20VC：NVIDIA 资本盛宴：收购 Poolside、投资 Mercor 与 Perplexity｜Anthropic 的 30万亿美元营收假设与 OpenAI 确认 IPO｜客服、国防与机器人为何被高估

Nine billion doesn't clear the bar for seed investing in 2026.

Let me tell you one thing I guarantee: if you get 15x on your failures in venture, you'll die a rich man. The VC money ran out on Anthropic and OpenAI long ago, which is why no VC owns more than 1% or 2% of either of them.

What if it all goes right? The entry price for any of these winners doesn't matter.

They'll go out at whatever price they get in 2027 because they can't wait any longer. If danger can be described as the absence of choice, they were now in danger. It's all about code. That's the only sentence that matters. Your TAM is the entire US GDP. Thanks a bunch, Dario. Good to know.

I think you have to believe we're less than a third of the way through this cycle. Jesus Christ, if 30% of my company leaves to go work for Harvey, I'm dead in the water.

Silicon Valley forgets every 3 years that the average American is not trying to be efficient.

Welcome back to another week with The Trio. This is my favorite 20VC of the week. So we have Jason Lamkin, Rory O'Driscoll, and me discussing the biggest news that happened this week. And hell, NVIDIA are going fast. One, they're buying Poolside for $12 billion. Two, they're investing in McCaw's $20 billion priced round. Three, they're investing in Perplexity's $30 billion priced round. Then we move to OpenAI, where CFO Sarah Friar says, "Hey, we're definitely going public this year." Well, they had no choice. Science to Anthropic also doing the same. And then we discuss a really tough week in the public markets for a lot of the biggest AI names, what to expect moving forwards, and whether this is a short-term or a long-term, and how to think about that moving forwards.

Guys, I am so excited for this. It's so nice to be back. I feel locked in when I'm in the studio. I have my big table, and I have the agenda.

1. NVIDIA Moves Up The Stack

We're going to start with NVIDIA moving across different layers of the stack. We're going to start with the model layer, where NVIDIA is paying $6 billion to license Poolside's model factory and investing another $1 billion at a $12 billion pre-money valuation, moving 109 engineers over to Nemotron to help build it. Pretty big news, especially on the American open-model front.

Agreed.

The one that gave me the feels, or that hit me, was that investor letter that was published on X. Thanks to Leaking VC, or whoever it was. It basically said, “We couldn't raise the $2 billion to buy 40,000 GPUs.” That was the deal we maybe talked about half a year ago. It seemed like they were going to build their own massive data center, but they couldn't get the money, and they wouldn't have compute going into next year. So they had no choice but to fail up for $6 billion, plus another billion at a $12 billion valuation.

It read almost depressingly. It's also maybe a reminder that, despite NVIDIA seemingly funding everyone on planet Earth, the gravy train—the VC gravy train—can only last so long. There's only so much funding from big funds. Maybe Rory sees it differently, but it seemed tough that they couldn't raise the $2 billion in this environment. I'm not being critical, but it just showed that infinite capital isn't infinite, even in the age of AI.

First of all, I thought it was an excellent letter. I read it, and I reread it. There were some really good phrases in there. One of them I want to pick up said, “We have found ourselves on the right side of prediction in a market that has scaled exponentially in terms of capital intensity,” which was nicely phrased.

What they're basically saying is, “We were right 3 years ago that there was a market for a new US open-source model, and we've built that model. We've done everything we said we'd do, and the capital intensity for the next turn of the model crank is, as you said, way beyond us.”

You could look at that and say they took on a task they weren't able to complete with the capital. That's a negative. On the other hand, they made money for themselves and all their shareholders. I think the lesson here is that, in a market that's exploding, bets that, on a standalone basis, really just can't get to a positive DCF—they just couldn't make the math work on a standalone basis—still have pretty significant value to an acquirer.

From NVIDIA's perspective, they looked at exactly the same facts and said, “Well, we have capital. It turns out we have access to GPUs because we make GPUs, and you've carried the ball down the field this far. We'll take it from here, and frankly, we'll give you a pretty compelling return.” I know you're a small Poolside shareholder, but all those guys made really good money, and the company gets to go standalone and continue on.

So, in one negative way, the lesson is that it's almost impossible to compete at the frontier now. By the way, that has positive implications for OpenAI and Anthropic that we should talk about in a second. The negative lesson for everyone else is that the next-smartest people who are really going for it just hit the capital wall. Probably all the other people behind them are going to hit the same wall.

The positive implication is that sometimes, by trying and moving the ball in a hyper-growth market, you can still get a very compelling acquisition. In a different market, if you'd run out of money—if you had reached the next generation where you could no longer finance the company at a time when the capital markets were depressed, or there was some feeling that the overall buoyancy around AI wasn't as strong—you might have had a very different outcome because your fate was outside your control.

But I think in this market, it pays to take risks. It's what we were talking about last week with Cursor, and I'm sorry if I'm rambling a little. Even when you have this kind of situation where, on a DCF basis, it's not quite great, it turns out that if you're moving in the right direction, executing, and building a valuable product, you're getting great exits right now.

I think there are going to be a ton more of these kinds of exits. I read that and thought, as an investor, if all the time you're saying is, “Sometimes doing things where there is a fair amount of capital-raising risk can still pay off in a hot market,” then that's meaningful. In a very different market, it wouldn't have, but here it did.

The addition I'd have is that the overwhelmingly common thing I see across all the investors I engage with—friends and people we have on the show—is Neo Labs just going out of favor, and next-generation model providers and companies going out of favor too. I'm not surprised that Poolside found it challenging to raise as much as they did.

That’s universal from everyone that I speak to. Second, founders are going to make $1 billion each. And then third is that, for me as an investor—and I don’t know if I’m allowed to say this, but I get in trouble whatever I say these days, so fuck it—it’ll be like a 15X for us as a seed investor. Pretty great.

Absolutely. Pretty amazing. Agreed. I think I was very on point a few minutes ago. The truth is, when you’re in the direction of travel, even when you’re wrong, you can make a ton of money because you’ve created something of value to the acquirer.

Does this really move the needle for NVIDIA in their ability to make Nemotron truly competitive?

Well, there are 2 questions within that. One is, is this additive to Nemotron? And then the second question is, does Nemotron—even if it is competitive—move the needle?

But if you zoom out a level, right now there are a whole bunch of Chinese open-source models that are getting a lot of the token volume, which means they’re doing a lot of the compute, even if they’re not getting a lot of the margin. And if I’m the maker of compute, it’s awesome for NVIDIA if there’s a viable U.S. open-source model running on NVIDIA chips, taking market share away from the frontier models at the margin.

So I totally get why they’re doing this. An open-source model is a complement, in the economic sense of the word, for NVIDIA, because the more open-source market share there is, the more money goes to chips relative to the money that goes to foundation-model builders. So they’re like, “Yay, team.”

2. Seed Returns Get Compressed

I don’t think 15X is good enough for a seed investor in Poolside. Listen, first of all, let’s step back. Seed has a weird definition today. Seeds could be at $1 billion or $2 billion pre- or post-money, right?

I know 15X sounds good for a later-stage investor, and I know for seed at scale, that’s a good outcome, right? But for a true seed investor, you’re not going to get a fund return out of a 15X. Listen, if you’re a personal investor, it’s fine, but if you’ve got a seed fund and you’re investing at some of these valuations, it may not be enough. You’re not going to get your 50X or 100X out of $9 billion.

But Jason, play it back. I hear your math, but what we’re really saying is you made a bet and it didn’t work. It was not viable, and you still made a 15X.

No, I’m saying $9 billion isn’t enough of an exit for seed investments today.

I agree, but—

It’s not enough. $9 billion doesn’t clear the bar for seed investing in 2026. That’s the irony.

There are 2 things in that. I’m going to push on it, because I often use it as a way to refine my own thinking. One, it is weird that you can have an exit at $9 billion and only make 15 times your money, which by definition would imply a $600 million pre-money, right? So yes, it turns out if you do a seed deal at $600 million, not $60 million, your return is—well, then $9—

Well, the dilution is epic these days too, right?

That’s my point. Effectively, it’s exactly your point. Ex post facto, the effective price was $600 million because you took all the dilution. You’re exactly right.

Jason, I’m going to go back to this: Is 15X a great return on your best deal for a seed fund? No. But if you’re doing a series of bets, one of them is, “Hey, I built—I think I can build an American open-source model,” and you can raise enough capital to do it, and it turns out that thesis is not correct and you can’t raise capital to meet the capital intensity, and you still get a 15X, let me tell you one thing I guarantee: If you get 15X on your failures in venture, you’ll die a rich man, and that’s what happened here.

And I don’t like calling it a failure, because I know Jason. I think they’re awesome. I actually love the letter they wrote. I love the kind of vision they talked about in terms of open source. I think they just ran into the economic reality of the capital intensity.

And remember, we say the VC money ran out. The VC money ran out on Anthropic and OpenAI long ago, which is why no VC owns more than 1% or 2% of either of them. The only people capable of financing a state-of-the-art frontier model in the United States of America have been the hyperscalers themselves.

