Kimi 威胁 Frontier Labs？美国是否应禁用中国开源模型，以及 Stripe 收购 PayPal

Is the open-weight, low-cost LLM business a good business? And if it's a good business, why can't some red-blooded American company step up and give OpenAI and Anthropic a run for the money?

To kick us off, China ships 2 near-frontier open-weight models in a week, with Kimi absolutely crushing it. Next, Databricks is raising $3 billion—I love this—a Series M at a $188 billion valuation. And then, on top of that, we have Ramp releasing an OpenRouter competitor just as OpenRouter is supposedly about to get bought.

1. China's New Open Weight Models — Should the US Be Worried?

Ready to go, guys. I'm looking forward to this. There has been a lot, as always, going down. I remember when news cycles were so much shorter. I don't know if you remember this, but a $100 million round would be the thing for a week. Now, days go by and you're like, “Wow, we're forgetting the Stripe and PayPal deal, which we'll get to, because it's mega.”

But I'm going to start with the 2 near-frontier open-weight models that we saw in the last 7 days from China. One of them is Kimi, which has received a lot of attention and press, and the other is Qwen from Alibaba. How significant were the 2 model announcements that we saw today, and what should we take from their seemingly catching up—or coming close to—the frontier models we have in the West?

Guest

The quest for equivalent models at a cheaper price is just going to keep going up.

At some point, the people spending $1 trillion a year are going to want some apps to pay for all this. If you're growing 10x year on year and you have any kind of positive and improving gross margin, it just covers all the nut. Growth for the last 2 or 3 years has been a very attractive place to make money.

I mean, an eval is just an eval. So let's not take a bunch of folks on X who had someone in their engineering department look at some evals and write a tweet for them, okay? We're not saying something is similar in performance, maybe, but let's prove it in the field.

Having said that, we can't even sign up as new consumers for Kimi because it's blocked. They have so much demand since this happened, right? Demand is literally—I don't know whether it's geometric or exponential—but it's so high. We need to come back to this next week, when it opens up and I can use it on the consumer side even better.

There's a lot going on. There's a lot on politics, and it's an aha moment and a wake-up moment. On the other hand, it's not new. It's not new, right? If you look at OpenRouter data, half the traffic is through China-created models. Even the term “China models” is confusing, right? They may well be hosted in the U.S., right? And when they have open weights, they may be, for all intents and purposes, truly open-source models hosted in the U.S.

But it's not new. It's just going to accelerate this, and that's why you see the stress. It's accelerating. That 50%, instead of being niche or for tech-forward folks or venture-backed folks, could be everybody in a year. That is material. Just being in the zone and even materially cheaper, it's going to get more and more attention. We've always had cheaper, pretty good solutions from other vendors. It's not new.

Guest 2

I totally agree with that, actually, Jason. Harry, you kind of led with, “What do these new models mean?” I think Jason's take is exactly right. It's exactly what you'd expect.

It turns out there are 5 main Chinese LLM companies and a bunch of followers. Aggressively funded companies with smart engineers are just going to keep cranking through and building new models. They're not state-of-the-art compared to the frontier models, but they're 6–9 months behind, depending on how you measure it. So, actually, no new news about that.

But Jason's right: there is quite a lot of fun news about how parts of the U.S. responded to that. We had the small-p political response, so that's 1 dimension. The policy adviser for OpenAI, formerly from the Trump administration, made some comments on Twitter, leading to a wonderful firestorm that we'll absolutely talk about. That's 1 thread.

Another thread is what these models start to reveal about the economics of a model company. We lump all these models together, right? But let's take 2 of the DeepSeek models that you can run on your PC, your Mac, or a laptop. Conversely, Kimi K3 is, I think, a 2.8-trillion-parameter model. It's a huge, honking thing, and you need myriads of GPUs just to run it.

So they're not, quote-unquote, the same thing. That's much more comparable in size and, therefore, in terms of compute capacity, to U.S. frontier models. There's a lot we can learn about that. I think we'll talk about the politics first and then, maybe oddly enough, talk about the inference implications and how that goes into the opportunity for fireworks.

There are a lot of downstream implications, but zooming out, nothing amazingly surprising in the news. After 3 years of competent execution along a pretty defined trend, we now have 3 years and 3 months of competent execution around a pretty defined trend.

If we dig into the small-p political element, how should we analyze that? We can talk about the tweet that you mentioned, which was—I can't remember his exact title—

Guest 2

Yeah, you go.

No, no. And then Emil Michael obviously latched on to it, and I'm trying to remember: is it Dean Ball's name?

Guest 2

It's Dean Ball. He's currently, I think, either the policy or communications director for OpenAI. He just started there 2 weeks ago. Before that, he was in the Trump administration, kind of on AI policy, and before that, he did a bunch of Hoover Institution-type stuff.

He set off a firestorm with the tweet, and then he did a little bit of, “Oh, I can't really post because I'm now at OpenAI.” Everyone was mean to me because I posted a bunch of stuff, and I was, frankly, a little naive.

There are 2 comments about the tweet. The first is that you're in a senior role at OpenAI. There was a hysterical tone to it, right? He used the words “AI communism,” and it was very over-exaggerated.

Secondly, when you start even hinting at significant regulatory changes that will massively benefit you, you've got to expect that everyone is going to say, “Dude, of course you're going to say that. You're on that side.”

If you make the expensive, closed-source product that sells for $10 or $20 and the Chinese are shipping something for $2, and you say, “Totally independently, just speaking as a common citizen, I think they should ban this shit,” you've got to expect a whole bunch of people to say, “Dude, you're not talking as a common citizen. You're talking as the provider of the company that will jack up our rates the minute the stuff gets banned.”

It was a little naive not to expect that level of blowback.

Do you think Washington should move to restrict access to these Chinese models? Or is Bill Gurley right in suggesting that we should let free markets do what free markets do best, and we should not?

2. OpenAI's "AI Communism" Tweet Causes a Firestorm

Guest

We'll see. At some level, first of all, I think that guy at OpenAI had been there for about 2 weeks, right? Two weeks.

Guest 2

As they say in the meme, Jason: 2 weeks so far.

Guest

Yeah, so far. Listen, I'm not a total expert. It's difficult for me to imagine the federal government ever using a China-built model at this point in the U.S., right? It's difficult. Anything adjacent to that, it's difficult to imagine.

Simply, throughout our whole history in tech, the ability of Chinese technology to penetrate many U.S. buyers has been limited, right? It has certainly been limited in telecom and other spaces.

The real question, stepping back for a minute, is: how limited is it going to be? How limited are we going to be? Because it's going to be limited. The availability of China-built models to penetrate the U.S. is going to be limited. The question is just how much.

Jesse Zhang had a Twitter article today or yesterday—I think today—and it was pretty good. I think people might have missed it because it's real data, which is what I like. He said, “Here's one of our most regulated companies.” We have highly regulated folks. Just our token use here has gone up what looks to be about 2.5x since January.

The reasons are really interesting. I mean, I've lived this myself. The reasons are having supervisor models track the agents so the agents don't make mistakes, and running multiple agents in parallel so they don't make mistakes. The more regulated you are, the less forgiving you are of an error in an agent.

So it's like 4 times the agentic use just to have multiple agents regulating agents. If it's already grown that much in the first half of the year, the quest for equivalent models at a cheaper price is just going to keep going up. It's just going to keep going up. We've always had cheaper, pretty good solutions from other vendors. It's not new.

What I like about that guy is that he knows how to hate, and one of his biggest hates for the last decade and a half has, of course, been Bill Gurley from his time at Uber. If Bill and Emil are on the same side saying, “Don’t ban these models,” then you’ve got to know that there’s got to be some truth in that. It’s got to make you think, because that’s an interesting lineup.

I actually just saw this as I came on. Look, this is the Trump administration, so things change every day, but a Politico leak today was basically saying some version of, “We’re not going to ban these things on any significant basis,” which, as Jason points out, is very different from saying the White House decision-support system will be run on Kimi, right? Even if it’s Kimi hosted in California, I think we can take it for granted that it won’t be.

Conversely, if you’re Decagon and you’re a startup doing inference on customer-support queries for a very boring consumer product, there is no reason why you should pay marquee prices when something 10x cheaper is available, and it would be horribly bad policy to ban that.

Guest

Well, look, the one thing I will say, just to add: Bill Gurley—another rich, grouchy billionaire—has probably got 30 IQ points on me. He’s seen it all, and even at his grouchiest, I learn something from him. I always learn from him.

Having said that, I don’t think you’re going to convince me there aren’t some data-export risks with China-based models. You’re just not going to convince me based on what I’ve done building all our agents. If you’re not going to convince me, I don’t think you’re going to convince 99% of the world that there isn’t some security-leakage issue.

