's still very much with Pat and Alfred—but Doug's back in the firm investing, which is very big news given that he is one of the OGs.

How do we read Doug being back in the trenches?

Jason Lemkin

Well, look, I don't know. Rory may have more thoughts. From a distance, it feels like something to calm the LPs.

Everyone is raising so much capital, and there's so much change there. I think you guys have even more experience than I do, but LPs are uncomfortable with change. LPs say that they're looking at the new generation in the vanguard, but they're comfortable when the old leadership is still actively involved in the fund. It does make LPs more comfortable, whether they're investing half the fund or in a few deals.

So, it struck me as that simple: you bring back someone who makes the LPs comfortable, and you get through this crazy amount of fundraising everybody's doing. But I could be wrong. I don't think it's just to get somebody on your Slack and get a little wisdom. You don't need to bring them back just to get an hour or 2 of insights on deals that you already have.

Rory O'Driscoll

It was a sensible move. I don't think it's an earth-shaking move. They've made the changes they've made already as a firm, and they all made sense. I think, at the margin, you're right: it helps on a bunch of different things. It just provides some continuity, which is important, I think, for LPs, for the firm even, and for entrepreneurs.

Also, let's not lose sight of the fact that he's a damn good investor. One of the questions we always ask when we're hiring someone—and in this case, you are effectively hiring someone—is, do you think the next check that Doug Leone writes will be better than a check that one of us will write? I think Doug Leone has proven that he can write pretty good checks. So, even at the margin, from a check-writing perspective, it kind of makes sense.

The transition—I mean, having made one transition to Roelof and having had to make another transition abruptly means the first transition wasn't that successful. I'm sure there's an element of scratching the itch, wanting to come back and make it work. They put a lot of his life into this firm. They've done an amazing job, and it just felt a little janky late last year when that transition happened. So, if a couple more years can help manage that transition and send a continuity message, why not do it?

With the greatest respect, I spent a lot of time with LPs—a lot. The insatiable appetite from LPs for Sequoia has never been more prominent. So, respectfully, I don't think it's LPs. I think it's actually in the face of increasing competition from Founders Fund, which has Anduril and SpaceX as tailwinds for founder brand, and Andreessen, which is more attractive than ever for founders.

You also question, how can we be more competitive? Doug is the ultimate winner of deals.

Jason Lemkin

Tell me, have the kids at YC heard of Doug Leone or even learned how to spell his last name? I doubt it.

Rory O'Driscoll

I'm telling you, when Doug Leone goes to that meeting with them—whether it's Christian Hacker at Trade Republic in Germany or the team at Wiz—he closes the deal.

Yeah, because maybe not the YC founder who's 24, but you're right. Across the venture and tech ecosystem, let's get real here: this is someone who's had wild success. Even in a meeting, he can bring knowledge to bear that would move the needle on a close.

Jason Lemkin

I agree.

Well, let's call it gravitas. Whether it's the founders or the LPs, it is adding gravitas back into Sequoia at a time of a lot of change. They needed more gravitas. That's just what it was. We need a little more gravitas, guys. Who can we bring in?

One of the things I admire about Sequoia, to be fair—I’ve always said this—is that even if they're winning on every round but 1, they'd be like, "Well, how do we win on that round as well?" It's never been more competitive. There are wildly talented, similar-sized firms. Why not, even if you have 10 great players, get an 11th?

6. YC Kicks Out Delve

Jason, you mentioned the youngest founders from YC. Some very young founders from YC obviously founded Delve, a SOC 2 compliance business. Sorry, don't get pissed at me, but they are very young—21-year-olds.

As everyone knows, Delve has been in the news for not providing the SOC 2 compliance product they said they were providing. There are a lot of problems around that. YC has since kicked them out of the YC community, which was announced this week—or leaked this week—from Bookface, YC's internal product, which obviously wasn't meant to be leaked.

Is this the ultimate sign of their guilt? Insight invested $32 million in the company within the last year. Should there have been more diligence from an investor perspective? How do we think about this?

Jason Lemkin

My guess is that a lot of things went wrong. One was making up a lot of audits with AI. We're going to find more portfolio companies did that. The second one was stealing from a fellow company—stealing IP, forking a fellow company.

Listen, I don't know how you manage it with YC, with thousands of companies, but there's a limit where you cross the bro code, the girl code, or the founder code with other folks at YC and portfolio companies. There's a line you just can't cross.

Whether they see it as open-source code theft or whatever happened, whether you're manipulating, you can't allow that within the core portfolio. I think they were ejected for the combination, and it wasn't just some young kids misusing AI. I think it was the second thing. I think it was breaking the code, and that's why they were out. There was no need to comment further: you broke the code, you're out. You're out of the team.

Rory O'Driscoll

I totally agree, Jason. Look, these things are going to happen. I was just running the math in my head: YC has 200 companies a quarter, so that's 800 or 900 a year, right? Step back: in the United States, we have 300 million people. We have approximately 3 million people incarcerated at any one point in time. So, roughly 1% between felons and misdemeanors across the whole population.

If you just index to that, that means out of the 800 YC founders a year, statistically, if they're no better or no worse than the rest of the country, there are 8 of them who, over the course of their lives, will commit some kind of crime. It's going to happen. You're going to have fraud. When you have a portfolio of 30 or 40 companies, as we do, mostly these are avoided, and every once in a while one VC gets unlucky. If you have 200 companies a year, it's going to happen to you a lot.