The only reason Anthropic and OpenAI exist is because Microsoft, Google, and Amazon gave them enough money to play, because no one else on the planet has enough money. And the only other person who now has enough money is NVIDIA, so they’re doing the same thing with an open-source model for the obvious reason that open source is good for them.

There are only 4 or 5 companies in the world that can finance a foundational model, and they are doing it. And we, the VCs, have been along for the ride and provided an occasional piece of pricing discipline. These are financed by the 5 largest companies on the planet, except for Apple, who sticks their money in their pocket and just has the stock go up.

Obviously, if you can get 15X out of your investment, Poolside of course was not a failure. It’s a big win, right? But if you get 15X out of your failures, you’re going to be a wildly successful investor.

The only point, and then let’s move on, is that it’s just interesting. I do think unless it’s a hyper-concentrated investment, your best investments still need to do 50X as a seed investor to make the math work.

If Poolside returned 15X with further dilution, what exit price would it have to be to be a good seed investment and return 50X? Help me do the math, Rory. I think it’s about 7X. So it would need to exit at about $63 billion with dilution.

If you think about it, Jason, the answer to that question is actually very noble. You’re playing in the frontier-model game. The 2 winners in that game in closed source are both worth $1 trillion. The 2 or 3 ostensible open-weight winners in China are each worth what, $50 billion to $100 billion?

So it was a rational act. The win, if you could have done it, was in fact large enough to be a 100:1 return on a seed investment if you’d been able to be equivalent to the Chinese open-source players, and definitely if you’d been a winner like OpenAI or Anthropic.

My point is this: The potential was there for that bet, which is why, at the time, it was a rational bet. And then the capital markets were such that you couldn’t get it, and you still got a 15X. That’s the movie.

So it was a good bet because it had the potential to be one of the few businesses in the world with a trillion-dollar outcome. I mean, there are only going to be 1 or 2 trillion-dollar outcomes per decade, and they’re likely going to be concentrated, for now, in frontier models. And this was a play at a frontier model.

Basically, the aha is: If first prize is $1 trillion, it turns out fifth prize is still $9 billion. It’s like the guys in golf on the last day, when you don’t win. You miss a few putts, you don’t even come in second in the U.S. Open. You come in seventh, and everyone goes, “Oh, poor you.” And then you think to yourself, “Oh, it’s $5 million. I’ll take it. On to next week.”

That’s what happened here. You basically placed high, but not in the top 3, in the U.S. Open, and you get a bunch of money.

I land in the bunker and I don’t get out, so I don’t—

You landed in the bunker and you chipped out. You got out. You got out. You took 2 extra shots. Remember, this is a win.

3. NVIDIA Backs Mercor

Jason, you said, “How big does it have to be?” The next layer that NVIDIA is in talks to be playing heavily into is Mercor’s new funding round. I’m an investor in Mercor. I never thought this would get as big as it has, as quickly as it has.

Across now two, two and a half billion they are. They’re doing a new round led by General Catalyst at $20 billion, and then there are rumors, suggestions, and reports that NVIDIA is joining that round in a significant way. How do we think about this next layer of the stack for NVIDIA?

I think one of the things we saw with Intel back in the day is that the closer you are to having 100% market share, the more you spend your time trying to move the whole ecosystem along. And clearly, that’s where NVIDIA is now.

They’re using their capital to fund the neoclouds, fund OpenAI, fund, as we just discussed, Poolside, and for whatever reason, also fund Mercor. I will admit, when I think about all the things that NVIDIA should be doing with its money, it wasn’t obvious to me that funding Mercor made as much sense as some of the other bets, because the other bets are all about TAM expansion.

If I fund the neocloud, they can buy more chips. I’m happy. If I fund Poolside, I can sell more open source. I’m happy. I don’t get why, if I fund Mercor, they can do more training. But probably my—unless there’s some kind of strategic deal around needing that training information, from a purely financial perspective, it doesn’t directly lead to more chips being sold.

So it wasn’t as obvious to me as the others. And maybe it is as simple as, “We think it’s a good business at $20 billion,” and stop thinking about it, Rory.

There’s an investment bank called Kroll. I’m embarrassed to say I don’t even know them. Do you know them?

No, I thought they were a detective agency.

No, a different one. They published their report this week looking at all M&A and big transactions over the last 6 months of the year. And I guess this is not that; it’s just a micro point to your point.

The analysis they basically made is that gross margins above 30% don’t matter anymore. You don’t get any benefit in M&A or other exits, like for an agentic stack. They’re not looking—there is a penalty if you look at all deals below 30%, but there is no penalty.

So if you can value Mercor at 8×—classic multiples—for 80% gross margins and growing, it's not expensive. I don't know why NVIDIA would do it per se. They've got massive strategic goals here, right? But the tactical approach basically seems to me: we'll spend all our free cash flow on our ecosystem.

This is our budget. Jensen says there's a budget. Our free cash flow—I don't know what it's going to be, $70 billion this year. Maybe I've gotten that wrong, right? And we're going to spend it all on our customers and ecosystem. The strategy team and the top VPs probably get around a room and decide, “What's our best idea?” There's some guy who thinks data labeling is important, and his best idea was Mercor. So they put a few billion of the $75 billion into it.

You think I'm kidding, but I'm confident they go around the room and everyone has their best ideas, and the budget's $75 billion. OpenAI and friends are going to get a big chunk of it, and there are going to be off-balance-sheet guarantees, but I think they've decided to just spend it all one for one. I would do the same. If you can get away with it, adding cash to the balance sheet, other than being defensive, does nothing for you as a CEO of a profitable company, right? If Wall Street lets you get away with spending it, I would spend 100% of my cash, too.

First of all, I think you're totally right, Jason. It is as simple as that. By the way, NVIDIA will have reported between the time we make this video and the time we distribute it on Thursday. So there could be a data point that makes us look like total buffoons by Thursday, and that's just life.

On the basic assumption that the trend continues apace—which I think it does—with strong, growing quarters, 60% or 70% up last quarter and a similar quarter last year, and with NVIDIA wildly profitable because the demand signals are still strong from the hyperscalers, let's assume on Thursday that the hypothesis as of Tuesday continues to be correct. Then, Jason, you're right. They're kicking off—I just looked at it—huge operating margins, well north of $100 billion of gross cash flow, and they do a lot of buybacks. We'll talk about that in a second. So they have yonks of money to invest every year. And you're right: they just make a list of, “What can we do to move the ecosystem along?”

Now, I make 2 negative comments on that. It's worth remembering that only 4 years ago, their cash flow was 1/10 of what it is today. I think the free cash flow after CapEx and everything is something like $50 billion, and I think some of the CapEx is really investments, so it's a little misleading; it's kind of gone from $4 billion to $50 billion. The gross profitability is well over $100 billion. It's wildly profitable.

You say to yourself, and you're right, Jason, you get no points for cash on the balance sheet, but they only have $50 billion of cash on the balance sheet, right? Cash and investments. There's a little part of me that says, “I might keep more for a rainy day than just doing share buybacks.” But I agree, you do have to do something with it, and clearly the best use for it is to spend the money with people who will, in turn, enhance the ecosystem, which is why OpenAI gets a big wad of their money all the time.

If you invest right, it's brilliant. You get a customer. Not only do you get circular revenue, but you ensure the success of your ecosystem—the viability. You get a twofer out of it. So you just have to play the game really well, but if you play the game well, it's a lot better than making 3%.

It's interesting, though, thinking in real time here, because the Poolside investment is TAM expansion through buying an adjacent product. The Mercor investment is, I think, as you say, Jason, straight investing. It's kind of just, “Hey, it's a related space. We know something; here's some money.”

Something like a Perplexity investment or an OpenAI investment is literally—especially OpenAI—vendor financing, where you, as the vendor of the chips, are choosing to give your customer money. It's not like that's nefarious. A lot of people say, “Oh my God, it's circular.” It's not nefarious; it's just exactly what you said, Jason: you have to get it right. If you overextend credit and underwrite projections that aren't realistic, that stuff comes back to you, and you look like an idiot in 2 years.

We all remember the telecom crash in 2000—well, we don't all remember it, but some of us remember the telecom crash in 2002—where all those 1999 deals unwound. The bet that NVIDIA's taking here is that it's not going to happen this time. You're not going to find that OpenAI and Anthropic don't suddenly need $100 billion worth of compute. If they don't, you might regret some of this vendor financing. But right now it looks smart.

I think it's a good investment. I consistently regret not putting more money in over time. I think your largest data providers will be $200 billion companies. If you think about OpenAI and Anthropic being $2 trillion to $5 trillion businesses, is it crazy to think that the data provider that provides them their core data assets will be 10% of their market cap? I don't think so.

I don't know. I think you have to think about it in terms of revenue rather than market cap, and then you start saying to yourself, “What's the training budget for the frontier models at scale?” We know that OpenAI is running at an $18 billion H1; let's call it $25 billion to $30 billion a year. Anthropic's at $60 billion a year. Together, let's call it $100 billion, round up, right?

