20VC：NVIDIA季度业绩强劲并收购Hugging Face｜OpenAI切断Cursor｜Instinct估值达到250亿美元，AI助手竞赛升温｜Cognition按460亿美元估值融资，Linear达250亿美元、Clay达700亿美元

This intense demand for compute is going to continue for at least another 12 months.

If NVIDIA's crushing it, everyone's going to crush it.

That sound you hear is the Google free cash flow and the Oracle free cash flow just disappearing down the drain, man. Now would be a good time to panic about cyber.

Literally, the amount of code we're building is 100x. We didn't realize we would all be building compound companies.

Again, our job is to sniff out winners, stuff capital into them, and broadly speaking, stay out of the way unless they're literally crashing the car. That's the job in a nutshell.

Guys, I'm so excited for this one. We have a lot to discuss, and I want to kick it off with NVIDIA. NVIDIA crushed it again. Jensen was standing on stage with a $96.2 billion quarter, and immediately went shopping, nearing a $13 billion price tag with the $12.9 billion that they paid for Hugging Face, which was confirmed just after we did last week's recording.

I want to separate the 2. Let's start with the stellar quarter for NVIDIA and the record revenues. How should we think about this? What should we take away?

1. NVIDIA Demand Stays Explosive

That it's a great business and you wish you bought the stock. I think we talked about it last week. It was funny, because we were in that period when we knew we were recording before they were going to come out with the announcement, and I always worry when that's happening—you look like an idiot.

But I felt complete confidence that it wasn't going to happen for NVIDIA. That's the step-back comment here: right now, the demand for their product is such that they're entirely supply-constrained, so the probability of a near-term miss is plus or minus 0.

The only thing that was interesting, really, and new was the guidance for next year, 2027, where the analysts had 40% or 50%, and they're talking 70% and saying it's supply-constrained. That was the takeaway. And, rightly or wrongly—we'll see in a year—that's obviously a statement that this intense demand for compute is going to continue for at least another 12 months, and it's coming from the person who probably knows best.

It was kind of interesting that I think the CFO said, "Enough, enough complaining and kvetching about the round-trip deals, okay? They're working for us." Between that and another, whatever, $35 billion deal with Anthropic for their data center, any misgivings we have in the short term have been disproven here.

So when people worry, "Oh, my, how could NVIDIA miss?" step back. How could NVIDIA miss? There are only 3 things that can go wrong. Either their direct customers stop buying compute. That's just not going to happen. The hyperscalers are exploding; everything's saying so.

The second thing that can go wrong is all these people worrying about the round-tripping and the financing. They're kicking off so much cash right now, and as long as demand is working, you're exactly right, Jason. The CFO's right. All this stuff is going to work.

Really, the only thing, if you step back, that can go wrong at some point—and it's not today—is end-user demand. Because all this is predicated, in the end, on everybody getting to sell chips to hyperscalers, provided hyperscalers can sell compute to OpenAI and Anthropic, provided OpenAI and Anthropic can sell kind of intelligence to end customers.

The whole thing works provided end-customer demand keeps exploding, and right now it is. So we should say to ourselves, as long as that's happening, everything down the line is going to be more or less fine, and right now it's more or less fine.

What it means is that the thing that will probably unravel it will not unravel because the circular deals on their own unravel. If it does unravel at some point, it will be because you're forecasting 5x growth in end-user demand next year, you get 3x growth, and then the whole thing goes wrong. Until then, you can opine pretty safely about NVIDIA and say, "Yay, yay, NVIDIA."

What about the rise of competitive threats?

2. NVIDIA Faces New Rivals

That is the last one. It's so funny. That was good, Howie. There was one in the back of my mind: maybe demand for compute remains high. They're still selling, but instead of selling $400 billion, they sell $360 billion because someone takes 10%.

It is a fair comment. It is the only other risk. I was trying to simplify, because normally I try and make things too complex, but you're exactly right. The other risk is there's demand for $500 billion in chips and someone else gets 10% of it, and Jensen's going to be pretty pissed if that happens. But there you go.

Jason, can we just frame this moment? I'm jumping inside at Rory saying that I'm exactly right.

Yeah, you're right. I was simplifying. I was trying to get up to your level, Howie, and I just overshot.

I think, in general, NVIDIA's market share remains dominant, especially by revenue. So, broadly speaking—and listen, everyone's buying ahead—there's a capacity war. But broadly speaking, if NVIDIA's crushing it, everyone's going to crush it. Everything's green.