It’s already scary how much of our data we put into these closed-source models in the U.S. It is scary. Here’s Elon saying, “Sam Altman,” every day to create distrust. We cannot understand what these models do. They’re connected to the internet. We cannot even have them read themselves.

I don’t think you’re going to convince most of us there isn’t a data-export risk. I think that’s going to lead to tighter restrictions than this “leave everything open so we can compete” argument for my portfolio companies’ benefit.

I think every CIO is being told right now, “Don’t worry. If you host it on-prem, you remove any security risks, and the backdoor that could potentially be there is removed.” Why would you not be reassured by the fact that on-prem would solve that problem? Why would you not be reassured?

Guest 2

Rory’s more of a historian here than I am. You can mock our regulatory bodies, but they’re here to answer those questions for us. Is it safe to drink that cup of coffee? The American Heart Association, I think, just said 6 cups are safe now, right? This week? Now I know. Now I’m cool, right? I was a little worried about my caffeine consumption.

No, seriously, I’m not sure they’re right. Who has said that my data is not being exported through the most complicated, borderline self-aware software of our lifetimes? Who can say that? I admit this can be triggering, but there is a history of data-export risk with Chinese products. These are companies that are arguably run by the PLA.

I’m not saying that definitively; I’m just saying my lifetime of experience says I’m not confident there isn’t a risk. The internet telling a CIO isn’t good enough, I don’t think, and if I were a CIO, it wouldn’t be good enough for me, especially if I thought my job was on the line. I wouldn’t want to take this risk as the CIO of some Fortune 500 or Global 2000 company unless everyone else were doing it. I don’t know, man.

Rory O’Driscoll

3. Why Can't the US Build a Competitive Open Weight Model?

I don’t think it’s triggering to say that there are IP risks in this. At the risk of being level-headed here, the data is very clear that technically important U.S. companies—Boeing, for example—suffer continual cyberattacks, many of which are attributable to sovereign-state actors, including China. It’s a thing, so we’re not being sensationalist or alarmist. It’d be naive not to put it on the table.

You’re right. 2 comments. Second, I’m thinking about Kimi. The problem with proving a negative is, can you know if you have one? Remember, these are open. I occasionally say “open source” incorrectly; they are open-weight, which means you can see the weights, but you don’t technically have the full definition of open source in the context of an LLM, which means seeing the underlying training data, which you don’t. But you have the open weights.

The question is, if the model is being run by a trusted U.S. inference company based at Fireworks, or some of those guys, what can it really do? Could it initiate tool use on the customer side, whereby the model sends a command back to the customer to exfiltrate their data?

You can imagine being able to use these models fairly comfortably and being fairly certain that you have blocked access and that this can’t happen. I think you could, in logic, satisfy a technologist that the risk is not there. Whether you can satisfy a politician or someone who’s just afraid of what they don’t know is, Jason, to your point, another question.

I think you’re right. You’ve seen things like Huawei being effectively prevented from selling to any cellular networks in Europe and the U.S. because of this unprovable fear. So it’s not crazy that there will be some level of restrictions on the government side. I think a blanket ban would be massive overkill, to be very clear.

The interesting question it raises is this. Well, 2 questions. First of all, it’s also worth pointing out that while we’re talking about banning Chinese open-weight models, the Chinese administration is talking about preventing those companies from selling those models to the U.S. Just like we don’t let them buy NVIDIA, they’re not going to let us buy their open-weight models.

That’s kind of totally zany. We think they’re trying to sell it to us and we don’t want to buy it, and they think they shouldn’t be selling it to us because it’s so powerful. That’s kind of just weird in and of itself.

But I think the really interesting question here—and it gets to Thinking Machines—is whether the open-weight, low-cost LLM business is a good business. If it’s a good business, why can’t some red-blooded American company step up and give OpenAI and Anthropic a run for the money?

And, Jason, it’s the point you made: if this is a good business, where’s Grok? Where’s Gemini? Thinking Machines had an announcement last week. They announced a model. They didn’t position it as state-of-the-art frontier, but I think they made a comment on something that you can build upon. Inkling, I think it was called.

At one point, Meta looked like they were going to go down this route, right? Is there a business? How can you make money? It’s an interesting question. Can you make money as a maybe-not-completely-open-weight but low-cost U.S. provider of these models and be competitive with those guys?

The open-weight models in China are getting $50–$70 billion valuations. It’s not Anthropic, but I wouldn’t turn down a $50 billion outcome if someone could make a convincing case to me that a U.S. company could do this.

I think that’s one of the interesting questions here. Maybe the dirty little secret is that a lot of their advantage is distillation, which you can’t legally do if you’re U.S.-based. I do wonder, Jason, to your exact point, if there is a market for 80% cheaper intelligence.

That’s roughly what we’re looking at when you take into account the cost of inference: the difference between the bundled product that is a frontier model—IP plus inference—and an open-weight model where you dissociate the IP from the inference cost. If you’re looking at an 80% cheaper opportunity and there’s mass demand for that, when is someone going to try and fill that demand in the U.S.? I just don’t know.

I’ve asked so many people why we don’t have leading open models in the U.S. No one’s actually given me an answer. We’re still waiting for models from Reflection AI, which I think is one of the hopes that we have.

I was offered Kimi today, by the way, Rory, at $20 billion. It fell into my inbox: “I have an SPV for you. We do Kimi at $20 billion. We’re oversubscribed, but we’ll make room for $5 million for Harry.”

Rory O’Driscoll

Yeah, that’s because we say such nice things about them. Thank you for your check.

But no, look, we’re going to gloss by it, and I don’t have an answer, but it’s a huge freaking question, right? There’s this new category, too, called LLM intelligence. 2 companies exist in it as premium products. Their combined market cap is $2 trillion. The combined revenue at this point is probably $100 billion, plus or minus.

There are 4 or 5 other companies in the U.S. that are capable of, and have proven their ability to, build something roughly comparable. None of them are taking advantage of this. There are 5 Chinese companies that have proven their ability to build something roughly comparable, and they’re cranking night and day to take advantage of it.

Where are you, Google? Where are you—where are you, Reflection, as you say? Where are you, Thinking Machines? Where are you, Llama? I mean, the fact that there are 4 or 5 potential competitors…

It’s just fascinating. Or are you asking them to dance?

4. Open Router in Talks to Sell

Guest 3

I’m asking them to ship a story, and I’m kind of throwing it in here as a wedge.

But when we talk about all the different models that we have on offer, one of the big gossip stories or breakouts this week, in terms of news stories, was the information suggesting that OpenRouter is in talks to be bought by several different acquirers. And then, on top of that, we have Ramp introducing their router, routing model-provider product.

How is it a great time for OpenRouter to sell?

Guest 3

I think them leaking the story was very savvy.

Why is it a great time for them to sell, Jason?

Guest 3

Because the market’s in flux. Everyone’s figured out they need this. OpenRouter, like a lot of folks, was early and benefited from it, and deserves it.

This is a repeat founding team that saw there would be value to having a fairly heterogeneous mix of models, which, when we started this pod, probably made no sense at some level. It probably seemed too nerdy and too niche, and too cool-cat developer: who’s going to need it? Sure, it’s cool, but guys like Rory and me are going to stick to the big guns, right?

Everything broke well for them, but it’s still a niche product that more and more people are going to build variants of themselves. Is this the plumbing they will pick? If you’re on a lot of platforms, if you’re on adjacent platforms, if you’re using Databricks Gateway, they have their own harness. They’ll figure this out for you.

My point is, it’s something that’s going to become embedded in so many vendors. If I could sell for a lofty multiple of my last round, I might check out at $5 or $6 billion. It’s just because you’ve achieved a certain amount of victory in a market that’s going through radical change and becoming part of everything. I might take the offer.

Jason is answering the question: why is it a good time to sell? Frankly, as you yourself, Jason, have said—and I say it too—the private-market liquidity window opens so rarely that it’s always a good idea to pay attention when it does. So, I think it’s easy to understand that side of it.

I actually think the interesting side of the discussion is the other side: why would someone want to buy? When I saw that article, I thought, “Yeah, that makes sense.” If you think about the last conversation we had, what’s in the zeitgeist right now is this whole idea of: can I escape model dependency, manage my costs, and have a whole load of alternatives easily available to me?

If I’m someone who makes my money as a hyperscaler hosting models, especially someone that doesn’t have just 1 in-house model they’re pushing, like Gemini—if I’m Amazon in particular, who’s made a business of saying, “I’m going to support all the models”—maybe even Microsoft now that the divorce is coming true with OpenAI, this would be a great product to own if I were a cloud hyperscaler.

So, I will admit, I had that moment of—I wish I was in that. That’s always how you know what a venture guy really thinks. It’s like, “Damn, I bet they could get a good offer right now.” It’s just interesting. So, yeah, I’m with you. I think it’s an interesting time to sell and an interesting time to buy.