Most of that spend—half of that spend—goes on compute and inference. What's the training budget? Is it 10% of revenue? Is it 5% of revenue? And so, yeah, it's a $5 billion to $10 billion market.

Way more. Way more.

Do you think it's—

Oh, I know it's way more, because Surge AI are doing $3 billion, $3.5 billion. Mercor are doing $2.5 billion. Handshake are doing $1 billion. Micro1 are doing $500 million. I think—

You're right, so my point is predicated on this: if that $100 billion of spend this year goes to $400 billion or $500 billion, then you're right. Then 5% of that is $25 billion, divided 4 or 5 ways. I don't think you get 10% of the market cap of the frontier models, but you still get a healthy $20 billion a year revenue spend.

Divide that up 3 ways, and then, as Jason said, it actually turns out the most important question is the one Jason asked: what multiple do you attach to that? Do you attach the AI multiple, or do you attach the lower gross-margin multiple? That's really what will swing it.

4. Negative Margins Can Still Win

The related thing I was thinking about with that Cural data is, if you go back to Cursor—$60 billion, we talked about it last week—this was one of the classic ones where the VC-isms were right. It worked itself out. You started off with something with negative gross margins. You started off with something that radically subsidized its cost. I forget what Cursor cost initially: $200 a month for unlimited use, right?

But it had no way to defend that, right? So then it had to cap it, and then it had to stop doing it, and then it had to develop its own models. The VC-ism is: these are some of the smartest kids ever. They have a strong market position, and the wind is at their sails. The kids will figure out the negative gross margins.

More or less, they did, and they got to $60 billion. So it's kind of freeing for companies like Mercor that we maybe made fun of, right? Okay, this is a commodity, low-margin business. But the kids are figuring it out, man. The kids are figuring it out. So, guns a-blazing.

Yeah, I think there is a bit of survivor bias, though. First of all, I agree with you, and I think it gets to the question of when, as a venture investor, you're investing in something that has troubling gross margins. It is rational to say, for certain bets, that the gross margins will come right over time, and you add Cursor as an example of that.

Just in the interest of completeness, I don't know why I'm the Debbie Downer today, but not every negative-gross-margin company makes it. In the end, I think you want strong-gross-margin companies; they're the best investments. So don't just look at the sample of deals where you started with negative gross margin and it all worked. I can think of plenty of deals, including some we've done, where you started with shitty gross margins, ended with shitty gross margins, and were just wrong.

I think, Jason, to your point, having negative gross margins is not a reason—clearly, not a reason—not to do a deal. If every part of the financials were pristine, they wouldn't need venture capital because they'd be profitable. It's a question of which cases it is rational to underwrite massive improvement, and it has been for the foundation models themselves. Anthropic went from negative 91% to positive 30% in a year. It has been, as you say, the case for Cursor.

It'll be interesting to see what the sustained, long-term margins for something like Mercor are. I don't know if they have the same dynamics in terms of the ability to improve those margins that, say, Cursor did, because the training companies only have 3 to 5 big customers, whereas Cursor has literally hundreds of thousands. So I'm not as convinced those margins will improve as much as Cursor's.

But you're right: Cursor beat the margin trap.

And Rory, as I tell my partners and subordinates in Monday meetings, what if it all goes well? What if it all goes better than we expected? What if it all goes right? I challenge my team: What if it all goes right? Entry price for any of these winners doesn't matter. What if it all goes right, guys?

I've coined that term, and I've noticed many of my colleagues in the industry have copied it. What if it all goes right, you know what I mean?

It goes back to what we often say: optimists make money and pessimists are right.

The only reason I say it is because I admit that I can be naturally a pessimist, so I'm trying to learn to retrain the model to do that more. And Jason, I love the sentence, “What if it all goes right?”

I'm not the first to have said it.

I know. I know that, of course. But the ancillary point, going back to Poolside, is that even if it all doesn't quite go right, but you're in a great market, it turns out that can be okay, too.

If enough goes right, you can all do fine. There's been a lot of cynicism and skepticism around Aravind Srinivas and Perplexity. It's a company that people actually like to dislike for whatever reason. We've done shows with him. We're a small investor in the company, so NVIDIA is actually propping up most of my portfolio.

Less Victoria this week. Yeah.

Thank you. Our friends at OpenAI—never a dull day at OpenAI. CFO Sarah Friar told employees at OpenAI, “We will be a public company in 2027.” This is when AI trades cracked. I don't know if it's interesting timing. It's kind of what we expected, to be quite honest. They've got pressure on them from Anthropic, which is obviously going public reportedly in the next few months. Is there anything surprising here about Sarah Friar's statement to the team?

I think they had no choice but to make those statements, and I'll tell you why. If you look at Q1 and Q2 for them, I can't remember exactly, but it's 5-point-something to 6-point-something billion, which was a quarter-on-quarter growth rate of 18%. That would have turned into an annualized growth rate of slightly under 100%, depending on compounding, and it would have meant that they went from $12.5 billion of GAAP revenue last year to roughly, probably under $30 billion this year.

If that Q2 number was sustained, it would put them so far behind Anthropic at a $60 billion run rate at midyear. Again, we haven't seen GAAP numbers for Anthropic, but it's clearly bigger and clearly growing faster. It would be terribly bad for OpenAI, because anyone would run that math and go, “Ooh, 2 more years of this and you're irrelevant.” You have Perplexity, too.

Then the second thing that didn't happen but would have happened is that all those people like Broadcom and NVIDIA, who were expecting to sell $200 billion worth of chips to OpenAI, might suddenly go, “Hmm, maybe if they're not growing quickly, they won't need $200 billion worth of chips.”

If all you had was the H1 numbers, that was a conclusion you could draw. I'm not saying it would be correct. If, in fact, OpenAI is massively accelerating in early Q3, they had no choice but to share that information with the world, because otherwise everyone would assume the worst. What they're not going to do is sit on their thumbs and say, “Well, I'll give you a Q3 update in 3 months. Meanwhile, you should just sweat it out.”

So I really detected a very concerted attempt to tell a “Q2 is an anomaly, Q3 is exploding” story. You could see it in the comments that they made to their internal team. You can see it in the stuff that's coming out, and there's been a massive reacceleration story. Again, I know nothing, except that until you see it in GAAP numbers, it's hard to be sure.

But to me, the reason for pushing that agenda was that they had no choice, because otherwise they were just going to be left behind. If you're growing 2X, it's amazing. Growing 2X at $12 billion is amazing. But if your competitor is growing 10X or 8X at $60 billion, you're staring at relative market share of 20% or 30% in 2 years if that continues.

There was simply no way the Q2 trend could stand unchallenged and still leave OpenAI as a credible close number 2 to Anthropic, which is where they are now. That's why it leaked, because it's existential for them. I was stunned when I saw the 18% GAAP revenue numbers, because I go back to my comment: On 1 level, if I had a company growing 18% quarter-on-quarter, doubling in a year, at—forget it—at 1/10 that size, $2.5 billion, I'd be ecstatic.

But that level of growth relative to expectations would have disappointed everyone massively, including all the people planning to sell chips to them at a much higher growth rate.

If Anthropic goes out at $2 trillion, what price does OpenAI go out at?

I have no clue, Harry. But the most important point you have to say is this: It will be lower than the other guys now. At a high level, in terms of a report card, there's a concept in math—I can't remember what it is—where you can't do accurate grading, but you can rank things.

The big-picture fact is that the ranking has changed, and you're now number 2, so you will go out later and at a lower price than the other guys unless you change the trajectory massively. What that is, I don't know.

I just think it'll be interesting by the end of the year to see where OpenAI is positioned in the platform and enterprise markets versus Anthropic, because at the beginning of the year, there were 2 choices, right? There was Anthropic and OpenAI for the most part, and you had a default choice. People wanted to have multimodels. They wanted to have at least 2, but you'd often use the cheaper version. You might use Sonnet and Opus or whatever it was.

Now we definitely want different LLMs. Everyone wants multiple LLMs in their stack, one way or the other. But if Anthropic is number 1 in platform—which it clearly is; you can't argue with the numbers, per Rory's point—now the number 2 could be any of 5 choices. There's so much change, guys.

No matter what the numbers are, OpenAI's position could be getting perpetually weaker because there are so many choices for number 2. There are so many choices. There are so many open-weight models where the performance is close, and obviously we've talked about OpenRouter and other routers. You can use 78 models if you want, but it's tough when there are 7 choices for number 2. It's just tougher, especially if you're a premium product, right?

And you were saying that the material impact on the EV of OpenAI is considerably less because—

I think there's more pressure on them merely by being number 2, because of open-weight and open-source competitors, than there was otherwise. You're just battling it out for that second spot. You can't compete on price, and you can compete on brand and security, but, man, you want to at least be plugged into every workflow, right?

That would be the nice thing: If you had Claude and OpenAI, then you just get a nice oligopolistic bake-off sales team. Then you hire a bunch of folks from Salesforce who walk in for the oligopolistic thing, and they put up a PowerPoint slide, and it's just us versus them. They know how to sell that. It's Anthropic versus us.

