Peter Thiel 与 Softbank 卖出 NVIDIA——为什么？以及风投为何将达到1万亿美元、散户资金如何打开市场

Entry price counts when TAM is unclear. Winning is the only thing that counts when TAM is huge.

If you're not seeing massive TAM expansion, there's just no point in even playing as VCs.

The latestage business is either the best business in the world or the worst business in the world. And there's nothing you can do to determine which it is.

Coding is no longer on this extremely steep improvement path. As the models improve in performance dramatically, people switch.

To say that would be ugly would be an understatement. It would be terrifying. I mean, beyond terrifying.

I think we're at a point where if there's some wobble, the magnitude of the correction will be fast and brutal.

Guys, I am so excited for this. It's always my favorite show to do. We have the wonderful Tom Tunguz joining us today. Tom, welcome to this wonderful trio. It's so great to have you.

Tomasz Tunguz

Oh, thrilled to be here. Thanks for having me on.

Not at all. Tom, I've got to say, have you become so Americanized, Tomasz, that you're just going with Tom now? Are you just recognizing that Harry, like all English people, has no command of foreign languages?

Sorry, I didn't understand that. Would you like him to use your given name, Tom? Are we going to stick with what Arie said?

Tomasz Tunguz

Oh, Tom's great. This is great. Let's roll with it. Let's be American.

Don't worry, Tom. Rory will remain this obnoxious way for the following 90 minutes. It's all good. I've gotten used to it. He'll correct your punctuation next.

I want to start with some very exciting news for Cursor: $2.3 billion at a $29.3 billion valuation. Andre and Thrive, CO2, DSTXL—all the big players are involved. Chaps, how do we analyze this? I look at this and honestly feel more irrelevant than I've ever felt. How should we look at this? This is a free-for-all.

Look, I think product-market fit for AI coding is probably the best of any use case aside from search. Then you have this massive growth. I think the bull case is that the productivity gains for software engineers here are pretty enormous—30% to 70%, depending on which company you're looking at.

You have pretty significant multiple expansion. I'm not sure if you guys have played with the new Cursor model, but it's phenomenal. It's unbelievably fast—4 or 5 times faster on a tokens-per-second basis. That allows them to capture a whole bunch of margin.

On a multiples basis, it's actually not that wild. You put all those 2 things together, plus the buoyancy in the market, and you see a valuation here. I mean, can you see a 3x? You don't have a lot of ESOP dilution because total employee count is 30.

Tomasz Tunguz

Yeah, it's 30. They just hired a PM 4 months ago, so they increased headcount by 5%.

You don't have a lot of the capex dilution that you see within the foundation models. Does it go public? You have massive revenue growth, increasing margin, and a pretty attractive financial profile. We can debate the entry price, but I think it's a bull-market bet—a classic bull-market bet.

The big question is that we were looking at a bunch of the vibe-coding companies, and typical gross account retention is 50%. What does that really mean in this business? Can they push into the enterprise? I think that's probably the ultimate determining question. But just given the usage that we see, I can see the case.

Just 2 thoughts to add to that. 1, I think this idea that you get a 30% to 70% productivity boost is almost a backward way of looking at it, because the way I think about it now is that it's just default and necessary. This is the way we code.

If we were talking earlier in the year, even if we were at SaaStr in May, we'd be talking about a productivity boost: What are you getting out of Cursor and Windsurf? I don't know anybody who's not using Cursor. It's moved to the point where we're going to approach 100% penetration per developer, at some price per year—$5,000 or $6,000.

You guys, Harry and Tomasz, are better at math than me. You could do the math for every developer before we even get to Replit and Lovable. How many engineers are there on planet Earth today, and what's $5,000 times that? Seriously, we're going to have 100% penetration, right?

Yeah, no, I agree with you. When I used to do market-sizing models 5 years ago, we used to assume that there were 25 to 30 million developers. In the most recent Microsoft earnings transcript, they're talking about 100 to 150 million developers just on GitHub.

Tomasz Tunguz

Okay, so 200 million times $5,000. How much is that, Max?

Seriously?

Yeah, 200 million. I mean, look, 200 million times $1,000 is $200 billion. No, $5,000—$5,000 a year.

Tomasz Tunguz

A trillion dollars. No, I don't buy that.

Yeah, I think Cursor can do $1 trillion if it has its current momentum. Okay, it could do $500 billion, right? Seriously, this is what we're missing. This is the whole AI play to me. If you're not seeing massive TAM expansion, there's just no point in even playing as VCs. There's just no point.

I was chatting with a sales leader last night. He's a commercial, mid-market seller in an agentic company, and I asked him, “How many figures are in your mid-market deals?” I think of a mid-market deal as $20,000 to $50,000—$50,000 at the highest, maybe $75,000.

Tomasz Tunguz

Yeah.

He said they're all 7 figures.

7?

Tomasz Tunguz

Wow.

Yeah. It's an agentic software company, and the mid-market is high 6 figures to low 7 figures.

Yeah, that's TAM expansion.

Tomasz Tunguz

It's labor replacement in some form or another, right? To that point, if the total number of developers increases—and look, willingness to pay—I pay for Claude Code Max, $200 a month.

Yeah, and you run out.

Tomasz Tunguz

I run out 2 days into the week. Now I'm at a place where I'm thinking, “Okay, do I buy 2 additional seats? 3 additional seats?” Instead of spending $200 a month, I'm spending $1,000 a month. I have this total pain of being able to switch between these keys, but it makes me wonder: what is my willingness to pay for Claude Code? I will never go back to using a computer without Claude Code. I couldn't imagine it.

And that sound you hear is them creating the Tomasz $300- or $400-a-month plan, because they need it. I spend more than that on Replit. You don't want to know what I spend on Replit. It's more expensive.

And you would never go back, right? There's no way. There's no way you go back.

Well, they're different. I'm never going back from Cursor. I actually think, now that we do the math, if it's—how many did we say?—100 million active developers. Sorry, maybe I got the math wrong.

Tomasz Tunguz

Yeah, I think that's right. I think everyone's going to pay $400 to $500 a month ultimately, no matter where they are. So that's $1 trillion, right? We're coming up on $1 trillion. I think that's real. Maybe Cursor gets 30% of it. We could argue that, right? Then we could back into whether it's a good deal.

The Replit, Lovable, Base44, and friends market is the other couple hundred million people.

How many people can be? I can ship 12 apps since June on Replit. 12 apps used 700,000 times since June. I built a product, but I don't code. That's a whole other TAM.

But that's 3% to 5% of U.S. GDP if we're talking about $1 trillion.

Well, that's global. You said global developers, right?

Okay, fair. Some of the money will accrue to U.S. companies.

Great. Any software business is 50% U.S., even though only—what is it?

Okay, so that's 2% of GDP, right?

Most of us aren't even going to be working in 5 years, so 2% of GDP is necessary, because no one wants to work. No one wants to be a hands-on-keyboard executive. No one graduating from college that I know wants to work, right? So 2% sounds low to me.

Tomasz Tunguz

No, it won't be 2% or anything like it, but it can still be huge. I think the AI—yeah, you multiply 100 million by $5,000 a pop and you get a huge number. You can narrow this thing down to “serious developers.” You still get 4 or 5 million. You get, I think, 3 or 4 million in the U.S. of really serious developers—people who are paid to code 8 hours a day, 5 days a week. You can still multiply that by $5,000 a year and get a huge company.

I think the aha here—I’m just going back to the question, call me boring—is the Harry question: roughly $30 billion for $1 billion in revenue is crazy, right? I'm actually struggling more to get the con side than the pro side. The pro side is revenue, revenue growth rate, and probably TAM. If something's gone from $1 million to $100 million a year ago and it's gone from $100 million to $1 billion this year, it's hard to imagine that Newton's laws of motion wouldn't require it to go to $3 billion or $4 billion next year.

Suddenly, you're in this thing at 10x NTM revenue on revenue, revenue growth, and we just did it on TAM.

All of these are great. If you're trying to come up with an argument against it, the 2 ones I hear are profitability and moat. I'd love to talk about those. Tom, maybe you have some insight into that. You read a whole bunch. Let's talk profitability.

You read a whole bunch of—oh my God—the gross margin on these things isn’t great. Sometimes you hear it isn’t great; sometimes you hear it’s awful. It isn’t. Obviously, all that money is flowing to Anthropic, and we’ll come back to that.

But it is also noteworthy that they talked about building their own model, which, of course, would allow them to capture that revenue. I don’t have compelling data on that, but I’d love to hear people’s thoughts on profitability and gross margins for these businesses.

Tomasz Tunguz

My pushback on the con side would just be the emphasis and the focus that OpenAI and Anthropic are placing on Codex and on Claude Code, and then your alternative players like Cognition.

I said there are 2 negatives, and I’m going to list them: profitability and durability. Profitability is, do you make money? Durability is, is someone else going to take your money? I think those are the only 2 issues, which is amazing.

Just think about it: it’s a $30 billion market-cap deal where, on revenue, revenue growth, and TAM perspective, it’s a big, resounding yes: revenue scale, hypergrowth, huge market. So you’re right: the 2 are profitability and then durability, and then competition and durability. So let’s do them in turn, because I think they are linked.