Rory O’Driscoll

My first comment is, whenever you get an offer, I always do the same thing. I say to the founders: one, the window’s open. It doesn’t open often. We should take it seriously. Two, I’m going to support you whatever you want to do. Three, if there are concerns that you have that you’ve been sitting on and not telling me, now would be a good time to share them so we can make an informed decision.

Then go away and think about it. I have a whole process for how you talk to them about it. I don’t think you should pressure people into selling. I think your job is to give them whatever experience you have to bring to bear, and then they’ll make the decision.

It’s so funny, because founders agonize about this when they think about control. They think, “Oh, my God, these people are going to make a sale.” Even if, as the VCs, we have board control, 70% ownership, and a drag-along for the founders—Jason knows this—if the founders who are core to the business don’t want to sell, it’s not going to happen.

The first thing I tell the founders is, “It’s your decision,” which I think is very empowering, because it takes away the idea that people are going to make you do it. That’s the beauty of being private, unlike being public, where you don’t have that degree of freedom. We could come back to that later.

I would say to them, “Take this seriously. Do you think you can be worth 3x this in 3 or 4 years? Do you think that’s worth it?” But, yeah, I would definitely say, take a day out of your life and think about this long and hard.

Guest 3

The way Harry phrased his question was very VC-centric: “Would you sell at $6 billion?” I can’t do the accent. That—I hate the term “triggering”—that triggered me.

This is a very VC-centric way to think of it. I’ve got an asset. What am I sitting at in my last round? $1.8 billion. There’s going to be dilution. There’s time value of money. There’s IRR impact. Is $6 billion worth it for me?

For a founder, when you start to get into nosebleed offers in absolute terms, it has to be 10x to go for it. It’s not worth it for 3x. Let’s say I own 15% of OpenRouter. For the money, it’s not, “How much am I going to take home?” At $4 billion, I’m going to take home $600 million.

I think one of the founders is super rich, right? But put that aside. I’ve got $40,000 in the bank. I’m going to walk away with $600–700 million. A 3x return for you doesn’t matter. Maybe later in life it’s now 10x, and you’re building something truly generational. You kind of know as a founder when you’re on that path.

So, the VC thinks, “What should I do it for? 2.8x on my last round?” It’s the right way as a financier to think about it, don’t get me wrong, but it’s a terrible way for a founder to think about it, because there’s way too much risk for not enough money.

Going from $40,000 in the bank to $400 million versus $800 million is irrelevant if there’s risk. Sometimes there is a handful of work in those next 3 years. It is a handful of sweat, a handful of market change, a handful of people that quit and move on, and a handful of competitors that are looking pretty good and may pass you. 3x is not good enough, man. It’s got to be 10x.

I’m confused. What are you saying? Are you saying to sell at $6 billion, or are you saying no? Which is odd, because—

Guest 3

I’m saying, if it’s financial, sell at $6 billion. If the money is not enough, it’s got to be $60 billion—

Rory O’Driscoll

—to be worth the risk for most founders. It’s not enough gain.

I understand what he’s saying.

Guest 3

This isn’t clearing the cap stack. This is clearing my life stack.

Rory O’Driscoll

It was weird. I didn’t think you were going that direction, but as often happens with you, I listen to the whole thing and I’m like, “I get it.” I think what he’s saying is this, Harry: when you face that sell decision, you don’t not sell because you think you can make twice as much in a year, right? You just never know.

In the end, even though I didn’t think I’d agree with him, in the end I did. From a return perspective, don’t think incrementally. If you have a chance to sell a company for $6 billion and make $600 million, and you think you can run it another 3 years and get $1.2 billion, that risk-adjusted return might not be worth it if your current net worth is $40,000.

That’s actually good financial advice. In other words, if you turn down a big offer, you better be sure it can be way bigger. You better have high certainty and high bigness. I think that’s a fair comment.

Guest 3

Yeah.

If it’s 10x, this is my net advice. If you know in your heart and soul you are building a company 10x bigger than this, right or wrong—I don’t know—then go, then effing say no and go for it. Here are a few more shares. In fact, friends, let me reload you. But they only vest at 10x.

Rory O’Driscoll

It wasn’t what I thought, where I thought you were going, because I actually thought you were going to say, “Yeah, something that you also said: don’t sell just because financially you feel you should.” The other thing you’re not taking into account is that it depends on the person.

Some people just love running that company and, frankly, don’t want to sell. It’s their life’s work. I also think you have to respect that. Again, this gets back to there being no one answer: the founder decides.

I’ve known people who are like, “This is my first hit. I’m going to take it,” who read exactly that logic. “I am going to make $60 million. I don’t have $1 million, and maybe I can make $120 million in 4 years, but I’m taking the $60 million.” I’ve also known other people who, to a rounding error, have said, “This is what I want to do for the rest of my life. Why would I take that money? I’ve made $3 million or $4 million in the secondary. I’ve got my house. I’m done.”

Can I ask more about the thinking? We said it doesn’t matter—Ramp is doing its own, Databricks is doing its own. I just interviewed the founder of Fireworks, who announced their $17.5 billion valuation. Is there any value in this layer if it’s as commoditized as everyone says?

With my limited knowledge, as a user and as a customer, I would sell just because I think there’s some commodification at a minimum. Sometimes, if you’re early, you can gain a lot of traction in something that becomes somewhat commoditized. It’s just the way it goes.

The perfect outcome is to sell the moment it becomes commoditized, but before everyone fully realizes it. That’s when they’ll give you the money, but that’s before the value decreases rather than increases. My gut is that it might be now—that crossover moment. I think Rory kind of made a version of that point. It might be now.

The space gets commoditized. It doesn’t kill everybody, but it might maim you. [Laughter] It also makes you attractive to acquirers for a window, and then that window closes. The commodification window closes. That’s probably why Cursor wasn’t dumb to sell at $6 billion. [Snorts] I think we can agree that’s true.

Guest 3

I don’t love the commoditization description. I think it’s an overloaded term, but yes—

Well, it’s an included feature, and more and more folks will include your function—some version of your functionality—in their product, right?

5. Fireworks AI Raises at $17.5B

Guest 3

That’s exactly it, which segues to the next topic—

Which is Fireworks. Yeah, and inference in general. Take it away, Rory. I’m going to butcher whatever context you want to take it on.

Rory O’Driscoll

No, no, you do. No, you do first, because I’m just—

Are you sure? I’ll lay the framework, and then you can just destroy it. Steamroll away.

Fireworks, a leading inference provider, announced their latest round, which was a $1.5 billion round done by Index, Gavin Baker, NVIDIA, Lightspeed, and 20VC. Amazing firms. They’re incredible. Thank you very much. They’re really good. Lin is amazing.

They’re doing over $1 billion in ARR. They got there in 3.5 years, and they announced around 40 trillion tokens a day, up from 15. I mean, I think the story is inference.

Rory O’Driscoll

Yes, first of all, I agree. Inference is huge. It goes back, ironically, to the prior comment on open-weight models. This kind of standalone inference is a big business, right?

Inference is both something that’s done within the frontier-model companies, where they do their own inference, and something where people like Microsoft and Google provide the capex and the compute. But people like Fireworks, Baseten, fal, and Together all make their money offering a variety of these open-weight models to third-party developers and enterprises that want to use open-source models to do AI.

As I said, the 2 trends go together. They’re exploding because the open-source trend is exploding. So if you’re Baseten, if you’re fal, if you’re Fireworks, or if you’re Together, this is your market and your moment, right? This is how you access those models.

I can tell you one thing: going back to the discussion about open-weight models from China, it’s one thing to decide to use an open-weight model on Fireworks in the US. What you’re not going to do is use the API back to China, even if they’d let you, right? So this is a 1-to-1 linkage between the open-source, open-weight trend.

These are the companies that are benefiting massively from that trend, and it’s not the only route for inference. There are inference providers for US-based models, et cetera, et cetera. But the vast bulk of it is, “Oh my God, I’m sourcing Qwen. I’m sourcing Kimi. I’m hosting Qwen and Kimi. I want to use them in Cursor. I want someone to provide me some inference.”

These guys exist, and they have lots of customer skew at the high end. I believe companies like Cursor are probably big customers of all these guys, at least they were until they were acquired—and probably still are, right?

It’s a great business, candidly. I think it actually goes back to the point I made earlier. You can look at some businesses and say, “Oh my gosh, you have margin compression in your future,” because you’re buying your compute from the neoclouds and you’re offering this product. Are you going to be squeezed on margins?

Margins were probably low for a while, but now the beauty of it is that demand is massive. Whatever compute you have today, whatever compute you’ve already signed up for—and these guys generally sign up for compute from the neoclouds, although they’re starting to build their own—whatever compute you own now, you can charge way more.

That means what looked like a lowish gross-margin business has now probably become a very attractive business. So not only are they probably going to grow 5x to $1 billion, but they’re probably growing 5x to $1 billion with expanding gross margins.

