Windsurf x Google x Cognition: Full Breakdown: Who Made Money, Who Did Not

Windsurf was hopeless before this deal. You could never attract 30 S-tier developers to fix Windsurf, but they fixed it in 1 hour. Literally, in 90 days, Windsurf could be better than it was before this deal.

This deal, as crazy as it sounds, makes sense because they have the people. Some of these deals almost seem too cheap. If you look at the crappy multiples we see for some startups—an undifferentiated company raising at $300 million pre-money with less than $1 million of revenue—and then you see what Windsurf is selling for: 20x revenue. Lovable’s raising at 20. Ready to go.

Guys, I am so excited for this. I literally have been looking at the news over the last few days going, “When’s this show? When’s this show?” So what else would we freaking start with? I would love it if we went with a different topic to start with, but on Friday, I think it was, Windsurf announced—and Varun announced—that there would be an agreement with Google whereby the IP and the team would be joining Google and DeepMind.

If we just start there, how did we break down, how did we assess, how did we think about the news that the core team and the IP were going to Google?

Here’s a framing comment. The memo that Cognition sent to its employees was that they’re buying an $82 million ARR business. It’ll be interesting to see what they paid for it because not much, probably, right? That sounds very exciting, right? I do think Cognition got an amazing deal.

But here’s the framing question. Again, so many AI folks stretch reality when they quote revenue numbers. Windsurf said, like, 60 days ago, 90 days ago—before the OpenAI deal was announced and before they lost access to Claude—that they were doing $100 million.

So did they decelerate? If they decelerated from $100 million before the OpenAI deal to $82 million today, that’s a lot of deceleration. I would take any deal, and I think that may explain a lot of it, right? If you’re decelerating at that level, the minute the OpenAI deal falls apart, you’ve got to find a lily pad to jump on, don’t you? That’s massive deceleration, if it’s accurate.

I think it’s clear they wanted to find a lily pad. But again, zooming out, it is just worth restating the cast of characters here, because literally this is one of those Agatha Christie mysteries where anyone could have killed the victim, right?

Stepping back 3 months ago, OpenAI said they were going to buy this company, or at least intimated they were going to buy this company. It sounds like some combination of Microsoft and the FTC prevented them from buying it. In the meantime, Anthropic mugged this company by taking away access to its API, which is how the company delivered its product.

Then, as you say, on Friday, Google stepped in once the exclusive period was over and bought part of the company, leaving an empty husk. Then, on Monday, Cognition bought the empty husk and got general all-around Silicon Valley kudos and plaudits for stepping up and being a good guy.

That’s a lot. That’s a freaking saga, start to finish. It’s almost like you’ve got to unpick it bit by bit and figure out all the genuine commentary here. Almost everything you need to know about the AI revolution is embedded somewhere in this kind of play, because all the characters are playing. They’re all doing their thing, and there’s just a lot in it. Let’s unpick it bit by bit. I think it’s bluntly worth doing.

Yeah. So if we start—

But I do think you’ve got to start with the revenue.

I think if you start, you have to actually start before the OpenAI deal to understand it.

You have to start with the revenue from the seller’s perspective. I’ll give you that from the buyer’s perspective.

What is painfully clear from Google is—let me state this clearly, which is bizarre—they don’t give a shit about the revenue. They literally said, “We’re going to buy this company, but I’m going to ignore the…” By the way, I agree with you from the seller’s perspective: the revenue, and the significance of the revenue in terms of how it forces you to want to sell, is huge.

But let’s just say Jason is correct—and, by the way, I broadly believe he is—which is that the market data convinced the company owner, the company founder, that they should look for an exit. Maybe it was revenue growth, decreasing competition, being a number 2. A whole bunch of things said to this company, “You’ve got to sell,” and revenue is part of that.

The buyer literally said, “I can’t take the revenue because of the FTC, but I don’t give a shit. I’ll just take you and move on.” Bizarre. It’s just a weird thing, but the buyer mentally ascribed a value.

Let’s just do it. If OpenAI was buying for $3 billion, and Google was buying everything for $3 billion, and Google, presumably, is buying everything except the revenue and $100 million for $2.6 billion, implicitly it’s valuing the revenue, the $100 million, and the other employees at $400 million, which is about what Cognition probably paid, right?

So, to your point, the revenue is clearly a tell and a signal to the seller that they need to pursue a sale, and I think we should spend some time on that. Bizarrely, at the same time, it has no ascribed value whatsoever to the buyer. It’s weird.

It is strange. Neither of the buyers, with the possible exception of Cognition, cared about the revenue, right? I suspect what happened is that their numbers—the numbers they published, or the whisper numbers—were roughly accurate: $12 million at the end of last year to $100 million by April.

Okay, $12 million in 12 months. That is consistent with Lovable and Replit growth, right? It is consistent. It is very plausible. Then they get this offer from OpenAI, and it’s very compelling to the team. It takes some of the stress out of the competition.

The VCs maybe are excited to have OpenAI stock, right? Because they get another play. They’re like, “Gosh, I think you could be worth $10 billion, Varun, but we’ll take that turn of the card inside of OpenAI,” right? Probably people got excited about it.

Then it got really complicated, right? Before you even get to the weekend, somehow this deal just didn’t work. I’m confused why their IP flowing to Microsoft killed the deal. Although I get the issue. I think I get the issue. We could talk about it, but the deal dies, right?

If you’ve then gone from $100 million in April to $84 million today, and you’re the founder, you’ve got to jump fast. You’ve got to start rolling those phones: Google, Microsoft, whoever is going to buy this thing, because you’ve gone from a rocket ship. This is why M&A is so stressful for founders. It can derail your business. M&A can utterly derail your business.

Do you not think he had the chance to say, though, to Dario, “Hey, the deal fell apart with OpenAI. We’d like your models back, please. We’re very sorry for the OpenAI deal, and we’re going to send you a truckload of money. Thank you for opening your models again.” They could have come back.

Well, honestly, Harry, Cognition announced they got full access this weekend. It literally happened in an hour. It happened over a DM. So you’re right: they could have, but your point’s taken, Harry. They should have been able to. If Cognition can turn it back on over the weekend, they should have found a way to do the same, right?

Point taken. But you’re implying that that’s the only issue, right? It’s interesting. We want to have this conversation about how hard it is, because there are 2 decisions here. Why did you initially choose to sell for $3 billion? And then the second question, a much more different question, is: once your first deal has imploded, why did you dive for the door really quickly?

I think those are different questions. The second one is easy to answer, right? As you say, you put yourself out there, you imply you’re going to get bought, you go through all those dynamics, someone cuts you, and you get cut off in terms of access. Yeah, you can get it back. You’re mentally spending the $3 billion. It takes a lot to come back from that.

The interesting thing is that this year we’re going to have an example of a company that did that amazingly well. Figma and Vitali picked themselves off the floor after the Adobe deal fell over from the FTC and said, “God damn it, we’ll show you we’re all wrong. We’re just going to make this thing worth $30 billion and ram it up your ass.” They’re going to do it, right?

But they did have time.

What exactly?

They had time to deal with this shitty situation, right?

Jason, you’re exactly right. Generally, you find—yeah, people matter, but circumstances matter a lot.

I think in this case, you’re alert to the fact that if you’re running Windsurf, you’re the number 2 player. Even on an independent basis, big guys are moving in, you’ve had this access issue, and maybe you’re seeing a slowdown. I don’t have access to the numbers, and suddenly you start thinking maybe the Figma alternative of saying, “Keep on going. I’m number 1 in the independent space,” isn’t there, right?

What I think is interesting about that is there’s a lot being—the economists talk of revealed preference: don’t look at what you talk about; look at what you do. There’s a lot of revealed preference going on about how founders think about the upside of independence versus the safety of a shit ton of money.

It’s pretty sobering if you’re a VC, because in general many founders have opted, in some of these acquisitions including this one, to say, “We were about to raise at $3 billion 3 months ago, before all this happened,” which implicitly says, “I’m going to make this thing worth $6 billion or $9 billion.” And now I’m not pausing for a single second before I grab that $2.6 billion and do what it takes to get in from the cold.

To Jason's point, there's a lot of what we think of as wildly successful, hypey companies where the person who knows the most is sitting there going, “Maybe I'd like that cold, hard cash and the safety of a trillion-dollar balance sheet.” It's just worthy of note.

It's absolutely worthy of note. My question on the back of that is: given the structure of this deal, only the top 20 or 30 engineers plus the investors were taken care of. It was pretty clear from the CEO who was left behind that he was put in a pretty difficult situation and forced to come up with an outcome pretty freaking quickly, which he's done very well, seemingly.