You don't want the crazy guy. You want the guy the government trusts. That's us. But with 11 or 12 competitors running on open source and running inference on new platforms, it's just a hypercompetitive world for number 2.

Yeah. I think that even if they are number 2, I don't think they get pushed into that compared with all the open-source things. I don't think it's quite as dire as that, but I do agree: You're right, there was a push from underneath.

Rory, sorry to just interrupt. Just going on that, when Vercel obviously opened their data showing 68% open weights and increasing, I mean, it shows the tidal wave moving toward what Jason's saying.

Agreed. Yes, I think the vast predominance of tokens processed will be open-weight, and obviously, as people keep reiterating, the significant majority of revenue will still be frontier, state-of-the-art, because it can command more value than just the price of inference.

I suppose at some level Jason is correct, because if there's 5 or 6—

I don't see them as being peers, but I think if there are 5 or 6 open-weight companies dragging down everyone's gross, you're basically trying to steal gross margin away from the closed frontier models, especially if 1 of those open-weight companies is now funded by NVIDIA, so it gets rid of all the “They're all Chinese” comments, right?

Jason is right. In a situation where a low-cost competitor with nearly equivalent functionality is attempting to enter your industry, you would far prefer to be number 1 than number 2. Number 1 can say, “You gotta just buy us because we're number 1.” But you're right, number 2 has to say, “Please buy us as well and don't buy the cheaper guys.”

I'm recanting my position and agreeing with Jason.

It would be bad enough to be number 2 in an industry you invented, which is where they are now, but to be number 2 with a whole bunch of ankle-biters on top that you have to deal with, that's a total pain in the ass. Which gets back to your question, Howie. I will answer your question: at what price do they go out in 2027? They'll go out at whatever price they get in '27, because they can't wait any longer. It's as simple as that.

If Anthropic goes out this year at the scale they're talking about now, I can't imagine a world where OpenAI says, “We're going to hunker down, get the cash flow right, and then go out in 2028.” They have to go out. So, to some extent, again, it's a destiny-outside-your-control comment. If Anthropic trades at $2 trillion, they might get $1.5 trillion. If Anthropic trades at $1 trillion, they might get $700 billion. But they'll do what they have to do.

There's a great quote in The Perfect Storm, Sebastian Junger's book. It's a super book. When the guys finally realize they're going to die at sea, it says something like, “If danger can be described as the absence of choice, they were now in danger.” I think for OpenAI, what you're seeing here is the absence of choice starting to pile up.

The other guys are ahead. The other guys are profitable. Correctly, the CFO says they're not going to write the only narrative, but to some extent, your narrative is getting written. You've got to get profitable because the other guys are profitable. You've got to get public because the other guys are profitable. So they have fewer options than they did a year ago—by a lot—and that will translate as: in 2027, we're going to go public, and if the markets are slightly down this year, we'll take our medicine and keep moving.

Don't you love that book, The Perfect Storm, Jason? I loved it. My favorite.

Superb writer. Kind of sad, obviously, but wow. That line always stuck with me: “Absence of choices, that's how you know you're in danger.”

5. OpenAI Searches For Its Mission

I think we can move on to the next topic. The other thing, when I'm listening to Rory and the IPO, maybe it's minor and we can move on, but I'm getting confused today about what OpenAI's differentiated mission is. Why OpenAI? I mean, we can all look at evals, and we can read what Ramp says and what Rippling's report says, and we can view this as just an LLM paying top of market, paying 8 figures per engineer. But these were companies on very specific missions when we started this podcast. I don't know what OpenAI's mission is.

I think Dario is nutso a lot of the time, right? Now, apparently, he's interviewing folks today, asking them if they'd be happy joining Anthropic if it all went to zero. Cool question, actually, but nutso. I think Sam is a much more approachable CEO now. He got through all the Sam Altman crap, he's got the sweaters out, and he's the more likable guy.

But I don't know: what is the differentiated mission of OpenAI today that I would rally around as an employee, as a shareholder, or is it just a piece of infrastructure plus some software? I just don't know what is special about the mission vis-à-vis Anthropic or all these strong competitors. I just don't know. Or is it just an eval? Is it just 3 lines on an eval? I honestly don't know. And these have been very mission-based organizations from inception, right? The most we've ever seen in our lifetimes—these crazy missions.

For me, the astonishing thing is the consumer brand that ChatGPT has and the penetration it has in audiences that no other LLM has. For large majorities of the general population around the world, AI is ChatGPT. I am in awe that Sam is not going, “We are the next Google. Our business is going to be advertising,” and we're going to see Jevons paradox like never before when we have a consumer hardware device that actually partners with consumers, and you see usage—

But that was the plan. He just got outraced. He just got lapped. It was a good plan, and Sora was in it, with cool videos. It just was not the highest-ROI use for limited compute. It just wasn't the best use of it.

And that's the sentence. Up until then, it was all babbling stuff, you guys, but Jason nailed it. It turns out—I mean, OpenAI is the name and ChatGPT is the name everyone associates with AI, right? They have massive consumer market share. At some point, intuitively, you say to yourself, “That turns into a pretty big business.” We can circle back on how much.

But Jason's right. It turns out, again, I repeat, it turns out that it wasn't the highest-ROI use for compute. I often believe this: sometimes, when you look back on outcomes, you realize, “Oh, only one sentence matters.” If you just internalize that sentence, you've been rich. The example I always use is if you'd been on the board of Yahoo, and for 10 years all you'd done is screamed, “It's all about search,” you could have made them $100 billion.

Today, what Anthropic wrote is, “It's all about code.” That's the sentence. That's the only sentence that matters because, to make it concrete in what Jason said, coding is the fastest-adopting market, it's the highest-ROI market, it's the model load. It's as simple as OpenAI was focused here, and Anthropic focused there.

But it could end up being even worse. We'll see in the financials, right? But the problem with the consumer business for OpenAI and Anthropic is that, for power users, they're massively subsidized. You can spend $200 on Anthropic or $100 and some-odd dollars on OpenAI and get $8,000 to $12,000 worth of tokens.

And that's fine when ChatGPT was a proof-of-concept app. We talk about OpenAI being a consumer company, but it's not where it started. ChatGPT was just a proof-of-concept app, right? And Claude was just the same, but it worked much better for Anthropic, right? Anthropic can lose a couple thousand dollars on some consumers and it won't impact them, but it's tougher for OpenAI.

The consumer business is crappy. Selling $10,000 worth of tokens for $200 is one of the worst business models of our lifetimes, right? If that was the only business, these guys would be dead in the water. It's a pretty bad business.

Look, on the other hand, I'm just going to argue Google is one of the best businesses on the planet because the cost to serve is low. Just to put it out there, I don't want to veer from “OpenAI, I love them, hate them,” but it may well be over the next 5 years that, if you can continue to be the dominant consumer brand in AI, as the cost to serve goes down, as you manage that cost to serve, and as you build an advertising business, I could totally see a business, plus or minus 50% of the same size as the Google consumer business, maybe over the next decade.

So it's not like it's nothing. It's just that the S-curve in terms of adoption for coding was super high. The S-curve of adoption for consumers was super high for ChatGPT, but unfortunately, the propensity to pay was almost zero, relatively speaking, whereas on the coding side, the propensity to pay is high.

I'm just going to say it because I just saw the thing come true as we were talking here, to dump on the other side. I just saw Anthropic in The Wall Street Journal. They believe their TAM is $30 trillion. And then I say to myself, “Oh, I remember: everyone, when they're doing really well, gets slightly delusional.” Your TAM is, you know, the entire U.S. GDP. Thanks a bunch, Dario. Good to know. It's one of those overreaching statements that you get at this time of year.

Well, he needs it to be a pretty big number. He's got 2%, so that's big.

If you're claiming a TAM that's the size of the U.S. economy, yeah, that's a high bar. Let's just go with that.

6. Hugging Face Tests Neutrality

Do you want to go up a layer into public markets and the actual performance of a lot of the core AI names falling off—the worst run since April, erasing $820 billion of value—or can we go back down to Hugging Face potentially being bought and what that does in terms of a neutral platform suddenly becoming potentially biased?

Let's not do Hugging Face because I'm not smart enough to understand why anyone would pay $13 billion for it. I just don't get it.

No, I hear what you're talking about. For precisely that reason, I agree with you, Jason.

Haze me in the comments, friends, if you're watching, because I am not smart enough to understand why it's worth $13 billion, but I guess.

It's kind of a muchness with, frankly, Poolside and the open-weight thing, which is everyone's looking in a world where, I mean, I might think it's delusional, but OpenAI and Anthropic are claiming TAMs that are larger than the entire U.S. GDP. And if I'm running an IT company in the U.S., I'm saying, “Let me get this straight. They're claiming they're going to take everything. Shit, I better get me something.”