So, first of all, no drama there. I saw all this, "YC is bad because this guy is a fraud." Well, dude, when you have this number of companies, statistically, it's just going to happen. That's the first comment.

The second comment, Jason, I love what you said. You're exactly right. What do you do if you're running YC? You can't stop this up front, especially when a lot of your value-add to entrepreneurs is the community. That is what you're selling, and you do business with each other.

Anyone who did business with these guys was, at the very least, discombobulated and embarrassed because you rely on this for SOC 2 compliance, and then it wasn't true. On top of that, you stole from another YC bro. You're exactly right. It's like in the Old West, when there wasn't much law: you have to take the law into your own hands and hang the cattle thieves. This is the same thing. Dude, you broke the code of the West, you're out.

From an enforcement perspective, I can totally see why they did it. Now, you can talk about whether other people should have known, or whether you should really buy compliance software from 21-year-olds. That's an interesting comment. Fundamentally, I think you're exactly right, Jason. You're going to have this thing, and the only way you can deal with it is not a priori policing. When you break the bro code—or whatever the non-sex-loaded term for "bro" is—when you break that code, you just have to be ruthless about it.

So, yeah.

Jason Lemkin

I think it was part 2 that did it. You know, there was stealing—taking a customer, Sim Studio, which is also a YC company, maybe even a batchmate, taking their open-source software, not attributing it back, and claiming it's your own software to your own batchmate or your own customer. We've all thrown a few things in the cloud and pretended we did the work. All 3 of us have done that. But this one breaks the code.

You took the open-source code from your batchmate and said it was your own software. They were your customer. You can't hand-wave that one away.

7. The Rise of Open Router

All right. You can't hand-wave that. Moving on. OpenRouter is a very well-known company, for those that don't know—a marketplace for LLMs, so to speak—priced at a $1.3 billion valuation and at $50 million of ARR, up from $10 million in October. So, obviously, $10 million to $50 million in 6 to 7 months feels quite cheap for an AI leader. Jason, I'm intrigued to hear your thoughts specifically on this one.

Jason Lemkin

I love OpenRouter. I use it. It's a very simple way to dynamically pick which LLM to use, right? Going to our conversation from last week, sometimes it doesn't matter if you're not price-sensitive for certain workloads. Sometimes, not only does it matter, but it's incredibly helpful not to have to do all this work yourself: “Oh my God, which model should I pick? How should I do it?”

OpenRouter lets you do it dynamically, or you can pick different LLMs for different use cases, and it just makes it elegant. What I love about it is that it's also really cheap, right? It's quite cheap. I suspect the cheapness is why it's not worth $10 billion, right? When you have such a low take rate from such high GMV, you do naturally get a little nervous about addressing the true TAM, even though we've given up on TAM.

That would be my guess. But they've become the market leader in this space. It's cheap, it works, and it adds a lot of value. You've got to love it, right? It's just that the flip side is something that you—

One of the reasons Anthropic got $20 billion, right, is that a really good Anthropic call at the API level is a buck. It's a buck. Here's my simplification: you can do so much on your $20-a-month Claude subscription or $200, but I can tell you, on all the apps I've built, the complex stuff is a dollar. That scales massively.

If you're taking 1% to 5% of a subset of that, there is, in theory, a ceiling if you don't expand it. What I like is, if you get market leadership in this kind of thing and you're not that expensive, there's no reason to switch. It's not worth switching for a tiny amount more in basis points. It ain't worth it, right?

Just for listeners' context, what the company does is, if you're building an application, OpenRouter acts as an interface between you—the builder of whatever software product you're building—and 50 to 60 different LLMs, such that it can dynamically pick, in real time, which LLM is the right one for whichever call you're making.

It charges about 5% to 5.5% of the money you pay the ultimate model provider. So, if you're building an app and spending $100,000 a year on LLM calls, using these guys, you pay 5% to them. In return, instead of having to access each separate LLM separately, you get access to them all through one kind of API call.

It totally makes sense to me. It's kind of in that Stripe-Twilio business model of an interface. Twilio was an interface between an app builder and all the complexities of telco, and these guys are an interface between an app builder and all the complexities of LLMs.

Twilio has gross margins because they account for everything; it goes for 20%, 30%, 40% plus. They're doing pretty darn well. Whereas, in this case, they're only booking the net revenue at 5%, so maybe there is actually room for margin expansion there over time. It's an interesting business. The world needs it.

The other thing that's interesting about it, which gets to the wider question, is that a number of folks have backed into figuring out what the most common models are. You see a lot of the Chinese open-source models now, right? I always think I'd love to spend time thinking about it, and I just haven't: OpenRouter must have a pretty good sense of which things need state-of-the-art models and which things you can do easily on much cheaper open-source models, right?

Right now, amazingly, all these open-source models are primarily Chinese open-source models. Until either Llama—until Meta reintroduces an up-to-date open-source model—or someone like Reflection ships one, there'll be a US equivalent. But right now, ironically, the Chinese Communist Party is effectively subsidizing the American small independent software vendor by providing cheap open-source models. God bless them, right?

If you look at the winner list on OpenRouter, it's all Qwen, Kimi, and all the other open-source products. It can even turn it on for you. That's one of the reasons I think OpenRouter is so clever. If you want, they will just decide which model to use for a workload. You don't even have to figure it out.