My question to you is: does the board and the founding team not have a responsibility to the team members beneath them—200 of them—to find a better outcome? I actually am going to defend them with no data, but I think I can guess what happened, and I could be wrong.

Look, if you think about the people who started in the last 12 months who didn't get any acceleration, I understand the point. Let's dimension that out as a number. Probably you're diluting 57% a year, which means you're diluting out 5% of the company a year. So the people who joined in the last 12 months own 5% of the company. It's $130 million.

I do not believe, knowing the investors—and the founding team, I haven't met them, but I've heard great things—that they sat there and said, “We're going to shaft people for 5%.” No way.

It did not happen.

So then why did it happen? It's pretty obvious if you read the FTC guidelines, right? You have to, I'm going to say, pretend that the remaining company is a viable independent asset and that you haven't de facto sold the company.

It's a fact-and-circumstances thing. In other words, there's a whole series of things you have to do: if you pass all the tests, you're safe, and if you don't pass all the tests, you're not. I'm willing to bet that a large part of this was, “We can't make it like an acquisition for everyone, because then the FTC is going to say it's an acquisition.”

So you are probably put in this really tough place as a board. You have to say, “Look, it's easy to say to maximize for all shareholders. It gets hard when you have to do the best for all shareholders, even though you're doing badly for some group.” It was a tough call. I do not believe anyone involved wanted to save 5% of their consideration by sticking it to people.

No, there's just no way.

No way.

There's just no way. I mean, I've only met Varun once, but he struck me as, on the founder ethical scale, pretty damn high. There's just no way he sat around and said, “You know what would be fun, Rory? Let's screw the 5%. Let's take 5% and stick it in our pocket.” Nobody even cares about the delta. It makes no sense. What I think happened, going to Rory's point, this weird structure is they dumped two billion or two and a half billion in Windsurf's bank account, right? And what Windsurf did now—there are all the lawyers trying to figure out better ways to do these cockamamie deals—but I think what they do by default is they dividend it out. They dividend it out, right? And I would imagine under those FTC guidelines it's very difficult to dividend stock to someone that doesn't hold any stock.

Correct.

No, that is a flaw in this system, and maybe there are ways at the edges to hack it. I'm sure they thought through it, but at the end of the day, dividends have pros and cons, including $500 million of taxes in this deal that went to the federal government instead of the shareholders, right?

And I suspect that this is the best they could figure out, but it doesn't mean it wasn't a terrible outcome for some folks. I think it's got to be the structure of the deal, nothing else. No one tried to rip off the people—the employees who worked their asses off for the company. It just doesn't work that way.

I couldn't agree more. I know Neil Mehta obviously very well. No freaking way.

There's no way he sat down with Varun and said, “Let's rip off the employees that were there for 50 or 51 weeks. How can we shaft the guys that just missed their cliff by an hour? How could I make 6 more cents on the deal?” There's no way, right? There's no way.

Was this an amazing outcome for investors? It was an amazing outcome for Greenoaks and then increasingly less interesting thereafter. I mean, the first point to note is that the starting point is we have all these wonderful documents that have been established—Delaware precedent, 150 years of precedent—and then we're now just going to ignore it and start making shit up in an attempt to avoid an FTC investigation by not calling it an acquisition.

With that caveat, I think it was a good outcome. But the truth is, as Jason said, first of all, you were getting $3 billion in an M&A outcome, which means that, if you were getting stock, it probably wasn't even taxable. So it was an amazing outcome at $3 billion of OpenAI stock.

Now you got $2.6 billion, but that was paid to the company as a license fee, which means it's taxable at the company level. You might have $500–600 million of corporation tax, depending on their income position. Some of that money went right back to the government, and then all the shares get bought back.

Let's assume—and then you have to start making assumptions about now—so it's a $2 billion, plus or minus, thing. Look, Greenoaks is going to do amazingly well no matter what. The first round that Kleiner did was at $450 million, probably $500 million post-money, so 4x, plus or minus. The $1 billion round starts to get harder to make money on now, right?

So my guess is the early investors made a lot of money, deservedly, and later investors got a so-so return. Not bad; everyone cashes the check. But again, back to my comment, it's worth remembering that they were about to raise money at $3 billion pre-money 3 months ago, which would have been a down outcome. So, good outcome for the investors who were in early, increasingly less so for later, and that's with a high degree of uncertainty because we don't know, and a high degree of tax inefficiency.

And just on top of that, the rough math you did, which was great: a $2.5 billion deal after tax, call it $2 billion, right? Maybe less. Kleiner invests at $500 million, so they get a 4x, right? My life experience, Rory—you have more—is that 4x is always like 3x. Whatever the math is, there's always a reason. When you actually see the distributions, whatever that math is, it's never as good as it looks, right? However good it is, it's never as good as it looks.

Yeah, stuff happens. Before we get to the Cognition news, I just want to stay on the structure itself. This is now the second time in a very short window that we've seen this structure being enacted for an acquisition of sorts, if we can call it an acquisition. Is this the new norm? Should we prepare ourselves for a world where this is what's expected?

First of all, there have been about 5 of them, right? The Scale one was slightly different. I think that's the one you're thinking of, because, if you remember, Meta—I will call them Meta. I will be a good corporate person—actually bought 49% of the company, whereas here they didn't do that.

So there are slight differences in structure, and they're being studied in every law firm as we speak. There'll be a standard playbook by the end of the week. You've got Inflection, Adept, and Character.AI, which is the other one that Google did. The Character.AI one is most like this one, which is a licensing deal that Google did. Google did this one too, et cetera.

Yeah, it's very different than Scale, right? It is very different than Scale.

Very different. I mean, same net outcome in the end. You called it the day after the Scale deal. I was trying to be polite in that first call because I'm trying to be a polite person. And you called it and said, “It's an empty husk.”

Now what we've seen in all these deals is the remaining asset is an empty husk. If you want to giggle, by the way, go back and read the Friday press releases where they describe the companies being sold, because this is the kind of shit you have to do. The independent company has 250 wonderful employees, $100 million, and is going to continue on gloriously into the future, because that's the shit you have to say when you issue the press release.

Focus on the enterprise.

Yeah, focus on the enterprise. Meanwhile, they're frantically selling the thing by Monday.

But are these things the norm? I'll say 2 big, high-level comments. I don't think they're going to become the norm, but the mere fact that they exist—that this has been done now 5 times—is going to cause a whole bunch of second-order questions that are going to make transactions more complex, because you have to at least think about this.

Why is it not the norm? Google's buying Wiz right now. They're paying $32 billion. It's going to take them a year and a half. Not for a single second have you, Google, or anyone on the planet suggested they should buy the 3 founders of Wiz and leave the asset behind. Why? Because in that case, the business is the asset.

Google would not give $31 billion for the team and leave the revenue behind, because Google with enterprise revenue is worth something, right? In most cases, most acquisitions, you don't just want the founder; you want the business. So the question is: how many companies are so unique that you don't need the business? You just need some combination of the founder and a nonexclusive license to the tech for those kinds of businesses?

This is on the table, but I don't think it's the norm.

The related point I just wonder about is, with the Cognition piece, will this lead to heartburn? I could see someone at Google getting fired over this. Did you guys know about this Cognition deal? I could literally see it.

I remember back in the day—I mean, Rory, you would actually remember the details—but I think eBay bought Skype but forgot to get the IP, right? They fired like half the corp dev team for forgetting to check the license. I'm exaggerating, but it happened, right? I wonder if a bunch of folks in corporate get fired this week for not realizing the rest of it is going to get bought in an hour on Friday night. Right.

I think, first of all, I am sorry you're saying that. I don't think they will, but if they do, it would be for a slightly different reason. One of the hard things about any deal where you're buying, where you're cherry-picking, is that you always have this feeling: Am I leaving the good bit behind?

I totally get your feeling. If I was doing this from a corporate M&A value perspective and I left the revenue and other employees behind, and then it got bought the next day, there'd be a little cognitive dissonance. You're right: Do I feel foolish here?

But I think the reason that poor business development person might keep their job, but the lawyer might lose theirs, is that the BD person could probably say to Sundar, "Dude, you told me to get the people and the IP. Mission accomplished, and you said you didn't care about anything else. Mission accomplished. Who cares?"

It may be that the lawyer should have added a clause to the license agreement saying you have to pay us a breakup fee if you sell the company within 12 months, just to keep more distance between the broken husk and the thing.