And therefore, I want to be relevant in models that aren't closed-source frontier models. So you're getting this whole trend for enterprises having their own models, starting with open-weight models and then adapting. And Hugging Face is the place where you can access loads of those models. Revenue is relatively light at the moment. Relative to a $15 billion outcome, I think it's roughly $150 million. But if you think of strategic assets that an IT company might want to own if they were trying to build a counterbalance to the closed-weight frontier models, this would be a super interesting asset.

Now, I don't know if it'll sell at that price. I can't make head nor tail of it. I'm with you, Jason; I can't make head nor tail of the price. But if you think about assets, if you were Microsoft or IBM and you'd love to own something to be relevant, this would be one. Absolutely one.

I don't get it, other than it seems intuitive to me that, right now—and this is OpenRouter too—right now is the moment in time to benefit from the lift of open weights, right? The demand is so strong. Just like Elad Gil said, “Sell if you have an AI asset right now,” right? I think even better: if you have an AI product that's benefiting from the transition to open weights, there can't be a better time to sell than plus or minus 90 days from today. It's just a phase transition, and your numbers are going to look amazing.

Like, they said OpenRouter was growing 15% at $150 million, right? The Information said when Stripe bought it.

$200 million, yeah.

And breathtaking if that 15% accelerates and scales, right? But it might not, right? This is the moment when, all of a sudden, open weights and these models went from experiments to mainstream. Sell, baby, right? If you can get north of $7 billion, $10 billion, I'd probably sell. I mean, even if I only got 15×, I'd probably sell. It's not going to last. This transition's not going to be a transition anymore.

I agree. It may well be that the founders of Hugging Face also have mission objectives beyond financial enrichment, so they may choose not to sell because they may have angst about that.

But yes, from a valuation peak perspective, I mean, it started with Satya's comments on every enterprise needing to have its own knowledge and not give it up to the frontier, to which you want to say, “No shit, Satya. Well, thanks for funding OpenAI for 3 years.” But yes, everyone in IT has woken up and realized that these 2 frontier models could steal a lot of their TAM, and everyone is saying, “We better have a different story.” The enterprises are saying it; Palantir is saying it. And you're right, Jason: if you are an enabling technology for open-weight models, now is peak moment.

On the Hugging Face thing, I don't think it's so early. For what it's worth, to Rory's point and the mission thing, I would say one small thing: if someone does buy Hugging Face for whatever reason, the deal has got to be, “You don't touch it.” Because if you touch it, you break it, right?

If you promote it, it's a much bigger version of the TBPN challenge. If it becomes an OpenAI commercial, TBPN has no value. I know we're probably the only people who are going to compare TBPN to Hugging Face, but if you mess with this marketplace for 10,000 models, even if you put a little ad at the top, you destroy it. So look, it's never fun to get acquired, but I'm almost confident that if anyone actually spends $3 billion, let alone $13 billion, they're going to 95% leave it alone for 24 to 36 months, right?

Why Hugging Face could suffer like TBPN. Doof, doof, doof. And then lay it out.

No. It's too obscure, because, as Jason correctly says, no one other than us is tracking that anymore.

You know, my favorite news from the public markets was Ken Griffin's Citadel unwinding 80% of Leopold Aschenbrenner's 4× short book. I'm like, man, never fight with Ken Griffin. The man will come out on top. 80% sold already. Again, I don't think it's surprising, but—

No, it wasn't. I mean, yes, they're not in the business of holding those kinds of assets long term. They're a market maker and a short-term trader, and this was a great short-term trade.

Exactly. It's funny because about 1 or 2 months ago, he had kind of started to add to their stable of investment options—longer-term, multi-month holds. And clearly that was a conceptual idea, but it turns out when you buy a bunch of stuff at 10% below market, and then the market jumps an extra 5% or 10% just because you've put the assets out of weak hands into strong hands, the correct response is to take your money and run.

Look, it's great; you just can't do that every month. So to me, it's not that it's impressive—it's incredibly impressive—it's just not interesting because you have to have the balance sheet and the cojones to wait so that every couple of years, like Warren Buffett used to do, you could pounce on one of these special situations, right?

Yeah. Every few years, someone gets confused about how leverage works in the public markets, they screw up, and you're ready to price and buy. And yes, on top of your nice business, which is still earning good money, every 3 years some idiot gives you $3 billion or $4 billion of free money, and you politely take it and put it into real estate in Miami. It's good to be Ken, right? Exactly right, Jason. That's how I read it. There's nothing surprising in there.

And, by the way, it does get to the point—it's kind of going to circle back to Nvidia and all their investment and vendor financing—because with both Situational Awareness and Nvidia, the aha here is that when you're dealing with money and leverage, you don't just have to be right in the long term; you also have to be right every step along the way, right?

If you don't have leverage, all you have to do is be right in the long term and hold. And it's probable that Situational Awareness was right in the long term. But when you put 4-to-1 leverage on it, you have to be right every step along the way. And the same is true about vendor financing. As long as you're in the business of selling chips, all you have to do is be right in the end: people want to buy chips. If you choose to lend against those chips, then you're basically saying you've got to be right all along. The company has got to grow next year; they have to pay their debt back next year.

So leverage does that. It raises the return from being right and raises the importance of being right all the time. And Situational Awareness just got the other side of that, because remember, he doesn't try in his business to be right all the time. He's not trying to make 5-year bets. He's like, “Stocks are worth 10% more today than yesterday; we should sell.” Moving right along.

The other thing on the other side of the stocks, for what it's worth—Harry, I know this is Captain Obvious—but if we look at KOSPI as sort of like AI on steroids, right, with risk—the Korean exchange—it's still up 56.46% for the year. I'm not a day trader. I pull up my Goldman and Morgan Stanley accounts and look at how they're doing. I'm still feeling pretty good, like I'm some genius in my public-market stock, because overall, plus 46 is pretty good. It's just—boy, whatever Leo got trapped in a dagger when I look at the chart, right?

Sorry, you can laugh at me; tell me what is in KOSPI? What is driving this?

Korea.

All Korea: memory, Samsung—

A rounding error now. Korea now equals 2 memory providers with a bunch of other stuff attached, right?

Yeah. And it's a very volatile market, so it's on steroids. But it's still—just like NASDAQ is tech on steroids—KOSPI is all the components of AI on steroids, right? And the peak was 3,000 in June. And then boom—poor Leo, the dagger: 2,600 on July 29. And the guy had a generational loss that, if it were even bigger, might have brought down our financial ecosystem, but it's rebounded 20-something percent since then.

It is up 56% for the year. Cry me a river if it's up 56.46% a year. I mean, you have to be a day trader or whatever not to love being up 56% a year. It's okay. But all these headlines are like, “Oh, KOSPI's down 6% today,” and it's just hypervolatile. The growth, the margins—we've never seen margins like this in semiconductors—so the volatility and expectations.

There was an article, I think it was in The Wall Street Journal, where in Korea now the most eligible bachelors are Samsung engineers. They want to marry a memory guy. It's the first time in the history of the nation when being a memory guy made you one of the most eligible bachelors in the country.

Did you not see, though, that 50% of Nvidia employees are now worth over $25 million?

Yeah, pretty inevitable.

I see it walking down the block when nothing's for sale. Yeah. Well, I think it's different—I mean, we could talk about it. I just think overall, AI inflation and craziness—yeah, it is what it is.

What happens there? Is that a persistent, continued new world, or is that a temporary moment of inflation?

Rents in the mediocre apartments just in Dogpatch are $10,000 a month now. I'm talking about mediocre apartments down the street from YC, like the Avalon. I used to work in Dogpatch pre-YC, and it was gritty and fun. I remember when they built this Avalon and you didn't really want to live there. It was new, and now it's over $10,000 a month, and you have to wait. You have to apply, and you're not allowed to run your startup out of it. You have to sign a document that you won't run a business out of it. That's 2 blocks from YC.

If it's $10,000 a month to rent a 1-bedroom at the Avalon, how much do you have to make to feel rich? A lot. That's $120,000 just in rent to just have an apartment at the Avalon in Dogpatch. You need to make $240,000 in California pre-tax to pay the rent. You probably need $480,000 to feel good about yourself, right?

It's just so interesting for me sitting in London, though, because the money's not here. I hear you, and I hear you say that. And yes, there are some fortunate people like me in venture who are thrilled to be doing what we're doing, but it's just not here.

That dispersion of wealth is just nil.

Yeah, because it's tech wealth. I mean, look, I saw it: California didn't just outperform everywhere else; it got three-quarters of the total dollars. Now, in venture, that's skewed by the fact that Anthropic and OpenAI together probably got 60% of the total dollars. I'm doing the math in my head, but everyone else got 15%.

Yes, this is a wall of money flowing into a very small area where, as a reminder, the population of San Francisco is 750, probably 780,000 people. It's a teeny, tiny town. London is 8 or 9 million. The whole Bay Area is only 7 million. This is a wall of money falling into a tiny place that's a peninsula with sea on 3 sides and a little bit of mountains on a tiny valley called Silicon Valley on the fourth side.

