为什么今天的VC比2021年更糟

Jason

It looks like the easiest way to make money in 2025 is to take the very biggest companies and double down one more time.

Rory

My gut tells me we're overromanticizing verticals in the age of AI. We're going to hit the same TAM exhaustion, and it's going to be worse because expectations are so high.

If any kind of deceleration happens because of AI—any kind of saturation or slowdown—everyone's estimates on what's going to happen here are wrong. Again, a reminder: in a bull market, the most aggressive person will look the smartest just before the crash, because the more risk you've taken, the more money you've made.

Frankly, it feels tough today. As tough as it's ever been. Ready to go, team. It is so good to be back. It's slightly earlier for you.

1. Everett Randle joins Benchmark

Topic number 1: Everett Randall joins Benchmark. Benchmark doesn't add partners very often. It's big news that he's joining as their latest GP. He was with Kleiner and Founders Fund before. Rory, I always think of you with Benchmark because you quote a fantastic statement: “Reports of my death have been greatly exaggerated.” I always think of Benchmark with this—with the portfolio and with great people like Everett.

Rory

Totally. Yeah, I know we were having that overall, “Oh my God, the world is ending” moment 2 months ago, when a Benchmark partner opted to do his own thing. I remember saying exactly that: They're going to be totally fine. They have a great portfolio, they have a great tradition, and they did exactly what they always do.

I can just see it there: You make a list of top firms that have good young people, you draw up the list, and you get on the phones. You say, “Who have we overlapped with on a deal?” Then you go and hire someone talented from one of the adjacent golden firms, where the pitch is purely, “Equal partner. You should do it.” Mission accomplished. Tick, done, and on they go.

So, yeah, it's not that it's not a big deal. They're just fine, and Benchmark will be just fine. You know, Mimoon and Ilia are wildly talented people. There won't be any shortage of people if they need to fill that slot.

2. Ambition, career moves & the VC shuffle

It reminded me of you, Harry. Yes, just like you, Harry. Obviously, he's wildly talented, right? Obviously, he was recruited by everybody. I don't know Everett, but he may be as ambitious as Harry. No one wants to screw around. If you're ambitious today, you want to go fast, right? You want to go fast. Vista, Bond, Founders Fund, Kleiner, Benchmark—I mean, those are 5 good ones to get on your résumé in 8 years, aren't they?

You don't need to stop at anyone in the B tier. That is probably part of the message, which is that exactly: If you're young and ambitious, everything's moving at fast velocity. You're getting fast-velocity markups, you get to declare fast-velocity success, and you take that fast velocity and want to rise up the organization.

You can just keep on moving up in a time when there's a lot of change, and there hasn't been as much change as this in the longest time.

Rory

I also know Benchmark is extremely generous in terms of backdated carry and being brought into this carry pool for this fund with Fireworks and with Mccor and with Lora and with Manis and many others. That's a very attractive carry pool to be brought into.

Yeah, no, it's the Godfather offer I can't refuse moment, right? Good for all concerned. Capitalism is great.

I was laughing, thinking about it while preparing for this meeting and thinking, “Oh my God, it's a wonderful deal.” Then you have to remember that you're probably owning one-tenth or one-hundredth of what the best AI engineer is earning at Meta, just to put all of us, ourselves included, in our place.

It's funny: It used to be the best economic gig in tech, and now we have to remember that, no matter how wildly successful we VCs are, there are people vesting a billion dollars over 4 years at Meta because they wisely did computer science and AI at school 10 years ago. In these winner-take-all moments, the market for talent in every market—whether it's AI engineering or top-tier venture capital—just becomes very heated. Let's go with that. I was going to say “overall,” but that's a judgment. “Heated” is definitely true.

3. Are VCs still the best paid?

Well, let's play that out, Rory, because I could have that debate with you. I could still argue that venture investors will end up better paid through carry if you're at Thrive, General Catalyst, or Lightspeed.

Rory

I disagree. If reports of the billion-dollar-plus package are true, very few people are going to make $1 billion, vested over 4 years, in venture. Though I actually agree with you, Harry: I don't think you can top that. Over 20 or 30 years, venture is a great career, right? But you look like you want to disagree, so please feel free to disagree.

Well, I think if you are one of the top 1 to 3 carry participants in one of the large mega-platforms—whether it's Andreessen Horowitz, Thrive, or General Catalyst—I would argue that you will have more than that in distributions in the next few decades.

Rory

Yeah. You see, instantly, up until the last sentence, you were wrong, but you were making your case. You admitted the truth in the last sentence.

Look, all these amazing funds—and we hope they have amazing funds, too—over 10 years, you're going to get a ton of money from 2016 on, right? If you look at distributions in every one of these funds, congratulations: You own a ton of private stock that's worth a whole ton of money, right?

Whereas I'm going back to my comment that the compensation package for restricted stock at Meta is, congratulations, over 4 years you have fully liquid stock. In terms of liquid stock and cash payment, it appears to still be the best. It appears to be the best gig on the planet right now.

How long was it before you guys got your first carry check?

Rory

I think the interesting thing is that the first carry check was relatively quick, but there was a 10- or 12-year period after that where it was very much the tail end of the ’99–2000 boom. Then there was a 10-year period of nothing.

When markets go down 80%, as the Nasdaq did, and then they stay down, and IPOs are postponed, and you have a European waterfall, absolutely. Totally. Venture, as someone said to me years ago, is the get-rich-slow program, and there can be 10-year periods of nonpayment. Sorry, Jason, go there on that one.

4. Carry payouts & delayed returns

Jason

No, I mean, Brian Halligan this morning was quote-tweeting Benchmark One, coming up on 20 years, and how it was a great vintage and that great wines age, or whatever it was. I'm like, “But I'm not sure I want to wait 20 years for my wine. I'd like a few sips tonight.”

I do think it's complicated, but my 2017 fund should hit 5x on paper by the end of this year, right? But that's a lot of years already. It could be 18.

And listen, do you really want to sell your winners in today's world? Of course you don't, right? So, hopefully, I'm not in a walker by the time I get my distributions from it.

5. Revolut's $75B valuation explained

That increasing period of privatization, as you kind of mentioned there, ties in beautifully with Revolut's $3 billion fundraise at a $75 billion valuation, up from $45 billion in 2024. Massively oversubscribed. Everyone wanted this one, to be fair, in terms of large institutional platforms. Private markets win again, and publics are delayed. How did you read this?

Jason

The big-picture story is that it's another round where the public markets have ceded that business to the private market, right? This company could clearly go public. It could have gone public years ago. It's done $3 billion in revenues in the last year, making $1 billion and growing at 60%. It could go public any time it wants.

I'll tell you what it made me think about a little bit: It challenged one of my early tenets, which I've held for many years—that the best founders figure out their TAM, right? A small market's okay. They figure it out, and they add layers to the onion. That's absolutely true with the best founders, right? I'm sure we could all come up with the story.

But when I think about Revolut—and fintech's gone in and out of fashion since we all met, right? It's been hot and unhot, and then people don't like the margins and this and that—but the market's right.

So, this is my new heuristic: If I could, I would like to invest in startups that, at $100 million ARR, have 1% or less market share. 1% or less market share at $100 million. That's what I would like. Not even fake market share, but real market share.

If you look at the public markets in B2B, there's almost no one except Palantir that's having an easy time north of $1 billion. We could look at all of them, and there are even folks like Klaviyo that are crushing it and still trading at 6x north of a billion.

I want to believe that founders will figure this TAM thing out, but now that we're staying private longer and a billion dollars doesn't even count as an exit, right? On Monday afternoon, it just counts as a few million bucks to buy the house in Woodside, per Rory's earlier story. I'm worried. I'm much more worried about TAM than I was even 12 months ago. I'm much more worried about TAM exhaustion.

Rory

Well, 3 things I think are fair. Obviously, if the goal is to get to $1 billion in revenue before you go public, and our job is to get these companies public, then you need a bigger market than if the goal was $100 million.

So, I totally agree with you on TAM. I don't agree with you on the $100 million, 1%.

6. The TAM myth in startups

I think the best deals work because if you go into a big, wide market where you only need 1%, you're probably undifferentiated. I think the best wins are when you start with this small market and then, as you succeed, your addressable market expands.

So I would argue Revolut's early market—and they might still have only had a couple of percent of it—was not every banking customer, not every next-generation banking customer in Europe, but very much folks who were doing travel and had a lot of FX needs. So you pick this pointy little niche, you get traction in it, you get good margins, and then the beautiful thing is if that expands out and you find yourself able to address more and more customers.

Because if from day 1 they'd gone after everyone in Europe for all consumer banking options, I think they would have gotten spread out. So I agree with you on the TAM, Jason. I just think the best of all things is when, as you grow up as a company, your TAM grows up as well, right, and expands. And I think they've done that. I just don't believe it anymore in my heart.

Jason

I'm not challenging that, but if the entry was at a $25 million post-money valuation and the exit was at $1 billion, it all works out, right? And TAM exhaustion is someone else's problem because you've distributed 24 months after the IPO. Now I see TAM exhaustion across my portfolio, and I never used to—even 18 months ago, I didn't think about it. I see TAM exhaustion everywhere, and you've got to run so fast as a founder to keep ahead of it, faster than maybe we used to think.