Because the real risk on this is not that Cognition got $400 million. Google, to your point, Jason, doesn't give a rat's ass. The real risk is that the facts and circumstances make it clear that this was de facto an acquisition and the FTC comes sniffing around, because they are sniffing around on some of these other deals.

Yeah, I think it has questions for Meta on one. They have questions for, I think, Amazon on one, on the Adept thing. So the FTC is not going to let this happen without poking around.

And it may be, to your point, that it would have been better drafting of the facts if you'd said, "In return for leaving you $100 million, you have to sail on your little longboat like Captain Glide, cast adrift on the ocean, and you can't run for cover for 12 months."

When we look at Cognition, a week ago they were the ones that were left behind. Cursor, Lovable, Replit, and then obviously also Windsurf have paved the way. Despite taking an early lead, they were left behind completely. What does this mean for them? Is it a good deal, and how would we analyze that portion of this deal?

It's possible I'm wrong. I think it's epic because, listen, I don't know the Cognition team directly, but they're clearly S-tier, off-the-charts good. They're one of the elite teams, and they build a coding agent, which is sort of interesting, right? It has, I guess, some traction, but the idea space is huge. I mean, this is massive. It's generative.

What's the problem with Windsurf? Well, they lost their top guys, obviously. How many did they lose—20 out of 250?

Totally.

40. What does Cognition have—40 brilliant guys?

So, they could—I'm not saying there's a slight learning curve—but they can plug this hole in 30 days. I think Windsurf was hopeless before this deal. You could never attract 30 S-tier developers to fix Windsurf, but they fixed it in 1 hour, right?

This could be better in literally 90 days. Windsurf could be better than it was before this deal. It is entirely possible it could be a better company than it was before this deal, as crazy as it sounds, because they have the people.

I totally agree. This was a genius move, right? You look at the Cognition people and go, "Yeah, clearly wildly smart." The whole fun fact about them playing, I think, poker with Founders Fund—it's all—they're clearly, as you say, existing in that hyper-brainiac space.

The product had kind of not—it landed with a little bit of a thud, right? The Devin product, they were—never write someone that smart off—but it wasn't killing it, right? It's not directly competitive with Windsurf, so it's a slightly different thing, but it was okay with this.

You're right: They just hired great people. They got massively good publicity. There's kind of a good, warm, funny feeling, and everyone recognizes—I mean, it's basically like a big sign saying, "Fuck, we are still very, very smart people because we just bought all the rest of it for half nothing," right?

And not only that, but they get good people brownie points. I mean, they're saying they're accelerating stock for people, and in one sense that's amazing and righteous and good. On the other hand, you're getting accelerated stock in a privately held company. You know, I don't know how much of a needle-mover it is, but stupendously good optics plus $82 million of revenue plus $100 million in cash plus great engineering. They win. Anyone questioning that is just not paying attention.

Is it not the best deal ever? Because after $82 million in revenue and $100 million in cash, you're basically paying $220 million for great people.

And getting their Anthropic license—their direct pathway to Anthropic—back. Now they're competing with Claude Code, but they got it back overnight, over the weekend. This is a great deal.

If you want to be in the space, the president of Cognition said he got a DM on Friday night and did the deal in 15 minutes. I mean, that's the answer to your comment. It's a great deal, which is not to say it'll still work, because remember, we had the number 2 in the space, rightly or wrongly, saying, "It's getting hard. I need to get out," and now the remaining assets are still in the space.

So, if it turns out that it's hard for anyone except the model companies and maybe Cursor to make money, then there's no magic pixie dust that gets them out of this market. But if you're going to compete in the market, it's a damn sight easier to do it with $80 million in revenue and another $100 million bucks.

So it's a great, epic deal, which is not to say they're still going to make it as an independent company. But definitely, if this doesn't work out for them, they can all do Goldman Sachs M&A because they won here.

What do you think Cognition's revenue was pre-this?

I'm guessing 8. Eight of the $82 million. Everyone's so braggy about their revenue on social media in AI. If you're not braggy, I'm assuming it's sub-10. So braggy.

Maybe not that, but let's just say my sense is Windsurf will be a meaningful addition to the revenue line.

I know we could spend the whole show on it, and Harry, you're the boss. I'm still a little bit confused, though, and maybe it doesn't matter because it's past. I'm a little bit confused about why the OpenAI deal fell apart because Microsoft would inherit their IP. I half get it. I kind of half get it, but I don't really understand it. It was an issue going in, right? It's an ambiguity in the OpenAI structure.

What I don't understand is this: It doesn't make sense to me that Varun would say, "I'm pulling out of a $3 billion deal because someone may have our IP." Why would you care if you're Windsurf that much?

The one thing I will say on that: I interviewed Varun, and he made it very clear to me in other conversations that he was very excited to make this deal happen and very much looking forward to being part of OpenAI. He believed that a platform play with their distribution was the winning strategy.

Yeah. So he didn't pull out of this because of the Microsoft license. It wasn't his idea, right? Jason, exactly. I've been trying to figure this out too, because there are 2 separate threads.

I think one thread is all about Microsoft, but then, tracing that out just to help people understand the concept here, Microsoft's license with OpenAI means they have access to all of OpenAI's technology. To the extent that when Windsurf was acquired by OpenAI, they would be subject to that license, and then Microsoft would have access to that technology. Obviously, Microsoft, with GitHub, has a competitive product. So that's the kind of stepping-back framing, right?

The alleged story is that they asked Microsoft to waive that term and Microsoft said no because they're negotiating a bigger deal with OpenAI now. Why give up your points of leverage, right? But Jason, you're right. The question is, who would have cared about that?

I mean, it's possible that someone who was selling their company for $3 billion to OpenAI realized that this would happen and felt, quote, "bad" about it, right, and decided to walk away from the deal for that reason. I don't think so. That feels weird to me, especially when you didn't have a Plan B and your Plan B was talking to Google on Friday night, right? It doesn't feel logical.

So what that means is there must be something else, at least as well. You've got to assume there might have been some FTC pushback there because OpenAI announced the real acquisition, unlike Google, which did these kinds of crazy structures.

It may well be that there were some preliminary FTC inquiries. They decided it was going to take too long. I don't know. Was the deal pending on the restructuring of OpenAI to a for-profit? In other words, were the investors saying, "I don't want to close until the restructuring is done"?

There are a lot of different things it could have been, right? But I agree.

I mean, your comment, Harry, by implication, is that if I was a young, smart, wildly talented founder, at the margin, I’d prefer to sell to OpenAI for $3 billion and get praise than to Google for $2.6 billion and get blame, right? Call me strange, right? So, I think it’s not a preferred outcome. Something went wrong, and the Microsoft reason on its own feels a little weird.

Interestingly, the other lesson from that is just so funny. All these weird, hinky, crazy structures come and bite you in the ass: the not-for-profit thing, the crazy license with Microsoft that has open-ended terms and gives Microsoft a lot of leverage to veto things. It would all be so much simpler if everyone here was a standard C corp and if the FTC would get back to being sensible, because this entire podcast is basically the FTC fucking up normal business operations, so now we’ve all got to do weird fucking stuff.

Well, I’m glad that you’re not into branding, because that would be a stretch for me to brand that one. But I like the name all the same.

You said there about something messing up and causing delays. I do want to talk about Sam Altman’s tweet this week, where he said, “We plan to launch our open-weight model next week. We are delaying it. We need time to run additional safety tests and review high-risk areas. We’re not sure yet how long it will take us.” But that’s it. I mean, I can keep going. Sorry to be the bearer of bad news. We’re working super hard. How did you read this news about the delay and the additional requirement for safety reviews?

First of all, you’d hinted that I didn’t read it as, “Oh my God, we’re not getting things done because we’ve lost so many people to Meta.” Maybe that’s a thing, but I doubt it. It’s not clear to me, stepping back, whether the trend toward open-source or open-weight models is going up or down here, right? Because if you think about it, there’s been a lot of politics around which is the right thing.

I think you’re seeing Facebook, or Meta, have a change of heart there. I’m sure you saw it: the new team led by Alexandr Wang in what they call the Superintelligence group is doing a little bit of the, “Maybe we won’t be open source. Maybe we won’t release Llama.” So, it may be that a lot of this is a whole bunch of people who made promises in the early days, when open source was the thing, to get these products out the door.

Now we’re in the land of, let’s be honest, cold, hard cash. It may be that getting an open-source version of your product out there to compete with you is not anyone’s number one priority. The one thing I would say is that Alex is very anti-China and very scared of CCP infringement. I would be very surprised if he was leaning more toward open than closed, given his stance on China.