Jason Lemkin

Yeah. At some point, the people spending $30 million a year on Anthropic—corporate IT—are going to wake up and say, “Do I have to spend all this money on Anthropic, or can I pass some of these calls to a cheaper model?”

Given the size of spend that OpenAI and Anthropic are getting, there is at least the opportunity for corporate purchasing to think about whether any of this is doable on a cheaper model. I'm a superfan, right? Superfan of OpenRouter—great software, super easy to use, super easy to deploy. It's like ElevenLabs: just super easy to use, super easy to deploy. I give it a 10 out of 10.

What I've learned from another investment we can chat about is, okay, so they're at $50 million ARR, they said. The nominal take rate is 5%, but some folks probably pay less, right? In some cases, you don't have to pay anything. So, they might be needing to manage $2 billion in inference just to get to $50 million in revenue.

It's easy to see how you get to a couple hundred million in revenue, right? In today's world, what I worry about with companies like OpenRouter is, how do you get to $1 billion in revenue? Do you just wave your hands and say, “These are great founders, they're at the heart of AI,” or do you say, “Oh my God, even if Anthropic keeps growing, some folks won't use it because they'll get big enough and they'll do their own things”? How the hell does this get 20 times bigger when it's already managing $2 billion of inference?

I'm going to give you the argument, which I'm not sure I believe. Look, if you believe in a world where you just look at the OpenAI and Anthropic projections, which cumulatively add up to north of $450 billion in 2029, according to estimates—$450 billion plus—let's call it $500 billion in API across both companies, right? If you take out the ChatGPT consumer business, maybe $300–400 billion of enterprise API calls across OpenAI and Anthropic.

I don't know, if 10% or 20% of that went open-source, that's $40–80 billion. And $40 billion at 5% is only $2 billion.

100% of the market, right?

Jason Lemkin

Yeah, that's fair. You've got to get 100%. I mean, maybe there's $40–80 billion of value going to open-source LLMs, and maybe you can get 5% of that.

The other question, to your point, Jason, is that right now, amazingly, all these open-source models are primarily Chinese open-source models. Until either Meta reintroduces an up-to-date open-source Llama model or someone like Reflection ships one, there'll be a US equivalent.

Jason Lemkin

What I think about OpenRouter, just for investing, is that small takes of large TAMs are intellectually confusing. Harry and I are both investors in a company called RevenueCat. I was the first investor, and they have about 50% market share in managing mobile subscriptions. If you have a paid mobile app, there's a 50% chance they have RevenueCat deployed, okay?

It's competitive, and their net take rate is like half a percent up to 1%, right? Even with all that, there's only so much bigger than OpenRouter they can be. They grew 40% last month because of AI. It's great, and I love the company—I love it. They have a clear path to $1 billion of revenue now, but my learning from that is that sometimes it's hard to do the math intuitively.

If your product is very cheap in a large-ish market but you don't get all of it, OpenRouter could be one of the greatest $200 million ARR companies, right? It's just a risk that I think about more than I used to.

That's fair, but I will give you the counterpoint, which is that the 2 best financial businesses on the planet are Visa and Mastercard. I sit literally 30 yards away from the Visa headquarters. They don't even get 2.5% because most of that goes to the banks. They get 15–20 basis points, but on every dollar every human spends on the planet. It turns out to be a remarkable business.

Jason Lemkin

I'm with you. I'm just—I guess my personal intellectual limitation is that the notional basis-points math doesn't always translate to the real-world basis-points math, right? That's the thing.

RevenueCat and Visa sound great, but niche products that seem mass-scale are sometimes more niche in practice. If your product is $200,000 a year or $100,000 a year, who cares, right? You figure it out later. If your product is dirt cheap, you really, really have to own everything. Own everything when it's dirt cheap.

I think we're all making a lot of AI mistakes here, and our investments are being flattered by high ACVs right now.

The ones that have high ACVs all seem to be doing great because they're getting $50K to $100K per check. Getting to where ElevenLabs got from the early days is much harder than a lot of Lovables and Harveys suggest, just because the large ACV flatters the inputs and the outputs needed to achieve that scale.

Agreed. I'm not sure I could trace it back to OpenRouter, but I agree.

I'm just nervous. I'm personally, as an investor—and this may be one of my many flaws; it's a long list—nervous about exciting AI investments that have very low ACVs right now.

Jason Lemkin

I think their actual TAMs may end up being smaller than they look, despite the epic numbers when we started this conversation, despite Anthropic getting to $30 billion in 5 years. The little tiny crumbs we get out of this $30 billion may not make a whole loaf of bread sometimes. It's a bad analogy, but there's some truth to that.

So rather than shoot from the hip when it's a $7 million post back in the old days, if I've got to shoot from the hip at a $100 million post in the pre-seed, maybe I've got to really believe that that small ACV will scale up. Do you think OpenRouter will be a $10 billion company?

Jason Lemkin

It's always a weird question, because if I knew for certain, I'd go do the deal and not sit here talking to you, right? I mean, that's what they're paying me to do.

I think we're in a world right now where everyone is just doing the build-out as quickly as possible. What that means is everyone on that journey can attract some capital and get some revenue. If you're solving a problem that's in the way, that's a rate-limiting step in terms of getting the AI build-out done, you can get revenue and grow quickly, and I think OpenRouter's an example of that.

You can put on your intellectual MBA hat and say, "In the end, when things settle out, maybe a lot of these businesses get commoditized." You can worry about that. A certain amount of that worry is legitimate.