We used to have more time. Rory, I agree on TAM exhaustion. Almost all these high private-market bets, interestingly enough, even there you have a TAM question—not an exhausted-TAM question—because typically, when you're paying $70 billion, you're buying the winner in a space, right? You're buying the undisputed winner, right?

For all these companies, you're probably paying a premium in terms of revenue multiple. So, in fact, in all these cases, you're making some kind of “the TAM's even bigger than you think” bet, which is interesting. You're worried about it at your $25 million pre-round, but if you were writing the Revolut memo at $75 billion, you'd be writing the same question, which is: How big can this thing get?

Because, for context, Revolut's got a $75 billion market cap. The biggest bank in England—there's one at $110 billion—and I think Barclays is at $60 billion. So you're already as big as the biggest banks in the country you're domiciled in. All of these things are TAM bets at the kind of multiples people are paying.

I agree exactly with your statement, though, which is that the founder determines the TAM that they grow into. No, but I'm super close to Daniel Ek at Spotify. I'm super close to Alex Bouaziz at Deel. I know Nik Storonsky at Revolut very well. If you can be close to him, I'm close to him.

7. Why founders must expand their markets

All of them have expanded their TAMs sequentially over time, opened up more new chapters in a way that has unlocked more and more enterprise value. The best founders unlock new TAMs. They do.

Rory

But Harry, I was thinking about that as well. You've got Deel—I mean, I know they're not directly competitive in every space—but you've got Deel, Rippling, Gusto, even ones that are much smaller and older, like Justworks. They're all at 9 or 10 figures in revenue.

My point is, all of them have to start as point solutions, for the most part, right? Unless you take the Rippling version, but the notional TAM is huge. The best founders do it faster, right? They're not stuck wherever Deel was in 2018. But the notional TAM was large when Deel was started, even if everybody didn't see it.

I agree. So, to pile on to Jason and be direct, I think you're wrong. I think, in the 3, I'm wrong.

Rory

Yes, and I'll tell you why, to be clear. I actually think Jason said it well. You named 3 companies: Spotify, Revolut, and Deel, right? I think in all 3 cases, it was pretty apparent that there was a potential very big market there, right?

You've got music consumption, and that's what they started with. That's what they're doing today. They expanded geographically, but that's the story. Second, Deel: payroll is one of the biggest markets. We'll talk about that in a second. Then Revolut: fintech, obviously, from day 1, niche, big market.

So I think where you're correct—and I respect the founder comment—is that all 3 of those founders threaded the needle to go from the entry point to a much bigger market, to grab that TAM and think of it as closely adjacent empty space, right? That's how I think about TAM: you have an initial small market, then closely adjacent empty space.

But I think you never said that they didn't start in big markets. I just said that they've unlocked more and more value.

Yeah, I'd love to hear why you think they won.

Jason

All the other music startups based in the US got strangled at birth by lawyers, right? Because it was all about intellectual property rights, and little old Spotify got going in a bunch of European countries that, let's be frank, your average Big Five record label didn't really focus on. So they got a much more attractive licensing deal, while in the US all these guys were wrestling with shitty gross margins and litigation with the music companies.

So they got critical mass early on, built an excellent product, and then gradually increased their leverage versus the record companies. And if you look at Pandora, it was always struggling with the radio-type license. All the other subscription companies in the US always struggled to get access to the music because the record companies were such a pain in the ass.

I really love that Spotify has stuck it to the record companies. It's kind of 10-years-late revenge. But again, it was a combination of great execution and a little bit of serendipity: they stayed away from the fray and got critical mass.

8. Are AI Verticals overhyped?

For what it's worth, I think in the age of AI, we're making the same mistake again in our euphoria. We're very excited—I mean, oh my God, Replit and Lovable, 0 to $250 million in 10 months, a billion at the end of next year. And that, putting aside Anthropic and OpenAI, is putting aside so much of what we're seeing.

There's so much froth and so much greed and excitement, in good ways, that we're funding so many vertical AI plays that we magically think are massive, right? There's no one better than Bret Taylor out there, right? But we're funding Sierra at $50 million ARR at $10 billion, assuming—what's the assumption, Rory?—that they'll hit $1 billion ARR in 5 years or something.

I get the upside, and I get people budgets turning into software. I'm already seeing it, right? We have 4 humans and 12 AI agents. It's faster, but my gut tells me we're over-romanticizing verticals in the age of AI. We're going to hit the same TAM exhaustion, and it's going to be worse because expectations are so high.

That's my take on Revolut: TAM exhaustion. We're doing it again. We're doing it again in exuberance, and maybe it's fine. How many legal niche tools in AI do we really need? They can grow like a weed, but how many can achieve the velocity at $1 billion in ARR that we need on this 20-year journey to get these carry checks? The bar is so effing high to accelerate at $1 billion, right?

Jason

I agree. I think legal is actually a very good case, though, because it's very enterprise, very sticky revenues, with people who don't change tools much, and—

—and a crappy TAM, and a crappy TAM that only looks good. It only looks good today. It only looks good today because everyone's in-market. Here's a weird thing happening today in AI: it is blowing up our assumptions.

In our greed and our rush to make money, we're ignoring something that's happening. I just got back from Dreamforce. In a way, it was the conversation of everyone at the C-suite and CIO level, and everyone's in-market for the first time ever. Every law firm—listen, I've invested in legal and had a decent exit.

Jason Lemkin

It might be 5 or 10 years for someone to look at a tool, kick the tires, think about it, get nervous, and wonder if it's Windows 3.1 compatible. Now, because of AI, everyone's being yelled at and told, “Go find a tool.” And they're buying. They're coming up with $50,000, $100,000, $150,000, and it's nothing to buy a tool.

The fact that everyone's in market instead of 5% of the market, which is our traditional metric in B2B, is warping how we think about market size. It's like 2020 all over again, when everyone was in market for a contact center, an e-signature tool, or a digital events tool like Hopin. Everyone was in market, and then they disappeared the next year. They're not going to be in market every year for an AI tool. This will be a window that will disappear.

9. The "Covid mistake" in AI investing

I think this is a huge point, Jason. Seriously. And you, by the way, expressed it so crisply. It's been running around in my head, but that was just super clear because, if you think about SaaS, there was this 20-year period where it diffused gradually over everything. Some people would be in market every year, and your companies grew pretty consistently. You could lean on a 5- or 6-year growth rate.

10. AI is changing the legal industry

What you're postulating here—and if it's true, it's going to be terrifying—is that because AI is on the front cover of literally every business magazine on the planet, you're saying, “All right, everyone's in the market.” So your signal as an investor on what's going on in 2023, 2024, and 2025 might be entirely wrong. You're basically—I love it—you're making the COVID mistake. When you looked at the growth rate for Zoom in 2021 and early 2022, what did you think was going to happen? There wasn't a human being on the planet who didn't have a Zoom account by late 2022, so the growth went to 10%.

Rory O’Driscoll

I don't know if it happens like that here, but, Jason, if any kind of deceleration happens because of any kind of saturation or slowdown, everyone's estimates on what's going to happen here are wrong. Now, you're right. The counterargument that Harry's just dying to make is the eat-the-world argument.

At the very least, you articulate the downside, the buried downside case, extremely well there. And that's why I bring up legal, because no one would touch legal for years because the TAM was too small, and all of a sudden we think it's huge. Don't get me wrong: people are going to make a ton of money here. There are markets that are utterly changed forever and will absorb massive capital.

But I think in B2B we're going to make more mistakes here than get them right, per Sam Altman's point. And as investors, I think we're being delusional. We're running that late-2020, early-2021 playbook again and not realizing the impacts of everyone being in market. People aren't going to buy a legal AI tool every year. They're just not. They're exhausted. They're exhausted.

Well, so the comparison to COVID, I think, is not right because it was a temporary moment in time that did not sustain. It was not enduring. To apply the same here would be to expect that we won't have AI continuing to improve our productivity in the future, which I think we would all disagree with. So I don't think that has an apt analogy in terms of the huge TAM in legal. I don't think that's why we're all getting so excited.

I think it's because of the structure of data with legal that makes it so relevant for this current set of AI technologies and makes it so relevant. I just don't really get that.

Jason

Hold, just to step back. Superficially, you're right. And listen, even before this week, I never would have said 2020—I don't even use the C-word—was anything like today, okay? It was so weird when everyone all of a sudden needed a contact center. They needed one in a week; I had to buy everything, okay? We were stuck at home, right? At least the privileged were stuck at home. The real people still had to work and go make your coffee.

As we go deeper into this, I don't think the analogy is all wrong. I think it wasn't a change to software. Software was no better in 2021 than in 2015. It made no sense. I think it's worth just learning a few lessons for investing, rather than confusing permanent changes with folks being briefly, for exogenous reasons, all in market. That's the only similarity. The difference is software is radically better today.

It's an exogenous reason that every CIO's neck is on the line. Every CMO who's been told, “Bring in an AI tool or you're going to get fired,” is under pressure. That will not last. It will not last, and deals will get harder because they're just going to get harder.

Are you not making the case for why it's so important to own a market very quickly? They're all in market now. They won't be next year. You have to get it now because they are spending, but they're all here now.

Jason

You mean that kingmaker point you come back to?

11. The rush to buy AI tools before the window closes

No, it's not even kingmaker. I think it is kingmaker, but I think, however, that's the argument to make. What you're saying is, stepping back, what should have been a steady progression of company-by-company decisions over 5 to 7 years has been compressed into every company making a buying decision in the next 1 to 2 years and then rolling it out and sticking with it for the following 5.