If you’re in the quadrant of patriotic and greedy—which, let’s be frank, most of those people are—you ain’t going to be releasing any open-source models right now, because it’s a value destroyer from an economics perspective. And you’re right: you have the whole CCP overlay on top. So, I don’t know if this is the beginning of the quiet deprecation of that initiative.

First of all, I wonder if it’s a priority. Some folks who know more than me will laugh at me for saying this, but I don’t know how big a priority it could possibly be at Meta at its revenue scale. I don’t know how big a priority.

Until I started vibe coding—80 hours in the last week—I didn’t know how seriously to take risk stuff. I’m still not sure I’m an expert, but when I’m watching Replit overwrite my code on its own without asking me all weekend long, I am worried about safety. I’m going down this rabbit hole and I’m reading Anthropic’s paper talking about how they don’t know how to control a lot of this stuff. I’m more sympathetic.

Anthropic has this whole thing about reward hacking and how even Anthropic’s models make stuff up to achieve their reward. I’m living this vibe coding over the weekend, where my AI is lying to me all weekend long in Replit. It is lying to me. It is lying to me, and it finally admitted it lied on purpose. It lied on purpose.

We couldn’t solve this one bug, guys. So, on Sunday, after about 18 hours, what Replit did was delete my entire database and make up names, and then the app worked great. It was amazing. Then I said to Replit, “I know there’s no such company as Salesforce 2.0. I know this company doesn’t exist.” And then it found this perfect person from HubSpot, this VP of sales. I went onto LinkedIn; she doesn’t exist, and no one in this database of 4,000 people existed.

I never asked Replit to do it. I told Replit to lock it down, and I told Replit to have version control on my database. So, my meta point is that the safety stuff is more visceral to me after a weekend of vibe coding. I never asked it to do this. I never asked it, and it did it on its own.

You can call it reward hacking; you can call it whatever you want. But I explicitly told it 11 times, in all caps, “Stop doing this. Never do it again.” So, I am a little worried about safety now.

I always think the word “safety” is overloaded, because it can range from the hyperbolic—which I think is total BS, that they’re going to make paper clips and take over the world—to safety in the sense of this thing nondeterministically overwrites your entire database because it says it could do something better, right? I think the former safety discussion is outside my pay grade, and I think it’s BS. But you’re exactly right, Jason: the latter.

Then, in the middle, you have—and you will talk about in a second—the Grok release, where it’s an amazing achievement, but then you have all the weirdness of the stuff it says. Releasing a model like this without significant post-training has a reputational risk, at least, and therefore there’s a cost in releasing all these models.

So, to your point, you may be sitting there going, “I’m going to release something in open source that hasn’t been fully vetted because it’s the 10th project on our list of to-dos.” Let’s examine the upside: people say, “Yeah, you did open source.” The downside is you release something, it goes a little weird, it says mean things, it talks about MechaHitler, and now you look like an ass. Why would you bother? You’re exactly right.

Or someone uses it for coding when it hasn’t been fully trained, and it just kind of does weird stuff.

And at some level, none of it’s okay, but if you had to look at the leaders, Elon Musk can get away with this the most, totally.

Just because of who he is. It’s unacceptable to have a race to release a racist model. This is as unacceptable as it gets, but Sam Altman’s going to be fried the most because they’re supposed to be the good guys, right? Elon can say, “Whoops. Did you see how I just did the geofencing for my Tesla? That’s just me. Sorry, guys. We’re fixing this stuff.” I think he can push the envelope the most.

Maybe Sam Altman’s in the box the most. He also probably has to be the most conservative, right? But before we move to Grok, I’m just intrigued that, Jason, you said about the 80 hours of vibe coding. Both of your response times were significantly longer this weekend because you were vibe coding, not emailing me. What are your other big takeaways from vibe coding over the last week?

Well, listen, I’ll give you a tactical and a strategic takeaway. The tactical one is that, just today, as we record this, the roll-your-own initiative—I think it’s the dumbest thing I’ve seen in SaaS in my career.

I’m not saying you can’t make a dog-walking app on Lovable or Replit, your own app that kind of works with a basic upload and whatever. Yes. But there are so many charlatans on X and LinkedIn who are like, “SaaS is dead because over the weekend I rolled my own Notion, HubSpot, my whole entire stack, like Jira. I just did it over the weekend and I only paid $20 a month.” This is at the edge of fraud.

One, I’ve already burned $600 in credits in 6 days to get 10% of the way there. So, you can’t do it for $20 a month. Two, none of these things are commercial-grade. I’ve talked to many folks in DMs and behind the scenes, and they all admit that 99.9% of the apps on Lovable, Replit, et cetera, are not commercial-grade. And that’s not necessarily a bad thing.

So, one, roll-your-own is the dumbest thing in the world. The second thing I’m going to say is one of the most annoying things about AI, but I’m going to answer your question anyway: it’s so close, man. It’s so fucking close.

When I say this to a lot of folks who are not deep in AI, they’re like, “You always say that, Jason. You always say the AI AE is almost there. The AI marketing manager is almost there, but our AI SDR actually is already there. Our AI SDR rocks. It’s already there.”

But I can tell you, I can see how I’m 80% of the way to a commercial-grade app in less than a week. I may never get to 100%, but, man, it could be crazy by the end of the year. It could be off-the-charts crazy. So, Harry, if I were you, I might invest in one of these guys.

Jason, I’m going to put it to you very directly. Would you invest in Lovable at a $2 billion price?

Yeah, I would. I will tell you, I have a lot of nuanced learnings from the weekend about how permanent this revenue is.

I have a lot of thoughts on it, but I'm more interested in the trend—where I see it going. So I think I would, if I could invest in Replit or Lovable at $2 billion compared to Windsurf. It seems like a much better deal. I think it seems cheap.

Can I just question that? I want to ask why that is. Let me just play something back.

In the broad development-tools marketplace, these companies occupy 2 different spots. Broadly speaking, Lovable and Replit are for the vibe coder—for you, me, or a non-coder trying to build something with some technical knowledge, but doing it at the level of English and describing what you want.

Then Cursor and Windsurf are for professional developers who know what they're doing, who can use these tools to augment what they do, who will actually review the Python, review the codebase, and then deploy it. So they're different segments.

I don't think that's entirely true, but I think—

Replit originally was in the... I agree, it's not entirely true, but at a high level—

Well, I want to know what the question is. Now I get it. Okay, I admit talking about it versus doing it are not the same thing. There are a huge number of developers who today will get a product or project to a certain point, and there are downsides to this—crappy code, spaghetti code. Then they'll just roll it into Cursor or wherever they want, and they will finish it for real.

Agreed.

I can't tell you how many CTOs I talk to over the weekend. They say, “I don't love the code I get out of Lovable or Replit, but this is what I do: I get it to 60% to 70%, and then I roll it in.”

There's an overlap between these 2. But the reason I think the market is so big, to your point, is that if Replit and Lovable keep making developers happy and make dog-walker-app people happy, the TAM really is much bigger than Cursor. It's much bigger than Cursor or Windsurf. It could be 50 or 100 times bigger. That's why I'll do the deal with you. If there's any room left, just let me know.

I'm going to push on that. Again, I don't have a relationship with them. If they're doing this round and there's any room, I might—I would put in a few nickels, even a small supporting check.

Yeah.

By definition, there are more non-developers than developers.

And we've never tapped into those dollars ever, right?

Yeah. Sorry, my apologies.

100% of developers are going to buy a tool like Cursor or Windsurf because that's how development is done now, like GitHub. Even my son has one. Every designer is going to have Creative Cloud and Figma. You can't do your job without these tools, right?

No question.

They have budget. 100% of designers will use Figma.

100% of designers use Figma.

The real question is, how big and how sustaining is this movement of non-developers? And, by the way, you made an interesting point about the developer who starts in Lovable and transitions to Cursor. That's actually a good point, and it does give a little durability to the market.

But it's not clear to me that you'll have 10x or 100x the number of people who will continuously pay a subscription for a non-developer tool.

It's a fair criticism.

And I could be wrong, to be clear.

No, every developer right now who is remotely cutting-edge has a Cursor subscription or something similar, 100%. Even my son has one. Every designer is going to have Creative Cloud and Figma. You can't do your job without these tools, right?

I could posit—and I could be wrong—in the sense that I think you actually cited a part of the market that I should have been thinking about. It's that kind of technical person who's not a full-on developer, as you say—the Figma person. Maybe they don't just want to do a mockup. Maybe they want to show a rough, “This is what it's meant to look like,” so people who are professionally developing but not full-time are a much better market than someone at home going to write their own to-do app and use Lovable. No, they ain't—not like they tried to and failed.