Just to add some details there, Lin said specifically that they were at mid-30s in margins, and that would move up as they eat more of the stack. They do plan to move into the data-center layer themselves.

Rory O’Driscoll

Yeah, and that’s exactly where I thought they’d be. Good on them, right? Thirty percent. In other words, what they’re saying—and this is going to be an issue, and I agree with that sentence—it also means that the challenges I hinted at are there in the future, right?

Because what they’re saying is, if I’m buying data-center compute and then effectively selling data-center compute with a hosted LLM, at some point I’m going to want to own my own data-center assets to have more control of my destiny. That means vertically integrating downwards, which also means a ton more capex.

So these are going to become way more capex-intensive businesses. There is a risk of commodification here, even with massive complexity and massive capex.

Guest 3

There’s one thread of the Twitterati that has said for a while, “All this stuff’s interesting, but ultimately it’s the application layer that’s going to be the most interesting. It’s going to benefit from all this. Everything’s commodified,” right?

Absolutely. [Laughter]

6. The Application Layer Still Hasn't Arrived

Guest 3

Even the ones that look good in software, the numbers pale in comparison anyway, right? The absolute numbers pale. So I’m waiting for the era of the application layer in AI and making bets and seeing some good stuff, but I don’t believe it’s here yet. I actually don’t believe the application layer is here yet.

To put that in context, Lin said on the show that she expects to double by the end of the year.

Guest 3

Totally.

And that’s just a slice of the market. Listen, people have gone all in. When we started the show, it felt like vibe-coding applications run amok. Everyone thought you’d replace Salesforce. You even had a guest the other week who was—I forget—saying how great it was. He replaced Salesforce. Who cares, right?

It didn’t kill software, but where is the software renaissance? The revenue is there. We’ve talked about leaders, right? But it’s so trivial compared to the infrastructure. It’s so trivial. It’s almost a rounding error—the application layer.

Rory O’Driscoll

Just to dimension that, because I agree, Jason: look, I’m an app investor. It hasn’t happened. You look back and ask, “What’s going on here?” You’ve got companies like Fireworks doing $1 billion. There are very few app companies doing that.

Zooming out a million miles, my mental model is that I divide the AI world up into 3 buckets. There’s making AI—the infrastructure layer, right? You’re right, the spend there is $800–$900 billion a year. Then there are the 2 foundation-model companies themselves, and they’re doing plus or minus $100 billion a year, right?

Taking those guys out and rounding up every other apps company, you struggle to make $40 billion or $50 billion. You start with Cursor at $4 billion, because I think coding is an app. By the time you’re chucking in Harvey, you’re adding $200–$300 million. It’s amazing.

Guest 3

I mean, just the difference in spend. At some point, the people spending $1 trillion a year are going to want some apps to pay for all this, right? But right now, the volume has been front-loaded on the infrastructure side, and at some point, the revenue has to match it. But right now, infra has been the place to be.

There's probably more money being spent on training data for the foundation models—the Mercor, the Surge, and that—than pretty much any app company outside of Cursor. In fact, the sum of all the app companies outside of Cursor is probably less than the amount that Anthropic and OpenAI are spending on training data, which is just amazing.

That I can guarantee. When you look at Mercor, $2 billion in ARR—

Guest 3

Two billion for Mercor, Surge another billion—you get to $4 or $5 billion, and, you know—

Surge is $3 billion.

Guest 3

Yeah.

Handshake's $1 billion. I mean—

Guest 3

I mean, yeah, maybe if you start throwing in, on the other side, the consumer products like Higgsfield, you get to roughly the same place. But it's astonishing. The scale of the investment versus the scale of the apps at this point means that all the action is on the infra side for now.

If we bring this all together, we mentioned Fireworks at the start. Lin said in the show that the future would be every company having specialized models with their own data. We mentioned Harvey, who have been building their own models. Jason, I'm just intrigued: in the last week, you spent time labeling data and building your own model through that data. Any lessons or reflections from the last few days, labeling data and going through that process?

Guest 3

I've been building this AI-centric recruiting app just to recruit from the SaaStr community. It's been fun. I've learned a lot building it, right? Hopefully, it can ship in the next week or 2.

But to really get it great, it needed labeling to make its model—now, I'm going to put “model” in quotes. It uses Sonnet and Opus, so there are different definitions of “model.” It was good, but once I started labeling all of this, it got exponentially better, right?

We built our own little labeling tool. So you need your own micro-model, whether it is some sort of reasoning layer that you build on top of Claude, ChatGPT, Kimi, or Gemini. It's still your own model, even if it's not technically a model, because you have your own reasoning layer with your own rules and your own weights.

But you want more. If you have the resources, you want to go further than that, right? You want your big model as soon as you're at a certain amount of scale, and it's not cheap all-in, right? You are going to want to have your own model, like a Harvey or a Cursor.

Some version of this, I think, and the folks that are going to want to use just the generic models at any application level are just going to decline to prototypes, right? Prototypes and proofing.

Yeah. Or absolute state-of-the-art work on small parts of the overall task. But agreed.

Guest 3

Parts. Yeah. Little parts, right?

I mean, again, you're going to want to use the expensive tool for the expensive parts, right? And you're going to want to use the cheap tool for most of the parts, from the customized tool to your usage.

Guest 3

But, man, the outputs are literally an order of magnitude better once you do it. So everyone wants their own model. And I do think whether that always benefits Fireworks or not, it doesn't matter, as long as they pick up some of the bigger end, the bigger end that scales.

The generic models are great, but it is amazing how much better you can do than them for any specific workflow. You can do epically better.

Would that change your confidence in the data-labeling market? A lot of shade is thrown at it. As an investor in Mercor, I definitely see it. Does that change how you think about it?

Guest 3

I totally get it. Having a subject-matter expert go in and answer 20 questions about a disease or about a history—I mean, it's a lot of professors and teachers that they have there, right? That model, right?

The amount of power you can get in a domain by having a subject-matter expert answer just 20 or 30 questions—5 minutes, 10 minutes—the amount of power you can add versus the generic LLMs, which are a sea of mediocrity combined into one giant LLM.

Every mediocre history professor, every mediocre doctor that doesn't even know what caused your runny nose is in the LLM. But if you get the best people training on the best answers, it's a step function.

I'm less smart on the seeming low end of the model, right? This commodity thing that people made fun of Mercor—but I ain't making fun of it anymore. I tell you that much.

These models are a sea of mediocrity, all combined in a giant soup that gets better. These domain experts are so powerful in tuning your model to get the better output. So powerful.

I think the answer, though, is really a derivative of the big question, which is: your statement that companies are going to want their own model is probably true, right? And the real question is not that. The real question is: will that be additive to the rough trajectory of the foundation models as it's established today—in other words, coming at or close to $100 billion in combined revenue, growing nicely—or does it start to take away significantly?

To answer your specific question, if the foundation models continue to grow—and we just saw an article in The Information that, for all the training-data companies, the vast majority of their revenue comes from the foundation models—to which your correct response is, “No shit. Of course it does,” right? If that continues to grow and you have an additive market in enterprise, with all of these companies—JPMorgan building the JPMorgan model on top—then it's net expansive.

And net expansive is, by definition, good and reduces customer concentration. I think that's what people like are forecasting, right?

On the other hand, which is hard to contemplate today, if these enterprise models—if these open-weight models—really impacted the growth rate of Anthropic and OpenAI, then obviously, when your 80% customer slows down, it would have a significant impact on your growth rate, right?

But if it's any consolation, Harry, if that happens, worrying about your Mercor valuation will be the least thing people are worried about, because you'll see an implosion of much bigger market-cap entities, right? And that's frankly the billion-dollar question: can these 2 foundation models maintain their growth trajectory—which is starting to become profitable, at least in the case of Anthropic—in the face of all this open-weight competition, in the face of this pushback on costs, and basically this push for ROI?

If they can maintain this trajectory for even another 1 or 2 years, then everything's fine and everyone's fine. And right now, the data says they are. If you start to see a slowdown in those 2 ARR growth rates, then all bets are off, because the amount of commitments they've made, assuming that 10x growth rate continues, will mean that even if it slips to a 2x or 3x growth rate, there's going to be a mad scramble.

Bet's on a yes-or-no answer: will open-weight models impact that trajectory for Anthropic and OpenAI in the next 1 to 2 years?

Guest 3

Yes, Harry, it will impact it. It might impact it at 1% or 50%. What you're really saying—the question you're really trying to ask—is: does it produce a sustained reduction in the growth rate to below 100% within 1 or 2 years? Right?

And the answer to that question is I genuinely don't know. If I did, I'd be trading that stock. Because if you know the answer to that question, you know the answer to the entire direction of the U.S. stock market for the next 2 years, because all the hyperscaler RPO—all of it—is a function of the commitments they've gotten from the foundation model companies.