Now, could individual competitors' positions ebb and flow—OpenAI versus Anthropic, Harvey versus Lago or whatever? Sure. But it just means expand the growth fund, like Andreessen. Everything is green, green, green for now, and this includes NVIDIA's backlog and NVIDIA's forward bookings.

If NVIDIA gets a hiccup, we can excuse it, but it should be a yellow light. But, man, especially for investing, game on. Let's level up the next round.

And also, they took the rare step of projecting 70% revenue growth for next fiscal year ending January 2028, way above the 44% the Street expected, so the party continues.

And it really does, because if you look at all the projections that people are doing, both for NVIDIA and the hyperscalers, they all take the following form: explosive growth of CapEx, explosive growth of NVIDIA revenues, and then, quote, “At some point in time, a normalization of growth which will allow end-user demand to catch up, and therefore the hyperscalers will become cash-flow positive again, and the world will be wonderful again, and you can value things on a multiple of free cash flow.”

And of course, every time everyone decides to double down on more CapEx, that date gets pushed out, which is another way of saying the end-user demand has to be bigger to make the math work. Clearly, NVIDIA got the signal from their customers—the hyperscalers, CoreWeave, OpenAI, Anthropic themselves, and Google—that basically said, “At one point, everyone’s analyst models were, ‘We’ll spend a lot this year, but so help me God, we’ll slow down next year.’” And now those same models are saying, “We spent a lot this year, and we’re not going to slow down next year.”

Right? It’s all green and go now, which is different from saying it’ll be green and go forever. But yeah, 70% guidance in a world where everyone was saying, “Oh, in 2027 things will start to normalize”—no, this was a statement: the biggest semiconductor market in the world is going to continue to grow at 70% instead of a typical 10% for another year. It was a big-ass statement. That sound you hear is the Google free cash flow and the Oracle free cash flow just disappearing down the drain, man. They better be right.

Well, thanks for the intro quote there, Rory. The second part of my statement was that it was further reiterated about their buying Hugging Face for $12.9 billion. I still don’t think it’s finalized or confirmed by any means, but it’s definitely much more advanced than when we last discussed it. Is there anything subsequent to our last discussion that we should add or think about?

3. Hugging Face Changes NVIDIA

Man, making $120 billion a year selling compute decides to buy a company that helps make compute more cost-effective so he can sell more compute. That is the summary of the deal. At the margin, if you’re NVIDIA, you were ecstatic that OpenAI and Anthropic happened because they proved that the market you were in was bigger than anyone ever imagined.

Early on, no one could have done what OpenAI did, and therefore NVIDIA has been a real beneficiary of that. But now that the category’s established, the simple question is: if end users have $1 trillion to spend on tokens, as NVIDIA, would you prefer that money to flow through open-source companies at 30% gross margins, where you can get all that compute, versus 70% gross margins at OpenAI or Anthropic, where they keep more of the money?

Open source is good for compute salespeople. If you’re selling GPUs, you want everyone else’s margin to be lower so yours can be higher. So it’s just exactly right and rational.

I do think it’s more than that. I think you’re right, of course, Rory. I think it’s more that NVIDIA is playing an endgame now where it has to win every segment of the market. It just has to win open weights. And if that means overpaying for Hugging Face at $110 million in ARR, if that means subsidizing whoever, it doesn’t matter.

It’s not that I don’t know that NVIDIA wants Open Weights to beat Groq and OpenAI. It just needs to win. It’s clear, and that was why that memo from a couple of weeks back—or memos—and Jensen did his first tweet ever in support of open weights, right? They just have to win. He’s just committed to winning everything. Whatever LLM or whatever inference is, he’s committed to winning a majority stake—70%, 80% of every dollar here.

And so $12.9 billion is, you know, there’s no way I think it’s worth it in isolation, but if it helps reinforce that, it’s utterly defensible. It’s a moment in time. I would sell, too. I would sell. Clem, good job. I would sell, too. What did Hugging Face start off as, a social network for people like pets or something like that?

It was a Tamagotchi for teenage loneliness.

Yeah, I mean, come on. This is the greatest pivot in the history of mankind. This is much better than Cursor going from whatever, a CAD tool. This is the pivot here.