There's a whole bunch of markets. I'll give you more: there's the labeling marketplace, there's the inference marketplace, and there's products like OpenRouter, where you say, "When things settle down and people start getting more efficient, then all these businesses will get scrunched a little bit."

That's true intellectually, but my advice—and I say it internally—is please don't overthink it. While that is true, at the same time, in the short term, this explosive lift in demand gives you a chance to be relevant. It's your job to add products on top of that such that when the great crunch does come—and it will come in a couple of years—you've just delivered enough value. You see what I mean?

In other words, do I think OpenRouter will get to $10 billion in value on just what they do today? No. If they just keep doing what they're doing today—no more than the inference guys, no more than the labeling guys—when things slow down, all these businesses will get crunched, right, when people start to optimize.

But you have a chance to parlay. You're building relationships with a whole bunch of app developers in OpenRouter's case. Your job is to find the add-on products on top of this that, over the next 2 or 3 years, give you value, or do the adjacent acquisitions that give you value. Maybe you start doing inference. Maybe you start hosting stuff on top that allows you to extract more value from those customers, such that when the thing slows down, you're the survivor. I mean, we saw it.

Yeah, I think that's the challenge with these investments. The one that, if I'm running it, is a dream: I'm at 50 people, right, at $50 million in revenue at the center of this. If I was a founder, this would be a dream job, right?

But I think my learning is Rory's point. The reality is you have to go truly multi-product earlier in this type of situation. Not just a little feature, not just a little enhancement, but you literally probably have to build 5 distinct products to get to that billion. Not all founders are actually up for that. They say they are, but you need a very distinctive founder to run the AI Rippling playbook and say, "Hey, I want to break something up in some ways that's crushing it with 50 people, if they have it."

I mean, again, my dream job. You say, "We're going to do 5 of these. We're not going to wait 2 years. We're not going to just focus, focus, focus." If they're up for it, I would hold my stock here. You probably have no choice.

8. The Mercor Hack and AI Cyber Threats Moving Forward

If you see this sort of Stewart Butterfield-esque reluctance to go multi-product, which was very rational at that time and place, then I'd be less excited to hold stock. I think you've got to run these businesses right now like you're in this insane period of time when money is just raining down on everyone. All the time, you should be saying to yourself, "At some point, the music will stop and 2/3 of the people will have to go. How do I make sure I'm the 1/3 that make it?"

Jason Lemkin

That's what the smart inference providers are doing. That's what the smart people in the space should be doing. How do I lock in?

Look, the truth is, even when the crunch comes, the foundation model companies make it because they're on top of the heap and they have the high intellectual property asset. They're going to make it.

9. Supabase Targeting $10B Valuation

Everyone else one level down has got to be saying to themselves, "When people sober up, they're going to say, 'Oh my God, this is a commodity. There's a bunch of adjacencies. How do I make sure I win in that world?'"

Speaking of whether this will become a commodity in a future world, we've seen the need and the explosion of databases. We've seen some people, like your Lovable and your Replit, incorporate them and build them themselves. Some people outsource to Supabase.

Supabase at $10 billion—Jason, you're the man for this, the man who's used more Replit instances than anyone else. Is Supabase at $10 billion a good buy? How do we feel?

Jason Lemkin

I think I like it. I do think it's an interesting buy. First of all, huge credit to the team. This is one I call an AI tailwind to the maximum.

Supabase was founded, I think, in 2020, right? This is pre-AI, and they're like, "Oh, well, we'll do another fork of Postgres, which is open source and free, and we'll make it easier to use and easier to deploy."

I know it was a hot YC company, which a lot of folks want to say is the unhot ones that take off, but sometimes it is the hot ones. I certainly wouldn't have thought it was obvious in 2020 that we needed another forked version of an open-source database.

Everyone was having issues with Postgres at the low end and the high end. Folks were having to shard it, and it got complicated for big companies, and it was reasonably difficult to deploy at the low end.

That was their idea, but then it just worked with agents. They built a product that could basically self-deploy a Postgres database, and it's what every agentic product needed, right? They needed to spool up a database without humans, and they leaned the hell into it.

They didn't get Replit. Replit went with Neon, which Databricks bought, but everyone else standardized on Supabase, right? They supported them, and then they let everyone—Lovable, Emergent, and all these other ones—white-label it a couple of months ago.

I don't have the exact numbers, but I know more databases are being created by agents than humans. That's the trend you're betting on, all right? Database is a fundamental category of software. It always has been.

The number of databases we're creating is an order of magnitude greater than it was 12 months ago. So why the hell wouldn't you want to bet on the leader in that trend? Every app needs a database.

What's interesting now is, I'm not sure if this is true of Lovable or v0, but Replit changed it a little while ago where every single app has a database, whether you use it or not. They found that enough of them are using databases no matter what they build, that it's not worth adding a database later.

Whether you even realize you have a database, all the millions and millions and millions of AI-coded apps have a database in the background. With Supabase, they get to monetize them all. They're charging these guys for every single database.

I do like this one. This is one where the agents are so far ahead of humans now. Everyone's talking about what the world will be like in 4 years, right? Database is a world where already the agents are creating more databases than humans. We've already crossed that line, so why wouldn't you want to invest in the leader?