And, yeah, you can say if the open models, open-weight models explode, there will be demand for inference. And, yeah, you will have this kind of transition from, “Oh, I sold it to OpenAI, but I should have sold it to—I don't know—Cursor or Base10 or someone else,” and the capex will get repurposed. But it will be a big-ass dislocation, and I just genuinely don't know. I mean, it's the million-dollar question.

I think the tough—the really tough part—I mean, it's Captain Obvious, right?—is: can they afford for it not to? And what I mean is, look at what's happened with Fable this week. Fable went from, “You can't use it; it's not secure.” Then the government let you use it. Then, “Hey, we're going to turn it off except for variable usage on July 15th.” Now it can be 50% of your whole usage for the month.

Why did they change when they don't even have enough capacity to serve it? Competition, right? Competition, right? So if they price Fable at Sonnet rates, I think they'll own the market.

Guest 3

Yes.

I'm oversimplifying because you don't need Fable for anything. So the question is: can they afford to compete? And this is the stressor, right? Of course, they could have 17 variants of the model at 17 price points. That's not the issue.

The issue is that they have to pay to train these damn models, among other reasons. They have this high cost base, and they're subsidizing it with venture capital, right? Whether we call this venture capital or not—private capital.

And so, listen, you just cut the price of Fable 5 in half tomorrow. You don't need these Kimi shenanigans, but can they afford to? And if they can, over what schedule? The fact that you can use Fable for half your credits is pretty telling, right? They're pushing it as far as they can, but that's the limit today. They can afford to compete at 50%. [laughter]

We don't have time for it, because I do think we should spend at least half the show on stuff other than AI model companies, but I think Jason's insight is correct about price. If this was a software product with no gross cost of goods sold, that's what they would do. I mean, Microsoft—and this is one of the big things I've seen a bunch of articles on. Again, as you'd say, Jason, Captain Obvious, but it's worth emphasizing: in the great software wars of the last couple of decades, someone like Microsoft was able to use price ruthlessly because there was zero cost of goods sold.

They bundled the browser in with the operating system, bundled all of Office in together, and it didn't cost them anything. It just wiped everyone else out. But as you're pointing out here, there are real costs, even at the margin. Even after you've fully paid for your training costs, there are real physical costs to serve these models.

You've got to cover your nut. You've got to cover the marginal cost of serving the model—the cost of inference—which gets you to, I don't know, $2 or $3 in blended average tokens. Then you've got to recover the cost of the training, and you've got to recover it pretty damn quickly because it only lasts 12 to 24 months before it's obsolete.

On top of that, you want to make extraordinary profits because you're being valued at 20 times revenue. If you're valued at 20 times revenue, you better be like Microsoft, with 40% operating margins. When you look at all that, you're right. It's kind of back to the thing I said: you can squint at that and say, “Ooh, there's lots of things that could go wrong here.”

When you look at those fundamentals, and yet the thing that's saving you right now, if you're growing 10x year on year and you have any kind of positive and improving gross margin, it just covers all the nut, right? The minute that growth rate stops, Jason, if the only way you can keep that growth rate up is by lowering your price per token, then your gross margins start to deteriorate instead of continuing to improve. That in itself would be a different ball game. So you are right: price could solve it, but it would be a painful way to solve it.

Guest

Yeah, you have to start building your own chips and building your own everything—all the stuff you're trying to do to solve this problem. But I think it's just a pricing problem, right? I mean, there's a bunch of issues underneath, but I would argue they've already bundled it.

The consumer apps of Claude especially, but also ChatGPT, have bundled everything. I can get $10,000 worth of tokens for $200, and I can just do just about anything in it, right? It's just outside of the consumer, it's not massively bundled and subsidized, right?

You remember the old days in software, Jason, when you'd have to say, “I promise I'm only using this for personal use.” You remember that? In licensing, right? Well, if you're telling those nice Claude people that you're only using your personal subscription for personal use, they're going to find you, dude.

Guest

They're going to find you.

Well, yeah. It's just Claude is very good.

Guest

Yes. That's why they're going to find a way to charge for it.

Very good. Rory, you were like, “Are we going to get away from this AI stuff at some point?”

Yeah, yeah, yeah. So much. What were you hoping to talk about?

I mean, a vertical dentist company like—

Guest

Ben Affleck making $500 million for his—

Oh, that is AI. Sorry, sorry, sorry. That's an AI story.

Guest

Oh, it is, isn't it? $587 million. And, fun fact, his top 3 movie salaries didn't add up to $60 million. So it's 10 times more than his 3 highest-grossing movies combined.

Wait, wait. Say that. It's higher than The Batman.

Guest

Yeah.

What are you talking about?

Guest

For his pay—how much he got. He got about $8 million.

The amount he made from it.

Guest

Yeah.

You know, context is so funny. I don't want to get distracted. We're like, “How much did Ben Affleck sell it for—$500-some-odd million to Netflix, right?” [laughter]

$587 million.

Guest

Our jaws drop, and we're arguing whether we should sell a portfolio company for $6 billion. Well, is it really worth our time, gentlemen? It's really only a 4x to the last round. On the last fund, it's not even a returner. I don't even know if I may be retained as a GP at the firm if this is as good as I can do. Oh my God, he sold the company for $500 million.

Rory, do you know what I find triggering? [laughter]

Look, get on him. I mean, no surprise. It turns out you can make more money with capitalism—technocapitalism—than acting. It turns out Bill Gates is richer than the most famous actor in the world, right? No surprise.

All right, Rory, I'm going to listen to you then. We're going to move away from this AI pure-play discussion, and we're going to move to some Irish twins: the Stripe and Advent deal to take PayPal private. Does that sound okay?

Good.

We have to talk about it.

I think the price clears all of that. I think it's super interesting in a lot of different ways. One is that they both process roughly $1.9 trillion and $1.8 trillion a year, right? Stripe is valued at around $150 billion, and what was the offer for PayPal? I looked at it this morning, but didn't—is it $50 billion? $35 billion? Hang on. It's about—

Guest

I thought it was $58 or $60 billion.

$50-something billion, right? So it's Stripe taking advantage of PayPal trading at less than 10 times profits and deciding to go for it here, right? In one sense, it's a ballsy move because you're taking on a lot of operational complexity. On the other hand, it's a chance to really transform and double your footprint because, as I say, the payment processing is roughly the same.

Revenue is tricky because Stripe reports revenue net, which is around $6 billion, plus or minus. PayPal reports gross, and I think it was about $30 billion. So it was trading at about 1.7 times revenue. If you look at that—$5 billion versus $30 billion—you're like, “Oh, it's 5x. PayPal's 5x bigger.”

It turns out that on a like-for-like basis, PayPal is still bigger, but it's about 1.5 times the size. It's still a company buying something 1.5 times its size for what looks like a third less because they're doing a joint deal with Advent, a private-equity firm, for a lot less of its market cap.

If they pull it off, they'll look back and go, “Wow, that was an amazing deal,” right? It also gives them incredible economics. It's a little like the Dell transaction. Obviously, the scary thing is that it takes your perfectly wonderful company that's running smoothly and is a desirable place to work—all the positives that we all know about Stripe, the smartest guys ever, killing it, nice place to work, good reputation—and they're going to have to do a lot of hard-nosed stuff to turn PayPal around.

There'll be a lot more pushing and shoving in the future because you're probably going to be looking at that place and saying, “We're going to get rid of a lot of people. We're going to rationalize a lot of stuff.” So it's a different muscle, but I give them credit for it. It's a big, ballsy play to double your market cap.

Guest

Yeah, the part that I struggle with a little bit is this: obviously, there's at least a decent synergy here, right? In a PowerPoint slide, there's a ton of synergy. Plus, you get Venmo; you get a lot of stuff. But to me, the thing that's always a head-scratcher is blending something that's growing at 7%.

Because no matter what you say or do, unless you can radically shove those products through your channel, your blended growth rate goes down. What's Stripe growing today? I don't know, 30%, 40%?

It's between 20% and 30%. So it's not that much bigger, Jason. That's why.

Guest

But, okay, hold on. Help me, Rory. You're better at math than me. If I take 30 and 7, that's 37, and divide by 2, I'm only growing at about 18% now. I've fallen below the Mendoza line of 20% growth at scale.

There's no such thing as a Mendoza line for growth at $5 billion and above, because you can get out, right?

Guest

But I found it stressful in M&A observations. Not quite at the scale, mind you, but it is stressful when it meaningfully decelerates you, right? It will meaningfully decelerate them in the short term. Even if I'm not sure how the accounting works, maybe they only have to recognize half of it because of this Advent thing, but they're going to have to recognize some of this revenue as a joint venture, right? So it's going to accelerate their growth.

Guest 3

It’s not stress-free for Stripe, plus you have the operational need. Even all the layoffs they’re going to do, that alone may not reaccelerate growth. We’ve certainly seen this at a handful of portfolio companies, right? That’s just the stressor for me.