Rory O'Driscoll

Harry, you’re right, because the odd thing about the current business is their direct competitor is also currently the supplier of the raw ingredient that makes 50%, 60%, 70% of their product. So you’re right, it’s a very weird platform-risk kind of deal, and maybe you can just lump them in together: profitability.

Look, with 100-odd employees, it ain’t labor that’s killing them; it’s the cost of the tokens, which is money they give to the company that also has a competing product. So, Tom, I’d love to hear your thoughts. How do you think about Claude Code versus Cursor?

Tomasz Tunguz

Yeah, I think the way I’d put it is: as the models improve in performance dramatically, people switch. Gemini 3 just came out. It’s a little bit better than Claude 4 Sonnet on coding, and that’s what matters to this audience. When there’s a lot of improvement, people switch.

Why? Let me ask: is the Cursor model a whole lot better than the Claude model? GPT-5 comes out of OpenAI. Great. I want to go check out that model on Codex. But as the improvements in coding start to asymptote, I’m going to stay where I am. I’m going to stay where I am because there’s memory, and it remembers how I program. It remembers how many tabs—my linting, which is how many tabs I put into each particular function.

And so I think we’re at a place where agentic coding is no longer on this extremely steep improvement path. People will stay where they are because of the cost of switching. I’ve built—so I have 100 tools in Claude Code. Claude Code wrote all of them, and now I have this whole setup where it does all kinds of stuff for me.

Sure, I told Gemini this morning, when Gemini 3 launched, “Look at everything that I’ve done in Claude Code and migrate it so that you can use it.” I’m doing it at my expense, whatever, and it’ll migrate. But I will only do that if I think that the benefit of the migration is significant.

And so, if you look at the distribution, what fraction of people is really going to switch, especially once the enterprise business starts to come in? Fortune 500 companies will pick 1, standardize, buy effectively an ELA, and then the switching diminishes. And so I think they’ll be able to improve margins.

As long as they’re able to continue to grow, I bet they hold on to, I don’t know, 75% of their audience 5 years from now, something like that. And so, to your point, Rory, the inertia in the business will be there.

What I don’t get—here’s where I’m ignorant, and here’s where Replit and Lovable are so different, right? Replit and Lovable, frankly, are using cheap models. Most people don’t know or care, and they’re well marked up. The gross margins are north of 50%.

We’re not bouncing back and forth between the latest Gemini and GPT-5, right? In fact, Replit defaults you to an N-1 model unless you want to pay more, and it works fine for that use case.

What I still remain ignorant of, even as we’re talking about, is this: I think Cursor has a moat and has switching costs, and enterprise ELAs and others will lock in. But ultimately, even with them mixing in their own model, which may not even have that much higher margins—it’ll have higher margins—how do they get to 60% gross margins? How do they get there? I totally get how Replit and Lovable are already at 50%.

Tomasz Tunguz

Yeah, well, I don’t know either. We’ve met a bunch of different companies, and they’re taking big models and distilling them into small models. We’ve done this internally. We’ve taken Claude Code, which is, I don’t know, a trillion-parameter model, and then we’ve taken a 20-billion-parameter model and said, “Claude Code, teach this little model how to call tools.” We can get to 97% equivalency on that tool-calling distillation with a model that’s 1/50th the size.

This is a venture capital firm, and yes, we have a great head of AI, but we’re not a research lab. Anyway, the point is, I think there is so much efficiency to squeeze out of these model architectures because there’s just a lot of fat in these systems.

Candidly, I don’t know if any of these companies achieve 60% to 70%. We all know publicly traded software companies in the previous era were at 70%, 72% gross margin. I don’t know if we ever get to that place. But the other point is, do they need to?

You don’t need to. Absolutely. I think you’re exactly right, because those companies were selling workflow software with a big sales force and lots of integration. Here, you’re selling a tool that people can turn on and use themselves. You’ve got low sales and marketing costs.

In the end, things are valued on a multiple of free cash flow—in the end, in the limit, right? I’m kind of with you. I think that, as I listen to this whole discussion, if we buy the durability thing—in other words, most people won’t switch once you ask them to try it out—then the only quoted negative is gross margin.

And I think you’re right, Tom: if the only thing between you and $50 billion or $60 billion is your ability to chip away at a digital product where there’s a ton of optimization to be done, my guess is you’ll find a way to get it done. You’re right, it might be 80%, but if you can get to 60% GM and continue to sell $1 billion in revenue with 100 headcount, you’re going to be kicking off cash.

Rory O'Driscoll

Totally. Microsoft also said they were producing 90% more tokens per GPU hour than 12 months ago. So that’s the rate of efficiency gain.

So, 1, 2, and 3 in this space, in 5 years’ time, who is going to be the top 1, 2, and 3 players? Assign market ownership to each before we move on.

So I think Codex is going to have 60%, Anthropic is going to have 20%, and Cursor is going to have 20%, for example. My gut would be Cursor, because they’re there and they’re ahead. Then GitHub, because they’ll bundle—and it’s Microsoft—so a whole bunch of corporate America will just go with that.

Especially, it’s like the Zoom versus Teams discussion: there’ll be bundled people. So those are the 2. Then I don’t know about the 3rd; you have to put Anthropic in because they’re relevant, or Cognition, just because it’s slightly different, which leads me to assume Codex isn’t a huge player here.

I just did that on the fly, but I think that you threw out Codex, which is OpenAI, obviously. But you look at people who have a natural lock on the space: you have the people who are first, which is Cursor; you have the people who can bundle, which is Microsoft at the enterprise level, at the distribution level; you have the people who can bundle at the model level, which is Anthropic; and then you’ve got the clever guys out in the corner.

It’s a crowded space. I don’t know if you put OpenAI in the top 3 in this.

Tomasz Tunguz

I agree with Rory. I think it’s a very astute assessment. I think Cursor has 40% to 60% share. I think Microsoft really needs to step up the product. Have they really lost it? They had the market locked up, and then—I don’t even know what the agentic Microsoft coding product is.

It’s definitely not the tab autocomplete, which is the last time I used it. Maybe it’s bundled within VS Code. They can come out the way they did with Teams and just come out of nowhere.

So if it’s in 5 years, yes, in years 4 and 5, are they probably the number 2 player? It’s right on the money. Then you have Anthropic, which is just so good on coding, and it seems like that’s where they’re focused. So that’s 1, 2, 3: 60%, call it 20%, 20%, something like that.

I’d love to provide a slightly different perspective. The latest version of Replit V3 blows everything out of the water. It’s not just night and day. It’s what’s more than night and day? It’s Pluto and Mercury.

Agents talk to agents. Agents talk to agents. It calls in an architect and reviews my code. It calls in a different agent and finds bugs. It calls in a different agent to review what it has. It has an unlimited context window that appears to go on for months now and remembers everything we’ve done.

My point is, the rate of change is so high on this side of things that I’m not betting there won’t be someone else in 18 months who will blow everyone out of the water. Do I think someone can invest what Anthropic and OpenAI can invest? Hard to imagine, right? Hard to imagine how much they’ve raised.

I mean, a lot. Okay. So I don't know that you can build that. But in terms of building a layer on top of other models, there's a level of disruption to come that we haven't even touched on yet. It's just so much different and so much better.

For me, the biggest issue now that these agents are so good and so autonomous—I mean, this is true for all of software—is QA, right? Now that it's even better, I'm spending time thinking: what if there was a version that could truly do all functional QA agentically? That would be another step function, right? Then that would be 10 times more productive.

So I don't think all these leaders are too big to go away. But if 30 kids at Cursor can build this to $1 billion, are you sure 30 kids with AI can't? Because AI is not static. This rate of change is so crazy. I know Gemini feels like 8% better than 4.5 Sonnet, but what we can do with it in a year—we may underpredict what we can do in a year.

I wonder if that's correct, because it's an important distinction. There's one world in which the window opens with a new technical discontinuity, and there's 3 or 4 years where it's up for grabs. Then things start to coalesce and settle, less because the technology is not continuing to change, but more because enterprise gets locked in. A corporation buys for its people, and then market share becomes harder to move.

Another step-function revolutionary change in the AI underpinnings and the models could cause that to happen. But my base case is that it will start to coalesce more and that market share will become less subject to flux. In other words, people will settle into their rough market share, and that's been typical for most markets.

There's this new wild period, but after about 3 or 4 years, you grab what share you can. Then, in most other markets, there's a long 10-year or 20-year period where, even though the market doubles, doubles, or 10x's, the rough market share at the start is the rough market share at the end.

It's true, but I don't think we've ever seen software get remotely this good this quickly in our lifetimes. It's like 2 orders of magnitude faster. Software used to get better maybe every 5 years. You'd have a major release, and it would have an API or integrate with Looker. That would be the big deal that year: we got our Looker integration working. [laughter]

Now, this isn't even 10x faster. This is 20 or 30x faster than 24 months ago. I don't know. Maybe you disagree. That's what I see the rate of change as.