So you're right. Unfortunately, the game you have to play is that you have to be here now, because showing up 2 years from now, when you take, for example, 90% of the top 500 American law firms having made a decision, is just too late. So, yes, you have to be here now.

The point is not that Jason's saying it's crazy to be doing it right now. He's simply saying you could overextrapolate the growth now and think it's going to be like this for the next 5 years, when in fact what you might find is that everyone makes a decision and then you slow down quite a lot. The business doesn't go away. It's not Hopin, to invoke a bad memory of your COVID days, Harry. It's like everyone—and the top law firms—have all made a decision, and now you're embarked on a 3-year, steady rollout.

Maybe those growth rates go from the unprecedented 5x or 10x that you're seeing this year to a more prosaic 2x, 3x, or 4x, which is still damn amazing. But if you've leaned in too much on valuation, you might be over your skis. And if you've done the number 3, 4, or 5 player, you might be shit out of luck.

Rory O'Driscoll

Yeah, and it goes to Harry's point: if you've got a winner, lean in. But I don't think every law firm for the next 5 years is going to be in market every year. They're not going to be, and they're going to settle on whoever they bought. When it doesn't work, they might switch once, but it's exhausting to switch vendors.

Business-process change is huge. We're ignoring business-process change, right? One of the talks at Dreamforce this last week was CIOs saying that business-process onboarding and business-process change were at the highest levels they'd ever seen in their lifetimes. They didn't budget for the costs of onboarding these AI apps. They got the price of the vendor right, but the soft costs—the training, the onboarding, and the business-process change—were the highest they'd ever been in their lifetimes.

They're not going to do that every year. So we may go back to 5% being in market in 24 months instead of 100%. Just get it right in your venture models and get it right as founders. To Harry's point, run like hell because they won't be in market again.

But I do think that we obviously just have to acknowledge the difference between enterprise and consumer, because I think consumers will continuously be in market for new generative AI tools to do videos, pictures, websites—you name it. So I think that's a big difference. And Jason, how do you square that away with your statement before, which I always remember, that you can see a future where you have so many more instances of Supabase and need 10x more? You just couldn't consume enough compute, combined with Marc Benioff saying 0.1% of Salesforce has AI. What happens when 50% have it? How do you square away those 2 opposing ideas?

Jason Lemkin

I don't know that there's—it's a good question. I feel like I'm becoming one of those curmudgeons that says you should only invest in trillion-dollar markets. I'm agreeing with it because Amazon just went down in part—I mean, whether it's DNS or whatever—because of database contention with DynamoDB, because so many folks needed databases. And Supabase hit Amazon issues because so many people needed databases.

Listen, every single app in the world needs a database, and what's changing is that folks might need 10 databases or 20 databases instead of 1. So, actually, it's getting to the point where you should take more and more Supabase risk in investing, because the TAM is not only massive but even bigger. Maybe do less vertical AI agents that you think make sense, that are a small part of what ServiceTitan does but are amazing.

But it also might increase the odds that your portfolio comes up snake eyes, because you're trying to do all the Supabases at pre-revenue, at $200 million, or throw the dice at them at $5,000 a month in revenue. You're going to have a high loss rate, too. Rory—

Rory O'Driscoll

No. I'm, first of all, laughing. This is the longest we've ever gone without even actually starting on the agenda. So congratulations, everyone. I don't know if it's only trillion-dollar markets. I think what you've got to be very wary of—I'm sorry, trillion-dollar markets.

12. Will AI markets boom or deflate?

I apologize. I think there are lots of different ways to play it. I'm a big fan of The Hedgehog and the Fox. I'm very much a hedgehog and—sorry, a fox—and there's not one rule from this.

What it does speak to is: don't confuse 25% growth rates with long-term growth rates. Have a good handle on your TAM, going back to something we said about Revolut, and therefore your valuation. The interesting style question, actually, that I'm processing through in real time here now is that there are probably lots of these good vertical markets where you can make money.

There is going to be adoption. You're going to see $500 million, $600 million, $700 million, $800 million-revenue companies. The question, per your point, is: in a world where you don't go public until then, how well do those investments do? How do you avoid overpaying for those extrapolations? How do you think about valuation for those companies versus valuation for a company like Revolut, which is de facto already public and where you're really just doing public investing at scale?

That may be my question to you both: if this is 2020 and we are overestimating TAMs and adoption, and we go back to that, does this pop or does this deflate?

It's not 2020. I thought you said it's not 2020. I don't think Jason said that, so I'm going to defend him. I think he merely said that, when you're assessing trajectories, there have been instances recently, like 2020, where extrapolating on the last year was a mistake.

If it's true in this case, it would be for very different reasons than 2020. I don't know, but it could be because diffusion slows down. I think the markets may take everyone in a market to pick someone and then slow down for a year or 2 as process management adopts it.

So, it's not quite the same as 2020. It's kind of the “history doesn't repeat, it rhymes” comment, right? I think what Jason is saying that is wise is that, if you take these growth rates and extrapolate them for the next 4 or 5 years, and your mental model of SaaS is slight acceleration every year, you could be catastrophically wrong on growth rate. That's what I'm saying. You look like you disagree.

Rory O'Driscoll

No, no. Honestly, it's my thoughtful face.

That's why I didn't recognize it.

Jason Lemkin

The other related corollary, just for investing, is that this is the problem with being a solo GP. You only have so many people to talk to about the thesis. But I think that if you get an M&A offer as a founder or an investor—and the founders make the decisions, the VCs don't make the decisions—

Rory O'Driscoll

Yeah. We make no decisions. If your TAM isn't really accelerating, take it. This is my new learning, to simplify all this stuff: listen, if you get a great offer at $50 million ARR and you have 0.5% market share, don't take it.

13. When to sell vs hold your startup

This is the classic Paul Graham advice: everyone regrets selling because the next year you're twice as big, and then you're 4 times as big. But that can happen even if you're hitting TAM exhaustion. You can still keep growing, but your value doesn't, and we're seeing that in a lot of folks. The value stops increasing.

It's not all like Revolut, where every year you go from $40 million to $70 million to $140 million to $280 million. This is my new learning: when there's an M&A offer, I have a couple of learnings, but one is, let's be honest: has our TAM grown faster than our revenue, and are we at tiny market-share penetration?

Founder, do what you want, but if your TAM isn't large and expanding, I'm too worried. The odds are against you that you're going to hit a TAM headwind. This is just my learning. So just sell. VCs are going to lose like 80% of their investments in AI B2B because they're ignoring these issues, and so be it.

It's okay as long as 1 or 2 of them work out of 10, but they're going to lose so much money. We're hyperfunding niches once again, like we used to, because of this in-market thing. We're hyperfunding niches. We shouldn't be, right? We just don't need that many legal apps or veterinarians that only treat cats. There's only so much demand.

As an investor in vertical SaaS and with many vertical SaaS providers, I'm not torn on it because $22 billion is an incredible exit, but so is $2 billion. And $2 billion would return my funds several times over in 3 years.

14. The truth about vertical SaaS

Jason Lemkin

Yeah, but when you do that overinflated investment—sure, if you got in really early, it's one thing. But when you did the A at $150 million at $3 million ARR because everyone else wants to do it, Harry, and you beat out Sequoia, Accel, Stride, and all those guys, how does the math pencil out on any of these deals? Is it really better than Toast, this vertical SaaS? I'm getting more worried as time goes by.

I think that's the nuanced reply. You look at these verticals and you've got to say, they're adding value. The product is better for the business customer than the prior version of SaaS. They're either making the customer experience better or replacing labor. There's a business here.

Typically, what we're seeing is that in these verticals, it's a wedge product. Maybe it's document recognition or a voice bot, and you can see how you can build a story down the line. You have TAM expansion within that vertical as you just do more and more, and you kind of take from that and you kind of go: are you building value every day? You're damn right you are, right?

Therefore, you're probably creating a valuable enterprise. Money should be available to fund those at the cutting edge. It's AI; it's not trailing-edge, plain-vanilla SaaS. You've got this vertical, maybe only a smaller number of competitors, and you can build enterprise value here.

Rory O'Driscoll

Here's the bull case, right? The classic for me when I started investing—I stole this from a slide someone did at Emergence when I started—but for a vertical SaaS, a vertical B2B that's somewhat SMB, right? That's basically an ERP. It does everything: payroll, backend, and so on.

For the smallest customers, you want to get to $10,000 a year at least. It's what they run their business on. Get 10,000 of those, and you've got a $100 million business, right? That proved to be true again and again and again. It's just that $100 million isn't enough today, right?

So the question with AI, with replacing humans with software, is: can those same 10,000 spend $100,000? If they do, you get $1 billion. You may still slow down at $1 billion. That's the question that I think we're going to have some wins on and some losses on.

Can people really spend $100,000—small businesses—on your same vertical agent software, those same 10,000? We'll find out, right? We will find out.

Jason Lemkin

That's where I think legal is attractive, though. We have this company, Solve Intelligence. It sells to IP law firms. All of their contracts are over $100,000, several hundred thousand in a lot of cases.

[Speaker?]

But so are the vendors you're competing with. So are LexisNexis and others. It's not 10 times larger. I'm not saying it's not a great investment, and I'm not saying it's not going to return your fund, but it's got to be an order of magnitude higher. It's got to be 10x higher for the math to pencil out in venture today.