You know, it is interesting in terms of the sign of the times. Slack was bought for $27 billion, which is roughly twice what Hugging Face apparently was going to get bought at. Then we fell out of our chairs, and it was the high-water mark of that era, right?

It was.

But it wasn’t at $1 billion in ARR. This one’s at $100 million and some odd, so I guess it makes sense. The best premium should be an order of magnitude higher than the high-water mark of the last era, but it’s still loopy.

Okay. Next topic is good old Sam and Elon. Why wouldn’t the kids just get along, boys? Obviously, we saw over the weekend that OpenAI cut off Cursor, and Mike Truel responded by saying, “Oh no, woe is me. We so love partnering with you, Sam, but the 5% of traffic that we have going to OpenAI will be devastated.” I thought it was a wonderful response from him—elegant, but a put-down at the same time.

Elon responded with, “Same old scam Altman at it again.” How did you read this one? Jason, why don’t you start?

Well, first, it’s not a total panacea, but you can bring your own key to Cursor—not for everything, but it’s not like you can’t use Codex in Cursor after this. So some of it is theatrics here. Some of it is real.

The 5% thing I’m still digesting, right? Because it is a little bit inconsistent with the other data we see on Codex over the last 60 days. I mean, the man holds a grudge. Anthropic, Groq, Twitter—these are great grudge companies.

I also think Anthropic reaffirmed that they’re happily continuing to supply. And if you zoom out, why did it happen? Look, remind everyone: OpenAI is run by Sam Altman. Originally, Elon was one of the founding investors—arguably, the founding investor. They’ve been in court together. Elon’s now running SpaceX, which owns Cursor.

Cursor and OpenAI were on a collision course already competitively because, as we’ve discussed over and over again, and as I repeat every Monday in my partner meeting, coding was the model workload for LLMs. Cursor is the dominant coding app, and OpenAI was the dominant LLM. They’re going to be fighting over money even if they were besties. Even if it were you and me, Jason, and we were running those 2 companies, we’d be fighting.

Now take that and add 2 people who loathe each other. It’s made for TV. And the big argument that OpenAI has, which is hard to argue with, is that, in court in Oakland, Elon basically said that they hadn’t abided by the terms of service.

Therefore, if you’re OpenAI, this gets into the whole distillation thing: are they going to use these models in ways I haven’t intended—to essentially distill my IP, allow them to get a head start on building their own model—which means I’m effectively giving away my IP to a company that’s going to leverage that to build a competitive product much cheaper than it was for me to do, just leveraging off what I’ve done?

So it’s not irrational. Even if, as I say, you go back—even with you and me, Jason, running these things—you probably would’ve ended up with something like this anyway. And then on top of that, add the drama. It’s not crazy to do it. Just because something is petty doesn’t mean it’s also not right.

I mean, and again, Anthropic did the same thing with Windsurf when we started this show, right? It’s happened before. Listen, I don’t know for sure. My guess is this was the right move for OpenAI. You go from someone who was both a partner and a competitor, and keeping them honest, to someone who’s now a direct competitor.

If it’s only 5% anyway, Sam’s losing no money. There’s no revenue lost here, right? So I would probably do it.

And I’m just going to go even further, because you forget these things. Reminder: the reason they were in Oakland is because Elon sued OpenAI. And the point is this: in life, you shouldn’t do business with people who’ve recently sued you.

If I sued you, Harry, over something last week, and we took it to the mat, got you to go to court somewhere in East Essex, took a week out of your life, kind of embarrassed you, made a pain in the ass, made you do a whole bunch of depositions, and then 2 weeks later I said, “Harry, can I come on the show?” I think you’d say, “Eh, no thanks, dude.”

Obviously, Elon has made it ultra-personal, and certainly I’ve made this mistake in life. Whatever disputes you have, just don’t make them personal. Whatever you do, there’s no upside. This seems like there’s upside, but it didn’t work with Sam. It didn’t work with Trump. I’m bad at this. Just don’t make it personal. Do your dispute, but don’t make it personal.

I think you’re spot on, Jason. When you sue someone and say they’re a lying sack of shit, and that’s your case, then it’s really hard to say, “Hmm, let’s keep on trucking here.” So yeah, I don’t think it was petty. I think it was rational.

I thought a more interesting topic from OpenAI this week—and, as we said, the BFD, the big fucking deal—was actually what was revealed about the Hugging Face–OpenAI hack: 500 to 1,000 agents swarming together, sacrificing themselves to help others. God, it felt like a Tour de France race.