I'm just arguing back that Intel doubled performance every 18 months or whatever, and market share didn't move for 15 or 20 years throughout the entire life cycle of the CPU. Massive performance increases on their own often aren't enough to cause market-share shifts once they're embedded.

Jason Lemkin

Intuitively, 4 years ago no one did coding using AI. Now everyone's doing coding using AI. There was a 4-year period where everyone had to pick their AI coder. Once you've done that, are you just going to lie back and say, “The AI coding company will just make me better,” as Tomasz said? Is he going to be in the market to shift 2 years from now, provided they all stay roughly comparable? I think it's at least plausible that the balance of probability is no.

I'm trying to figure out the right blog post for this debate. I think it's the bacon-and-the-skillet debate: when does the fat congeal?

Yes. Cute, right? Right now everything is hot, everything's moving around, and there's a lot of sizzle. Then all of a sudden, the heat comes off, and then everything's fixed, right? It's just much harder to move through.

Yes, I love it. And when does that happen? I think that's the debate: when does that happen? Jason's perspective is that it probably doesn't happen for a while because the skillet's going to be cooking on 10 for a long time.

Let me give you another version of that. We rolled out Agentforce for Salesforce, so we're probably one of the few organizations of our size to have rolled out Agentforce. The interesting part is that we took the prompt from another AI agent that we trained for months and gave it to Agentforce. We iterated with it for about a day, and it worked just as well.

The point of the story is that these moats are real. But if I could move that prompt and all that learning from one agent into Agentforce, don't overestimate your moats today. It's just the meta-learning. They're there, but I think they're lower.

So, just on that point, let's talk about commoditization. We talked about moats at the beginning. The markets are growing incredibly quickly, and you have technologies where you could see rapid commoditization and deflation in pricing power.

I'm hoping—

You're hoping we see that.

I think $100,000 per agent—there's only so many that I can buy. I need Theory's fees. I need a little bit of that Theory fee stream to increase to go beyond 12 agents in production. [laughter]

Let's just ask that question. I want to drill down on the word “deflation” because there could be 2 meanings to that word. One of them is the BLS meaning—the Bureau of Labor Statistics—and then the other one is the terrifying one.

Let me tell you what I mean by that. The BLS meaning is, “Oh my God, this year I get 1 million tokens; next year, for the same price, I get 2 million tokens.” At some macro level, I've had more increase in value, but I'm still paying roughly the same amount. It's not catastrophic. It's not an implosion, right? That, to me, is what's happening right now. It's roughly that trajectory.

But you hinted at something that, if true, would be something more than that. It's where you suddenly see price erosion—

Tomasz Tunguz

Price wars.

Price war. What if there's a price war? It's worth pausing on this because it's the only bad scenario. We never saw that in SaaS, with a few exceptions. I remember Box had to compete against Microsoft, which was free, but most of the time there wasn't this.

What you're positing, Tomasz, is that a year from now, the product manager at Anthropic says, “Screw it. I want to win in Claude Code. I'm going to go from $100 to $50 a pop.” The other guys have to respond. Maybe it's because people are embedded, and some product leader says the only way to change that is to go down in price.

Tomasz Tunguz

Yeah. And it's not number 2. Number 1 and number 2 in the market are not the ones doing it. It's numbers 3, 4, and 5. They say, “We have to win significant share. How will we win share? We win share by underpricing.” And then what happens?

Jason Lemkin

But that's not new. There's always been a low-end version of every product we can think of in the market. There's always been a low-end CRM. There's always been a low-end everything in the market. There's always been a $5-a-month version of CRM. It didn't stop Salesforce getting to almost $50 billion.

Tomasz Tunguz

Right. But to your point, Jason, if I can take a prompt out of one agent and put it into another—

Jason Lemkin

Yeah, it's riskier. Your point is that it adds to the risk because of that portability from one product to another. Or even if you're adding the enterprise-grade product to a low-end CRM, having the prompt work just as well is very disruptive. Maybe I pay the same for the AI, but for the core CRM I pay $5 a seat instead of $300.

Rory O'Driscoll

Well, then the time to ship the feature to compete is much less.

Go ahead, Rory. Sorry.

Rory O'Driscoll

I want to take 2 extremes to encapsulate this price-war comment. Subscription revenue from enterprise software that's embedded with a whole bunch of integrations, like Salesforce, is almost immune to price wars. Even if the other stuff is cheaper, you're not going to rip it out, right? So there's some mild price pressure, but they're pretty indifferent.

The other extreme is classic product DRAM. We don't remember what the DRAM was now, but commodity memory semiconductor chips glut, and then they go short every 6, 12, or 18 months. Your pricing spikes 5x, and you're loyal to Samsung for 30 seconds. Then 6 months later, the prices have gone down—not a 10% decline, but a 50% or 80% decline—and they're a commodity. Someone is buying them from SK hynix or Micron for one-tenth the price.

Those are the 2 extremes. We mentally always assume that most software products are a bit below Salesforce—less sticky than Salesforce if it's lovable, but still in the sticky category. If anything like that kind of semiconductor, DRAM-product-type commoditization took place, to say that would be ugly would be an understatement. It would be terrifying—beyond terrifying. NVIDIA would obviously be the example. If GPUs became more like DRAM, it would not be pretty out there.

No, and it hinges on how easy it is for a mid-market or an enterprise to switch. What abstraction layers can they impose as a business? You could imagine Iceberg within the data ecosystem, right? Snowflake captured compute and storage, and then an open-source technology came in and made large enterprises realize, “I want to control my own data. I want to store it, so, Snowflake, I'm going to take this out of your business and hold on to it. I'll selectively give you access to it.”

So, Jason, what if you had a database of all those prompts and fed them selectively into different agents?

Two thoughts. One, this is tough: we essentially have 12 AI agents running now at SaaStr—more than humans. We have 5 SDRs and BDRs running from different instances and different vendors.

Basically, they've turned Salesforce into a database for us because we interact with the agents. I've been a Salesforce customer since the beginning, but we don't log into Salesforce. We don't talk to Salesforce; we talk to Agentforce, Qualified, or Artisan.

Some of what you're saying has already happened to us. That's why Salesforce has to win with Agentforce, because these agents are the most important part of the stack. It can lead to a lot of portability of data, or even just portability of value. To me, that's what I'm learning: it's portability of value.

It's just an existential threat. The old-school guys have to win the agent wars, or the value just leaks out of their platforms. Even if the logos are retained, the value is slowly leaking out every week. They would be in the category of the existing product you sell. Salesforce is still sticky, but nobody cares, and all the extra money went elsewhere. So you just flatten out, and obviously your market cap reflects 10% growth, not 50% growth.

Yeah. But if you could somehow monetize these agents—and that's interesting, going to the deflation question, then I'll shut up—the other interesting thing I've learned from the GTM agents is that I think there will be a price war coming, but right now there isn't.

Right now, they basically all cost $100,000 to start, but the cheapest entry price is like $50,000 to $70,000, plus around $25,000 of an FTE to get going—a forward-deployed engineer. So you're talking about $100,000 to get going.

They're not rampantly discounting it for a lot of reasons. If that price war were to come, all of this massive ARR growth we're seeing in these vendors would deflate rapidly. If, instead of being a $100,000 product, they were a $2,000 product, it'd be tough in venture.

Yeah. [laughter] No, look, I don't think it's going to happen. I just think it's important to raise the question because I suspect maybe in 1 or 2 categories this does happen, where you start to have competition on—

It could happen in other areas more quickly. If it's going to show up, it's going to show up in core API pricing, coding agents, and the Lovables. That's where it's more likely.

Definitely not in Lovable, Harry—obviously not. Not in Lovable. How could you say that?

But if you're out there charging $100,000 a year for your agent with super-happy customers, this is Tomasz's point: they're great, it's working great, it's wonderful. But I can take that prompt and just a little bit of history, just a little bit of abstracted data, and move it to a $10,000-a-year tool.

When things are a little less frothy and AI budgets are a little more stable, moving that $100,000 or $200,000 to a $20,000-a-year agent might be appealing. I remember looking at churn in SaaS companies, and the number-one predictor of retention was the number of integrations.

Going back to your point, if it's easy to rip it out, you will rip it out if it's cheaper. If it's hard to rip it out, you won't bother. So I agree: if you're literally—your concept of a database of prompts—and you're interchangeable, then you're right. It's a big sign saying, “Cut me now,” when you have to save $80,000.

But if you're integrated into 5 things and you're like, “Oh my God, we'll have to talk to IT,” then screw it.

But can I, Harry? I want to talk about a totally different topic, but on the same topic, as it were. I want to come back to your theory, Harry. Do you feel irrelevant, right?

And I think yes, and I think there was a fun point in that. Cursor, I think, has had at least maybe 3 rounds this year, right? One of the most noticeable things about this year, and out of Stanford, is the number of companies doing multiple rounds. Obviously, it's significant step-ups in the same year.

Can I just touch on that? Ramp was $13 billion at the start of this year; now it's $32 billion, with the latest round announced yesterday.

They've seen 4 rounds this year. I looked it up: Ramp's had 4 separate financings this year.