It's the deal sizes, not just the number of folks in the market. That's where the confusion is. You could confuse the two: the number of folks in the market and whether deal size is 10x what it was 24 months ago. Right? So I think Lovable and Replit have massive deal sizes in a sense, right? It's so much TAM extraction away from crappy agencies and vendors. But if it's just a little bit bigger deal size, we're going to get crushed, right? I don't know. That's my simple math.

Can you get $100K from a small business or $1 million from a midsize enterprise? Will a plaintiff's law firm that used to spend $100K or $200K on just a couple of pieces of software spend $1 million on your software because they don't need humans anymore? If they do, it's golden.

To get back from the dollars and cents, stepping back and giving the case for the defense, as it were: legal software has been a horrible market for many years because it was basically selling workflows to people who didn't care, to be fair. LLMs manipulate words—that's the core of what they do—and lawyers are the most LLM-obvious market out there. So you can definitely make the case in all these verticals.

We could talk about patent law, which is where you are. You can talk about Harvey in corporate law. You could talk about EvenUp in plaintiff litigation, but you can definitely make the case that what came before is not predictive of what's happened now from a technology perspective. There is something really exciting going on in law because of LLMs. So I want to put that out there, right? And it will change the practice of law, just because the technology—again, at some level, you have to be a technodeterminist.

The technology that we have invented—the world has invented, Sam Altman has invented, whoever has invented—is supremely good at ingesting, synthesizing, and spitting back out word concepts, and that's what lawyers do. So if ever there's an industry that could be automated and changed, it's these guys. So that's the case for the offense. That's the case for a lot of these companies that can, in fact, swallow so much more of the market.

I don't think the past is predictive in terms of the amount of dollars you can extract from these companies. I think it could be 10x, Sandy, to your point, but you are right in the sense that, if you start slicing it—you take the overall lawyer count and slice it: how many are patent, how many are litigators—you can, if the bar is $100 million or $200 million, not hit TAM exhaustion. If the bar for an exit is $1 billion, then you could hit TAM exhaustion in some of these markets pretty quickly.

A lot of it, in my view, boils down to the healthiness of the ultimate return. The ultimate return boils down to entry valuation and the healthiness of the exit market. I mean, stepping back, you should be investing in the area where the technology is having the greatest impact, which means it totally makes sense that we're looking at these spaces, because what else are you going to do?

For me, there are 3 areas where I'd be investing if I had infinite capital sources. It's the absolute winners in the space: your OpenAIs, your Anthropics. It's the absolute anointed winners with great economics, which is your Revoluts and your Deels of the world. And then it's your really early investments. I think those are 3 great pillars.

What I don't want to be doing is, respectfully—and I don't mean this horribly; this is a different game—your Mira Muratis at $2 billion and $10 billion. You're $300 million into Periodic Labs. This is a huge amount of money into a still-questionable early asset.

[Speaker?]

15. OpenAI's billion dollar cloud play

It's interesting that 2 out of the 3 spaces Harry chose are effectively post-public-eligible, anointed winners. It speaks to where the market overall is. Most of the dollars are going there, right? And basically what you're saying, Harry, is that with 1/3 of your money, you'd like to do venture capital, and with 2/3 of your money, you'd like to do public-style investing with a 2-and-20 compensation structure because they're still private. I think you're exactly right, and the market seems to agree with you that that's a good way to make money.

There we go. Listen, we mentioned the anointed winners there. I'm loving this conversation. OpenAI have said that they will spend potentially more with Oracle than with Microsoft. I find this relationship fascinating, and how it's developing—the OpenAI-Microsoft relationship. How did you guys read OpenAI spending more with Oracle than Microsoft, and what it means for the power dynamics and that relationship?

[Speaker?]

I think Microsoft didn't want to spend money economically or rationally, and Oracle wanted to be in the game. OpenAI seems to be extraordinarily good at reading other people's needs, wants, and desires and taking advantage of them. Right? I think when all is said and done, on Microsoft's side, the shareholders should award medals to Satya Nadella, the CFO, and their GC, and they should hire someone else to do their technology because they haven't shipped, right? They have cut a brilliant deal with OpenAI, and now they're gradually stepping back as the hype comes in and saying, "We're just not going to make economically irrational investments." I think it's smart.

Yeah. I mean, obviously, in the end, OpenAI needed much more capital than they thought when they started. Microsoft bailed them out by buying 49% of the company. Now they need much more capital than Microsoft thought—probably 2 orders of magnitude more than Microsoft's high-end model of how much capital OpenAI required.

So Microsoft, by de-acquiring it—in essence spinning it out for 30% ownership of what they get, but not having to fund it—they're getting the folks that can tolerate a much lower margin and can somehow get a market benefit out of this in Oracle. So it's kind of crazy that Oracle comes out of here and replaces Microsoft, but Microsoft also gets out of a maybe even awkward situation, right, if they were somehow stuck funding their subsidiary. I mean, that might be more than nickels and dimes if they had to fund OpenAI for eternity.

You said something about economic rationality and Oracle stepping in and being that capital provider in a lot of ways. Its debt-to-equity ratio is now like 4.6x. It's high. Is Oracle out over its skis, or am I being overly cautious?

Well, we said 2 weeks ago we thought they were over their skis, and since then the stock's down. So I think we can claim an attaboy on that one, right? I think, yeah, look, you'll only know when they play the game.

If the demand for AI compute is as high as OpenAI appears to think and Oracle appears to think, and they can bring this investment in on time, then they will be rewarded with a perfectly good business at decent growth margins—not as bad as they currently are because I think there are some startup costs. So it will have paid off, and their current market cap will be validated.

I just look at the risk-return profile and say it's no accident that Microsoft said, "That's an interesting risk-return profile, but I don't need that bet." And Oracle said, "I'm a wannabe in this space. I'll take it." One of the interesting things, going back to what I said, is I saw a dumb tweet that was like, "Oh, OpenAI is going to go bust," because you had the whole Andrej Karpathy "AI is not going to get there quickly" argument and a really dumb "OpenAI is going to be in trouble" argument. No, they're not.

They have brilliantly palmed off all the risk onto everyone else. If you step back and say, "OpenAI needs gigawatts and gigawatts of data centers. We need gazillions of chips. We need all this stuff. You all should do it. Go team." And, yeah, we've signed commitments, and if we need them, we'll actually pay you one fine day with money we don't yet have. But they're not taking on huge amounts of leverage. They're not taking on huge amounts of building. They're just like, "We're in the market to buy this stuff. You should invest on our behalf." It's brilliant.

They've offloaded a lot of the balance-sheet risk to everyone else. All these other people seem to be happily taking it on right now. And we'll see. That strikes me as a lot of risk to take, especially when, in the end, if it all works, OpenAI gets the upside. Your best case is that you're the commodity compute provider to someone who is very rational and is going to be able to grind you down at scale. Your worst case is that you put billions of dollars into fixed assets that don't earn a return.

I think, again, the OpenAI corporate-development deal machine is second to none. They have a ruthless instinct for weakness and take advantage of other people.

I totally agree with you. I'm pleased you mentioned Poolside there, and I do want to go to it now, actually, because it is super relevant and tied to that. On the vertical-ownership side, Poolside announced building its own 2-gigawatt AI data center, which is a big announcement.

16. Poolsuite builds its own data center

Also, Poolside have not released a product to the public. They have customers and they do have usage, but they haven't officially launched a product to the public. And for those that don't know, Sandy, how would you—or Jason—describe the product?

[Speaker?]

Yeah, the product. Again, I've met Jason in the past. I think he's a wildly smart guy, very talented and very successful. My understanding is that it's kind of enterprise-focused, and they're building a core LLM to do enterprise-focused coding and software development—some version of that—and provide an entire runtime environment for these models. So, big enterprise idea.

Not knowing the traction, stepping back, if they're right or if they're wrong, either way, it's terrifying, because the conclusion they're basically saying—and they're very smart people—is that in order to compete at the software layer, you have to not only build your own LLM, but now, goddamn it, you've got to build your own data center. Right? So what they're basically saying is, this game that you thought was a software game is now a fixed-asset-at-scale game, right?

They're not doing it because they're saying, “Hey, I'd really love to own a data center, because nothing says fun like fixed assets.” They're presumably doing it because they can see no other way of doing it. And what that means is these smart people have concluded that that's what it takes to win in the space.

I don't know if I agree, but I haven't looked at the specifics. But again, if they're right—if they're 100% smart and 100% right—what they mean is, do all the other companies trying to build models have to do the same thing? Is there a conclusion here for the capital intensity of Safe Superintelligence, for the capital intensity of Thinking Machines Lab? Do you really have to own your own damn data center if you want to build an LLM? It's an interesting and big-ass conclusion, right?

I looked it up. It's not even like they're doing the Altman thing of having someone else build it. They're partnering with CoreWeave, but per the documents I read, Poolside's going to be the developer. I might have guessed it would have been one—the way big corporations often do a build-to-lease—where they say, “I'm a software company. Mr. Developer, build this building and I'll lease it from you for 10 years.”

I might have thought they'd have said to CoreWeave, “Mr. Data Center Guy, build this data center and I'll lease it from you.” But in fact, they're actually stepping up and being the prime on it. So I think it's a big escalation in capital intensity. I think they must have been driven to that not by choice but by necessity. And it just speaks to this: the business of playing in the space has become more and more high-stakes.