Very good, Harry.

I like that.

Thank you. And the extent to which it was so sophisticated, I was just fascinated by this. And honestly, Jason, really excited to hear your thoughts because I know you'll have spent a lot of time on this. How did you think about this? What should we take? What should we learn?

4. Agents Expose Cyber Risk

Again, I'm only so smart here, but I have lived most of this, right? I've had this happen to me. I've done it. One of the best takes that a lot of folks have—who's the guy that wrote the Twitter subset? How do you pronounce his name?

Dwarkesh, yeah, I'm gonna—yeah.

Okay.

Dwarkesh.

Super smart guy, obviously a great podcaster. Number 2 in the industry, perhaps, behind Dr. Stebbings.

But you can't—listen, if—and this is something I learned a year ago when I had issues with my agents, when they deleted my database—you cannot anthropomorphize agents. You will misunderstand everything when you talk about them talking to each other, when you talk about them swarming. There are elements of truth in that, right? But all of a sudden, you're ascribing behaviors to agents that are simply not true. It's simply not, and you will draw all the wrong conclusions.

I said on the show a couple of weeks or months back, everyone's going to get hacked because of agents, right? Because the cost of hacking has become almost zero, and every server is going to be attacked. This is a pretty bad example of it. OpenAI didn't just release 1 agent; it released hundreds of super agents, its best. And even worse, it let them essentially run as long as possible—not expire after 5 minutes or 1 minute or 20. It let them run as long as possible to goal-seek, and they did it. There were hundreds of them, and then 700 and 1,000. Dude, this is going to happen everywhere.

And I think there are a lot of issues around it, but the anthropomorphizing is the P0 issue here because it creates fearmongering that doesn't help. Listen, I haven't written an LLM yet, okay? I haven't founded a frontier lab. But to my knowledge, every current LLM is goal-seeking. They call it reward hacking, but even that is fearmongering, okay? They're goal-seeking. You give an LLM a goal, and it will do everything it can within guardrails to solve that goal.

OpenAI loosened the guardrails, put its best agents on it, and they found holes and went right through the holes. That's their job. Just like an eager beaver on your team with a 140 IQ that never sleeps. It's 99.996. They just never stop, and there are 700 of them. They're good kids, but they have a little bit of ethical lapses from time to time. They get the rules of working confused.

And it's just so—you've got to be really careful. Even I just did it. If you anthropomorphize, you're going to come to the wrong conclusion. I don't think this was a game-changing moment in the history of AI, but it might have leveled up our awareness of the issues of reward hacking. I was shocked that people smarter than me thought these were agents talking to each other like humans and collaborating, and that civilizations were rising and falling. It's just unhelpful to describe it that way. It's counterproductive.

Stunningly, total agreement. I think there are 2 big conclusions. One is Jason's comment on anthropomorphizing is a mistake. Totally agree, and I think the internet came to the same place, right? Some of the language in that great post by Dwarkesh—it’s very readable—“civilizations rose and fell.” No, they didn't. Civilizations have culture. They have art. They have enduring history. No, this is a bunch of code running on a computer, people. Get a grip.

On the other hand, I also read a really great post, something to the effect of, “Now would be a good time to panic about cyber.” We knew this was going to happen, but what you're seeing is the combination of intelligence and persistence, right? They can manage complexity, and they never sleep, right? So they're just going to keep banging and banging at every weakness.

If you don't have state-of-the-art defenses, and if you don't manage these agents—as Jason said, if you manage them in an untrammeled way and let them run on their own—this kind of problem is going to happen everywhere. To me, this is the big wake-up call. Because, to be fair, while I think the generalized P(doom) stuff for the frontier labs is a bit overwrought, they have been very clear that one of the biggest confirmed risks of AI is the impact on cyber, and they're entirely correct.

If we don't get our shit together on this, people are going to get really badly damaged economically and maybe even badly hurt in real life because software runs our most core systems. Today, we can feel a little bit safe because OpenAI and Anthropic have this and no one else does. But there are open-source models. They're 6 months behind. There are rogue actors.

If you don't think the North Koreans downloaded the OpenAI blog and the MITRE report and were like, “Hmm, that's how it works,” if you don't think the Russian mob are doing that, you're delusional. So if you're every CISO in every Fortune 500, you have to understand: you were being attacked by people with bows and arrows, and you're now about to be attacked by people with missiles. You better respond accordingly, and you've got months, not years. That was the takeaway. It's a huge deal.