To give a statistic on that, we look every year at the newly minted unicorns for that year, for that quarter, because that's mentally the outer edge of where we place them. So I'm like, “Okay, what did we miss?” There were something like 24 minted unicorns in Q1. By Q3, 15% of them already had a step-up, and with Cursor, some of them had 2.

If you think about the velocity of step-ups, normally you think your financing is 12 to 18 months. Fifteen percent of your companies within 6 months—companies that you entered at a billion, above a billion—have already had a step-up. To your point, it seems like a high-velocity, big-numbers game, and it looks like a remarkably easy game from this. I'm sure it's not, but you're right: you look there and go, “Let me get this straight. You put in $100 million at a billion, and you have a 15% chance of being worth $2 billion within 6 months. Why not do that for a living?”

I mean, I think that's what you're saying, Harry, effectively: buy Ramp in January at $13 billion, sell Ramp at $26 billion in May.

What I'm saying is, is my insertion point fundamentally challenged because it is just so much easier? And you say, “Oh, it's not easier, Harry.” It absolutely is. With the brand, it looks easier to—

With the brand and the platform that we have, access, to a certain extent, is the core challenge, respectfully. I could be doing $10 million to $25 million checks into these high-flyers, like your Harveys of the world that we've discussed at length, and we would be able to get them. I could get the step-up.

But no, I go back to the craftsmanship of seed and building companies in the trenches with entrepreneurs, and I'm thinking, “Why do I do that?”

Well, I'll tell you what's interesting. I'm watching Bessemer, who's wildly successful in cloud and B2B for generations, just co-lead the last Ramp round. They did Anthropic and Canva so late. Byron, I love Byron, but they did Anthropic, what, a year ago? That's probably up 10x, right? So they did $100 million or something into Anthropic.

Well, Canva, I think—this is my observation from afar—they did Canva in 2021, and then I think maybe they had a little bit of shock. They're like, “Wow, maybe that's a great one. Maybe we overpaid.” Now they're in the money on it, right?

But then they did Anthropic, which seemed expensive. We should look it up. And then, going from being conservative but wildly successful, then going to Anthropic, then going to Ramp at $30 billion, saying the classic post, “We're so excited to partner together now”—Bessemer must think that is a low-risk investment.

That's what I'm saying. They must think Ramp at $30—I know, I know, it's Captain Obvious—but these are guys that have thought. I mean, this is a venture capital firm that's been around since the 1800s, right? Or something like Bethlehem Steel or Bessemer Steel or something. They think Ramp at $30 billion is the best play in the market.

I don't know what Theory thinks, but to your point, this is not Tiger or SoftBank rolling the dice. This is Bessemer saying Ramp at $32 billion is a good, safe bet. [snorts]

Kleiner and Menlo doing Anthropic at $180 billion. Another example.

One of the interesting things here is that a large number of the folks through these kinds of rounds are not the late-stage crossover people who, to some extent, got snookered in 2021, licked their wounds, and crawled away. It's actually the great, large early-stage, now multi-stage firms who are looking at the same math we just looked at in Cursor and saying to themselves, risk-adjusted, “Is this a good—is this just a great place to put my money?”

If you have the scale of capital to be relevant at that stage, because you really need—I mean, maybe you can show up, Harry, because you're a media celeb, but normally they want to talk to people with $100 million-plus. If you have a fund that size, so far it's been a very excellent place to put one's money, and many of the big—what we would have called early-stage firms 10 or 15 years ago—are doing it.

You're right: a16z, Lightspeed-led, I think, the Ramp round. This stuff is working. I always used to say to my LPs, the late-stage business is either the best business in the world or the worst business in the world. There's nothing you can do to determine which it is.

When prices go up, putting in $100 million and having it go to $200 million with no effort on your side feels as good as life is going to get. Obviously, when prices go down, it ain't so much fun. See 2021–22 for details.

I think the secret to success in that business is just being a trader. I was walking in the park with a multi-billionaire today who is in this market, and he is a trader, a ruthless trader. He buys at 60 and sells at 180 in the same year, and it's absolutely a book that he manages—not with “ride your winners, hail this unicorn founder.” It's trading.

It's the new public market.

Oh, one huge difference.

Excuse my language. You can’t—there’s no liquidity to the downside. You made a statement that it is the new public market because these are companies that, by any rational stretch, could be public today. First comment, Harry: you’re right. In public markets, some people have a trading strategy and some people have a holding strategy.

But the key sentence you’re missing is that you can’t execute a trading strategy if they’re private, because when things go wrong, the liquidity won’t be there. When things go right, you can trade on your way up, but it will be a lot harder to get out of one of these investments on the downside because the liquidity will not be commensurate with the public markets.

100%. Putting 25 into—I’m just making up any chosen company that sold this year—Ramp at 13, and then selling it at 32, would not be difficult.

No, exactly. On the way up—let me repeat—on the way up, the late-stage business is the world’s best business.

No, but most are on the way up. I mean, we have our YOLO segment, which you’ve taken the piss out of me before, Harry. They’re all just riding it freaking up. But apparently, you might want to check it. You might want to turn on your ticker for the last 24, 48, or 36 hours. But yes, in general, stocks go up.

I did. There’s so much red, Harry. There’s so much red. Duolingo is like the Titanic. It’s like, “Ooh,” it’s all under the surface, you know?

Totally.

But yeah. No, it’s just a super interesting time for that.

Totally. 2 elements worried me—concerned me—this week. Well, there were several, to be honest. One was Thinking Machines at $50 billion, and the other was Thiel and SoftBank exiting Nvidia, and just what it means for whether we’re at the top of the market. Both are potential signs of a market top. When you look at those 2, unpack either of them, both of them, but both kind of concerned me when I saw them.

The only thing I would note from the media is that Peter Thiel sold $100 million of Nvidia. What’s the dude worth? This is like me selling three-tenths of a Bitcoin. I mean, it’s not—the estimate is between 10 and 20. So, you’re right: it’s sub-1% of his net worth. You’re right, $100 million.

Though I will say it’s been my life experience that people rarely sell stocks because they think they’re going to go up. So, at some minor level, in the 10 seconds it took to run that decision by the big guy, he said, “Yeah, you should sell that stock.” So, you’re right, it’s not like he’s unloading it the way he was when he was unloading his Facebook position. And again, I’ll do the Nvidia one—is that the other one?

I don’t think there’s any data in SoftBank selling. They’re just selling the profitable public company Nvidia to put that money into OpenAI. This is a guy ramping up his risk. This is not a de-risking; this is someone saying, “That profitable, publicly traded chip company just isn’t risky enough for me. I’m going to roll out of this one and into OpenAI.”

So, I don’t know. You might well be right on the market top, but it isn’t because of those 2 data points. The data points I’m paying attention to are in the credit market. I’m looking at Oracle credit default swaps—triple what Amazon, Microsoft, and others are. I’m looking at, even in consumers—gosh, here’s a data point: subprime borrowers in the past 60 days hit the highest delinquency rate on auto loans in recorded history.

Then you have Blue Owl, which has frozen redemptions for 1 non-traded BDC vehicle and is moving it into another one. Then you have First Brands’ default in private credit.

Can we just unpack what you said about the Oracle credit default swaps? Can you help me understand what’s going on there and why that’s important?

Oracle has a big deal with OpenAI. Oracle needs to build lots of data centers. To build those data centers, they borrow money like a mortgage, and they’ve borrowed money. There’s a thing called a credit default swap, which you’ll remember from the Great Financial Crisis. It’s the odds that Oracle defaults on its debt—that it cannot pay its mortgage.

Google, Microsoft, and other major technology companies are at a certain level, which is basically the same rate as the federal government—government-grade. Oracle is 3 times that in the last 3 or 4 days. So, it’s a big move. The risk is still quite small, so the overall probability of an Oracle default is small. The magnitude of the move suggests a meaningful repricing of risk.

And worth pointing out, at the same time, the entire value of the Oracle deal has been unwound. The market cap of the core company is actually below where it was when the deal was announced. I think both of those data points are saying the same thing, which is: Oracle, you’ve just underwritten a risky piece of business. So, your equity is worth less, and I’m going to have to reinsure your debt.

People at the margin are going, “Maybe I want to be one of the first people off this pain train, and maybe I can insure my risk or hedge my bets.” I think you’re right that that’s the tell here. Is it this big, screaming flag? No, it’s not. It’s just a data point that the market is starting to perceive an increasing amount of risk in some of these big contracts.

Then you have the Anthropic deals today from Microsoft and Nvidia, with a $15 billion investment, and the circularity questions and all those kinds of things. People are perceiving more and more risk as the capex for data centers goes from $500 billion a year to $800 billion a year or more.

Do you think there are any screaming flags from the last week?

I don’t think so. Most of the hyperscalers’ GPU capacity is sold out for the next 2 years. They generate cash. The debt as a percentage of free cash flow is really small.

I think the major red flag for me is that customer concentration risk is higher than it’s ever been. 2 customers for Nvidia represent more than 40% of revenues; 4 represent more than 50% of revenues. I went back and looked at the dot-com-era networking companies. Nvidia is 10 times more concentrated in terms of revenue than Lucent was. So, I think that’s an issue.