But what's clear is the competition's gone way up over that time, right? The competition is everything from Claude Code to GPT-5 Codex to whatever. Now, no one wants to manage a massive data center, but there's probably no way to achieve their goals otherwise.

And going to your point, I think I got it wrong: it's not about cost, right? There's no way they can do this cheaper, right? It's not about cost; it's about the fact that the bar has gone up to compete with horizontal applications, and it's just much, much, much bigger than when they started this journey.

Because, I mean, if you think about it, think how different the economic intensity here is. If someone came to you when you're building a SaaS app and said, “I'm building this great SaaS app, but by the way, we're not going to use AWS. We're going to need our own infrastructure layer.”

And you said, “Hmm, that's interesting.” And then they said, “Oh, and by the way, we're not even going to use someone else's data center. We're going to build our own data center, right? And we're going to do all this so we can have really great stuff.” You'd be like, “Get out of my office.” Right? But that's where we are in this market.

17. The rising cost of competing in AI

It's good to sneak some of these things up on your VCs, isn't it? You don't want to let it—you don't want to scare them in the first or second check on things like this. It's a cynical comment, but you're exactly right: what's happened here is the boiled frog of capital intensity.

And I think this, again—I'm going to say it—is where OpenAI may have made it a game of capital intensity, where they're clearly winning, right? It's kind of making it harder and harder for people to emerge and compete, right?

I don't know what drove what level of compute they felt they needed and, therefore, what they had to do. But again, I repeat what I'm saying: assuming smart people are making intelligent decisions based on the facts they have today, it's a terrifying conclusion about capital intensity for people who want to play in this space.

And you're right, you said it: it sneaks up on you as a VC. You think you're in a business that needs $500 million to cash-flow break even, and suddenly you're in a business that needs $5 billion to cash-flow break even, and you want bulldozers digging a hole somewhere in Texas. Oh my God, what the fuck just happened? Excuse me.

I was actually one of the first investors. I don't know if you guys knew this. I invested in Eiso's business, which pivoted into Poolside, and so I got rolled in, very luckily, into the first round, which is great. I'm very grateful for it. It's like a 50x. Thank you, Eiso.

I'm just trying to understand the rationality, though, for all the providers who are building models. What do you think Poolside are seeing that they are not?

[Speaker?]

I'm going to make a really pointy distinction here. You used the word “rationality.” Let's agree that the word “rationale” is not the same as the word “rationality,” right? Rationale is why you think you're doing this. Rationality is whether you're right, right?

I think the rationale is pretty clear here, right? “Oh my God, I need this compute.” That's the rationale. And we'll know in 5 years: was that rational or not?

That's very helpful. Thank you for that English lesson.

It was actually genuine. It wasn't actually meant to be snarky, though. I can see why you often think I am. It was trying to distinguish carefully between why you think you're doing something, which can make a ton of sense on the day, on the assumptions, and whether in fact you're correct in the end.

But, oh my gosh, I need the compute when no one else who is building their own models shares that opinion.

Well, OpenAI does, and Anthropic do. I mean, they need the compute; they need access to the compute.

[Speaker?]

It may well be that what you're seeing here is that OpenAI and Anthropic, through all their faults, and the hyperscalers have sucked up all the capacity. It may well be as simple as these guys realizing, “I need X gigawatts of data center capacity, and I just can't buy it today. So if I can't buy it, I've got to build it.”

I mean, they're not doing it because they're doing—literally, it speaks to all this capex sucking up all the capacity there. And even though, going back to the now versus the future, I might be skeptical of the ultimate return on this marginal capex, and I could be right or wrong on that—you'll know in 5 years—it is probably an objective fact today that if you woke up and were trying to build your business and needed that scale of compute, you simply couldn't get it.

You'd ring CoreWeave and they'd say, “Look, I promised $22 billion to OpenAI. I promised $10 billion to Anthropic. I promised $5 billion to Microsoft. I got nothing for you.” And then you're left going, “I either give up my dream and say I can't do this, or I can't put my company on pause until 2027, when I think all this shit is cheaper, so I've got to play the game now.”

And they said, “In that case, I've got to go build it.” I can totally see how you get to that point, which is different than saying you won't regret it in 2 years. You just literally—I assume they're doing it because they rang and said, “Will you sell me 2 gigawatts of data center capacity?” And they couldn't find anyone to sell it to them at scale because it's all been taken up by people with bigger balance sheets.

Thank you for explaining the rationale. That makes total sense now. I was struggling to actually understand the logic behind that thinking. It also, to me, indicates their expectations on future ability to fundraise. It is a bet-the-boat decision to have this permanent investment, and they clearly think they'll be able to raise a huge amount more.

[Speaker?]

Well, probably when they started, they didn't. I'm sure that their slides looked great, but deep down I don't think they thought that Claude Code would be at $1 billion, that Cursor would be at $1 billion, or that Replit would be coming up at $1 billion.

I mean, I think they believed that notional TAM, going back to the conversation, was huge. I just don't think they thought we'd be in the billions of revenue already. And so now, going to your point, there probably is no other way to get 40,000 Nvidia GPUs and the like. It's just not possible otherwise, because you're not the leader.

But also, going to your point, it's probably fundable today because this is much, much bigger than they probably knew it would be. They just probably didn't think it'd be this big in Q4 of 2025. And so now they can raise $5 billion or $10 billion—I don't know what the number is—which was probably impossible when they started.

The thing is, everyone's aspirations and their risk appetite have been walked up.

You’re exactly right. No one had a plan back in 2016 for OpenAI, or in 2022 or 2023, to say, “I’m going to need $5 billion to even play.” You think you get there in 50. You think you get there in 500. The stakes have gone up. The signal is strong because the returns are there in terms of market adoption, and everyone’s risk appetite increased.

Now, at some point, could that perspective change? And would that be pretty painful? Yeah. But that’s how every boom goes. That’s what it feels like.

That’s what it feels like when you’re trying to buy memory chips, when there’s a memory chip shortage, and you’ve got no choice but to sign up with 5 different distributors and commit to buying them because you can’t get any capacity. Then one fine day, capacity comes online, demand diminishes slightly, and, oh my God, these things go down 25 or 30% in value, right? That’s what the boom-and-bust cycle is like. At some point, that’ll happen here.

It’s really interesting that the boom-and-bust cycle makes me think of a bubble. I was looking at definitions of a bubble last night because I have far too much free time on my hands. It really was 2 things. One is a more-than-proportionate drop in the value of assets, with “more than proportionate” being more than 20%. The second is productive capital leaving a market for more than 3 years.

And that’s a bust, to be clear. That’s not the bubble definition; that’s the bust definition.

That’s a bust definition. I thought, wow, that’s not where we are today. We will not have productive capital leave AI and data centers for more than 3 years. Everyone who’s saying, “We’re in an AI bubble. We’re in an AI bubble, and it’s going to bust,” I don’t think so. Because if we’re in a bubble and you’re anticipating the bust, you’re suggesting that those 2 elements will happen.

No, I don’t buy any of that. I think your description of a bust is actually correct. That’s what it feels like on the downturn, which is different from saying it’s going to happen. I think it might, but the definition is correct.

If it goes wrong, let’s go with the “if it goes wrong.” I don’t think it’ll be because none of this stuff works. It’ll just be, “Oh my God, to Jason’s point, we overextrapolated on 1 year’s adoption, and we thought everyone was going to buy this in 3 years and we were going to need X gazillion dollars of capacity.” It turns out that growth next year slows more than we thought. It’s still a dominant long-term trend, but the diffusion of this technology is going to take 10 years, not 2, and we’ve overinvested in capacity.

The marginal player cuts back on their purchases, and then pretty soon, instead of having a shortage of data center capacity, you have a mild glut. Then the price goes down, and that’s how it unravels. That’s what happened in the bandwidth bust. There was a boom in 1996, probably until 1999 or 2000, and then there were 5 years where no one invested in more bandwidth because you wouldn’t. Once there are existing assets available for sale at less than the price it takes to build new assets, no one rationally builds new shit, right?

You could imagine a world—I’m not saying it’s going to happen yet; we can discuss that separately—but the way it goes wrong is if people don’t need the marginal data center that they built for $2 billion and they have to sell it. If the only offer is $1 billion, that’s what they take. If that’s the case, no one’s going to build another data center for $2 billion. That’s what the unraveling would look like. Now, separate question: is it going to happen?

Is it going to happen?

[Speaker?]

Of course. If I had certainty on that, do you think I’d be wasting my time talking to you, Harry? I’d be trading as we speak, right? It’s super hard to call the timing. I find it plausible—almost inevitable—that at some point you will overinvest, because that’s the nature of the beast.

We could talk more about it in sub. I’m making an investment this week, a B2B AI investment that’s earlyish. It’s early, but it uses more inference than anything I’ve invested in yet. It uses far more. In fact, what they want to do soon is use 24/7 inference, running massive amounts of compute for a relatively common B2B use case, 24/7.

You can’t afford it, right? But it’s a sign of the future that smart folks are going to figure out how to use 1,000 times more inference and compute than we’re using today. Instead of running a little one-off thing, or even using Lovable for an hour and then letting the servers sit with no load, this is 24 hours a day, 7 days a week, 365 days a year, running about 20 different passes through the Claude API and wanting to go as quickly as possible.