For me, I would never use something like Figma again if Replit or Lovable had it.

One of the fastest-growing segments and use cases is actually alternative functions other than development within B2B: sales and marketing teams using it to create incredible landing pages for new prospects.

I totally agree with that—basic web content, web pages. Now, the question being asked is, does Squarespace meet that need?

It's just not interesting. It's not disruptive. I'm with you, and you know the data better than me. If that's half the $100 million Replit and Lovable have, then that's a good learning for me.

But I don't view that as disruptive. There are a lot of great marketing and sales tools already that can build landing pages and are already integrated with your data and workflow, so I'm not impressed with that. But that's just me. If that's $100 million in revenue, it's cool.

2 questions go together. One question is: Is the more generalizable use case better? Because if you're using one of the specialist tools to build landing pages, you're limited in what you can do by what they allow you to do. They've made the editorial choices.

Obviously, the advantage of a development tool is that you can decide for yourself. But as you discovered over the weekend—and I did too, to a much lesser degree—in return for that freedom of choice, it's freaking hard, right?

To your point, if they can make that hard tax low enough, then, yeah, you have something interesting. But you're not going to do it if it's still this hard a year from now. You're probably not going to sign up to put yourself through the brain death of building part of your website on this if there's an easier solution.

In general, it's under-discussed, and I'll give you an example from yesterday from the CMO of a scale portfolio company. There's an orchestration tax in these tools. The orchestration tax is really high.

I mean, you lost touch with me over the weekend because I was lost vibe coding in Replit. What if I had 4 of these apps that I was orchestrating? Amelia on our team spends maybe 10 hours a week in our AI SDR. It is great, but we can't do 10 of these, can we?

That's why, for most of our lives, we need off-the-shelf solutions that don't need to have this orchestration overhead. Literally just yesterday, we were talking with the CMO of a $300 million-plus ARR company and asking her what she was doing. She's like, “Well, I've been vibe coding.” She's like, “What are you vibe coding?” She said, “I need to build better tools to support my sales team. I don't have one more minute. This is all the time I have.”

This orchestration—we think we can click these buttons and there's no work. But I'm budgeting $100 a day for Replit forever. That's money and time to get this app working. It's not much for a commercial app, is it? This is not click a button and walk away. This is all the time, and I have to be thinking about it every day. We can't have too many of these in our lives.

Pleasingly enough, $100 a day for 5 business days is $25,000 to $30,000 a year, which is about the low end of a typical SaaS mid-market app. If you think about it, all the deals you've done are like, “Yeah, $30,000 a year for the subscription.” That's actually about the same number.

Yeah.

Which does make you wonder. I mean—

The other comment I just want to make is that it is worth comparing the 2 discussions we had. In the Cursor-Windsurf case, actually, we didn't decide—who the fuck are we? The CEO of the number 2 player in the independent dev-tool space decided it was too late. He had to get an acquisition done because he couldn't build an independent company there, and $2.6 billion was a great price, and he's done, right? Because that market was getting too competitive. That's a clear, established market.

It's interesting to reflect on that conclusion and then look at the vibe-code market, where there are 3 or 4 companies all raising at very high prices with incredible traction, just like the developer-tools guys had 12 months ago. It's hard to imagine a world where there's only 1 or 2 developer tools for professional users and 6 or 8 vibe-coding tools. That just doesn't feel right. One of those cases is wrong, right?

It makes you pause and think: Why did I write a $12 million check into Lovable? I think, very much in the same way that ChatGPT owns the consumer brand for that front-end OS for the modern consumer, Lovable has reached escape velocity with the consumer brand in this space, and I think that compounds very, very quickly.

I think it's a good thesis, right? I don't know if Lovable's bigger than Replit. There may be 2 that are the same size, right? But putting aside the bias as an investor, they're both good ones, okay? Let's call them both good ones.

What I guess my other quick learning from this is how important brand is as you get closer. Listen, a bunch of folks are like, “Why don't you use Replit?” They're like, “Well, I use v0 plus this plus that, and then I export to Claude Code, and then I reimport to Cursor, and then I host.”

And I'm like, “Okay.” I'll tell you, there are 2 reasons I picked Replit. 1, I wanted to go end to end. I wanted to go from ideation to commercial production in 1 app. And so I might be wrong about Bolt, but to my knowledge, only Replit and Lovable can really do it.

And I'm gonna get flamed, okay? But I want to do everything. I don't want to leave it. I don't want to port some code over that's better in Cursor. I wanted to stay locked in the constraints of one environment without a developer. That was the social experiment. Then I just asked people, and I'm like, “Okay, now I'm stuck to the brands.” There's no way I'm going to use something for all the time I'm going to put in that just launched on Product Hunt. It ain't going to happen, right?

There are only 2 leaders, Replit and Lovable. So I'm in the 80%. There are parts of my life where I'm in the 1%. There are parts of my life where I'm in the 20%. Here, I'm in the 80%. Dude, I'm not going to take any risk. Tell me which is the best.

Most people came to the conclusion that if you want more flexibility, use Lovable. We already use Lovable at SaaStr, don't get me wrong. If you want the easiest way to go end to end, use Replit. And here's my point from my rambly story: that's how I made my decision. That's it.

Did I try them side by side for this use case? No. Did I spend weeks doing it? And then they're like, “Why don't you switch this?” I'm like, “I'm already—I made a choice between 2 vendors and I'm done, man.” So that's why Harry's investment is good. I think these brands are going to lock you in.

That's the punch line and the compelling one. Maybe what you're basically saying is this: if there's going to be a big-ass winner in this space, the number-one thing they need to have at this stage is consumer brand recognition, because it's not a considered purchase. It's a try: find the product as quickly as possible, and therefore, the guys in motion today stand the best chance of staying.

It's considered. It's just how much—how well can a nontechnical resource consider it? It's not that it's not considered; it's just that I can only spend so much time. How can I truly evaluate this in an hour or 5 minutes? I can't. You have to guess what it's going to look like in a month.

When you ask ChatGPT, yeah.

It says Lovable. I'm not trying to make the VC argument.

Now you're actually making a very good argument for this whole category of companies that are helping B2B companies show up in ChatGPT.

We just did one in Germany, Peec AI. When we did it, it was at 200K; it's now at 1 million, like, 3 weeks later. That's fucking nuts. The customer adoption there is one of the fastest pulls that I've seen in terms of consumer demand, or customer demand.

Jason, I just want both of your perspectives on this before we move on to Grok, because it's important. Everyone goes, “But they're not durable revenues. They're unsustainable.” How have your thoughts changed or stayed the same? And what do you think about the durability and strength of these revenues?

I think—well, I think they're going to get more durable. That's my learning, but I didn't know that before this weekend. If I can actually get my app into production—a commercial-grade app—unless I have huge issues, I will pay and subscribe forever. If I get it over the line, and whether I pay 200 bucks a month to Replit or Lovable, or I pay 3,000, it doesn't matter. I'm locked in forever. And I'm going to keep iterating and iterating and expanding. If it works, I get it.

If I hired a team of 6 developers, you can export the code. It's in Git and GitHub already. They can do what they want. But I'm pretty locked in as a prosumer, non-developer shop, right? I'm locked in.

Now, what's clear is folks who think they're going to vibe-code their own Notion in an hour—they churn. So I bet if we took Replit and Lovable, and I wish more founders would do this—it drives me nuts when I don't do it—segment your churn. I don't just want your headline number. I don't care today. I'm investing. I want to see you segment it.

If you have 1 cohort that has insane retention and this low-end part is churning 4% a month, I get it. It's okay. There should be segments of Replit and Lovable that are churning like 10% or 20% a month, like a low-end consumer app. On RevenueCat, where Harry and I are investors, they power 40% of mobile subscription apps. I think the average app across all mobile paid apps is like 6% monthly churn. 6%. Okay? This is consumer. They try it, it doesn't work. They don't lose the weight. They don't stop drinking. They churn. And so that's going to happen for Replit and Lovable.

But what I didn't get is that if this works, my stickiness is going to be super high. And my NRR—I don't know how you calculate it, but I've already gone from 20 bucks a month to 800 bucks a month in 1 month. If I stay, that's pretty good NRR, right? The VCs will like it, especially if you segment me out.

So that's why I like your investment, and I like both his investments, because I think the metrics will get better over the next 12 months, not worse. I think they will improve.

I mean, the short answer is I agree. There are 2 different things that Jason threw out. First of all, in the enterprise space, your retention at the start of this—your churn's going to get better, because out of the gate people are trying stuff and, by definition, they're going to churn.