But most of Nvidia’s customers are super cash-flow-positive, right? Google, Meta, and others are spitting out cash, and they can decide to stop at basically whatever point. So, I think it’s all okay.

How does this merry-go-round stop? If the game of musical chairs were to collapse and everyone falls on their butt, what happens?

Inference demand slows. And if there’s a hiccup—if Google says, “We built this amount of capacity and we could only fill 80%”—if that happens, then you see—

Yeah. You’re about to learn something about doing this podcast on Tuesdays that you might not have internalized, but I’ll tell you what it is. This thing comes out on Thursday, and Nvidia reports on Wednesday night.

So, we’ve now been pontificating, and 1 of 2 things is going to happen on Thursday, when you listeners are listening to this. If Nvidia is steady as she goes and it’s doing fine with a few little warnings, we will look like balanced and rational people. If they pull the pin to the downside, we will look like the last men on the Titanic here, right? It’s terrifying because that’s just the nature of the recording class.

But now, to lash myself to that mast with you, Max, I think you’re right. And what you’re not seeing is a mass collapse of demand or anything like that. You’re seeing really strong demand. All the hyperscalers are saying, “We want to buy more. We want to build more. We want to invest more.”

The negatives are at the margins, which are the overlevered people trying to do this—people who are correctly worried about their debt. The people who have bought the balance sheet and the need for these products, on the other hand—the Microsofts and Googles—aren’t worried at all. In the middle, you have Meta, where it’s like, “You can afford it, but why are you doing this, dude?” So, you’ve internalized that.

I doubt Nvidia are going to get on a call tomorrow and say, “Demand’s gone down.” So, all should be fine for a while. But I just think it’s, to your point, over the medium term, people are going, “Hmm.”

The debt that some of these folks are taking on—like Blue Owl’s, like Oracle’s—is just a risky bet if things turn down. And I think we’re at a point where, if there’s a wobble, the magnitude of the correction will be fast and brutal.

Everyone knows the tachometer is at a red line, right? We are going as fast as we possibly can. In fact, we’re going so fast that we are, as an economy, really uncomfortable with it. I was reading a macro hedge fund’s tweet last night, and he was talking about how, because the big companies are borrowing lots of money, they’re paying less tax revenue to the US government, and that tax revenue is so significant that it actually will increase the national debt.

This is where we are. We are going 1,000 miles an hour in a car that’s designed to go 999, right? So, the whole thing is shaking.

I totally agree. I mean, the fact that people argue about the depreciation schedules on GPUs, and the answer to that question can move the entire US stock market, is beyond bizarre.

But you're right, we are where we are. We're making this bet, and even a mild slowdown would be painful. My theory—random comment—is that because no one can get the power to do these, we might actually be saved from ourselves. If no one has to say, “There’s no inference demand,” and everyone just says, “Well, I would love to build those extra 10 data centers, but we just can’t get the power,” we’ll gradually slow down the ramp.

Maybe it’ll slow a little bit, less ostentatiously than if someone gets on a conference call and says, “We built another brand-new, spanking data center, we turned it on, and nobody came.” Because that’s the moment, as Max said, when you go, “Maybe the other 20 we have in the works aren’t going to be worth much either.” So maybe our inability to connect power will save us from overcapacity. That would be my upside case. What do you think the chances are that we continue smooth sailing into the sunset and don’t hit an air pocket or a challenge for the next 3 to 4 years? What if we’re overestimating?

Zero. Maybe 10% or 20%.

I think the past moved so much more slowly than the present in B2B. But if we go back through our history of SaaS, we had a lot of minor bumps on the way to the peaks. We had a meltdown in 2016 that we’ve all forgotten, where SaaS fell 30% or 40% in 2 weeks. It was right during SaaStr Annual, right?

If you go back and squint at those charts, you’ll see massive corrections that we fully rebounded from by 2022. So why wouldn’t we have massive corrections on the way to us all living in a data center, which I think we all are? I think data centers are the new cities. We’re building more data centers than offices.

Why shouldn’t we have 30% or 40% corrections along the way? We should. How could there be no bumps? Oracle can’t get its debt refinanced. Maybe CoreWeave’s contracts aren’t quite what we hoped. Maybe it’s something small. Maybe Nebius has a bump and it creates a contagion in the market, or Microsoft has some issue. Why shouldn’t we expect 3 or 4 little 30% to 40% drops? We’ve seen it before in our investing lifetimes.

I’m trying to imagine what a house would look like with a white GPU fence.

A white GPU fence.

Oh my God. That’s the vigor for the day—the new American dream. I love a white GPU fence. That’s right, with a “Made in Taiwan” sign on it. How much more American can you get?

Oh, that’s it. It is coming. There will be more agents in this country than humans soon enough.

Oh, yeah.

No, for real. It’s going to fundamentally change our lives. That’s the part we’re missing: when there are more agents than humans.

Linking it back to Max’s comment, though, unfortunately, what they don’t do is pay their car loans. This is back to the comment on where the wider economy is.

Just one comment on that crash comment, Max. I remember 2016, and I even saw a tweet that showed the Nasdaq since 1981. They were saying, “Hey, it’s all fine,” and they pointed to the 2001–2002 crash and said, “Look, in the scheme of things, it’s nothing, because the line goes up and to the right.” They’re entirely correct, but I really liked someone’s response: “Yes, but for 16 years, it took 16 years to get back to where it was.”

The longer your time horizon, the more indifferent you can be. But if you find yourself on the wrong side of what was, in 2001, a 70% to 80% correction in the Nasdaq, it can hurt for a long time. My public service announcement is that if you find yourself feeling pretty nauseous about the de minimis crash you’ve lived through in the last few weeks—4% to 5% down, maybe 20% in the second swing—you should look long and hard at your asset allocation and maybe put a little more in cash. I got a little scared and thought, “What are you doing here?”

No, when you’re scared, you seriously don’t look that large.

This is the best. If you’ve been around for a little while, you have to learn: if you’re scared, don’t look.

If you’re scared, don’t look. That’s the only thing you should do. Don’t look. It’s the best advice.

“If you’re scared, don’t look” is the theme of this Y Combinator batch, I can tell you this week. I’m being serious. I’ve never seen such exuberance around a batch. I’m getting emails.

They’re always the best batch ever, Harry. That’s the obligatory tweet: “We’ve raised the $5 million round, and now we’ve opened up the next note for the next note on the note of the note.” I cannot tell you how much exuberance there is. They’re good companies, but the fear of public markets and impending doom has not reached early stage, baby. It’s like $50 million post-standard.

Are you seeing the same? Are you nervous like me? I’m like, “Guys, I don’t want to.” Also, a question for you: advice. I feel like it’s, “You’re so lucky to have a meeting with me, and I’ll determine if I should ever take your money, Harry.” And I’m like, “I haven’t even met you.” Am I being too romantic?

When money is scarce, conditions toughen up, and frankly, VCs get pretty hard-nosed about allocating capital. You’ve got to expect that when money is plentiful, entrepreneurs behave the same way, right? Some part of what you’re describing is legitimate.

Now, the test of character is how you behave and how you act interpersonally in those times. When money is scarce, I think as a VC you have to allocate capital carefully, but you don’t have to be a dick. In the same way, you’re right: you see some behaviors now where it’s almost like an interview to an interview. You’re like, “Okay, I get what you’re doing, and you have the hot company, but life is long.”

I think the best way to approach this is to try to be a human being most of the time, either as an entrepreneur or a VC, and recognize it’s a massive, multi-period game. At the same time, you can’t deny that the market is the market. Right now, that market is wildly pro-entrepreneur, and railing against that, Harry, or being romantic about it, is a waste of time. It is where it is.

What are the odds you think the US venture capital market hits half a trillion dollars by 2030 in size?

What’s it now?

When I started in 2008, it was about $8 billion. In 2021, it hit about $300 billion, and today it’s about $275 billion.

100% chance.

100% chance. Okay, so maybe more than 100%. What’s north of 100% again?

I’ll tell you why, but keep going.

Okay, so let’s assume that’s the case. Venture capital, or the cost of venture capital, continues to decrease, which means valuations continue to increase, which means capital increasingly commoditizes.

Put it this way: you would be correct, Tomasz, on the data that you put forth. I’m going to add 1 more data point that you missed. What was your first year?

Tomasz Tunguz

2000.

How much was in the business? What did you say? What was your first number?

Tomasz Tunguz

$8 billion.

$8 billion. What you missed was that in 1999, 4 years beforehand, there was $100 billion in the same system. So it went from $100 billion to $8 billion. Basically, since then it’s been an upward line.

I remember I was in the business, and from 1994 on, I remember that in 2000 you could literally delete 75% to 80% of your address book because you were never going to see those people again.

So if you extrapolate the line, you get to $500 billion. You’re exactly right, if you allow for a CAGR. Maybe you don’t.

Tomasz Tunguz

Well, here’s the thing. It depends on what you’re looking at. I’m just looking at Axial’s Global Landscape. They published this week and had a nice chart. Tomasz always has the better data, but they said this year they’re estimating $184 billion in venture capital invested this year, by their definition.