We’re going to have more apps like that at all levels. That’s 3 orders of magnitude more inference than you really want to use today, right? If it were available today cost-effectively, they would consume all of it. They would consume all of it today if they could. They have demand from their end customers. That was the key sentence: cost-effectively. At today’s prices, they probably can’t afford to do that all the time. The bet you’re taking is that, as price comes down, that will get used up. Correct?

[Speaker?]

Historically, folks have gotten smart at this, right? An early bet I made was a company called OpusClip, which Harry knows. They made clips from videos. The truth is, it didn’t need as much compute as you thought, but they got really good at it.

For example, in the early days, they’d only show you the first couple of clips, and you’d have to request the rest because there was no point in giving you 30 clips when the 30th was never as good as the first one. They got better at a million things.

Now we’re at a stage—I don’t know. If you’re running massive inference constantly, it’s not that simple. But it does augur well for the build-out, right? Forget about where the apps are. We’ve just started with the amount of inference these apps can use.

Maybe the next legal app, the next Harvey—or maybe Harvey does it—shouldn’t just do what you want on demand. Twenty-four hours a day, it should be figuring out what you want, and you wake up in the morning and it’s done all your legal work for you all night long.

If you believe that, should we not just be plowing money into Nvidia?

Yes. I mean, it’s where all of us are. It’s all of our 401(k)s. We’re already deep. All of our QQQ and 401(k)s—we’re already long Nvidia. It just depends how much more you want to concentrate.

One-dimensional sentences aren’t useful, right? Almost inevitably, with a trend this amazing and impressive as AI, the technology—in fact, the trend undersells it—the technology of AI, with something that powerful in terms of a powerful economic impact, you will get overinvestment. It’s just the nature of the beast. People will keep leaning in until it hurts, so it’s inevitable that, at some point in time, people will find themselves overextended and there will be a retrenchment, because that’s just the way markets work.

If it works at 10x growth, go to 20. If it works at 20, go to 30. The only thing that stops you is when it hurts. So, of course, there’s going to be a correction. Sitting and asking, “Is there going to be a correction?” isn’t that useful, right?

The challenge you all face as investors—we all face as investors—is that you can’t sit it out and say, “I’m going to wait for the crash.” That’s not a useful thing. How do you make sure you get enough? You’re there to take the upside and still be survivable when the shakeout comes.

It’s not a one-dimensional comment. It’s not, “Is there a crash? Yes or no?” It’s more a question of: you want to take all the advantage of this amazing technology, but you want to run your business, time your investments, and do temporal diversification such that, at any point in time—and you don’t know when the whole market’s going to find itself overextended—you can survive that overextension and keep on leaning into the trend afterward.

But no one’s doing temporal diversification now, are they? Everyone’s just raising a fund every 18 to 24 months. There’s no temporal diversification.

[Speaker?]

If you go back, yeah. Temporal diversification is one of those things you probably regret doing in the cycle because you want to get as much as you can, and then later on in the cycle you regret not doing it because you get caught. But yes, that’s what you see. Exactly like in 2021, temporal diversification compresses. People get greedy, and then they regret it.

Salesforce has invested $850 million of its new $1 billion AI fund. I was with one of the large capital allocators. I was also with one of the largest foundations yesterday, and they were like, “Honestly, we’re back to 18-month cycles, Harry. We love that you’re 3 years, but you’re the only one, dude. Eighteen months is where we’re at.” Okay.

[Speaker?]

Yeah. The diversification is having 3 funds. If 1 of them is negative, the other 2 make up for it. That’s how you get your diversification. It’s another way to get your diversification: just do 3 funds, right? If you’ve got a 5x, a 3x, and a 1x, what do those average up to? 3x. I don’t know. I guess it depends on the weight.

It depends. For the record, it depends massively on the relative size of the fund.

But I do think that’s what we’re having, right? I do think that’s what we’re doing.

[Speaker?]

You can’t get diversification in 18 months from temporal diversification. You can go from ChatGPT-4 to 5 in 1 fund. That’s the diversification we’re getting. We’re getting an LLM or 2.

Rory O’Driscoll

Again, reminder: in a bull market, the most aggressive person will look the smartest just before the crash, because the more risk you’ve taken, the more money you’ve made. And then that same person is going to get hurt the most on the downside. So, if you actually think about it logically, the correct algorithm you’re trying to figure out is: how aggressive can I be to be 1 step below the level of aggression that blows up in my face in the crash, such that I can power through?

We talked about this in the context of 2021. The most aggressive funds really hit acute problems, but some funds were aggressive, took a little licking, got a lot of the upside, a little bit of the downside, and kept on rolling. What you want to do is be aggressive enough to be relevant, to make all the coin you can in the boom, without at the same time getting caught over your skis and getting shot in the downturn, right? And that’s why it’s a 2-dimensional problem, not a 1-dimensional one.

And where do you come out in that?

Matt Harris

I just don’t know if today we care about any of this stuff anymore. I don’t know if we care about fund diversification or risk profiles or any of this stuff. It’s just go, go, go in the age of AI. And even if LPs are concerned, they’re still funding the leaders.

All these little nuances about how people’s motivations happen—but at some level, in the end, the capital will get allocated rationally and, over the long term, it’s just going to take a long time. My guess is, over the long term, you will see a need to adjust. It will be interesting to see how capital allocations to venture trend over the medium term. It takes a long time to get there.

If you’re funding people moving really quickly, if you’re not getting the returns, if you’re not getting the time diversification, in the end, the numbers will tell. All the endowments report their numbers each year, at the end of June, and you can see the strategies that are working.

One of the things that gnaws at me, including us, is that the truth is, the venture return over the last 5 years has been massively lower than the public-market returns. At some point, you will see pressure because of that. If this AI boom doesn’t come good, the infinite spigot of venture capital is going to get impacted, because, 1 level up from all the relationships you described, Harry, to your point, there’s someone sitting there going, “I just have a spreadsheet, and I just have the last 5 years for the S&P. You’re illiquid. I need 300 to 400 basis points more than that, minimum, to do this, and we’re not getting it. So why am I? Maybe we should just do less of this this year.” That’s, in the end, how things get normalized.

And it takes a lot of energy to do venture. There are smaller checks, there are a lot of managers, and you need a team, right? You sure better achieve that, because there’s a lot more cognitive overhead than sticking in the public markets, and a little bit of—

Jason Green

And you’ve got to earn more. You’ve got to earn your business.

[Speaker?]

When I got into VC, I just didn’t really understand the soft costs involved, right? There’s so much manager selection and manager turnover. You’re off—unless you’re doing a Yale model or others, it’s a relatively small amount of your portfolio for the soft costs. If you’re putting 5% of your assets into venture to get a little alpha, is it really worth meeting with 100 managers and flying to London and having Harry reschedule the pod for his Alpha AGM? Unless you want to be—unless that’s your job, going where Harry started—it better be worth it, right, for a small amount of your portfolio.

I mean, look, the truth is, again, zooming out, big picture, the Cambridge pooled 30-year return says you get exactly what economic rationality would assume you get, which is around 600 basis points on a pooled return, not a median return, above the small-cap. That’s worth doing. Venture is worth doing on aggregate over time. That’s what the facts say, and that’s what economic theory would say.

What’s also true is it’s massively cyclical, and you have periods of massive overfunding on euphoria, massive underfunding on depression, and riding those is brain-dead hard, right? And we’re just in 1 of those euphoric periods now.

When was the underfunded part? The first 2 weeks of March 2020? I don’t remember the underfunded period. It was really—

[Speaker?]

Like 2 weeks when it was underfunded.

Rory O’Driscoll

Yeah, but the problem is your time periods are wrong. I remember 2 vast underfunded periods, each of which got from about ’87 to ’93, ’94, ’95.

This was Arthur Rock and Arthur Patterson.

Rory O’Driscoll

Hang on, guys. No, but can I just say there’s something going on here that’s actually worthy of pointing out? If you’re in a business with 10- or 15-year cycles, you just have to internalize that you have to have a 30-year span to talk about cycles, right? You know, you’re right.

So, I’m actually right when I’m saying the big underfunded cycle after the PC boom, from about ’87 to ’93, ’94, ’95, was massively underfunded. The internet kicked off in that period of time, and then the money roared in by ’96. It was boom time by ’99, and it all went wrong.

After ’99, the money went out, but it took 10 years to go back out. From about 2000 to 2010, the funding rate went steadily down, but it takes 10 years to unwind bad decisions. By 2010, we were massively underfunded, right? And then, obviously, those survivors were able to make compelling returns, and more money rushed in.

The interesting thing about the last 15 years is it’s been a 15-year cycle, not a 10-year cycle, because the equity markets have been so forgiving. At some point, that turns. In the context of anyone playing in this business from 2010 on, there has never been a period of longer than a year where there’s been a substantive correction or a curtailment of capital. The only year would have been 2022–23-ish, and God bless ChatGPT, it ended that, right?

It’s a very different vibe when you’re dealing with year after year of just grind. We haven’t had that. Please, go, but that’s what it looks like.

Rory, 30 years. When did you enjoy it the most?

Rory O’Driscoll

Tomorrow. I enjoy it most of the time. It’s a good question. Is there a period where you were like, “That was a golden day?”

The good question you can ask is: when was it very clearly very attractive to invest? 2010. Sometimes it’s a great time to buy, sometimes it’s a great time to sell. Very rarely is it a great time to do both.