Agreed. I think it is literally an excellent example of 2 things we've talked about. One is that kind of thing I just mentioned, which is you start with something and you have to parlay. The other thing, Jason, that you've talked about a lot is being a pre-AI company that brilliantly finds a way to co-attach.

These guys co-attached to the trend. As you say, they did the deals with many of the vibe-coding things. Now their job in the next 2 years is, before the music stops, to be perceived just as MongoDB was the right database for the SaaS era and for cloud. You want to be the right database for vibe-coding and agent apps in 2026, 2027, or 2028.

At some point, when things slow down enough for the Lovable, the Replit, and the other folks to say, "Hey, maybe we should just bake it in and do this ourselves," you want Supabase to be in a position to say, "No, every developer on the planet uses us."

Every agent framework supports us. Why would you do this? Your users will rebel. But the playbook is super clear. As I say, it is literally just like the SaaS and cloud playbook, but on super-fast speed.

This is all going to happen in 2 or 3 years. Make sure that, before things slow down, you are a lot more than you are today in the eyes of your users. Think about how hard it classically has been to deploy a database. Oracle is still massive, right? I've never deployed Oracle, but I can only imagine how difficult it is to deploy an Oracle database.

That is work. These products—and that was disruptive—these things are work. I even have a vector database product, and I deployed it for one of our apps. It only took a few hours, but it took head-scratching and headaches, and not everyone could do it. Supabase, you can do in 5 seconds. It's so disruptive.

All these databases start being easier than the prior alternative. I don't remember when relational databases started, because it was in the '70s, but I remember even in the early '90s—Arthur Rock used to talk about it, though. I remember the relational database days. But I do remember when MongoDB started, and it was the drop-dead-simple, cloud-based alternative to a lot of the other alternatives at the time. It wasn't so much directly competitive with relational databases, but it was for some of the newer use cases. And then they get more complex.

Jason Lemkin

For someone who could spend a month configuring it and getting it going, that is disruptive, right? Yeah, because it didn't take a month. It took a few hours. It was easy. It was JSON. It was whatever.

And you're right, now it's 5 minutes. So, it's just—now, I've no doubt—

Rory O'Driscoll

Actually, it's invisible. You don't even know you have a database until you need it. Here's what's interesting: You build an app without a developer, and you didn't even know you needed a database because you're not a developer, and it's already there and configured and has all your data. It's pretty cool.

That's true, but the odd point I was trying to make is that the tragedy is, that's great, but over the medium term, a white-label business to 5 or 6 vibe coders won't be enough. So, they're going to have to expand beyond that.

Ironically, over the next 5 years, that will mean adding complexity and adding functionality. In 10 years' time, someone—and it won't be me at that point—will be saying, "Oh my God, those legacy Supabase products are almost as bad as MongoDB." There'll be a new alternative at that point. But that's just the movie, and this is Supabase's time to crank. Good for them.

I think it's also a reminder that we've given up on worrying too much about intellectual durability in these investments, right? It's a winner, the growth is exciting, the NPS is high. We're not worried about the fact that everyone else may build their own Postgres databases or that other things may change. We don't even care anymore.

Rory O'Driscoll

I'd say differently, by the way, just to be clear. It's not that we don't care; it's that you just don't have the luxury of caring. There are very few things where you can say, "Oh, this is something that is highly differentiated. It has that level of defensibility." Arguably, LLMs themselves did, because there was only a small number of people who knew how to make the magic.

But you're right. Most of the time right now, I can regret the fact that there's not a lot of barriers to entry, or I can just accept that that's the reality that exists today. The barrier to entry is, as Brian from Andreessen said, speed. If you execute well, you create these barriers to entry over time.

But you're right, Jason. Right now, most of the deals you look at have low barriers to entry in the short term. What that means is, if you stumble, you lose, right? Because if there's 5 of you going out of the gates, 1 of them won't stumble, and they'll win. Over time, by winning, they'll be able to create—I believe downstream there will be barriers and moats created, second-order moats. But out of the gate, you're exactly right: It's a race.

I don't know who the Supabase competitor was, but they didn't get the 2 or 3 key white-label deals, and there you are.

Jason Lemkin

I think this is also why I view a lot of VCs today as enablers. It really bothers me, because Rory's point is accurate. If you stumble today, you may lose forever, right?

I see way too many—the classic VC thing is, "Guys, keep pushing. You've got time. Keep at it. A bad quarter or 2, falling behind the competition..." It's not that I don't think you should be supportive of your portfolio companies. Of course you have to be, right? And what choice do you have?

I just see too many VCs running a pre-AI enabler playbook, where, when folks do fall behind a tick or 2, you see kumbaya activity instead of code-red activity.

Rory O'Driscoll

I don't think I'm a kumbaya, but I was interested to see where you were going with that, because, Jason, I think it goes to your point of what you said before. There's no point in being difficult with founders or being opinionated in the way that you've been before, because they don't listen. You say it yourself: They don't listen. So why?

10. The $1.8B Two-Person Company

Jason Lemkin

No, but I'm making a slightly different point. For example, I'm an investor in a company that's crossed 9 figures in revenue, but it is hitting massive AI competition and issues, okay? It happens, right?

They have a new investor on the board who doesn't really know the space that well, doesn't really want to learn, and frankly isn't as close to some of the AI changes as we are. Every email and conversation is, "Great job, guys. Keep at it." Don't you understand the disruption in the space? Don't you understand the issues?