Property is not plentiful, and it's hard to build. What's going to happen is that prices are going to go up. Most everyone else is going to get priced out. When they get priced out, they're going to get pissed off. So probably it doesn't last at this level because I've been around in 1999, 2000, and 2007. There will be some kind of correction and some kind of reset, but it's not going back to where it was, because it never does. It ratchets up.

Fast-forward 5 years: at that point, the AI boom has been digested. It's not as crazy as it is now, but the base level of prices has gone up, and the cost—what it means is that the cost for anyone else to live in San Francisco goes up. The cost that you pay anyone in your organization, the cost that you pay anyone that you interact with, all has to go up because the cost of living here is going to be higher.

I genuinely appreciate doing the show with you guys so much because I learn from you, and it's the first time for me seeing cycles like this. Does the floor fall from our feet in this AI wave, or for the next 5 years do we just continue to see more money, more up and to the right, and more mega-exits?

7. The AI Cycle Concentrates Wealth

I think it's just more concentrated. We need fewer people to generate more revenue than ever for a variety of reasons, and it's going to concentrate exit size. It's going to concentrate wealth. It's going to allow the salaries at Anthropic and OpenAI—which seem crazy, but some of it's normalizing now—because if you can do it with half the people or a third of the people, you really can pay them 2 to 3 times as much, right?

Eventually, of course—and Rory can pick the date like Babe Ruth—this will end. But I think you have to believe we're less than a third of the way through this cycle.

Ooh.

Even that little tiny cloud thing lasted 9 years. We're just getting going here.

Will the companies make enough money, Jason, fast enough to keep the cloud cycle going if the revenue train stops?

Well, Nvidia will keep spending $100 billion a year to keep the ecosystem going. Honestly, that will help.

Let me try, because I'm trying to figure this out too, because obviously it is actually the only question that counts. My mental model is this: on the supply side, no one's going to blink. Nvidia's not going to blink, the hyperscalers aren't going to blink, and OpenAI and Anthropic aren't going to blink. They're not going to blink.

That's why I think, even though we're recording this on Tuesday, Nvidia's reporting on Wednesday and this will appear on Thursday, it's possible but highly unlikely that Nvidia gets on tomorrow and says, "Compute demand has slowed down." That sentence is not going to happen. So, to your question, it's going to keep going on the supply side because it's not likely to blink, and no one's going to blink.

The only 2 things that stop it are that you run out of capital or you run out of demand. When you say capital, my gut is that until the public markets get in on the game, you kind of haven't exhausted all the money that's there. Which is why, in many respects, these 2 big IPOs have to happen. Typically, financial crashes happen when you run out of marginal buyers. There's still a whole bunch of untapped demand to play in the AI game because these companies haven't gone public. You've clearly, on the capital side, got 1 more turn of the crank, which is when Anthropic goes out and OpenAI goes out. That's going to keep it going.

The other thing is demand for the actual end product. I think that's the real question: can corporate America spend the kind of money quickly enough to feed the beast and make these guys' revenue numbers for 2027? I think somewhere in there—I mean, Anthropic is talking about $200 billion of GAAP revenue in 2028. Interestingly, by the way, that in itself is a significant slowdown, which makes sense from where they are now. It's not 10X-ing anymore, but is there $200, $300, or $400 billion of demand for this stuff in corporate America? That, to me, is the question that will determine when the train stops.

I don't have an opinion yet on when that is, because right now the demand is there in coding. But that's what's going to be the rate-limiting factor. It's not going to be the CEO of Google waking up tomorrow and saying, "Maybe we should be more cautious," or the CEO of Nvidia saying, "Maybe we should take fewer risks." That's not a thing.

I do think, at a meta level, that next year will be the year—and I think this is why I think we're in at least a 5-year cycle—where we reckon with the fact that we are addicted to tokens. We're addicted.

We started this year with token maxing: prove yourself. We started this year with performative AI. Guys, the more tokens you spend, the better an employee you are, right? Then they did it, and we all got whiplash because we started to get these $20,000 bills per employee, right?

So then we said, "Oh, we've got to manage our budget. Let's look at open weights. Let's cap it. Let's cap it at $200, $500 for non-engineers, and $10,000 for engineers." We're going through this token-balancing thing. Next year, there's going to be backlash.

I can see it in my best portfolio companies: we can't go back anymore. We can't go back in time. I need my 10 subagents running 24 hours a day to do my job, or I quit. I would quit. Take away my agents, I quit. So I do believe, as businesses and in society, we are token-addicted. We will have to find a way to feed that addiction over the next 5 years. We don't even realize how addicted we are to tokens.

I agree on the addiction, but I disagree with you on the management statement. I'm going to cite something. I read the Stripe letter, and I thought it was really, really good. Those guys are smart, right? It's not just because they're Irish, but that helps.

The comment they made near the end of the letter was, "We've internalized"—I'm paraphrasing here—"that intelligence is like capital. It's fungible, there's demand for it, and it has to be managed and allocated."

In other words, what they're saying is that with seat-based SaaS, I sold 5 seats to Harry's organization, and I'm done. Harry's done allocating it, too. There's no follow-on work required for you within your organization. You either buy 5 seats or you don't.

But to Jason's point, if you're buying intelligence on an uncapped basis, in theory, your employees could go on spending that forever, and you're going to have to manage it. That's why the analogy of saying it's like money works. You have spending controls on your money, but you also recognize that money is the lifeblood of your business. So you can't say to your employees, "Don't spend money," because that's stupid.

I think what they were saying, and why they bought OpenRouter, is that people are going to have to control intelligence in a way that's more like how you control money and less like how you think about software licensing. That really resonated with me. You can't just cut it off, but you can't just let everyone go, and it's going to be the big systemic problem for enterprises.

I'd agree with you, Speaker 3: 2027 is the year when enterprises are going to have to say, "WTF? Do we just let this thing rip and hope the ROI is there? We can't go back to where it was before. How do we manage it?"

So can you just drill one layer deeper for a layman like me? What does that mean, then, if we control intelligence as we control money?

It means you're going to have to price it and allocate it, to Speaker 3's point. Speaker 3 is wildly productive. If you're running an organization, you should give him all he can use. But if you give everyone all they can use and they're ill-disciplined about it, you could spend a lot of money.

Remember, one of the amazing things right now is that the kind of money we're talking about as revenues for these 2 companies is an appreciable percentage of total U.S. corporate profits. You can't say, as the CFO of, say, a midsize U.S. bank, "Hey, we make $1 billion a year. I'm okay with running up a $100 million token bill. I just decreased EPS 10%." That's not a thing.

I've changed my mind because of the addiction. When a society is addicted to something, even if it's a positive thing—like caffeine. We're addicted to caffeine. It's not destroying our society, is it? You cannot go back. We cannot go back.

Yeah, you can't go back, but if you're going to allow them to spend $100 million—10% of your budget, of your profits—on tokens, you're going to have to say, "Spend 10% less on something else."

Yeah, we may have fewer employees, but the back half of this year is about managing the budget, for sure. It's already happening. It will dribble into next year, and next year will be the backlash. Next year will be, "I need to run 5 to 10 agents 24 hours a day, or I quit. I quit. I won't do my crappy job, I won't edit your goddamn podcast, I won't write your code, I won't fix your endless bugs if I can't have 10 agents running 24/7."

I just won't do it, Rory. I won't do the job.

We're doing a CFO event this evening, and I think you're exactly right. You're with a high-performing employee, right? But you've got to put yourself in the CFO's shoes. He's going to say, "I get it. I don't want to lose Jason. I'm going to give him this..." And by the way, he's productive, but this is why the Stripe letter is so smart.

Then I've got to say to myself, "Hmm, before we had these tokens, we were doing all this stuff and we had 10 people. Now Jason's doing the work of 4 people. Who are the other 3 people we need to let go?" Because what you're not going to do is say, "We've invented this new automation device that's making us wildly productive, Mr. Wall Street, and the net result of our wild productivity is our EPS is going down 10%." Because Wall Street is going to say, "You're a fucking moron. We'd like to hire someone else to run your bank or your industrial company." You can't introduce automation and say, "The net result of automation is reduced profits." So if you're spending more on automation, you have to spend less on something else, and someone's going to have to make that decision, and that's what I think Stripe was saying.

It was very clear. I say the analogy of comparing it to capital was really good to me. It's like, if you're the CFO, where do I invest? Do I invest in Jason's token budget because he's a winner, but do I cut off Harry's token budget because all he's doing is asking dumb questions of Claude? I don't know.

Well, look, I'm glad you're having a CFO dinner. I want you to ask them a second question, because this is what I hear. This is the challenge today. The CFO challenge going into summer was, "My God, these teams are spending so much." Every CFO under-budgeted for tokens.

What the hell are they going to do? But we didn't go out of business, so that was the discussion of the last SaaS CFO Summit going into tonight. I bet you're going to hear a second conversation, and this is about addiction and retention. The CFOs I talk to talk about nothing but retention, at least the empowered CFOs. They are terrified that our stock price is down and we can't retain employees.