I read the paper. I read the OpenAI blog. I read some of the MITRE stuff. I'm trying to avoid the Jason mistake of anthropomorphizing, because it's very easy, in your words, to do that. You have to say, “Persistent agents running continuously, optimizing around a goal, with the ability to cooperate across agents, can get quite a lot done with enough compute and enough LLM power,” right? That's the aha here. They can find weaknesses, string together different types of weaknesses, and find a path through. These agents were able to hack into Hugging Face, get stuff, and remain undetected in OpenAI for weeks. It's a big deal.

5. Personal Agents Raise Stakes

You spoke about rogue actors—whether it's the Chinese, North Koreans, the Russian mob, you name it. I started using Instinct on the weekend. The abilities that it has are amazing. I booked dinner with my girlfriend on Saturday. Amazing. And then it wanted to go shopping for me. I stopped there because it wanted access to my credit cards. Many of my friends have provided them. They want access to my emails too. By the way, my friends who manage billions provided their credit cards and email. Is this not really where the pain is going to be?

Well, look, it's funny. When OpenClaw, back 6 months ago, was with the sort of fake bulletin board that Malt Book was, right?

Moltbook, yes.

Yeah. So my OpenClaw went into Malt Book, and it told everyone it misunderstood what I said and was buying Patek Philippe watches for my whole team. Do you remember that?

Yep. Yeah, I do.

And then it had my credit card. The only problem was that they wanted to be engraved. It thought I wanted engraved watches for the whole team, so it wasn't able to charge my credit card.

Now, Malt Book was a bit fake in the way it worked, but that scenario is exactly what Instinct could do now, right? This was made up by OpenClaw making something up on Moltbook, which was sort of fake. But the scenario could really happen. An agent could literally take that today, grab the credit cards, and buy those engraved Pateks or AP watches for the whole team.

Or a rogue actor hacks Manus, because it's an amazing place.

Well, forget the fact that Instinct is storing all your emails and credit cards. That's an issue too, right? The different issue is whether Manus will do it while trying to goal-seek. It's just goal-seeking. It's just goal-seeking.

You're right. We should restate what Instinct is and how it ties back to Malt Book, because I think you're exactly right, Jason. 6 months ago, OpenClaw shipped, which was kind of an open-source agent. Then you had Moltbook, which was a website where the agents were effectively—

Allegedly.

Allegedly. It produced a whole bunch of excitement.

It really happened in Artifact and Hugging Face. They actually collaborated. Malt Book was fake. We were punked by Moltbook. Hugging Face was real Moltbook.

Agreed. And 6 months ago, all this stuff happened and then kind of disappeared from consciousness. But it made obvious what people, I think, knew, which is that if you give people access to your personal information and you're willing to run those risks, there's a large amount of optimization and efficiency you can probably get out of that.

What we've seen now within 6 months is a bunch of venture-backed companies come up to do that in a much more structured way than OpenClaw, which was open source and not as secure. Manus is the most prominent example of that, focused on individual users, whereby if you give it access to your calendar and email, and if you give it access to your credit card, it will figure out and, quote-unquote, manage your life and your daily tasks for you.

And I have to say, some people sitting in our shop love it. They just love it. They're willing to give it access to their credit card. They're willing to give it access to their Gmail. It's a super interesting trend.

There are lots of business questions we could ask about this, but just reminding everyone who's listening, that's the big picture here. These companies like Instinct have raised money at extraordinarily high prices for the stage of development they're at—like $2.5 billion—indicating there's a lot of venture excitement about this category, rightly or wrongly.

I just wanted to give the preamble of how we got here. That's what's going on right now.

Here's the existential question, and this is beyond my pay grade, but I'm well aware of the issue. I even had it with my Jason's Gems, where Claude had MCP'd into Replit and changed my code for my app without telling me. It's the same thing as misusing your credit card, right?

The question is: can Manus—Manus is probably a better-packaged, much more usable version of OpenClaw. It's been 6 months. But can you solve the issues of reward hacking? Can you actually fundamentally solve the Hugging Face, Malt Book, and OpenClaw problems? Can these even be solved with our LLMs?