He's become an enabler—an inadvertent enabler—by being a cheerleader, right? I just worry about it because so many folks are still hiding, and I don't think having enablers around the table, even if it feels good in a given month, is helpful today.

I think enablers can enable a death spiral that you feel good about as you approach the event horizon and your startup implodes.

Rory O'Driscoll

I was thinking about what you said and, frankly, just checking myself: Have I at times? One man's description of an enabler can be another person's description of being supportive, right? And at times, when things are tough, you want to be supportive.

So I was thinking about it, Jason, and I actually think it's a very important comment. What makes a difference is this: You have to be very clear-eyed with your companies on where the competition is and understand exactly what the other guys are doing, and therefore how well and how far behind you stack up, right?

That takes it away from enabling—it's a judgmental term. "Am I being..." Because being a jerk is also a judgmental term, right? What you're saying is correct: You're not a useful board member unless, A, you understand what the company does and how it compares to the direct competitors, ideally with hands-on experience of the products, and then, B, you find a way, without being a jerk, to keep the company honest about where they are relative to the competition.

What are they seeing? What do our competitors' products do? What do the adjacent products do? How do we think about that? Not in a defeatist kind of way, but how do you distinguish between, "Oh, these guys leapfrogged us for a month and we need to get our act together," versus, "We are a year, year and a half behind in the market. We may never catch up. We should sell while we can"?

I think that's threading the needle. You don't want to be an enabler, but you don't want to be a jerk. You want to be supportive. It's a slogan we have internally. It's not so much founder-friendly; it's founder-honest. But, actually, as I think about it, it has to be founder-fact-based.

The number one thing—and I'm even thinking about a couple of my deals where I have to do some work this week—is, do you really understand where your 2 direct competitors are, the strengths and weaknesses of your product, and what the last 3 wins and losses say about how you're really doing in the field? Because if you don't, you're just cosplaying a board member. And are you being honest about what has to change?

Yeah, agreed. Okay, chaps, there are many other topics that we can discuss. I often get chastised for my selection, so I'm going to—

Jason Lemkin

No, I want you to make the selection. You make the selection, Harry. I'm too tired to decide.

I think we will have to, at some point, address Mercor. I think someone in a comment put, "Oh, Harry, you often shill them and say you can't not talk about them when there's trouble." I don't like to do that.

For context, Mercor is obviously a data provider to some of the largest companies in the world, most notably Meta, which they have since reportedly lost, in part or entirely, as a customer for their data. It was also, unfortunately, at the same time that Forbes released its list of young billionaires, where the founders of Mercor are on that list. Very unfortunate timing there for them.

Jason, I'm sure you've got an opinion on this in terms of, bluntly, the Mercor hack and then losing Facebook as a customer.

Jason Lemkin

Well, just 2 thoughts. Not that long ago, I was with part of the management team of one of the leading hyperscalers. I was with him when there was a minor security issue with a third-party vendor.

Relatively minor, right? Not—not like this. Not like all of likely Mercor’s private data being exposed. What he said was, “There’s not much higher on our list with partners.” When this happens, there’s not much higher. We have no tolerance. It’s not worth it. There is no tolerance.

In this case, they got a pass because it was a relatively minor issue with a vendor that was honest, they fixed it, and it did not lead to any internal data issues. It was just an external issue. But it was crystal clear: there is no tolerance. It is not worth it for us.

The other thing—and I’m not an expert in the data-labeling or likely Mercor space—is that I don’t know how fungible the products are at some level. The more fungible it is, the more freedom you have to route that to what’s left at Scale, or whatever the guys at Handshake want to do. If it’s not fungible, you can bang your chest, but you’re still stuck using them.

One of the realities of the space is that all the largest customers are customers of all of them. They are generally heavily fungible, so I would be very worried, because this comment was chilling from this executive. It was, “This is the highest thing on our list with third-party partners: security. We’re exposing our applications to folks we’d rather not expose them to because there’s so much confidential stuff flowing through what we do at all levels. So we have no tolerance for anything that’s material.”

I just don’t see how you would come back from this if you’ve crossed that line. There’s just no coming back unless you have to. I think it’s potentially death.

The reason I bring it up is that, in the old days, through 2023 in our lifetimes, you always got a pass once as a vendor. Even for the worst breaches and the worst issues, unless your app is down for weeks, you get called into the CISO’s office and you get yelled at. It was brutal, but you always got at least one because it was just the reality of working with emerging vendors. But I just don’t know that you get a second one here. I just don’t know.

A lot of the outcome depends on how you handle it. I don’t know at this point: do we have any sense of whether it was a state actor, a malicious employee, or just a stupid configuration breach? I don’t have any sense of the forensics here.

Rory O'Driscoll

It was an organization—I think they called it likely Lapsus$—which basically holds you for ransom, and you have to pay to get it back. It’s commercial, like you see.

Jason Lemkin

Gotcha. It’s commercial—decent, decent, honest criminals.

Rory O'Driscoll

They’re very successful.

Jason Lemkin

I asked the team if we could invest in them. It seems this is one of many very successful—

Rory O'Driscoll

Actually, you’ll find, Harry, they don’t actually—

Very good. They’re very good, right?

Rory O'Driscoll

—they don’t need much outside capital.

Jason Lemkin

Right. That, in one sense, sucks. On the other hand, the good news is at least they’re going to be rational and you can buy them off.