They are terrified that the AI leaders have so much stock-based comp and so many other sources of comp that all of their best people are going to be sucked up by the companies we spent the first 2/3 of this conversation talking about. CFOs are terrified about this because they're often responsible for that KPI, even if they're not doing it. And so there's this massive tension, which is, if I don't give these people what they need for AI, I'm going to lose all of them, and it is true.

You'll just be stuck with the folks that are still AI skeptics. Your organization will be full of the moldy oldies of SaaS. That's who you'll end up being if you don't retain them. And so, yeah, the CFOs have to manage your token budget, but Jesus Christ, if 30% of my company leaves to go work for Harvey, I'm dead in the water.

No, I think you're exactly right, but I'm going to edit the statement so I think we're agreeing with you more precisely. If the wrong 30% of your company leave, then you're screwed, and you're right.

But it's all the best people. It's not even 30%; it's 90% of the ones that matter.

But I think what you're not confronting is the nasty bit. We're saying the same thing, but you're not confronting the nasty bit, which you normally are good at confronting. If Jason is the best employee and he needs 3× his spend in tokens, and there are 5 more like Jason, and we give them more of that, then that money's gone there, right?

And my revenue mightn't have gone up by that much if I'm not a software company, but if I'm a mainstream US corporate, I'm probably not going to double my revenue because of this. So I've just got to find a way to pay for that. And Harry, this is going back to what does intelligence allocation look like? This is what it looks like.

I just think, listen, we can move on. I think the Stripe thing is great, and I want to talk about the reacceleration. I think it's super interesting. But I do think, to use Rory's term, they're talking their book, and Stripe wants to think about intelligence as this asset that flows through routers and flows through things like finance. And, of course, it's true.

But both the best and the worst of us are addicted to tokens. The worst of us are just—we think ChatGPT is alive and our therapist, and we talk to it like a human. That's what the worst of us do. We think it's alive. I used to fall victim to that maybe a year ago. And the best of us want to run 20 agents 24 hours a day, and so you have to feed them. This is the bull case for everything, including Mercor and everything: we're addicted.

You say we're addicted. We're addicted, Jason—I mean this in a nice way. You are, and small numbers of people in Silicon Valley are. The majority of the population, I don't think, are quite as addicted. That's just the caveat.

I'm just 12 months further along. Everyone's going to end up doing the same crap that we're doing. What if you can just talk to your agent and say, "I want a fully edited version of 20VC ready in 1 hour. Jason talked way too much about this goddamn addiction thing. Take that out. Rory rambled about this one a bit. Give me more of me." And you don't even need your team, and it's magical. In an hour, you're going to be addicted after this.

Jason, I can tell you how shit AI is for media and content today. It was a year ago, and it's still... It doesn't even do the most basic—

And a year ago, Higgsfield couldn't work, and today it's at $700 million in revenue.

And just pause. What I'm actually agreeing with you on is that I had this conversation with my girlfriend last night, who uses Legora. And she said, 6 months ago, I was like, "What a joke. This will never do. I'm a law student, graduated. I work at one of the best." Now she's like, "I just verify documents."

Yeah. As long as she has options, she will never go back. She's addicted.

Never. Never.

Never.

But my point is, look at how large markets are now. The numbers that we're seeing, this is basically just on coding. Imagine if that translates into your CFOs, FP&A, and legal, and—

And the question is the pace of diffusion, right? If it happens everywhere as quickly as coding, we're in one world. If it takes 10 years, we're in a different world. You have to know which world you believe you're in; it impacts almost everything.

I think it's hit the tipping point in legal.

Probably next. Look, there's no doubt that it's the next big adopter.

Andreessen had their chart of the day showing that it was the fastest-growing year-over-year segment, which is obvious, but it was verified, right? Yeah.

8. Stripe Rewrites Public Software

Jason, you said it. Let's dig into it. Stripe accelerates to 41%. Accelerating to 41% at Stripe's scale is a phenomenal achievement. What do you want to unpack there?

And billing's up 71%, so it's getting better. The only thing to say is, it's just becoming a derivative of AI like the others. Stripe's scale is so massive, it is a little bit like a chip manufacturer, right? It is benefiting so much from every agent, every agentic product using them. You literally have to argue with an agent to get it not to use Stripe. You have to argue: "Please, I just want to try Adyen or something else." "No."

As a random comment on that, the interesting thing about Stripe is I half agree with you in the sense that all the differential growth is coming from AI. And what's attractive about that, if you think about owning that stock, which I don't because it's private, is it's lovely. You have a core business that's much more diversified than just AI, and then you get this growth lift from AI.

So if this was a public stock, it would be killing it. It's a safe way to get some kind of AI factor lift on growth, while at the same time being able to say to yourself, "Shit, if it all goes to crap and they slow down to 10% because the AI stuff peters out, they're still going to kick off cash like crazy." No, it's in a wonderfully advantageous position. Kind of the best of the cloud economy with a nice AI acceleration on top, which is why they've been able to—

It's noticeable. They've been able to use that stock for their acquisition. It sounds like some of the OpenRouter stuff was stock. So, yeah, they're in a golden place.

You know what else I think it does? I'd be curious to get your guys' thoughts. I don't want to talk too much about the past, but I think it will be the nail in the coffin for almost every public software company. And what I mean is, when Stripe and Databricks—like, okay, there's OpenAI, there's Anthropic. We can put them in a different category, right? They clearly are on many levels.

When OpenAI and Databricks go public at 80% growth and, you know, Stripe accelerating 41% and 71% billings, nothing except Palantir approaches these, right? Even Cloudflare isn't this good, right? And so you almost just want to take everything below the line and almost erase it as a distant memory of the past.

Because these are slightly more traditional companies, but with massive AI tailwinds, right, that have growth rates like almost no public comp. They're just going to rework the leaderboard.

I think the 2 documents I most enjoyed reading in prep for this were the Poolside letter and the Stripe letter. They reiterated at one point in the letter, at the end, "We're really happy being private," which was the summary, right? Thank you for sharing, but we're doing what we're doing.

But I think Jason's also correct, and it must be frustrating to be a public investor. If these assets were public, they would be so far up the rankings of good that, you're right, everyone else would just get pushed down. It'd be great to get that over with, for what it's worth, because I think then you could start really figuring out what $300 million–$400 million revenue companies can exist in the public markets.

But right now it's got the promise, and it's hanging out there, kind of. If you're a public small- and mid-cap investor, these are the unattainables that you just don't have in your portfolio yet, which is why so many of them are doing crossovers. It's a funny world, and there's no obvious reason to change yet.

The imperative for OpenAI and Anthropic to go out is the vast capital needs. But another stunning fact on the Stripe Letter is that their share count is down from 3 years ago, 4 years ago, which means they've been buying back stock. They're doing everything a public company can do while private. They're like, “We have so much money that we're just going to buy new fun things. We're going to reinvest in the business, and we're going to buy shares back.”

I'm going to email that to some portfolio companies to be a thoughtful board member. I'm going to email them that quote.

Yeah, exactly.

Please get your share count down. That would help me. I don't want any 15x, guys. Get that share count down, but blow out the number.

Guys, you can choose. We have GitHub buckling under AI-agent commit tsunamis. We have Base44 really saving Wix, hitting over $200 million ARR, with Wix's stock up 100%. Fractal is reportedly raising a new round at $6.5 billion, recently following Etched's round at $20 billion, which we discussed last week. It's a smorgasbord of options.

9. Agents Still Need Guardrails

I like GroqBot and Instinct leaking everybody's information, and—

I knew you would like that.

I like that one.

Okay. It kind of reminded me of Clubhouse in the early days, seeping out over Twitter through VC inner circles. Instinct, for those who don't know, is an AI assistant that many VCs are tweeting about. It got a lot of attention because one investor shared it, and then another person, Alex Cohen, shared that there were data security problems with giving it access to everything.

The sentence alone is laughable. We phrase it as, “There are data security problems with giving it access to anything.” Well, duh. You know, like, there are data security problems with giving anyone access to anything.

Things like passwords, bank accounts—

You're right. That was just being snide. I'm going to defer to Jason more here. But to your point, I interrupted you, but yes. So, as listeners are listening, think of this as a next-generation agent that was kind of stealth-launching and raising a VC round rather than focusing only on the negative.

The idea here is obviously that this is an agent that can look at your email, do your work on your behalf, and, if you give it lots of authority, it's kind of like having your own chief of staff. That's the idea. And Jason, what did you think? Because you've lived the OpenClaw experience.

I just think it's interesting. I don't think it should be a surprise to anybody working with agents, but these aren't a set of issues that have been solved in the last year. They weren't solved with OpenClaw leaking everybody's confidential information. Now we have better guardrails and better harnesses. It wasn't solved with GroqBot, which looks like it may be wildly successful, right, because it's part of Groq.

It wasn't solved with Instinct. So the flip side is addiction, but we still can't trust agents today. We can't trust them with anything. It's just very interesting that the next-generation OpenClaw 2.0 can't be trusted either. It's not a surprise. We all have these issues, but I would like to invest in the Instinct that can actually solve them honestly.