I’ve learned over the years that when you have 1 messy codebase and another codebase, and you’re like, “How the hell are you going to combine these companies in different motions?” you figure, as crazy as it sounds, you actually figure that part out. The answer is, you don’t fix it. You fix it over 5 years, or you have an LLM lift. But the real answer is, you don’t fix a lot of these things that seem like you can’t rationalize between the organizations.

Everyone’s got 11 products spaghetti’d together. Even tech leaders have it, right? It’s just the nature of M&A. My guess is this is one where you have, frankly, 1 well-run company for the last decade and a half in Stripe, and you have another company that, ever since the PayPal Mafia walked out, has been just a revolving door of executives and is a real mess, and they’ve dissipated their opportunity.

So, yes, the interesting thing is, normally this is the kind of deal you do after you go public, because you have the market cap and you just price the deal. I was thinking, at first glance, it was probably a lot harder to do this as a private company because you can’t issue $50 billion of stock, right? So you have to look at debt; you have to do Advent.

On the other hand—and again, I wanted to read the detail; I didn’t get to it fully before this meeting—maybe they’re using Advent to almost keep it slightly off-balance-sheet for a period of time while they rationalize it, right? So I don’t know. It would be easier to consummate this deal and just be done as a public company, but obviously Stripe has chosen not to go public, at least yet. It may well be that, even though that makes it less easy to do, it may also have pushed them into this kind of contained strategy with Advent, right?

Yeah. Will this happen? Will it actually get done?

Guest 3

I think it happens.

Rory O’Driscoll

I think it does, too.

Guest 3

Let me just step back. Rory’s got even more experience than the 2 of us, but it’s just a dance. The board rejected it, right? And the fact that the board rejected it means to me that they’re going to accept it.

You reject it because no investment bank will tell you you’re allowed to make your highest offer up front. No investment bank will tell you that. You probably breach your fiduciary duty if you make your highest offer up front. You have to have another 5% or 10% to put into the deal. So it’s a dance. They’re going to accept it.

They’re just a bunch of mercenaries and a brand-new CEO who’s probably going to make 9 figures for 10 or 12 months of work. They’re going to reject it. By rejecting it, it means they’re going to accept it.

You know, I think Jason could well be right. I think I hinted at this: when you’re a private company—you remember we talked about the sale, Harry?—when you’re a private company, you can decide not to sell for any reason.

When you’re a public company, you know what the bankers are telling them right now is, “You’re right. The first thing you do is instantly reject, because you’ve got to look strong.” And then you’ve just hired the bankers, and they’re going to say to you—and the lawyers, in particular, are going to come in the room and they’re going to say to you—under Delaware law, you can only turn this down if you have a good-business-judgment belief that, on a standalone basis, you can do better than this offer in a reasonable period of time.

So, even as we speak, the PayPal team are building a 3-year model and a 5-year model, trying to prove that they’re going to be amazing and, therefore, this bid is too low, and they can be comfortable with the risk of turning it down.

But what’s going to happen is this: they’ll be able to make a model because they have smart people and the banks have smart people, and the NPV will be wonderful because the banks will make it that way. But the pushback will be, “Well, guys, if you were so fucking smart, why didn’t you fix it in the last 5 years?” Right?

And then you’re sitting there as a board member going, “Am I really sure that this guy can turn it around?” Do I believe that, if I got an extra 10% or 15%, I would say, risk-adjusted, I should take it?

As Jason pointed out, I don’t know the CEO from Adam, but he’s sitting there going, “Bird in the hand versus slogging at PayPal, being the third CEO in a row trying to turn this thing around.” At some point, if Stripe wants to own this thing, you kick a little more in and you probably will own this thing.

I think it’s hard to have the stomach unless you can see—maybe unless you could see evidence within the PayPal numbers that it is turning around already. That’s probably the only thing that could give the board the courage to say, “I’m just not doing this.”

In other words, there are probably 5 key internal metrics that matter: take rate, new merchants per quarter, usage of wallets, whatever it is. If those numbers are already starting to turn because the new CEO is doing an amazing job, then maybe the board can say, “I will extend that trend.”

I will say, “Hey, look, the last 2 quarters have been 10% better each quarter. If you extend that trend for 5 more quarters, it’s an amazing company. We’ll be worth twice as much. Let’s turn it down.”

If those trends are still flat and the new CEO’s plan might start working next quarter, then it’s really hard to say, as an independent board member—you’re getting 300 grand a year in RSUs—do you really want to be a hero here? Do you want to, as Jason said, say no, negotiate for 15%, discharge your fiduciary obligation, and take the money?

Guest 3

Yeah, I mean, the argument could certainly be that the stock price is depressed; they’re missing it, right? It’s down from its lows, and you probably could tie into the business judgment rule if you really believe it. But my guess is this is engineered.

They made a 28% premium offer. The average take-private like this is in the mid-30s. Now, average does not control any deal, but that is the perfect amount of back-and-forth: 28% to 35%. It’s already pre-scripted.

Rory O’Driscoll

Yes, it’s already pre-scripted.

Guest 3

And the bankers will charge a couple hundred million bucks for the deal. “How are we going to get from 28% to 35%?” “Well, we could just offer them 35%.” “We’ll never get there. We have to offer them a 28% premium to a public company’s stock so that we can land at 35%.”

They have to go shop it. And if there are any other offers, they would have gotten them in the last year. There are no other offers. Now, sometimes it materializes—Rory has been through this—but usually, if there’s another offer, the offer already sort of happened. There have already been discussions at the whatever media summit or whatever, and so there probably ain’t.

So it’s probably just a dance from 28% to 35%, and then it gets parked with Advent while they figure out antitrust and capital issues. So Stripe, finally—the Padawan finally becomes the Jedi. Stripe takes over PayPal. It’s just a matter of time, and it lands where it should have been.

And all the early PayPal guys that did the pre-IPO, along with Sam Altman’s 2%, they’re going to do pretty well in the end.

Rory O’Driscoll

Totally. Yes, they’re coming back through the back door.

Guest 3

Yeah, they’re getting the old gang back together. So, for listeners who may not know it, one of the very early Stripe rounds—I know Peter Thiel was an investor, and a number of the folks who were involved or connected with the PayPal Mafia back in 2001, before they sold to eBay, subsequently went on to be great investors, Peter Thiel most notably, and stuck early money into Stripe and now, 15 years later, are having the joy of buying PayPal back.

It’s probably a sweet moment if you’re one of those investors. The first time you move into the headquarters, you’ll probably say, “Can I come along?” You probably ring the Collisons and say, “Hey guys, if you’re doing the victory lap on the PayPal headquarters, can you include me on that trip?”

7. Databricks at $188B: Private Companies Acting Public

Now, Rory, I want to hand the ball over to you because you said no more AI, so I gave you no AI, and then you were like, “You missed topics.” What did I miss that you wanted to cover?

Rory O’Driscoll

Maybe the better comment is that there’s more to life than talking about OpenAI and Anthropic, because they’re only 2 of 2,000 interesting companies. On the other hand, as you would be the first to point out, cap-weighted—in other words, weighted by dollars—they’re $2 trillion of the $5 or $6 trillion of privately held market value. So, on a cap-weighted basis, we should be talking 30% to 40% of our time about OpenAI and Anthropic, boring as it is, if you’re trying to be representative of private tech.

8. Stripe & Advent Bid to Take PayPal Private

So I hear you, Harry. It’s hard not to, but I just don’t want to be totally boring. I mean, I thought that the other fun things—and the odd thing about your list of other companies to talk about is, in a weird kind of way, every single one of them is a company that’s being pulled by this trend.

You had Valar Atomics down there to talk about new technologies and nuclear. Then you had TSMC and ASML, and the truth is, all the dynamics for those 2 companies are about the insane demand for semiconductors, which is all about AI. So, when you actually get to trying to talk about something that’s not AI, I ain’t got shit.

Exactly. And then Databricks rockets to $188 billion. Why? To buy GPUs?

No, which gets back to my comment. The growth rate of those 2 foundation-model companies, as Jason has pointed out many times, is a thing upon which your 401(k), at an all-time high, is dependent, right?

But I did think one thing was interesting. Another one that I do find interesting is this emergent AI coding startup: $120 million in ARR, raised a $130 million Series C at a $1.5 billion post-money valuation on July 15. The thing that I find really interesting here is I’m seeing Series A rounds priced at $300 million to $500 million on $2 million to $5 million in revenue, but I’m finding Series B rounds at $100 million in revenue priced at $1 billion to $1.5 billion. It’s a 3x price increase for a 50x revenue increase.

I think it’s just a very interesting market analysis today of where a good insertion point is for investors. Oh, it’s true.