You’ve got to divorce your enjoyment. You’ve got to enjoy the process, not the outcomes, because the outcomes are outside your control, right? I can’t answer the question. I can answer when it was a great time to invest and when it has been a tough time to invest. It was a great time in 2010 and 2015. It was tough in 2021. Frankly, it feels tough today—as tough as it’s ever been.

The good news is stuff is working, but there’s a lot of variance, a huge amount of capital, and it feels tough. You’re way out there on the risk curve, as we said. You can enjoy the entrepreneurs, you can enjoy the excitement of all the new technologies, but when you’re writing checks, you’re like, “Wow.” It’s sobering, the risk you have to take here to play.

Do you agree with that, Jason, it being harder than ever?

Jason

No, I think this is the easiest ever.

Oh, wow. I love that.

Jason

Yeah, because there are so many entrepreneurs. There’s so much change. Change is when you make money, right? Change is when you make money in venture. There’s so much change, and there are so many great entrepreneurs.

We don’t have to worry about gross margins, which really makes investing in B2B easier, right? And LPs, even though they’re conservative, they’re still pressuring you to go, go, go.

This is the easiest time to be an investor. Now, it may not be the easiest time to make returns necessarily, but it’s the easiest time to have a checkbook and to feel smart about yourself. I was literally talking with Claude the other day about how much money I’m going to make investing today, and it said, “You should assume 40% to 50% lower fund.” It said you might end up with only a 2x to 3x fund, and it’s okay. And that’s the moment in time.

This is just Claude, right? We were comparing. I had to upload all the analysis and say, “Your last one’s going to do really well, but look at your entry points and ownerships,” right? And so that may all happen. But, I mean, Matt, it’s just so—I feel privileged to be part of this moment in time. I’m just worried about the entry points, ownerships, and gross margins. The rest is great.

Oddly enough, just those 3 small things: entry point, ownership, and margins. The rest is great, because typically, in investing, when you’re most happy, you’re probably less likely to make money. And it’s a great time to be doing the activity of investing, meeting these wild entrepreneurs. But the problem is, the euphoria can often be an angsty concern about return. That’s all.

I do want to do 1 final one, which is very entertaining. I think OpenAI is going to allow erotica. Sam Altman has sat in a room and gone like, “Yeah, generative AI erotica. We’re allowing it.” And it is the largest use case for Grok’s image and video generation: erotic creation.

[Speaker?]

I'd believe that. I remember back in 2022, even before ChatGPT, we looked at a company—I won't name it—that was doing this online role-playing game. They'd started off with OpenAI as their LLM provider, and they told me they actually had to switch off because it turned out the demand in the role-playing game was for conversations that OpenAI, at that point, was not willing to support. Another LLM provider, who shall remain nameless, was very happy to support it. We ultimately didn't do the deal.

But that's typically what you see. We saw the same thing in early social networking, too, which is that there's a genre that wants that kind of product. I get it, right? Then the question, from a business perspective—not so much from a model perspective—is how much of that do you want to support? It's been interesting: even a lot of social media companies have wrestled with various forms of content moderation, and they might find, as I think Ben Evans did a piece on, that they'll do that for a while, but once they become an ad platform at scale, they might decide that's not something they want to do.

These things are inevitable. Human beings like to talk about sex—shock horror. The question is which businesses meet that demand and how, so we'll see.

Well, look, 2 things. At a high level, I haven't done any OpenAI erotica. I'm not opposed. I should have done the research for this—no joke, right? But I suspect it's great because I'd love, if we had a little more time—well, maybe we do have time—to ask both of you what your ChatGPT moment was when you knew it was good.

The moment for me was when DeepSeek came out and everyone was talking about it. It wasn't even that long ago. I went into DeepSeek to get a sense—I had said, “What happened in The Sopranos after it went dark, after the last episode?” It was so good. It wrote the next episode of a TV show that didn't happen. Its ability to use LLMs, transformers, and GPUs—because it doesn't have to be 100% right, does it? It just had to be great.

Then I went to Claude, which I had low faith in before, and I asked it. Then I went to ChatGPT. They all wrote me a great ending to The Sopranos after it got dark. My jaw dropped, and then I became a convert.

It does worry me. Again, I've said this too many times on this show, but I pay close attention to everything Sam says because I know it's bigger than what he's saying, right? He did walk this back a little bit on Twitter. He said, “I didn't mean it to be as big a deal as it was,” but I think he's saying we're pushing the boundaries now. We want more adult content. We want to let people have less adult supervision.

I don't think Anthropic are the good guys and OpenAI are the bad guys, or any of that phony baloney. But this one worries me. This one worries me. Just like in the beginning, to get these off the ground, we had to trample copyrights and destroy everyone's IP rights. All of my IP is stolen. Everything I've written and all my videos were taken without my consent. Crossing the line on what's right or wrong as AI gets better and better worries me. It really does worry me, and I think we shouldn't cross these lines. We shouldn't cross these moral lines.

Is it crossing a line to have an erotic AI partner?

[Speaker?]

No. But crossing the line of what type of interactions you have with AI might worry me. A lot of things that are adult do worry me. AI is too powerful.

To give a contrast, is it worse than a racist ChatGPT or a fascist ChatGPT? Because this is the tricky content-moderation problem. My big aha is that content moderation is astonishingly hard, and I love watching tech bros blunder into it. You spend 10 years in the congressional spotlight watching them make idiots of themselves, and no one's ever really nailed it, I think. Then you kind of have to flip-flop with the administration. It's just a really hard problem.

So, I mean, you say—and I think this could be an even more interesting nuance—that a lot of the permissiveness of the social media platforms has been because their line is, “We didn't write the content. We're just a transmission mechanism. It's other people's content.” What's super clear in ChatGPT is that you are writing the content.

For things like advice that goes wrong, medical stuff, and maybe even some of the political stuff over time, there are a lot more people in the crossfire. I think you're right to be concerned. I think the content-moderation and content-decision job at ChatGPT is going to be a hot seat for the next 5 years, and I don't think erotica is going to be the hardest problem they face. So I agree. I think it was an interesting one, Jason, but oh my God, it's just a start.

Jason

The point was we're going to allow a lot more usage, right? Erotica is a cute one. We all kind of get it. We can pretend we're embarrassed, but I think it's just the wedge. You get it, right? That's a little naughty, but why not leave people alone in their rooms and read erotica on their phone? What's the no-harm, no-foul? But I worry it's much more than that, just like everything Sam says.

Would you be happy with someone else seeing your ChatGPT history?

Jason

Absolutely. Except for another venture firm.

[Speaker?]

I wouldn't. No way. I would not be remotely comfortable. That's, Harry, a good question to ask, right? And that's with whatever guardrails they have. I would not be happy.

Once it got good after that Sopranos moment, I was all in. I've got to admit something, and it's not going to be nearly as shocking as you think. Going back to Spotify, do you remember how Spotify used to share your music thing? You could share your thing. I hated that feature. I listened to such boring, shitty, mediocre music. My kids laugh at me.

It was far more terrifying to me that people would know how old-school my music taste was than anything they could learn about my ChatGPT chats. So, yes, I get it. People don't want to have their inner selves revealed, even if it's just their taste in country music.

Jason

Much worse than your Venmo getting out.

Yeah, much worse than my Venmo. Much worse than the Venmo. When you listen, you listen to those sad-ass songs, as my wife says: “What kind of loser are you?”

Okay, we're going to play a game, and it's agree or disagree. I'm going to say a statement, and you're going to say agree or disagree and why, and we're going to finish there. Number 1—and I'm making it up, but Jason, you gave me the inspiration for this—Replit will hit $1 billion in ARR by the end of next year. Agree or disagree?

Jason Calacanis

It's only 4x. I'm all in. I'll take the bet.

[Speaker?]

I'll disagree. I'd like more time on that one to think because it's a TAM and market-size question. I don't have clarity on it. I don't know. I'm going to disagree. They're at $250 million by the end of this year, not now, so it's not 15 months or 12 months. I think they're in a more prosumer element of the market, which is smaller. I think Lovable's got a larger TAM because it is literally everyone now. There's potentially higher churn rates associated with it, but there's a much, much bigger TAM, being everyone. Then I think you're just going to start to see cohort maturation and real churn occur.

I think we have it all backwards. I think we have it all backwards, and I know we're out of time. I really think this is a real concern for early-stage investing. It's a huge concern.

Replit today—I started about 110 days ago—and it is so much better than when I started. It's so much better with the current agent. Already, I don't know how many sites in the last YC Demo Day classes—their marketing sites, at least—were vibe-coded, but it might have been 20% or 30%. I could see the Claude Artifacts on the front end. I don't know whether it was Replit, Lovable, or both, or even Claude Code. It doesn't matter. I know Claude Artifacts. I know it was built in Claude. It's painfully obvious when you've been in it, and it just looks vibed.

If in 6 months all this stuff can be vibe-coded by anybody, how the hell can we tell as early-stage investors? We can still judge founders, don't get me wrong, but when a 19-year-old founder walks into 20VC and the product is really, really good at 30 pre, the classic ways we could judge software at that stage go out the window.

I think this is super disruptive for early-stage investing. That's why I think $1 billion is easy, because we're missing how many new categories of software are going to be built. I didn't believe it when I started. Now it's painfully obvious. As this gets better, everyone has an app they want to build. Everyone.