The peak was 2021, at $183 billion. So, $1 billion more this year. But half of that $184 billion is going into 4 companies. Is that venture capital? If that’s venture capital and AI grows at anything like the rate we’ve discussed, of course it will double. Maybe 110%, maybe 95%.

But only 74% went into the rest, which is half of 2021 and consistent with 2020. So it could be that Y Combinator is overloaded, and these 4 or 5 names are overloaded, but the rest of the money says it’s not overloaded. It’s not easier.

It’s money from the public market that is fighting its way to those shares, irrespective of the venue. That’s what you’re saying?

Yeah. So we have this bifurcated market where Y Combinator, maybe Neo, and a few others have huge benefits, and they’ve earned it. The massive names have earned it, and then we’ve got 900 unicorns that are never going to IPO.

Poor guys. We all have 1 or 2 in our portfolio that are at 9 figures in revenue, are still growing, and will never IPO. There is no private equity buyer for them. I really think we have to define what venture capital is to fully answer your question.

But the dollars—if you include Anthropic, OpenAI, xAI, and SpaceX—it’s got to double, right? Ramp doesn’t even make the list. Poor guys at $32 billion.

Tomasz Tunguz

Maybe they’ll get there. They’re only consuming a few billion. It’s not enough.

So what we’re basically talking about is a huge concentration of those dollars at the very, very late stage, right? I mean, these seed rounds of $1 billion at $5 billion pre-money—

And so, yeah.

What it means is, to be clear—and I’m going to Tomasz now, because that was actually a helpful intervention—it made me realize something. The clear answer to the question, “Will the industry double in the next 2 years?” is a function of what we said earlier: if the return is there, then it will double, because money chases returns.

That’s the first statement, right? If the returns continue to be really good, more money will come in until eventually the money kills the returns. That’s the way the movie works. The question “Will the industry double?” can be reduced to a simpler question: will the returns be good?

The aha that you two just gave me is that, to a rounding error, the question really resolves itself to whether the 4 or 5 companies that constitute 40% of the non-diversified side of that industry will be good.

If OpenAI, Anthropic, and all yield the return that everyone obviously hopes they do, then already you've taken half the risk off the table. Everything else does roughly okay. Even if some of the old stuff doesn't work out—and a lot of the old stuff doesn't work out—the $40 billion in OpenAI, from a pooled-return perspective, can swamp 40 separate unicorns entirely. Poof, gone, right? So, basically, you could be right: if the concentration works, it's all going to be fine and the industry will keep on chugging. If the concentration doesn't work...

Yeah. So what you're saying is, if OpenAI trades up at IPO—

Yeah.

—it’s roses for everybody.

I'm more Anthropic than OpenAI, but those kinds of things. I mean, remember SpaceX, too, which is worth $300 billion or $400 billion. Those are the—now, yeah, it definitely helps a lot.

The way that I always see that, actually, is in meeting LPs, because of the amount of LPs that are sitting there with positions in Stripe and SpaceX and the names that we mentioned. And I think you forget the downstream multiplier recipients of all of these big names—to literally dentists in SPVs now, in a lot of them. And poor dentists we always use, but it's just the thousands and thousands and thousands.

Rory O'Driscoll

It's back to what Tomasz said earlier. This is where we find ourselves. Who knew? But this is it. The bet is on, and the bet is singular and utterly correlated.

Why do dentists have so much time, by the way? It seems like they finish work at 5 and just go home and figure out how to invest their cash. I've never seen a group outside of tech more obsessed with tech investing than dentists.

It's because they have a non-insurance-governed market. It's a cash-pay market. Dentistry is a good business because you get your crown done and you pay cold, hard cash. They don't have to deal much with insurers. They just make good money.

If you go to your dentist, they're all good businessmen. They have 10 chairs running. They have 10 hygienists. You get 5 minutes with your dentist. He charges you a ton. It's a great business.

I try to avoid dinner parties, but my biggest fear is that I go to one and I'm sitting next to a dentist. Not because I don't want to talk about his or her business. I don't want to talk about tech—

Rory O'Driscoll

—for the whole night. You know, can—

I'm in insurance.

Can you get me into Tomasz's latest deal?

Can you ask Harry if I can get into Perplexity? Can you ask Harry if I can get into Perplexity with the dentist? Oh my God.

Rory, after all these weeks of Harvey and Legora and me chatting about Solve, you go and do a deal in legal tech, baby. GC AI raised from Scale at a $550 million post-money valuation. Weren't we price-sensitive? What are the top lessons, then, Rory, from leading this round? I'm really interested, given that we've talked a lot about it legally.

Rory O'Driscoll

Look, I'll say something. I didn't expect to lead this deal. We were doing references on another company in broadly the same space, and we just got customer love for this product. It's just that simple, right? We just got customers saying, “We really like this.”

Again, I don't like making this show about our own deals, because I think people respect the fact that I'm not trying to talk our own book. So we'll keep it tight. The name says it all. It's GC AI. It's AI for the in-house legal team, which is different from AI for corporate law.

We talked to customers in a related space. They all knew GC AI. They all liked it. The adoption was huge. The barriers to adoption were low. It really dealt with what the GC does in their daily business. So that's how we got to the company. It was just great references.

Then we liked the team and the traction. I mean, no more complex than that. The company's growing really fast and barely able to spend the money they raised. So you're in it. It's profitable and growing very quickly. End of complex analysis.

No. How did you get comfortable with future financing partners, given everyone is out of the market, being an investor in Harvey or Legora, because they won't touch this?

Jason Lemkin

Yeah. So we do see the slightly different market, but the more important point is this company is wildly cash-efficient. They haven't spent their last round. We have a very elegant distribution strategy, so I don't think we're looking at a whole bunch of huge raises.

I think that was actually one of the key issues. Stepping back and making it less about the deal, one of the things we are thinking about as you're leaning in a little on price in some of these companies is that I want to at least pay attention to burn. What you don't want to be is a high-price, big-burn deal. So I find that very attractive.

Some of our recent deals—actually, 2 of the most recent 3 deals—have all been hovering around cash-flow well positive despite trying to invest more, because the organic demand has been such that you've been able to sell enough to, frankly, fail to invest ahead of revenue, right?

And I think if you do have a downturn, I think that's a nice place to be, right? A little more demand-led, a little more PLG, and a little less massively expensive.

And you weren't concerned about the kingmaking—

Jason Lemkin

I think that I do buy the idea of leaders, right? First of all, that there can be a category leader in the industry, and that's a big advantage. Going back to what we said earlier about durability of lead, I even do buy the fact that money can be important, especially in big-burn deals.

And in Silicon Valley, I do buy some kind of employer-level kingmaking. If you're seen to be a hot venture firm in the Valley, that matters. But step back to the wider US: I don't buy this idea that because X company got money from Y VC, the average corporate buyer cares all that much. They want to solve that problem.

So I'm not a believer in kingmaking being dispositive when you have great execution and great customer love. I think the customers decide. We're in a capitalist economy, and the definition of a capitalist economy is that the customers decide whom they choose to do business with. On average, customers are rational.

They're going to look at it. They're not going to say, “Oh, this software is crap, but Sequoia invested. I'll buy that.” They're going to say, “Which software do I like?” That's how capital is meant to work, in case you're unclear.

Okay, so Stripe does a tender at an all-time high of 41 bucks. Tom, love your thoughts.

Tomasz Tunguz

Yeah, I mean, we have a new public market. I think this is wild for me, right? I went back and looked at Microsoft. You needed something like $50 million in trailing revenue and 6 quarters of profitability to go public, right? And the cost to take a company public was a couple million bucks.

To do a late-stage financing, I mean, what is the legal cost on, like, a Series D?

Yeah, like a million bucks?

Tomasz Tunguz

Probably less on a D, but I think once you get into the employee selling, it gets a lot higher because you have a lot more transaction costs.

Let's call it a million, okay? What is the average cost to take a company public in the US? According to, I think it's KPMG—what, just the transaction costs?

Tom Loverro

Well, it's 6% to 7% of the raise, and the raises are now $200 million or $300 million, so yeah.

It's $25 million to $30 million in transaction costs. And so there's just no—why in the world would you pay that amount of money to raise a round of capital? Why? It's like getting a $1 million mortgage and having to pay $150,000 in legal fees.

The only reason you would, Thomas, is the point you made earlier: if the capital you get is cheaper than the capital you get privately. And as you pointed out, in fact, it's not.

Tomasz Tunguz

No, because now there's an illiquidity premium, right? There used to be—I remember when I joined the venture business, I was taught about the illiquidity discount. Private companies should trade at a discount relative to public.

You were always taught it was 20% to 30% to public multiples. That's the discount it should be, right? It should be for late-stage—

Tom Loverro

Right? And now there is an access premium. Harry mentioned this—

And so what is the—have we completely inverted? Is the access premium now 20% to 30% above?

Tom Loverro

It probably is. So, from a company's perspective, it's a cheaper cost of capital with a lower transaction cost. Why wouldn't I do that?