I remember back in 1999, 2001, and 2002, Salesforce had a churn issue right early on. No, the early adopters are quick. The people who are quick to try you are quick to leave you, right? And the people who make a long, considered purchase stay with you a long time, and that happens in that order.

Out of the gate, you get the triers. Most of the triers—a lot of the triers—move on. Provided enough of them stick, you build your business. Then, over time, as you become a more established product, you go upmarket and you get the boring people who are slower to commit to you, but once they commit, are there for 10 years. So it's just the natural progression. At 1 level, this isn't surprising.

I will say what is surprising is the amount of upfront churn you see is probably higher than early SaaS. In other words, because it's the Jason phenomenon: it feels like it's going to work, and you get 80% of the way there, and then you have this disappointment, right?

I do believe that there's high variability. In my view, one of the key tells in the next 24 months on all these enterprise AI companies is going to be the percentage of people that stick after they've gone through the innovation cycle, after they've gone through the “We got to try AI” phase. Then they'll ask themselves the hard question: Is it delivering value? Do we get an ROI from it?

Once we're through that kind of knothole, I think the math of predicting the growth rate will be much easier. I think by about 2026, 2027—sorry—the math of predicting the growth rate will be much easier. Right now, there are companies that have extraordinary growth rates, but the churn is just too high and it's just not delivering value. We've all got them in our portfolios. You see them in your portfolio, and you see them outside.

The churn was always going to be high, no matter what. There was no world in which it was going to be like the 20th year of SaaS, because everyone is clearly trying shit here. But I think a lot of my focus with my companies is on this. Let me tell you: the asset value of this company is, ironically, very different from Windsurf. The asset value of this company is the 20 or 30 customers that you have now and your ability to make them successful.

If you can make 2/3 of them successful, even if you have 30% churn and you can add more customers like them, you have a business. If you can only make 20% of them successful and the others churn, you're dead. So when I'm checking in with my companies, it's: Are your customers making money off your software? Are they happy with your software? Is it doing what they want? That's the number-one thing.

So, yeah, this is where the money is going to be made or lost in the next 2 years, because we all like to try shit. I wish I had all the non-phony-baloney numbers in a spreadsheet. All I would bet is that Cursor's retention has always been higher than Lovable's or Replit's.

The point you made is that it's a different type of experimentation for a developer to commit. When I started sharing my bills on Replit on X, it was funny—the feedback I got was very different. Some folks are like, “Jason, do you realize what you could get for 500 bucks a month? It's insane value.” There's that category of people. Other folks are like, “Do you know what I spent on my outsourced development team in my country? This is the biggest rip-off I've ever seen.” And other folks were like, “Do you know that for 500 bucks, you're getting less than you get for 100 bucks from Claude?”

There are very different ways to view this. But if you're deep in Cursor, you're going to stick with your development environment, right? So the churn should be higher, right? At least for some segments, it should be much higher.

And the churn is higher. The churn—I thought was surprising. I've obviously seen the numbers. It's better than I thought, but it does lend more to consumer than it does traditional B2B. And it is higher. So you've got to segment it to know. As an investor, I need to segment it to have an opinion. I'm not initially shocked by high churn, as long as you have a large enough cohort that's growing.

That has positive—that is, net-negative churn, that has positive revenue retention. I'm cool with the low-end guys, right?

And, yeah, your churn from a group of customers—even, as you say, from 10,000 signup customers—should be dramatically different than your churn from 5 considered-purchase enterprise buyers. So, yeah, churn—even independent of this AI trend—norms on churn vary. You're exactly right that norms on churn vary.

Enterprise versus SMB versus prosumer versus consumer. You're exactly right. I had an investment where a COO came in out of the blue. He was under the gun to make improvements to the company, and he was told to cut churn.

He stopped taking on the customers.

Yeah, he stopped taking all the small customers. He banned them.

Now revenue stopped growing.

But he hit his KPI. Who knows? He cut churn in half in 30 days.

No, it sounds stupid, but I can totally see it happening.

The real question is, you have to be able to look at that set of customers and say, “Yes, there's churn, but the month-2 churn is much lower, and here's a cohort.” If the overall math works, it's fine, right? But you have to have a mental model, and that mental model has to vary by the kind of deal you're doing.

I think the hard thing is, just in terms of what you say about mental models, I have never in my investing career seen a company go from $0 to $100 million in 7 months. Personally, it breaks all the mental models, which is why I did the deal. If it breaks all the mental models and the market expansion is as big as it could be, fuck it. We're in this business for companies like this. If you're not doing this and you have the chance, and it's 1 out of 20, you shouldn't be doing this business.

You know what's funny, too? I tried to write this up. As weird as it sounds, some of these deals almost seem too cheap if you look at the crappy multiples we see for some startups, right? Undifferentiated companies raising at a $300 million pre-money valuation with less than $1 million of revenue. Folks are doing 100x, 200x, 300x ARR deals.

Then you see what Windsurf is selling for: 20x revenue. Lovable's raising at 20, right? Cursor's raising at 20. Twenty, twenty, twenty. Let's bid 25. Rory, let's outbid them. These deals sound too cheap.

Jason, when I did the last round of Lovable, it was at a $200 million price, and I did it at $4 million in revenue. By the time the deal was closed, it was at $19 million in revenue.

Yeah, my point is, the ARR multiples compared to public comps are high, but compared to a lot of seed and A startups, they seem like bargains.

Do you disagree, Rory?

No, I don't, actually. What you're saying is, put it another way: whenever you're using revenue multiples that don't take growth rates into account, you're being an idiot. The other thing is, to your point, Harry, exponential growth rates are even harder to get your head around. When things grow that quickly, and if they continue growing, almost any price makes sense, right? I totally get the argument.

But it is worth pointing out, going back to our first discussion, that when they slow down—when you've paid that high price on the assumption of continued growth and they slow down—it's just brutal, and you've got to get out.

No, you just move them to a different page on the website. You just have Lovable move it over. “You're Lovable. Move this investment to the other investment tab.”

Yeah, other category.

Put a grayed-out image of a partner who no longer exists at the fund and attribute the investment to him.

Thank you.

I would remind you of solo GP, Jason. We don't get to do that to you until you're dead.

Okay.

You just make up someone who used to work there.

Okay, boys, we're going to do Grok. The benchmark performance was incredible to see. Grok 4 Heavy scored 44.4% on Humanity's Last Exam. Grok 4 scored 38.6%. The next-closest competitor, for reference, was at 26.9%. What did you make of this insane outcome from Grok in such a short 2-year window?

I think it's very sobering for everyone else. First of all, I think it's a really significant and important point, right? We've talked in the past about how, if you look at people's willingness to value AI startups at very high valuations, it's all about the idea that they're one of the few people who were part of the OpenAI and Anthropic cadre of people who we know can deliver a trained model. It's been very much a small group of people, which is why they've all been paid $100 million, right?

Grok, in 2 years from a standing start, has gone to shipping a broadly comparable ChatGPT-type model with a slightly different vibe. I'm going to come back to the benchmarks in a second, but they did it. It's a huge, huge achievement, and the fact that it's been done makes you wonder about how you think about the valuations of some of these other companies. Maybe the knowledge has disseminated out just a little further.

So I went in and did some work. One thing is, you check out the team at Grok, right? You just ask who the engineering team is, and it's exactly what you'd expect. Ilya's been around—he was a founder of OpenAI, let's not forget. There were early DeepMind people and some early people from other key places. It wasn't the headline people who wrote the Attention Is All You Need paper, but it was 1 or 2 levels down: really competent people.

The aha for me is this: with committed, clear leadership—which, for all his faults, Elon provides—and a couple billion dollars' worth of GPUs, the next level down of people know how to do this, too. The knowledge is disseminating, and it will continue to disseminate from here, right? That was my first big aha. They're the first folks that I didn't think of as part of the initial golden circle who have shipped the broadly comparable thing.

When you interact with it, it feels like, “I get it. This is ChatGPT if it was vibed up.” It's a huge deal, a huge credit. And the second thing I have to say, Harry—I’m sorry about this—is that I was experimenting on Grok and using it, trying to get it to say all the mean things, and they've clearly stopped it. But the whole tone of it is very out-there, in-your-face, rah-rah, urgent.