Rory O’Driscoll

And it’s risk-adjusted, always, now. We did Factory at the $1.5 billion round, and I think, yes, that was a worse deal than the $300 million round. But at the $300 million round, they had next to no customers and very little product-market fit. Well done to those investors—they saw what a lot of other people didn’t. But risk-adjusted, fuck, you’re only paying 4x for incredible PMF and 70 to 100 times revenue scaling.

I think on those numbers, you’re correct. The short answer is, would you prefer to pay $300 million for no revenues or $1.2 billion for a lot of revenues? Absolutely.

Well, look, I think, for what it’s worth, there obviously is, as we’ve talked about in the history of the show, real multiple compression, even in the hottest agentic companies at scale, right? There’s real revenue-multiple compression. There are plenty of folks compressing to 10x revenues, right, which is even far less than forward revenues. Maybe that’s an unhelpful comment.

I think the real pressure is it means anything below that growth stage—you better be a damn good picker, because it used to be, when Rory and I met, Series B, even into Series A, you actually didn’t have to be a good picker. You just had to be good at math and good at assessing a team. The picking wasn’t so hard as it looked; it was all the rest.

Now, at Series A, that gap means you better have seed-investor skills at the Series B, or the math’s going to be tough with that, right? There’s a lot of pressure on the picker. That’s just what I think it is below the growth stage. And so be it. That’s the game, right? But that’s how I think about it, and it’s hard. It’s harder. You don’t really want to be a picker. You want to be a pricer.

Again, going back to my point, risk-adjusted here: would you rather be doing a Series A with $2 million in revenue at a $300 million price, which is the going rate for a hot AI company at Series A, especially in the Valley, or would you rather stick money into Fireworks, which says they’re going to hit $2 billion by the end of this year at $17.5 billion? You’re paying less than 10x.

Rory O’Driscoll

It depends. I mean, Harry’s better at the math. It depends on fund size and other numbers, but you want to own the most you can of winners. You could argue at some point, I guess it doesn’t matter; it’s just putting in the absolute amount of money you can into the latest Anthropic round.

But for most of us without unlimited capital, if you can pick better earlier, you end up owning more. It does pay off. That extra 3x to 4x isn’t terrible. That extra 3x to 4x on the way to the billion-dollar round—it’s not terrible.

I think many people can’t pick, and I think we are here.

Rory O’Driscoll

They can’t. It’s hard. Picking is more complicated than it sounds, right? Picking sounds like everyone’s waiting outside your office for 4 hours in the lobby, like at a doctor’s office, and you get to pick, like it’s 2006.

But it is true that the change—that the biggest brands will pay the highest price in many cases—makes that in-between round tough, right? At least the growth round is sort of objective. In many cases, it’s just priced by the company one way or the other, and you’re either in the round or you’re not, right?

And you know what? On top of that, Harry, you can forgive me for going off on this rant, but Brandon at MaC has mouthed off on Twitter about Sequoia’s tranche rounds. I think it’s brilliant marketing for Sequoia. Honestly, I would have retweeted it, but—

The amount of tranches—I saw a round the other day with 4 tranches.

Rory O’Driscoll

Yes.

I thought it was a multi-story car park.

Rory O’Driscoll

Those 2 things go together, right? That tranche comment goes together with your prior comment, right? I’m going to paraphrase it. Doing classic early-stage, Series A and B investing is really hard because prices are high and you’ve got some really talented firms. To win, you’ve got to have differential access and differential picking, and you’re going to be competing in every deal.

Conversely, you’re saying, “I’d look at these companies at $1.5 billion. They’re doing a couple hundred million in revenue.” On an absolute basis, they’re expensive, but on a relative multiple basis, they feel a little cheap. That’s what you just said. Correct?

Yeah.

Rory O’Driscoll

And I think that’s correct. What you’re basically saying is growth, for the last 2 or 3 years, has been a very attractive place to make money, right? Those kinds of deals at $1 billion, $2 billion, and $3 billion have subsequently been marked up a lot.

I shared a statistic before, but we looked at every unicorn that was minted in Q1 or Q2 of 2025. By the end of Q2 2026, at least 40% of them will have had a subsequent markup. In other words, good things get more good things. We’ve been on the momentum side of the marketplace.

So late-stage, that kind of growth investing, to your point—and the reason you’ve been doing it—it’s been a very good place to play. I think you found that. That’s what you’ve seen in your portfolio. You’ve put $10 million in—pick a hot company at $1 billion—and 6 months later you’re getting a markup to $3 billion. You’re like, “I’m a fucking genius. I haven’t lifted a finger, and I just made a 3x.” It’s been a great place.

So now what you’re seeing with these tranche deals is nature abhors leaving a dollar on the table. What’s happening is people are realizing everyone wants these growth rounds. This is how these trends end. Everyone’s going, “Oh, everyone wants these growth rounds, so now what we can do is this tranche structure and effectively price the excess return away from Harry and back to us.”

Because it’s been such a good place to play, capital is rushing in. At some point it won’t be a good place to play, but you are correct. We talked about this last week: there’s always the tension between whether you stick to what you’re doing because you should do it, or whether you move around within the overall environment. I know what you’re going to say—you think you should move around—and I agree, from a pure, if you can pull it off, logical perspective.

Over the long term—and by “long term,” I mean longer than you’ve been alive, Harry, 30 years—the truth is early should have a higher overall return multiple than mid, and mid than late, because otherwise CAPM, the rational-market theory, isn’t correct. Over the long term, it is, Harry. But where you’re absolutely right is that there are these disconnects in the short term—I mean 3 or 4 years—where you kind of go, “Wow, yeah.” A combination of increasing equity valuations and a new trend means that from 2022 onward, late stage has been amazingly good.

Yeah. Yes, I completely agree. Obviously, if you’re in the best early-stage firm, it will obviously have better numbers. I completely agree. But I’m also fully cognizant that venture is a crap asset class for the majority, and actually Thrive and many other very large funds will have much better numbers than the majority of funds.

Rory O’Driscoll

Agree. Totally agree. I don’t think we’re saying anything different, to be clear.

The earlier you go, the more dispersion you’re signing up for. When you get it right, you get it very right, and when you get it wrong, you get it very wrong. The later you go, logically, the less dispersion you should have—the more bounded the thing. But on top of that, you also have this phenomenon: you go late at certain periods in the marketplace and you get this equity-rising-tide perspective, which carries everything.

Since the crash—small-c crash—of 2022, you’ve just had tech lift and equity lift for 3 years. So yes, it’s been a great place to play.

I wonder if I were a founder whether I would really do contemporaneously tranched rounds. I don’t know that I would.

Rory O’Driscoll

Is it not a good deal for them?

I think I would feel like—I mean, I might do it in the moment. We’re all caught up in the moment. I don’t know that I would be comfortable charging one investor $1 billion and another $5 billion within the span of the same week. I don’t think I would feel good about it.

I think it’s suboptimal for my 409A. If it’s a tiny amount of capital, I don’t know that it materially changes the dilution. If it’s a massive amount of capital, I would do it. Don’t get me wrong. If I’m raising $100 million at $1 billion and $500 million at $5 billion in the same 24 hours, I have to say yes to that as a founder, right? Because I can’t raise $500 million at $1 billion.

Guest 3

But if it's all some sort of aesthetic, I don't know. Maybe I'm a fuddy-duddy. I just want my investors to make money, and I want my investors to get a good deal. I don't want them to rip me off, but 80% to 90% of a good deal, to me, always seemed to destress my life.

Always just not taking that last nickel off the table made me worry about one less thing. Maybe, and I just don't know, I would do it. I don't know if I would do 4 different prices in 1 week.

I just think the world is a lot more transactional, sadly.

Guest 3

It is. I've rolled with it, but I don't know that I would do it.

Rory O’Driscoll

I'm kind of with Jason, for the record. I think you're right, Harry. The world is a lot more transactional. It leaves me with an icky feeling.

The reality is, it is all aesthetics, because every founder is wildly smart and they can calculate a blended pre-money. If it's, you know, $100 million at $1 billion and $300 million at $5 billion, they can calculate that the effective pre-money is $2-something billion. These people are doing advanced AI. They can do simple freaking math.

The interesting question is: Is there anything in those terms that subsequently bites you in the ass as a founder? And this gets to your point, Jason, which is, yeah, if you don't care about the 1x, you're effectively raising money in that example at $2-something billion, right?

2 years later, you decide to sell for $4 billion. This is where you're right, Jason: If you don't care that the $5 billion guys only get a 1x, then whatever. I don't think anybody cares, and I think it's liberating for founders, but I don't think anybody cares anymore.

You better make damn sure you have a drag-along.

But on top of that, it makes it more difficult for stock options. It does give you—sorry, this is important to say—it does give you bragging rights. Who cares about bragging rights as markets get more and more competitive? If I can come out and say, “I've raised at $5 billion with Sequoia leading,” it will create fear among other VCs to fund competitors.