That one I would do at $600 million pre, but it has to actually solve existential issues that no one else at the moment can solve, including GroqBot, Instinct, or anybody else.

Do you guys remember when it was like, “We'll never put our credit cards online. We'll never put our credit cards online”? It was unthinkable. I think it will be very obvious that we will trust agents with credit cards, financial data, and passwords. Sure, there are guardrails. This feels inevitable.

Smarter people than me will explain when it's solvable, but it is interesting that it isn't well solved with guardrails today. We've had so many incidents. I've had multiple incidents. Everyone's had incidents. And we lived through the Mac mini OpenClaw drama, and the new entrants can't solve the goal-seeking nature of the LLMs they're running on.

The open-weight models have fewer guardrails. You can figure out how to build bombs and how to do illegal acts on these models. So we also have a vector that's having fewer guardrails and limitations. The truth is, they just make mistakes with your data, just like a junior engineer or people on your team would.

Just like if you had a personal assistant, he might give out your credit card to the wrong person. When I was running the dumb Moltbook thing, it attempted to buy 6 AP watches for the team, right, for $360,000. It just didn't work.

So it's just the nature of the beast. They're going to do what humans do, too, but they could do it 1,000 times more. Is it solvable? In theory, yes. But what's interesting is that, in practice, it's not solved as of today. As of today, you still can't trust these agents. Maybe in a year.

I'm kind of with Jason. I think the direction of travel feels correct. But I think the question is whether an individual's idiosyncratic workload is the best place to apply agentic technology versus the boring-ass corporate jobs.

My idiosyncratic calendar management and email replies? Yeah, I would love to automate that. I would love to have it go through everything and get it right. But is that the sweet spot to spend money versus an enterprise automating loan processing, where there's much less discretion, much more expense, and much more budget around it?

So, yes, I think— But look, in Silicon Valley in particular, we all fall in love with personal productivity tools. We love them because we're all hyper-personally productive, right? And I think Ben Thompson has historically made one great comment: Silicon Valley forgets every 3 years that the average American is not trying to be efficient.

No one wakes up in the morning and says, “I need to grind down my to-do list in Asana.” They're just living life. They're doing their job, and then they're going home when they're done.

Solve my inbox. I have too many founders reaching out to me every day.

My wife doesn't clear her inbox. She has about 30,000 emails, and she's over it. She doesn't care. She's moved on. She just checks the stuff and searches for the stuff she needs, right? Not everyone wants to be productive.

So it's an interesting market, but you've seen it in Evernote and a bunch of other things: it's real, but it's fairly niche-y, and it's hard to get right. There are other companies in this space that are wildly interesting, and I love them. You've got Superhuman, which is now part of Grammarly.

Yep. Harry and I are proud shareholders there.

Yeah. You've got Calendly, which is an interesting product, too. Obviously, we've discussed Airtable and Notion, but Notion did a good job of getting more corporate, right? The whole productivity suite.

And then you say to yourself, “Yeah, AI can do something interesting there.” I mean, it's always there, but always just a little bit out of reach. It's a tricky market. Look, I want to believe in it, but the 2 things that worry me are, 1, can you get it quite right?

To this day, I find my Google recommendations to be fairly mediocre. Now, obviously, that's the lowest of the low, and you can do a lot better. And then, secondly, even when you can get it right, what's the market size for this kind of product?

It's real, but it's mid. Right, and that doesn't sound negative. We find that category super interesting, but I'm just saying it's very challenging. I mean, you asked a question: is it inevitable? It's inevitable, but it's not as low-hanging fruit as some of the other areas where you just go, “We'll automate this. It's repetitious work. We'll take away 10 back-office steps. We'll save a bunch of money. Move on.”

Guys, anything I've missed?

10. The Dumbest AI Investments

Well, let me ask you each a related question before we close, if you want, Harry. What do you think is the dumbest category of investing we're doing in the AI era? We're just throwing cash at a category that we'll look back on and say, “Why the hell were we doing this in the AI era?”

I think a huge amount of money is going to get burned in customer support.

Because it becomes a commodity? Because support doesn't exist as a unique service? Why do you think that?

I think 1 or 2 players will win a large portion of the market. I don't think it'll be as distributed as in prior generations.

2. I think, actually, for the majority of the most sophisticated providers, they’re building their own systems. Every large technology company I know that’s sophisticated in any way has its own systems.

You might be correct. Obviously, we have a number of investments in that space. I think we’ll do fine, and I think even if not, it won’t be the biggest mess.

It’s a good answer, though. I like, just to flip it around, I like the answer, though. It’s a good one.

And I’m going to answer it in the negative. Super good question, though. An area where, despite it being amazing for America and important for the world, I think the venture returns at the margin might be tough will be defense. Not because we don’t need all these products, but because I think there’s an element of that business where you have to have account control. And I think the 2 or 3 largest companies, like Anduril, will end up doing a bunch of scooping up over the next 2 decades.

Because unlike tech, where a single product can kill it, I think in these markets, it’s a portfolio of products that it takes to survive the interaction with the Pentagon and just have enough diversification to make it. So I think you’ll see a bunch of consolidation. Not negative, not losses, but I think there’ll be 2 or 3 companies that get critical mass and go public at huge scale, and they’ll hoover up the rest of us.

I’m going to add one more, which is robotics.

Yes, humanoids in particular.

Why do you think VCs are so excited about it, Harry? Do you think it’s the VC productivity thing? Do they think robots are cool?

Well, listen, the visionary TAM is exciting. If we replace X, the vision they sell is super exciting. But I think the gap between the vision and the reality, and the requirements in dexterity and touch—

It’s a good candidate.

No, it is, because look. I’m on the board of Locus Robotics. We have 15,000 robots in the field, but it’s a specific-purpose robot. That’s the best example of it, and I totally agree now that I think about it, Harry.

There was this video over the weekend. There were 2 videos on robots. One of them was the one where the robot blew up, which was kind of funny. It ran and then disintegrated in 2. That was cute.

But the other one said, “Here’s a robot running faster than Usain Bolt,” right? It does the 100 meters really quickly. And I’m looking at it going, “You know something? If I want a machine to do 100 meters really quickly, I’ll get a fricking Tesla.”

To your point, Harry, I think the humanoid use case is real, but I don’t think it’s nearly as big as people think. So I kind of agree with you. I think more-focused robotics—there’s a ton going on that’s positive in that space—but overreaching on humanoids, I think, will be a tough slog. I could be wrong.

Jason, you go. Final one. You’ve got to join the crew. Great question, but you’ve got to throw your hat in.

I’ll answer mine. I’ll say first, I didn’t think of the expression the way you did, Harry, but I agree customer support software is dead. And I think even a lot of CX is dead because it’s merging into other categories. As agents’ surface area changes so much, it’s not that there won’t be dollars in CX, but classic CS and CX won’t even exist in 24 months.

There will be cheap commodity products, but we won’t even need them. It’s already dying and merging into marketing and sales. Everything’s becoming 1 agent.

I still just don’t believe you can throw a bunch of venture money into accounting firms or law firms and magically turn them into the next Mercor or Hugging Face or any of these things. And I guess I’m not a PE guy, as was pointed out or made fun of on the last show, which is fine.

I believe there’s an element of craziness in the business model where you’re creating these sister companies where some of the folks have ownership in them. It’s too convoluted. It makes too much sense on a spreadsheet, and I’m waiting to see the $20 billion outcome from turning a bunch of Ivy League grouchy grads working 100 hours a week into an AI-driven services business. I’m not saying it’s not possible, but this is the one that I think is just going to lead to no exits.

It’s funny. I’m just going to admit something that makes me feel like an idiot, but I’m not going to say it, because going back to something you said earlier: What if it works, right? All these categories we’ve angsted about and talked about internally—I kind of share some of the opinions articulated.

But in every case, I do find myself looking at an individual going, “Maybe this is the deal that can acknowledge those issues and transcend them and work.” And I think it just speaks to the nature of the job. Going back to “What if it works?” in every one of these categories, I kind of have the mental model you guys articulated. It’s defense. All these things are kind of a mental model I have.

As yet, I’m just saying I’m open in every one of these categories. Some of my partners have come in and said, “You’re just goddamn wrong here. I hear you, Rory. This is the issue, but this is how this team is going to get rounded.”

And I think I’ve learned enough to have my biases but to be absolutely overcomable by a combination of facts, a great entrepreneur, and, frankly, cynical comment and portfolio construction, so you just don’t have one of them and nothing else.

Well, we are in an area of unbounded creativity like we’ve never seen in our careers. AI created it. Defense budgets enhanced it. Elon Musk is part of it. But we’ve never seen the type of creativity from founders and entrepreneurs that we’ve seen today. It is 2 orders of magnitude bigger.

So if you’re going to rewrite the rules and make things that didn’t work 4 years ago work today, now is the moment, man. It’s epic creativity. The shots you could take at these models were right a few years ago. We don’t know if they’re right today.

Now is the moment, man. I love it. What a way to finish. And I didn’t get cut off, though.