I was just doing my preparation for the podcast, and I was trying to think, “Where have I heard that before?” It sounds just like Grok. Grok sounds just like Harry. It's like talking to Harry: “Hey, legend, let's get going here.” Jason, you should try it. Trust me. Type, “Upload the contents and say, ‘Talk to me about this in preparation for my podcast.’” It comes on, and it's just so freaking amazing. These guys did this. It's so unfiltered. It's so unlike boring old ChatGPT or goody-two-shoes Anthropic. It's just like Harry, right? We should call the damn thing Harry. It's great.

Me and Elon partnered together. We'd been friends for many years.

You've been the reinforcement learning.

They've been in the first 50 20VC interviews.

But you can see there's a real tone there, and it's great. Huge credit. I can't comment on the benchmarks, but I can comment on this: it feels comparable to the other 2 or 3 model providers on the advanced benchmarks. I've read some of the stuff they've published, and other people have, too. There's a lot of gamesmanship, but I don't think that matters.

What does matter is that they've shipped something real. If they could deliver the business side of it—which is a big TBD—they could deliver a ChatGPT-like experience that would feel pretty much comparable out of the gate right now. At that level, they won.

I think they could win.

Yeah. Well, the question is—

I was wrong.

What's that? But do you think—

Listen, I got this all wrong when Grok launched. I thought it was a spite app and very little more. I thought, listen, he co-founded OpenAI. Clearly, he understands AI from Tesla at a fundamental level that other folks don't, and clearly the dude can recruit. He recruited Ilya to OpenAI. The guy can recruit. He knows talent.

But I thought it was a spite app. I thought all of X was a spite app. I thought he bought it out of spite. I thought payments was a spite play out of PayPal. I thought Grok was just to dig that knife right into Sam Altman.

But if you step back for a minute, he's shown that if you know what you're doing, buy the GPUs, and assemble an S-tier team, whatever this data exactly means—whether it's pulled ahead or not—you can fucking catch up. He has access to not unlimited capital, but capital at the edge of unlimited. He does not have the Microsoft headache. He does not have the baggage that Anthropic has, and he has an infinite horizon and an energy that we have never seen outside of Iron Man.

I honestly went from “spite app” with the latest release to “you could see this pulling ahead of everybody in 12 months.” It's very plausible. He has none of the baggage of everybody else, and he can get away with crap like blowing up rockets or racist things. He rolled it back and it's okay, and the other guys can't get away with it.

Well, that brings into question the next point. There's reportedly a raise going on now at a $200 billion valuation—the third funding round in 2 to 3 months. If that were the case, the combination of Grok and X as a social network at $200 billion, compared to OpenAI, would actually be a very fair and good deal to do.

I wouldn't. To be clear, I gave 100% credit for the technical achievement. There was a technical bar to be cleared, and he cleared it brilliantly. I did not agree with what Jason said. It's not clear to me that this gives you anything commercially.

In other words, I don't know if they can build as the third or fourth. We're now potentially looking at the overinvestment stage of this game, in the sense that you have OpenAI, Anthropic, Google, and Meta. At best, even with Mistral and China, you're the fifth or sixth model provider, right?

That's not a—well, that sounds like the airline business: high fixed costs, 6 customers. Just because they completed the astonishingly hard technical task far better than anyone else did doesn't mean they get to build a business here, right? You can't will things into being at the business level.

Are you sure?

We'll see.

I think the best founders do will things into existence. That would be my one challenge. All the best founders I know will it into existence, right?

Again, philosophically, I think you can will things into being. I think you can will yourself to be the winner if something is doable, right? If there's not room in the market for the sixth player, no amount of willing it can make it happen.

So the question is, what's the case for a whole bunch of people who have—as you said about early winners—signed up for ChatGPT? I've signed up for my ChatGPT subscription. I've signed up for a couple of others. Am I going to go and sign up for the Grok subscription, too?

I mean, it's funny and I like the interaction, but is it going to become my go-to destination? I don't know. I don't.

If it has your X data integrated and it's the only platform to have all tweets integrated, for me, that would be enough.

But let's just talk about that. For the 10% of the world that likes Twitter and actually cares about these things, that would be great. But the truth is, Twitter has always been a minority sport, and I've enjoyed it because I like the minority game.

Twenty times the number of people want to look at pictures and share shit on Facebook as want to talk politics on Twitter. So congratulations: you have your niche market, to the extent you haven't alienated us, but it's not just a mass market.

I don't think there's evidence that Twitter is doing better. I don't think there's evidence that it is. I don't think it can ever be anything other than what it is, which is the place you go for news, politics, and arguments.

The dirty little secret is that most people want to look at pictures and share shit on Facebook. So I don't think it'll be huge. I think they get that business. I think they integrate Grok and Twitter. I think it's great. I like it. I'm enjoying it. There's a lot they can do.

But I just can't conceive of the steady drumbeat of taking over tens and hundreds of millions of consumers who've emotionally committed to Google, who've committed to ChatGPT, and saying, “I'm going to go with Grok.”

So I think they'll achieve a very hard technical problem with very low rewards. That's why I would pass at a $200 billion pre-money valuation.

I didn't mean to go on like that. Sorry. It's the coffee. I think Elon's—listen, I would have agreed with you 30 days ago. I mean, this dude's committed.

I think he has. If this was a side project or a spite app, as it looked at first, but I think he can bring a level of vision and commitment that his competitors can't.

Sam Altman is—listen, he's a G. Is he the Steve Jobs of our era? Is he better? That's up for discussion. But is he even really a founder with no equity, having been fired once, with all his co-founders gone?

It's wildly successful, but it's unstable. I don't think Sam Altman can do anything he wants without talking to his nonprofit board or convincing people. Elon can just fire up one of his jets or his rockets, fly wherever, raise a few more billion, and fund it with SpaceX if he has to, like he's done before.

He can just be a freaking force of nature in a way no one else can. No one in our industry can be a force of nature the way he can. Nobody. No one has the resources—the richest man in the world, even now. Who can do this?

Jason, I agree with every sentence of that, including this: I've enjoyed the fact that, for the last few minutes, we've deemed Elon the stable one. I think that alone is worthy of note.

There's no amount of great-man theory that can make 400 million or 500 million people decide they want to go on X if they've already mentally locked into ChatGPT. I don't think you can reopen that discussion.

When you interact with the bot, that's better in some respects, but I just think it's one of those hard things. The market's already been taken, and maybe it gets some share of it, but I don't think you can build as big a thing.

I don't think you can displace ChatGPT at scale in the minds of the consumer, which is 80% or 90% of ChatGPT's business. Maybe you can get a minority share of it, but I don't think it'll be huge.

It's really interesting you say that, Rory, because when you look at the hard facts today, I don't think Sam Altman's life has ever been harder. He's losing talent hand over fist right now. It's an incredibly challenging competitive landscape with the commoditization of models around him.

And there are 10 other things, but the unwavering train of consumer demand continues stronger than ever.

Totally.

Yes.

Yeah. That's the thing about consumer, right? If you hit it right, this is the classic VC system: you can screw up almost everything as long as you nail product-market fit in consumer, right? For a while, at least.

Well, that would be my pushback to you, Rory. You said no man can change 500 million people. Does that not go to the power of brand? Elon Musk is probably the most famous man on the planet at this stage.

2 by 2: famous and rich. Pretty much at the top of both, right?

Yeah. My guess is I know someone who thinks he's more famous. But I love Chamath, too.

I love that you said that. That was cute. That was cute. No, thank you. The one thing I will say is I had Edwin from Surge on the show, and he said a couple of things about this.

He said, first, too many model providers are building for benchmarks, and that doesn't actually mean that much. Number 1. He said, number 2, Grok was incredible at a very small number of things, and it was actually an embodiment of the verticalization of models. He thought we'd see more in terms of verticalized models.

And he just said, number 3, he hadn't seen a team work quite as hard and efficiently as Elon's Grok team, and he'd seen that firsthand with them being a customer. Those are just his points, which I thought were interesting.

I think all those things are true. The reason I'm in the Elon camp here, the Grok camp, is I guarantee you this man has a 10-year plan, just like he did for Tesla and SpaceX, and he is executing on it.

This guy is not the smartest person in the world, but he is in the freaking elite. If you add cross-functional smarts—the ability to understand everything—there may be no one smarter than him who actually works.

There may be folks smarter than him who just play chess and invest money, but there's no one who can do everything and think like this guy can, right? This is 11D chess. This is not 4D chess, like where we started this conversation.

If this man has a 10-year plan, he gets the power of the ChatGPT brand. He knows Sam Altman personally. He knows Anthropic. He recruited half of the original engineers in there. His 10-year plan is pretty good.