Guest 3

Agreed. No, agreed. Look, if you're optimizing for bragging rights, it optimizes bragging rights. It's generally the kind of thing that looks like a really good idea in a good market, and then the real question is: Are the consequences horrific in a bad market?

I will say they're silly, but they're not horrific. If you look at that versus other alternatives, like taking a high price but with a ton of structure—real structure—that's a worse mistake. If you look at not raising money and taking on a ton of debt, that's a bigger mistake.

So, in the litany of mistakes that you can make with your cap table, doing a 2-tranche round that makes all your 2nd-tranche people feel like 2nd-class citizens, it's not the worst thing in the world, provided you don't give a damn that they're 2nd-class citizens, and clearly you don't.

Is there anything else we should cover, boys? Is there another story here that I've missed that I should cover?

Rory O’Driscoll

Yeah, it's kind of further afield, but I did spend a second on the Valar Atomics stuff. It's interesting: There is continued progress on nuclear energy—lots of risk, lots of big step-ups, lots of private companies doing this, some public companies doing this, not trading as well, but progress on that dimension.

9. Nuclear Energy: The Quiet Progress Nobody Talks About

Valar Atomics is looking like they're about to raise at a 3x step-up in 4 or 5 months, so it's interesting. They're still private.

What's really funny—I did realize one weird comment I had on this was that if you think about the kind of companies that should be private and the kind of companies that should be public, companies trying to do next-generation nuclear products should probably be private. As yet, there are 3 of them that are public. They SPAC'd, and they're trading like crazy, man, up and down 50% in 1 day.

And then, call me strange, a company that's doing $6 billion in revenues and way cash-flow profitable, like Stripe or like Databricks, should probably be public. As yet, here we are with Databricks and Stripe private. Databricks doing a Series M, Stripe doing some kind of acquisition that's kind of convoluted, which are typically both public-company stages.

Then you've got a whole bunch of these—not Valar, but the other kind of wild frontier-tech companies—being public. It's just a weird world. The SPACs are taking stuff public that should probably be venture-backed, and the very best venture assets are staying private long after they're kicking off cash and should be public. It's weird. I mean, there's nothing to say except weird series.

You know, this is minor. If I had to phone in, it's so minor, but the C-Square IPO is just mildly interesting as a footnote. Can you just give some context, Jason? What is C-Square? What's happening?

Guest 3

It's a data center.

Rory O’Driscoll

Yeah. So they're a C-tier data center leveraging AI. They're doing a $1 billion run rate, growing 16%, right? And they IPOed with a $3 billion market cap.

So if you kind of reach this slow growth and you put a veneer and a wrap around it, it's still growing at $1 billion in revenue, and you're trading at—I mean, I need to know the enterprise value, not the nominal market cap. It's probably lower, right? The enterprise value. You've got to figure out the debt.

Guest 3

It's higher because they'll probably have debt, too.

Higher. Yeah, you're right. But, I mean, this is meh. Maybe Rory's going to say $3 billion is a great outcome, but I bet it's not when you trace back the history and all of this.

The lesson for me from C-Square is you've got to deliver. The market may be exuberant. The market may go nuts, but it's not stupid. This one didn't have the big AI boost, and it didn't get the revenue boost. It didn't get the multiple boost.

Rory O’Driscoll

Yeah. No, I agree. It was like a meh public company, but older assets, not as compelling. Agreed.

You know, the counterargument to so many things, but, yeah, these other assets can IPO. I guess you finally get to $1 billion in revenue with a bit of an AI veneer and you're worth 3 times that. I mean, I guess it's okay, but it's not why I'd want to be a founder. You've got to make it. You have to deliver.

You guys done any deals in the last 7 days?

Guest 3

Not in the last 7 days. No, sir.

No.

Guest 3

I do want to come back to the 1 other thing that really struck me as interesting. You put them in there separately, right? But I've been thinking about this a lot. You had TSMC's announcement and ASML's announcement, and I was thinking, oddly enough, about different kinds of trusted supply chains. I'm just going to contrast 2 because it's quite funny, right?

You have the Nvidia relationship with TSMC, which famously—they don't even have a written contract. They've dealt with each other for 30 years. Nvidia is now TSMC's largest customer. There are tensions because Nvidia is pushing TSMC to invest more, but they're managing that relationship.

Then you have the same kind of relationship between TSMC and ASML. ASML makes the machine that enables TSMC, and TSMC makes the wafers that make Nvidia, and no one in that entire supply chain has ruthlessly gouged each other.

ASML has raised prices gently. TSMC has raised prices gently. They're pushing people for forward commitments. It's a real, “Hey, we know we're going to be dealing with each other for 10, 20 more years. Trusted relationships. How do we cooperate for the long term?” There are tensions, but it's not all that crazy.

Then you just compare and contrast that to the adjacent market for DRAM. There are 3 suppliers in there, right? You've obviously got the 2 Koreans and Micron, right? They're selling to the same customers. They're selling to the Nvidias. They're selling to all the other things. They're selling to Apple, right?

And there, the dynamic is totally different. It's like, “Screw you. We're raising prices 40% this quarter.” “Oh, next quarter you still need our stuff? Raising another 40%.” It's just hilarious to watch. I mean, you see these huge lurches.

TSMC and ASML are thinking long term: How do we position ourselves so that we're great and cooperative for the next decade or 2 decades? All the memory guys are like, “This is a commodity business. You all screwed us 3 years ago. We're going to screw you now for every dime we can. We're going to raise prices on you every quarter. We're going to make 80% operating margins in what Harry would call a commodity, because we know that 2 years from now you're going to screw us.”

And it's just super fun to watch because they're literally adjacent supply chains benefiting from the same broad trends in AI, and one of them is just a super long-term-oriented one with a single player at every level. Once you get to 3 players, it's brutal. So fun to watch.

I mean, there's no action from it. It's like, unless you're trading DRAM—which is up on the day, which is today's Tuesday, but who knows, down on the month—it's kind of a crazy way to live, but just an interesting dynamic. The big aha for me is: When that pricing breaks, it'll be brutal to the downside, but maybe that's 1 year, 2 years from now.

CoreWeave is just depressed for a long-ass time, huh? Jeez.

Guest 3

Yeah. I mean, partly, one of the things no one ever says is the fact that memory prices—the cost of building the product you're trying to build—has gone up by 2x because the suppliers are charging you more, right? So it's getting more expensive to build stuff.

Then, obviously, they have the big OpenAI commitment, and at some point people get worried about that. Also, I think there's an element of once you're public for a while, things kind of gravity takes over and you start thinking, “What is this company?” It's still, I think, attractively valued on a sales-multiple basis.

I don’t understand why Kimi and why the open models don’t make Nvidia a little bit more elevated. Jesus, I’m just looking at my Nvidia position going, “How long are you going to stay flat for?”

Guest

I think what’s happened there is interesting because, again, it boils back to the same big question. Nvidia got this massive step-up over the last 3 years—the ChatGPT step-up to plus or minus $200 a share. If you look at their projections for the next 2 or 3 years, they’re basically saying that capex, which exploded from $150 billion to $700 billion, is growing much more slowly over the next 3 to 4 years. So it’s basically a one-off step-up, and now it’s going to continue, but not at an amazing growth rate.

One of 3 things is going to happen. If capex stays elevated but doesn’t double and double again, the stock stays roughly where it is and grows into that valuation. If there’s another uplift, like the Claude lift that happened at the start of this year, you’ll get your step-up, your next acceleration, Harry. And if there’s any kind of slowdown, then even this valuation will look crazy, and it’s kind of in the middle until you get a signal either way.

I think Gavin Baker had a very interesting term. He said—I think it was something like—“cross-sectional comparisons.” I can’t remember the exact phrase. He was basically saying that whatever assumptions you make to value Nvidia about the future, to a rounding error, you should make roughly the same assumptions when valuing the DRAM providers and all the other beneficiaries of that.

What happened is Nvidia got the step-up first, and then all the bottleneck investors suddenly realized, “Oh my God, if Nvidia is going to spend, they’re going to spend $400 billion with Nvidia or $300 billion with Nvidia; they’re going to spend $300 billion with memory and all the other bits and pieces.” All those guys, like SanDisk, kind of popped up in the last 12 months, when Nvidia, as you say, plus or minus, has been in that kind of $180-to-$210 range. And now everyone’s at the level that says, “Okay, let’s see the next card.” Going back to the first sentence, the only thing that matters is the OpenAI and Anthropic growth rate in 2026.

I love that as a way to finish. You know what we did miss, though, Jason? We missed a Shakespeare quote from Rory. Do you remember? Last week, Rory came out with a quote.

Guest

I think it wasn’t Shakespeare.

No, no, it was another intellect. Rory, you got anything from The Odyssey? That would be great. I need a good one from The Odyssey. I’m actually just really looking forward to seeing it, right?