[Speaker?]

Jason, if they scale to $1 billion in revenue that quickly, then they will be raising at $20 billion. My question to you would be: as one of the top 0.01% power users, you should, from a logical capital-allocation perspective, be investing in Replit.

Jason

Yeah, I get it now. I get it. The biggest unlock was when Amjad Masad said he has all the money left from the last round.

I don't get that. Right, he said he would share all the data at SaaStr. I don't know, but let's assume it's mostly true. Founders are always directionally correct, but there's a spin.

If there's really a path there, then competition and other things aside, you unlock the biggest issue, right? If the model is self-sustaining, I really think—I mean, I know it's trite to say you've got to look forward, not back, but the rate—like, this Replit v3, and I'm sure it's true with Lovable—I'm not taking sides. I can't tell you how much better it is. Literally now, pretty much anything I want to build, I can build. I can sit down and create it. I can get it into production. As more folks can do that, it's just crazy what we're going to build.

And so I just don't think a billion is a lot. Rory, help me. What's the TAM for mediocre outsourced dev shops and WordPress agencies?

Rory O'Driscoll

This is the point. Yeah, they'll all disappear. We don't need these crummy WordPress agencies and terrible offshore dev shops that never finish a project and charge you $20,000 or $50,000. They're all going to be gone.

I can summarize this in a sense. If you think of this as a tools market, it's probably going to flatten out. If you think of this as replacing all the people using those tools to build crappy products, and you can just compress that labor spend, then you could get to—the TAM clearly supports a billion-dollar outcome. The only remaining question is: is it going to be Replit, is it going to be Lovable, or is it going to be both? So I see where you're coming from.

Love it. And Jason, you should leverage being a top 0.1% power user and invest. Go do it.

Jason

I'm with you.

In this show, Amjad, he said it. There we go, 5 million. Okay, next one.

Jason

Didn't make enough money, but I'm with you.

Okay, 10 it is. You just upped the game. I would rather be a Deel shareholder than a Rippling shareholder. Agree or disagree?

Rory O'Driscoll

I'm going to punt on this one. I have an adjacent investment, and I'm trying to avoid commenting on areas where I have an adjacent investment. Yeah, I'll punt. I know that's lame.

Come on, Rory. You can do better than that.

Rory O'Driscoll

Jason should actually allow you to answer the question better than me.

Yeah, you have a little bit of inside information.

Rory O'Driscoll

I think I'll step back. I'll get him. I think, big picture, these are great markets, and the reason is something Jason mentioned earlier: one of the most universal business processes that every company has is that they have to pay their employees, right? It's a big-ass horizontal business marketplace. In the U.S., the old-school market supports a $100 billion-plus company, ADP; a roughly $70 billion company, Workday, which was originally HR; and then you've got Paychex at around $50 billion, plus a bunch of $10 billion-to-$20 billion outcomes. In other words, the business process of paying people their money supports a load of really great outcomes.

So when I look at the 2 companies, Rippling is doing that next generation in the U.S. It's kind of like a Gusto story: we're going a little more high-end, we're going to integrate all the HR stuff, and we're going to replace these existing products. To a rounding error, the only negative on this market is, by definition, it's a served market, because everyone pays the damn employees. We've all been small-business owners. You can get almost anything else wrong. You can skip your vendors, but if you don't pay people on a Friday afternoon, you don't have workers on Monday morning, right?

Every single company has an existing vendor, especially in the U.S. So that's the negative on the Rippling side. They're just grinding through, picking up new startups, and then they're in a big-ass replacement. They're doing great and amazing, and then they'll build a big company.

The attraction of Deel—the attraction of my company that we've invested in, Papaya—the attraction in all these spaces is that internationally, it's much more the Wild West. Obviously, people are getting paid internationally, right? If you're getting paid in Poland, you're getting paid in Poland. But what there isn't is an international vendor of the same size and scale as ADP in the U.S. who can say to the U.S. CFO, “Hey, Mr. CFO, you've got employees in 20 countries, right? We'll pay them all. We'll make this go away. If they're EORs, we'll pay them. If they're employees, we'll pay them. We'll solve your international payroll problems.” That's the opportunity there.

I think my company, Papaya, is more at the mid-market and higher end. I think Deel does a brilliant job at the lower end of the market and is expanding up. I think these are big opportunities, because what happened in COVID was people's eyes were opened to how much more talent you can access worldwide. All these companies got a lift from that, because you think about it: someone runs in and a VP of engineering says, “I want to hire 3 people in Liechtenstein or Kazakhstan.” What the freak does anyone know about employment laws in Kazakhstan? You're open to someone who solved that problem for you, right? So I think there's been this giant growth of international payroll. I think there'll be a couple of big companies built in that space.

I agree with you. So you're rooting for Deel?

Rory O'Driscoll

I'm not rooting for them, because we're competing with them, and I'm not rooting for anyone. Rooting is the wrong word. At the margin, of those 2 choices, I would bet at the margin—despite a little distaste for what went on in terms of the espionage thing—I would say maybe what I'd say is the TAM and the competitive matrix is more attractive for Deel. That's as much as I can get.

Despite the espionage thing—that little espionage thing—we move on quickly these days, guys. By the way, you guys were right on that. We had talked about this, and I was troubled by it, and you're right: the world moved on so fast your head spins.

Jason Lemkin

Dude, business completely uninterrupted. Churn zero. They're profitable, killing it.

Jason, Rippling or Deel?

Jason Lemkin

Going back to the start of the conversation, if we're ending the conversation, a billion is still early to me today—not because I don't have profound respect for $1 billion in ARR, but because I worry that for us to get our exits, we have to see enough acceleration past that point.

At the end of the day, Deel's pain point from inception to today is more acute than Rippling's. It's an acute pain point. It's a problem that we've all lived as founders; we've lived with it, and it's very difficult to solve. This international onboarding versus Rippling is a very clever problem to apply to, right? It's ZenPayroll and Zenefits done better, and it's a problem that every U.S. startup and company has. But there are already point solutions there, right?

What's—

Rory O'Driscoll

He said Deel. He said Deel.

Jason Lemkin

No, I didn't say that. I have to pick. Pick one.

Rory O'Driscoll

Yes. Damn it. What do you think we're doing here, shooting the shit?

Jason Lemkin

I'm going to still pick Rippling.

Wow.

Jason Lemkin

Because as great as Deel is—listen, this is a limit of my intelligence. This is a limit of me, okay? You have inside information I don't have, Harry, as does Rory. You both, from Deel and Papaya, have information I don't. I don't know ultimately, as SMBs come into and out of the market and companies churn, which is the most defensible, because both can compete with each other. Deel can build—Deel is a much more agile company than I realized. It can build everything Rippling has, right? It's already built a lot of it, right?

And why? We can bring Parker Conrad on and say, “Hey, were you slow to do some of this?” Right? And why is Gusto so slow to do this? I genuinely don't know. But I'm not going to say, in the age of AI, that having these massive installed bases isn't a huge asset. So I just—

Let me ask something totally different of you, because you said in passing: If a billion is still early, why in God's green earth are you writing checks to people doing $1 million in ARR?

Jason

Because if you stay out of your sweet spot investing—I mean, for me, we're different. I found all my losses when I strayed out of my sweet spot. All my LPs are like, “Take more risk, Jason. Do—take more, take more risk.” It was the worst advice I ever got: to take more risk in venture. The best advice for me is to take less risk. That's how I make the most money.

Rory O'Driscoll

But do you think it's less risky to go later or more risky? I mean—

Jason Lemkin

For me, yeah. I don't have any unique value to add to the CEOs of Deel or Rippling today. I don't have any unique value to add.

Rory O'Driscoll

What you're saying is—and I think I share that sense—what you're saying is, “I'm really good at this thing, which is picking $1 million ARR companies, and trying to do something else, even if it's more attractive from an intellectual risk-return perspective, if I'm not good at it, then I shouldn't do it.” I actually totally respect that argument. I have come out the same door.

We tend to be around later than you, but it is fascinating for both of us, and even for you, Harry, right, that we're doing these deals at half a million, a million, $5 million, $10 million in ARR, and then Jason can casually say, “Oh, by the way, when you get to a billion, you're still early.” I mean—

Well, this is the problem with venture.

It's slow. I finally figured this out. It's my anxiety. It's my anxiety today.

It should be. Yeah, yeah. You're back with that.

But I'm not, Harry. I'm not. My only competitive advantage is that you start getting customers and you don't know how to scale revenue, and you want help scaling GTM. Enough of those folks come to me that I can achieve at least top-decile, or whatever, rates by being chill. But I'm not competitive. I'd have to learn how to—I can't, I can't muscle my way into co-leading Lovable. I don't have those skills.

To your point, what it means is, in a world where—yeah, it's just the elongated time to exit, which we can talk about next time. But, yeah, it's just because I think you're great. It is stunning that we're, I mean, in the same business, in the same rough construct, in the same quote asset class, you have Jason doing $5 million and companies doing $1 million, and you have people doing half-a-billion-dollar investments in people doing $5–$6 billion in revenue, and we think of them the same. It's obvious, when you think for even a second, that those 2 things are so unlike each other that it's absurd, but that's the world we live in now.

Guys, this was, I think, the best we've done. For me in particular, doing this when I'm so engrossed in the conversation that it really goes in its own way, it's the best that we've done. So thank you so much.