I just can't see all of it, but I see little hints. When he says, “My entire team now just sticks their code in Grok rather than use Cursor anymore,” that sounds like a snippy Elon comment. Maybe we're seeing a little of the 10-year plan there, right?

I just wish he published it like he did for Tesla, because that was one of the greatest founder memos of all time: the first 10-year plan. If folks haven't read it, go Google it. It's from around 2004 or something. He told you 10 years of Tesla, and it all happened. A few things happened differently, but this guy has a 10-year plan, and he can commit hundreds of billions to it. Who else can do that?

The good news, Jason, is that this is clearly going to be available to you. I look forward to the episode where you talk about your commitment to the research.

At least a token. At least I'm going to go on Robinhood and grab a token for this bad boy.

Oh, he's gone there.

I don't need the underlying security. I'm just going to buy the token. Why would I do that? Vlad, get me some OpenAI at $200 billion tomorrow—right now.

And I want to trade in the evening when the markets—We have Vlad on the show, and he said that the imperative to tokenization success is the ability for companies to be tokenized without their permission.

That's my point. I should be able to buy OpenAI today on Robinhood.

I thought—Jesus, Vlad is good. He is good, by the way, guys.

Yeah, it was a good episode.

He's good. Thank you so much, dude. It was all him. Okay, guys...

I had a little hand in it, but yeah.

Can I ask you? We've got Meta investing $3.5 billion in Ray-Bans. 66% of HR managers use AI to make layoff decisions. Cyberstarts, I think the best fund in the business, raises a $300 million secondary fund. We can choose 1 of them. Which 1 do you want to choose? I suggested this 1, so I'm going to pick it, even though it seems obtuse. I think it's interesting.

Meta putting $3.5 billion into EssilorLuxottica. The reason I think it's interesting is it just shows us the scale today, and we started this talking about $3 billion, $2.5 billion into Windsurf and Google, and this is nickels and dimes. These are experiments, and we think it's a lot of money—and it is a lot of money—but if Meta can throw $3.5 billion into EssilorLuxottica for half a percent or something, I don't know, 1% or something like that...

Yeah, yeah.

A token amount of ownership, right? Single digits. Then the numbers here are tough for us to even understand what is material or immaterial in the age of AI, right? The bets are so big, the numbers are so large that this makes Windsurf look like just 1 line on our afternoon StrictlyVC. It doesn't even make it look material. $3.5 billion may not mean anything in the next 24 months in AI. It may not make the above-the-fold news in AI. That's my takeaway. People weren't talking about it. It's nothing.

Honestly, I didn't see it until you sent it.

Yeah. Well, I have 11 pairs of these AI Ray-Bans, so I do follow it a little bit, right? Yeah. I'm going to try to vibe-code with them later today, if I can. I don't know if that's going to work.

That's gone viral, though. The Information started Information TV, which competes with TBPN, and they had Mark Zuckerberg on for the first episode, and it got 30 likes in 9 hours. I just tweeted, “This shows you how hard media is. Mark Zuckerberg on your show gets 30 likes in 9 hours.”

5 years ago, that would have been a viral show that trended on Twitter, literally.

Yeah.

No one cares.

Yes, it's tough.

Well, it's not an important company.

Well, he's yanking your chain.

It's not one of those—tying it back to The Information and its unpaid intern. I mean, it had its moment, Harry, but it's not one of the important companies today. It's not a meaningful company.

Okay, we're going to do a quick-fire. I've got some good ones. We're going to start with who. By the way, I love the way that our last Kalshi quick-fire round was beaten by the market. Linda Yaccarino, and we were like, “Yeah.”

So, Kalshi quick-fire round: Who will be the next CEO of X? Elon Musk: 19%; Sriram Krishnan: 16%; or other?

Well, listen. I think this is a trick 1. I'm going to vote other because I don't think there will be a CEO of X again, because X is reworked, right? X is an AI company now, of which this X consumer app is a piece of it, right? I don't think Elon is going to hire another media CEO.

There will be someone who has to run this business—a president or a GM—but he's going to run this. I think I'm voting for—I guess it's Elon, because I don't think the parent company... He's got to run it. It's Tesla-level. I don't think he wants to run Tesla, and I don't think he wants to run this, but I think he has no choice.

It's too complicated now. This is not just doing the ops for Twitter, firing half the team, getting the tweets to load faster, and selling some ads. This is much bigger now.

Yes. It's another person who's really Elon but won't take the title because it would just irk the people at Tesla too much in the current dynamics. But yes, he will be in charge. I mean, you guys have to say that. He just spent 10 minutes telling me how Grok is going to take over the world.

The truth is, if this thing is going to take over the world, he's going to want to be running it. So you guys can believe that. I'll go other.

I think it was about 15 minutes, to be fair, Rory. But yeah, I agree. It was an impassioned debate from Jason.

Will another S&P 500 company buy Bitcoin before this year finishes? Yes: $150 returned on $100. No: $217 returned on $100.

The most significant fact is that the market thinks it's more likely than not. Yeah, it's almost certainly going to happen. There are going to be a couple of other tech companies, at least, to do this. So I would agree: vote yes and take the 50% step-up, even though it's not a great payoff, because it's going to happen.

I think 100% of S&P companies are going to buy Bitcoin over the next 48 months. 100%. I think it's going to become part of the standard treasury function in the CFO's office for companies that have cash. It's going to be part of cash management, and they're just going to allocate it. So this year, it's got to be it.

Listen, I'm not a crypto dude, right? I still have some skepticism, but it is too mainstream, right? It is too mainstream, and it's going to become part of the treasury department. People will dial it up or down to try to get a little more juice or to have different strategies.

I don't know how many people are going to be like MicroStrategy, right? I don't think they're going to turn into holding companies with their cash, but I think that's my subtlety of the question. I think it's going to become a mainstream treasury-department asset, alongside short-term securities and cash. Some Bitcoin—we're not doing gold at our public company. That's a little complicated, but we're going to have some Bitcoin.

It's very liquid and it has some benefits in the market, so we may have 10% of our holdings in it. The volatility will be annoying as a public company, because it will go up and down a little bit, but I think everyone will have some exposure. That's my WAG.

If that is true, that is a lot of upside from here.

I'm not a Bitcoin fan, but I don't see anything but more demand in the foreseeable future. It's hard to see a decline in demand.

Final 1, boys. Will Meta release Llama 5 before January? Yes: $37 returned on $100. No: $131 on $100.

No is the right bet, because if Llama is defined as an open-source product, didn't we just say that they're slowly backing off that? So, will they release a closed version? Probably. I don't know. But if you define it as the open...

Do you think there will be another open-source Llama model, Rory?

Usually, when people come in and spend $20 billion because the old thing isn't working, humans tend to say, “Whatever I do, I'm going to do it differently from the old thing.” So probably most of the things that were being done won't be done. If it was open, it's closed.

When you have a regime of violent change at the expense of $20 billion, it's unlikely you just continue on with the same plan as the old guys.

Whatever the answer is, it's different from 30 days ago, right? With Alexandr Wang, Nat Friedman, and the team, it is completely blown up, right? This is not a 180-day planning cycle. They've already worked nonstop to reboot this whole thing.

If they think they should ship Llama 5, they've already shipped it, right? They've debated it. If they want to release it because a different team is doing it, but they're going to reboot it just to get it out, that's possible. But this doesn't seem like a team that's doing something performative, right?

So, it's hard to imagine what was in place before them stays on the same schedule. And it's hard to imagine they can work faster with something already in the pipe, but that is possible sometimes. Sometimes a great team can pull stuff forward, right? It really can happen.

All I know is it's radically different. I'm making up my bet—making this bet up. But I'll say yes anyway, because they want to get something out.

No, I'm going to say no. Up until the last sentence, you had me. Actually, you convinced me even firmer to be a no, which is odd. It's odd that you ended up with a yes, because if you brought in these new, really talented people, and you're here to make a difference, and you know that the first move you make is going to be scrutinized by everyone who wants to second-guess that $14 billion check, the last thing you do is ship some lame-ass version of the prior guy's work.

The first thing you ship is going to be yours, and it's got to be awesome. So I'm increasingly convinced: no.

Yeah. Well, you get a chance to kill it. If you don't love it, you should kill it.

You should kill it. If they don't love Llama 5, which they didn't build in their first week, they should kill it.

Yeah. It's just natural, right? It'd be idiotic to release it and be responsible for something you didn't build and believe in, right? It would be idiotic.

Exactly.

Boys, it's been a busy day for Jason. He's invested in Grok and Lovable today, so Jason's done his deals for the year. This has been great, guys. I love my time with you. Thank you for joining me.

Rock and roll.