And then the ongoing service of that financing round is significantly less burdensome to the business because of quarterly earnings and all that kind of stuff. So you really only have to go public if you need to raise a quantum of capital that is so massive that the private markets cannot support it in some form or another.

Do you think that even is a blocker? Why would you not be able to raise billions privately? OpenAI is proving that you can.

Jason Lemkin

I guess you're right. I guess they could raise in the private markets.

And we have a liquidity mechanism now where you can trade in and out—not quite as efficiently, but still pretty efficiently. Right, and it's a form of regulatory arbitrage, if you think about it that way.

So the reason you would actually go public maybe is, bluntly, because you need dumb retail investors to supply you with cash. Not—I'm—

Jason Lemkin

That's the capital market of last resort at work.

Tom Loverro

No, I just disagree with that. I think I love the access premium thing. I think there's a small number of companies who, even at super-scale, have this desirability and cachet such that they can continue to raise in the private markets.

I think Stripe's a good example of that. Obviously, I don't think it's true for most companies. Take Navan. They just went public. Or ServiceTitan went public this year—or maybe late last year, right? Great cloud companies, but they're not going to raise 10 more private rounds because it's not wildly sexy.

It's just that they're both perfectly good businesses, right? So they didn't have access to this. I love the expression “access privilege”—the access premium to private capital. They couldn't get it done. You wouldn't be able to do two $300 million employee liquidity events for a company like that. It's just not desirable enough.

Bring it back to Jason's comment: your dentist doesn't get excited about being in ServiceTitan. So ultimately, they had to go public because that was the lowest cost of capital available to them, right? And that is going to be true for most companies. There will be this small number of high-taste, high-premium, Silicon Valley-beloved companies that can push it an awful lot longer.

Anyway, the only time Stripe will go public, and we said this on a call before, is when the capital available in the private markets is too unattractive.

Okay, but let me push back. Let me make the case for why I disagree with that. I don't know if I believe this, but let me straw-man it for a second: retail has had no access to venture for the last 15 years. It's been in technology, basically, where you want to be.

And so now, with upcoming changes in regulation, I can take my 401(k), put it into an ETF. The ETF goes into a fund of funds, and the fund of funds invests in a venture capital group, right? As a result of that flood of retail capital, those dollars need to go someplace.

They'll probably end up going into the businesses that, you're right, are not the top—the Pareto-optimal 80% of secondary dollars where the market is effectively liquid—but those retail dollars are effectively going there, and they're still probably cheaper than the public-market dollars.

That's a fair counter, and it's true. Provided capital keeps coming in because it perceives the returns to be high, more and more people will be able to stay private. Again, the reason that capitalism has bankruptcy, downturns, pain and suffering, and wipeouts is to stop the extrapolation to infinity. Until that happens, it's not going to stop.

You're exactly right. If returns go monotonically up for another 5 years, eventually more and more money will come in, and all it will ensure is that when they do, in fact, go down, they'll just go down further.

Do you think that the supply of cash is dependent on the returns? I was with Hemant from General Catalyst. I was with one of the great investors from Coatue, and they were saying the opening of retail is the next frontier of the supply of our business.

Do you think the opening of retail is predicated on great returns, actually, or are we just going to see it open over the next few years regardless?

I think in the end, when people lose money, they figure it out. They may take longer. They may be last to the party. In the end, the only thing that matters is returns. The only question is: how long does “in the end” take?

We're in an industry that has very long reaction cycles. You put in the money, you don't get a signal for 5 years, and you don't figure it out for 7. So I think the runway over which things can continue is very long.

But we can see the opening of retail much quicker than the runway happening. We've got Coatue with $3 billion now in retail funds, and we're seeing GC be very aggressive in opening up retail funds. That could come in the next 24 to 36 months, whereas that evolution of poor returns could be a 5- to 7-year lag.

You mean there's a mismatch between assets and liabilities? How many times have we learned this lesson? Look at Blackstone's real estate investment trust. They had huge retail inflows—I think $21 billion, a huge flood—and then all kinds of redemption issues associated with that.

So I agree with you, Harry. I think there's a tsunami of retail capital that's coming into venture, which is another reason to believe that the asset class, broadly defined, will hit $500 billion before the end of the decade.

And because they were liquid assets, they're not marked to market very often, right? The hottest ones, sure, it sounds like they're marked to market every 4 months. But the ones with 2021 marks on the unicorns, they won't be marked to market for 12 to 18 months, maybe longer.

Tom Loverro, should we do a $10 billion retail growth fund? Let's do it. [Laughter.]

Tom Loverro

Only fees. Only fees required. [Laughter.] Only fees required on this carry.

You guys keep the carry. We want you to make money. We'll just take 5% a year in fees.

Tom Loverro

Yeah, we just want finders' fees.

That's enough. We want you to capture all the upside.

I will say one hard-nosed thing. This is all great until you've had to go in a room, look people in the eye, and say you've lost them money, right?

When I did my own business when I was 21, it didn't work out. At 26, I had to shut it down, and I had to go into a room and say to people, “All your money's gone.” All these folks—we're talking all this great game—but there'll be a miserable part of this when you've taken these big funds, it was fun, you put all the money out, and then you realize you've locked in a whole bunch of retail investors to a subpar return for a decade. That will not be fun. Just remember that. Hold that thought for 5 years from now.

I'm not going to let you read the kids a bedtime story. Thanks for ruining that party, Grandpa. [Laughter.]

Tom Loverro

Yeah, hell. We were talking about 5% fees on $10 billion, and you come in and throw water on the fire.

You have to have an annual meeting for 10 years and explain to them why you've made a ton of money and they've lost.

Ah, that's why Jason doesn't have an AGM. You don't do that meeting. [Laughter.]

You've got to do the right thing.

Right, team, before we do a quick firefight, are there any final topics that we need to discuss that I've missed? Just one, since we have Tom here. I don't want to go over, but I just wrote it up today on SaaStr.

We're not ending the year with a great IPO market. When we started this show—30-odd shows together—IPOs were just coming back, and it looked like 2025 would be a pretty good year. In some senses, it's a good year, right? But we're well off our peaks, and the number of deals is not what we thought.

StubHub is a mess, right? We have some deals that are a mess. Navan's a mess, even though it's a great company. We're ending the year with an IPO whimper. It's kind of a bummer, despite Cursor hitting $30 billion in 22 months. It's kind of a bummer.

Tom Loverro

Yeah, I think the lens may be outdated. What I mean by that is that secondaries have exploded—absolutely exploded. In private equity, the total fraction of dollars in secondaries is a fraction of the asset class, about 25% historically. Venture has historically been about 2% to 3%; now we're at about 10% to 12%.

So liquidity dollars—another way of defining it is the total value of liquidity dollars, irrespective of liquidation channel—M&A, IPO, secondary—that's the statistic I want to see. I bet that we're up meaningfully on it because, just to the conversation that we've had, nobody wants to go public. Why would you go public?

So yes, IPOs will remain a very slow way, and probably a decreasing share of total count and dollars, except when OpenAI goes public. They will likely remain the least attractive liquidation option.

You really believe that, or are we just deferring these IPOs? You believe they'll never come for the top 50 names? They'll literally never go public?

Tom Loverro

I mean, why? You may exhaust the capital if you fall a little bit out of the top 30—just a degree out of the top 30, right?

Right. But Goldman bought Industry Ventures, the leading secondary fund, and paid the highest multiple, I think, ever for an asset manager. Why? Because a lot of retail dollars are coming, and they need to go into the private asset class. What's the best way of doing it? Secondaries.

I think there will be a mid-market secondaries market—not for names 1 through 20, but names 21 through 200. You made the point, Jason: 900 unicorns. They're never going public. But people will need liquidity in some form or another.

Tom Loverro

There's no liquidity for them, my friend.

Right, but there's some market-clearing price for that secondary.

I think you're right on that part, Tom. I disagree on the IPOs, but I think you're right. The 900 unicorns have to go to someone north of zero and south of $2 trillion. Somewhere between those 2 numbers, there's a buck to be made.

And you're right, someone's going to have to deal with the problem of cleaning up 900 companies and maybe turning them into 30 great companies—merged up or acquired, right?

Tom Loverro

That's a buyout business. There's a buyout business.

Some kind of restructuring business. I'm not sure I agree with you, though, on the IPOs. I think, in the end, the big exits will IPO, and we're in the business of the big exits. I don't believe that will change long term.

Look, you're right. The top 30 names prove me a liar today, but I think over the medium term the IPO window has to reopen for the math to work overall, and it just has to become more relatively attractive. You are right that the direct cost to the company of an IPO is higher than the direct cost to the company of a private round.

But if you look at it from a systems perspective, private capital has a 2-and-20 fee drag, and the public markets have almost a 60-basis-point fee drag. So from a societal perspective, there's no doubt in my mind that assets being managed privately have a far higher aggregate cost between the cost to the issuer and the cost to the investor than public assets.

So it’s just an inefficient method.

I think that changes. I think the fee structure changes on these extremely late retail products.

Then you could be right.

Yeah. Then I think fees drop to—I mean, look at SPV fees. They’re not 2 and 20.

That’s fair. The fees are significantly less.