Rory O'Driscoll

I think—is it fatal? Hopefully not. In these things, you generally pay a pretty significant penalty. How you handle it is a key part of it. Were you straightforward and honest with your suppliers and your customers, and did you let them know what happened? Was it entirely your fault? Were you crassly stupid? Was it bad luck? Yeah, so you can manage through it.

I think it is hard when you have 4 or 5 vendors of roughly the same thing. On the other hand, cynically, there also appears to be an insatiable demand for labeling data. I think the Meta statement didn’t say they had ended the relationship; it said they had paused it, which makes sense. Everyone’s going to pause, right?

My guess is that the likely outcome is you lose a fair amount of revenue and a fair amount of time. You’re going to have to spend a ton of money to bolster your defenses, but if you do the right thing, you’ll be able to earn your way back slowly and over time. That’s the likely outcome here—not, hopefully, fatal.

Jason Lemkin

I think Sam Altman, to tie it back to the start, said something this week: that massive cybersecurity attacks from AI are coming. I genuinely think that most B2B companies are going to get hit worse than likely Mercor.

This likely LiteLLM incident was one of the weaknesses that they had. How strong and state-of-the-art are the security teams at most B2B companies? They’re not. They’re relying on a hotchpotch of open-source and other products that are barely monitored in many cases. They’re busy, they’re under pressure for profitability, and they’re managing their teams.

Likely Mercor is probably a small company, so they probably don’t have a huge team. But my point is that it’s going to be really tough on a lot of startups and scale-ups because they just don’t have the teams to deal with the levels of threats that are coming from AI.

This is a start. It’s going to get worse. This likely LiteLLM incident is going to happen to everybody. As soon as you figure out that you can hold all these B2B companies and others hostage, they and their network of thousands of affiliates are going to do it.

I don’t know how the average pretty-good-to-mediocre company is going to keep up when they don’t even really have a security team. Remember when Gainsight was offline for a month? Drift was permanently destroyed. And this was before all this craziness.

There was an old saying my CTO told me back in the day: “The only reason we haven’t been hacked is that no one cares about us.” That has resonated in my ears for years. With AI, you can hack anybody you want.

Rory O’Driscoll

I think, Jason, you’re exactly right, because in general, security purchases tend to lag. New stuff gets deployed. People should think about security up front, but they tend not to. They deploy, bad stuff happens, and then they panic and think about security. That’s the way it’s always been in every cycle.

I think there are 2 separate vectors. First of all, the AI apps themselves have security needs that are different from pre-AI apps. You’ve got the whole prompt-injection issue. But I think the bigger issue is Jason’s point, which is everything else: using AI, the bad guys can now automate attacks, automate phishing, automate voice duplication, and all that kind of stuff.

I think Anthropic referenced this. It’s the Red Army quote that I often use: “Quantity has a quality all its own.” With AI, you can make a quantity of fake people. You can make a quantity of attacks. You can do the thing.

I think it’s not so much that you’re going to get attacked because of your AI apps. It’s much more that AI apps are going to attack you. A lot of the AI doomerism I kind of discard, but this is a legitimate and big issue: the ability of AI to escalate the velocity and ferocity of attacks.

That’s why I think the whole response of security stocks going down in response to the Anthropic announcement was absurd. Anyone who’s cutting back on the security budget in 2026 is missing the point. Not only are the attacks more ferocious, but, thinking again, Harry, to your point, the second statement is really Captain Obvious, but it’s worth saying: these things get more important every year because more and more of our stuff is online.

The percentage of things we do that are done online just continues to escalate. It’s stupid stuff—it’s all the stuff in your house, all the stuff in your financial life. There’s nothing that matters that isn’t done online at this point, which means that attacks there can attack more and more of our accounts.

You should be seeing a real acceleration of investment in a different class of security to cope with a different class of threat. It’s a fatal error if you don’t. There are only really 2 things that can destroy one of these companies: the app goes down for a long period of time, or the app gets hacked grievously. Those are the 2 red-card fatal errors. That’s where you’re going to have to put the money.

There’s a wave of the second tier that’s coming. It’s just massive, right? Even Supabase, which I’m a superfan of, a lot of times folks turn off the default security, and we’ve seen these issues. In the old days, no one would find it because they didn’t care about our little Rory, Harry, and Jason app. Now AI will find it in seconds on the internet.

It will use every possible way to penetrate that, steal the data, create malicious acts—everything that is possible on the face of the earth that can be delivered with software. Maybe there’s nothing we can do, but I think we’ve all underinvested in the attacks to come.

It used to be hacker farms of people in, fill in the blanks, the Philippines or Russia. Now it’s going to be hacker farms of AI agents cranking 24/7. It’s going to be miserable.

Guys, you can choose 1 more, okay? Jason, Rory doesn’t feel like choosing, so he’s delegated to you.

Jason Lemkin

Absolutely.

Okay. That’s the 2-person company. Again, for people who use this as their news on tech: a 2-person company that uses AI intelligently, scaled to $1.8 billion in revenue selling GLP-1s.

Interesting. A lot to unpack there. Wix buying back 31.6% of its shares, given its low stock price. Oracle getting rid of 20,000 to 30,000 employees via a 6:00 a.m. email. Jason, any that stand out, baby?

Jason Lemkin

Obviously, at some level, people are so excited about the one- or two-person billion-dollar company. Facts were glossed over, points were missed, and all that, right? Oversimplified. They used deepfakes. They made representations about doctors they shouldn't have. They did all the wrong things. They did all the crappy affiliate marketing stuff people have been doing for 20 years, and they did it at scale with AI and built a big business on it, right?