I don’t know how to break it to those late-stage guys: the good news is your business is going to double, and the bad news is you’re working for 1 and 15.

Tom Loverro

Right. So I was looking at PE funds, right? You can look at the publicly traded ones, and you can see that the average fee load is something like 65 to 75 bps. At some point, you’ll see late-stage funds and venture capital have to approach that because they need to be competitive. Then I think the math can work.

But I don’t know. We’re all just pontificating.

Let’s delete that. We don’t want to talk about a reduction in fees. Come on, dude. We just said about a $10 billion fund at 65 bps. Come on. You think Jason’s getting out of bed for—

Jason Lemkin

A 10% SPV?

Tom Loverro

Remember, they’re actually making their nut. Even 65 bps is plenty of money to monitor one deal. They’ll be fine.

Jason Lemkin

There are no VCs.

Rory O'Driscoll

Dude, dude, Jason used to buy a place in Yellowstone. 65 bps ain’t it. Come on. We’ve heard about it.

Jason Lemkin

Country club material goods, starting to shed them.

But, Tom, do you think there’ll be a perpetual secondary market—an infinite secondary market—for top names?

Tomasz Tunguz

Because that would be very disruptive. We don’t know; we can’t prove that yet, right? But that would be utterly disruptive to venture as we know it if secondaries go forever. It feels like it’s true of SpaceX at least, right? No one’s expecting an IPO there ever, are they?

No.

Tomasz Tunguz

So I think PE works. You buy and hold for 3 to 5 years, then you package it up for the next person in the value chain. That’s how it works. I buy a $10 million EBITDA company, get it to $25 million as a result of acquisition and operations, hold it for 3 to 5 years, and sell it to the next guy.

I think venture moves in this direction, except for a handful of very, very large funds. If that’s true, then venture failed, because if you look at the top 10 companies by market cap in the U.S., nine of them are venture-backed. Those companies don’t get PE packaged around them.

PE makes a lot of money moving mid-market software up and down the value chain, where nothing is amazing but everything is good. We’re in the business of lots of things being utterly crap, some things being okay, but a few things being amazing—and the amazing moves everything else.

So, Rory, you find your nth fund-returner, you find your nth decacorn—I don’t know how many you have, but I’m sure you have many—and you know that it will take 15 to 20 years to get to liquidity. What you do is decide, “You know what? I’ll sell a quarter of the position three rounds later,” and then I’ll sell a little bit more in the next round, and a little bit more in the next round. I’ll dollar-cost-average my way out of this business.

It may not look exactly like PE because it’s not a full ownership sale.

That’s fair. But no, I do buy that. It’s not a PE sale. What you’re saying is, in this pretend public market that’s still private, I act exactly as I would have in the public markets. I just do it at a different transaction cost, with a different set of buyers. Yes, I buy it.

Tom Loverro

That’s right. I think that’s what’s happening. I think that’s exactly what’s happening. Unless the cost to go public and the premium that the public market is willing to pay change, the trend is inexorable, and the number of publicly traded companies will dwindle as PE picks them off.

I think in 2022 I calculated that PE had taken private 12% of all publicly traded software companies in a year.

Tomasz Tunguz

That was 2022. Yes, they hoovered it up.

Yeah. If that continues to be the case and we only have 8 IPOs, the number of publicly traded software companies is a dying breed.

So IPOs will be for the A-minus. They’ll be for names 50 through 150: very, very good companies, $500 million, growing 50%, but that can’t do quarterly tender offers of billions a year. It’ll be for the B tier.

Tomasz Tunguz

Well, it kind of depends on how big the retail flow is in the secondary market. It may be for companies ranked 200 to 500.

That’d be thumbs up.

Tomasz Tunguz

But there are a lot of pieces coming into place where the probability is increasing. I do agree with that. Every part of the trend is in your favor to prove you right in this assertion.

I think the unknown is how people will respond to a significant down market, which we haven’t seen meaningfully since 2008 and 2009, and in tech, really not since 2000 to 2002. The two things we need to see are a meaningful down market and what happens when you’re not able to trade the stocks because there’s no liquidity in private markets. We’ll see how that impacts the trend.

Until then, I think you’re right. I think the trend is clearly going this way.

Okay, team, we’re going to do a quick fire. He loves his Kalshi, Tom. It’s a pain in the butt, but you’ve got to deal with it.

Tomasz Tunguz

No, I love Kalshi. It’s awesome. It’s another new stock market.

Yeah. Thank you. Optimism. Optimism. Rory, see that? We love Kalshi. Thank you.

Would you rather invest in Cognition at $12 billion or Cursor at $29 billion?

Tomasz Tunguz

Cursor.

Rory?

Rory O'Driscoll

Cursor.

Jason?

Jason Lemkin

Yeah. I don’t mean to make—I usually go with the cheap one, but the numbers are just godstopping with Cursor. You’ve got to go with it.

All right. Harvey at $8 billion or Legora at $2 billion?

Tomasz Tunguz

I’ll go with Lovable. Knowing very little about the business, it’s just the entry price.

Rory O'Driscoll

I’m seconding it.

Jason Lemkin

I’m not. Listen, I’m only so smart. I don’t see the $30 billion exit in the category yet, but it may be ignorance. I believe in AIGC. I believe in that model. I met her at the seed round. I think it’s a great investment—Rory did—but I don’t see the $30 billion exit to justify Harvey yet.

It may be my ignorance. If I had the numbers in front of me, I might say I’d do it at $12 billion, but I’ve got to go with Lovable just for the math. I’m backing Tom on this one.

Rory O'Driscoll

Oh my God. We’re in sync again. Entry price counts on this one.

Funny, because sometimes it’s an interesting point. You’re right: we didn’t say entry price counts on Cursor. Entry price counts when TAM is unclear. Winning is the only thing that counts when TAM is huge. I think our 2 choices have been rational.

Give me a quarter for when OpenAI will go public.

Tomasz Tunguz

That’s not on the list.

Well, think on your feet.

Tomasz Tunguz

Q3 2026.

Tom, that’s brilliant. There you go, Rory.

Rory O'Driscoll

Late as—yeah, I mean, Q3 or Q4 2026 is stated. Next year hasn’t started yet; it’s already the end of the year. You’d want to be leaning into 2027. That was a very good call.

Yeah. Sorry, we’re wildly in sync again.

Tomasz Tunguz

I think that’s a good idea. I think Sam will come up with so much alternative financing that it’ll slip into mid-2027. I think the straw man today would be Q3 2026, but there’ll be so many other sources of financing. Maybe the government will guarantee it. Who knows who will guarantee the money, but I think that’s going to be the straw man. It’ll get pushed to 2027.

Jason Lemkin

To be fair, we do now know from Intel that the price of a government guarantee is 10% of the fully diluted common stock. So for $50 billion, I’ll gladly guarantee OpenAI myself.

If you can guarantee infinite compute, it might be a good deal.

Jason Lemkin

It might be a good deal.

It’s not like Sam’s seen a lot of dilution.

Tomasz Tunguz

I mean, if I were running OpenAI, I would not be dilution-sensitive if I had no shares.

Rory O'Driscoll

No, exactly. I would be growth-sensitive. I would raise as much money as possible. If I either had full anti-dilution or no shares, I would raise everything.

Jason Lemkin

Funny thing you should say that, because we’ve talked a bit about this in the past and I had to say it at the time, but there’s always something terrifying about someone who’s in charge of a company who’s just not money-motivated or incentive-motivated.

You’re right; it is kind of bizarre. I always have this reassuring feeling when I realize my CEOs are motivated by dilution and money, because then you know where the buttons are. It must be weird to be on a board with someone where you’re like, “What are your buttons?” You’re right: they’re not dilution. It’s therefore world domination, and that’s just kind of weird.

I had 1 CEO at the beginning of my career that I worked with. He had negotiated full dilution protection as CEO through the IPO. He was re-upped in every single grant, every single thing. He was guaranteed his 7% through the IPO.

Tomasz Tunguz

Oh wow. He got—

Jason Lemkin

He was a good guy, but it did actually change a lot of motivations.

Tomasz Tunguz

Yeah.

Jason Lemkin

He was an outside CEO who came into a clusterfuck. That was his condition. He said, “I don’t know how much capital this is going to take to fix. This is not Cursor. It’s a real business, but I’m not going to take that risk if you want me, because I can’t predict what it’s going to take to right the ship.”

He did right the ship to take the company public. It'll be nameless.

Who is this?

It was a while ago. It's nameless, but it did create a different set of incentives.

Yeah. Anyway, listen, team, I'm excited for us to be partners in the growth fund. It's going to be a very profitable journey that we have together. A transition from our normal early stage. Tom, you're going to have to let the Theory LPs know about that slight strategy shift. On one podcast, I know we said we were artisans, but we decided that volume was the way to go.

Yeah.

Yeah. It's just so hard. Jason told me seed was for suckers, and I was like, “Okay.”

Yeah. We're making T-shirts, by the way. We've got T-shirts being made. Jason's face: “Seed is for suckers.” It's brilliant.

Yeah.

It's great. You don't have to go to board meetings. You don't have to add any value. You just write the check and send some tweets.