I get it. But what is interesting about it is, let's not over-glamorize this company that maybe is at the edge of fraud in many ways in its marketing tactics. The fact that they could use AI to scale this with 2 people and maybe some consultants and stuff on the side—it is—we are seeing the future. Why? AI is completely changing marketing. Marketing has not gone away. Dario and Sam are everywhere because marketing matters as much now as ever. And this is a different version of how marketing is changing. If you don't adapt, right? Maybe buying TBPN's a bad idea, but you got to adapt to the new world of marketing.

There are a number of startups that have tried to automate all this marketing at scale with AI. Most of them are terrible, right? They don't quite work. They create boring assets. They look like awful art. They look kind of fake. They're that bad, a lot of the assets. They don't have context. They use too much stock art. But there's no reason that in a year everyone shouldn't have the AI power to blanket the entire internet with the best hyper-personalized marketing in the world. Hopefully, in a year, all of us will have the power to do some of what Medvi has done. And I think, if we step back, it's a chance to see 12 months into the future, when all of us will have more of this power.

In other words, just because they're illegally selling weight-loss drugs that are off-label to people without the appropriate FDA safeguards doesn't mean they're not great marketers. I talked to so many CMOs who are still struggling to run the 2023 playbook and falling further and further behind. You have to adapt, and crappy automation doesn't work anymore, right? But these guys that were able to do that at mass scale and target everybody, this is the future that we all should know. Marketing, I guess—the code is marketing—appears to be more powerful than ever in the age of AI. Humans have to control it, but if you don't leverage how marketing is being done in the future, you're going to be stuck in the past. You're going to be killed by folks running the old playbook.

Rory O’Driscoll

I should be getting the GLP-1 ad specifically targeted to you. Jason, you don't technically need it, but I noticed your jawline on 20VC could be a little bit tighter. What if we just microdosed you today so you could get some of that T-shirt look that Harry has? I'd click on that in 2 seconds.

You would, you know.

Jason Lemkin

I shouldn't get that, versus stock art of some grandpa running with his golden retriever on a beach, right? So, I both think this is fraud and overhyped and the future. We're watching the future.

I think you're right. I would argue that a huge percentage of it is not on the fraud side, but on the hairy edge of the regulatory side. But I actually agree, upon reflection, that what you're saying is correct. It's funny. It's a little like the last comment. What we're basically saying is the most entrepreneurial people on the planet out there right now are the hackers and the kind of dodgy marketers that are on the front line of pushing things, but the tactics that they're using and the way they're leveraging AI, everyone's going to be doing within a year or 2, or they're going to be left behind. I could go along with that.

Jason Lemkin

Yeah, there was an era in the old days when the best affiliate marketers knew things nobody else did and built billion-dollar companies out of it, right? Then there was an era, which is now faded, when some of the best companies used SEO in a way nobody had used before: millions of pages, right? Even things—you know, maybe it's only worth $10 billion—DigitalOcean was built entirely on SEO farms. So was Zapier and others. There was a group of folks that knew how to do this, and it worked.

The next group of folks will know how to do this mass personalization at scale that really works to millions and millions of people, and they will win like the prior affiliate marketers. They will crush folks, and the rest of the world just won't get it. They'll think it's dark arts, and they need to become agentic marketing experts.

Rory O’Driscoll

But in the same way that the great affiliate marketers actually largely originated from porn and Viagra—if you want the dark arts.

Jason Lemkin

Yeah, yeah, agreed.

And here, I think probably some of the best next-gen AI marketers will spawn from GLP-1s or anything at the borderlines of next-generation e-commerce.

Rory O’Driscoll

Yeah, but I think that Jason's insight—I agree with that, by the way—is profound. I mean, there's no doubt that criminals—I mean, like many of these new technologies—are adopted by crime, adopted by porn, and dodgy marketers. But I think Jason's insight is a profound one, which is that the marketing tactics that are used by them and look very kind of dark-artsy, over the next 10 years, tend to be adopted by everyone.

And now, as you say, everyone in corporate America is an SEO expert, whereas 20 years ago it was a dark art. And I think you're right, Jason: in the next 2 or 3 years, if you're not adopting agentic marketing as a digital marketer, you're just going to be left way behind. Personalized marketing and leveraging the technology—I knew it already intellectually, but actually, I will say, you kind of crystallized it in my mind because I've lived the last year.

I remember when SEO marketing was literally something that only the leading companies did, and there was a 2- or 3-year period where they had hypergrowth, $100 million businesses kicking off 30% cash. And then those businesses went away as normies adopted the technology, and the correct play was to be the software provider helping the normies tool up. I think the same is true here in agentic marketing. So, that's my stolen insight from you for the day, Jason. Now I'll go look for those companies.

Jason Lemkin

Well, all Medvy has to do to make $400 is drive a GLP-1 lead to someone that buys a product that, just like tokens, the world can't consume enough of. Most of these folks are selling the exact same product, and they're fungible, right? So you're getting $300 to $400 for delivering a customer that already wants to buy your product, right? It's a moment in time where it's a great marketing arbitrage if you're excellent at this. You get $300 to $400 for delivering a customer from a Facebook ad. Man, you want to run this.

You do what it takes, boys. A lot to discuss this week. Thank you for being so good. It's so good to see you.