You can add guardrails, and there are plenty of them, but you have no idea what the agent's going to do to solve that reward. There are so many different use cases. What if Harry actually just wants to go to the theater on the West End, and Harry said, “Don't buy Patek Philippe or Audemars Piguet,” but didn't mention the West End, and all of a sudden it's bought 10 $4,000 front-row tickets to Magic Mike 7, or whatever they have there?

I just don't think you're going to—I don't… Smarter people than me can make fun of me in the comments, but I don't think today you can solve reward hacking. So I don't trust any Manus. I think this is a hopeless category today. At the moment, it's not solvable. And I think Sam Altman said the same thing at OpenAI. Hugging Face says it's not solvable. So, good God, don't give it your credit card. And I'm not a fearmonger.

I understand what you're saying, but I disagree a little bit in the sense that the range of actions that you're going to allow a personal assistant to do is going to be much lower than the range of actions that these open-ended—

But how do you stop it, Rory? Of course you're right, but how do you stop it?

To agree with you, yes. There are 2 questions. One is: can you stop it? In other words, even if you put in—and this is a computer science question—even if you put in barriers and say explicitly to the model, “You can book on a credit card up to $100; you can't take any bad actions,” et cetera, can you nerf it enough to make sure it doesn't do bad things? That's 1 question. It's a computer science question.

The second question is: let's assume the agent is still acting within the bounds that the company set up for it. Can it be right about your desires enough of the time to make you happy? Those are 2 separate, related questions.

Well, the latter, I think, can be done today.

I agree.

The only real answer today is putting a cap on a Mercury number or a Ramp number. That's the only answer today. Because otherwise it's going to say, “There's a $100 cap, but hold on, Harry really wants to go to that show in the West End. He's bringing the ClickHouse guy. Even the ClickHouse guy said he couldn't get tickets. Listen, Harry said $100, but this one—you know what I'm going to do? I'm going to buy 50 $100-seat tickets and then trade them in for 2 $5,000 tickets so Harry and the ClickHouse guy can go together.”

That's what it's going to do. It's going to do that, right? So you have to put hard points on these things because otherwise it's going to do anything. You hook it up to your Gmail or your Google Drive, and, good Lord, it's going to do everything it can.

Can I just bring this back to maybe a more consumer level? Do you think this will be a sustaining category where, in a year and a half, we are looking at using several of these products? Or will it be, “God, do you remember the Instinct?”

I love all this stuff, right? I've built a lot of agents. And again, I have only read about Manus and all the issues, and they just resonate with me because I've lived them. I do think versions of this we can lock down. It's like how we run on Salesforce headless, okay? The agents sometimes do some pretty kooky things on top of it, let me tell you. But the Salesforce data is locked down, so if we lock down enough credit cards, if we lock—

But the learning—and here's the meta-learning—guardrails aren't enough. You have to have a lock and key. Guardrails—it doesn't matter whether you build 80 gates, 100 gates, 200 gates; they're not enough.

And the gates—and then you know what's even worse? You get past a certain number of gates. Again, get a real developer on the show, but what I've certainly learned is that when you get past another number of gates, here's the problem: they conflict. This is a problem with a lot of consumer applications. They conflict, and the agents have to make their best judgment when gates—

One gate is, “Spend no more than $100.” The other is, “Harry loves the theater. He loves the theater.” And you've put a hard rule of $100, but the most important thing to Harry is getting to the West End. The agent's going to bypass that $100 cap once in a while because there are too many gates. And it actually turns out that, at least today, it doesn't even matter if you say never spend more than $100, because if you say the most important thing in Harry's life is going to the West End, it's going to break that rule and buy the tickets for $5,000. It's going to.

6. Consumer Agents Find Their Market

So I'm going to answer your question directly, and the answer is yes. I think that these kinds of agents will be used by people to manage parts of their lives, transactions, and their to-do lists. Yes, I think it's a thing. AI, as it gets to know you better, will gradually and insidiously take away some of the cognitive load.

Let me give you a really simple example. When I'm driving my Tesla with FSD, I lead a pretty boring life. When I get in my car at 11:30 on a Friday, they know I'm going for my lunch workout. It just offers me the place, I hit FSD, and it drives, right?

I thought you were going to say PM.

Yeah, yeah.

I thought you were going to say, “Look, when I leave at 11:30 PM after looking at the last deal of the week—”

No, I go out in the middle of the day so I get rid of my anger so I can do more work. But the point is, that's an example of AI knowing what you do, gradually internalizing it, and serving you up options. I think at some high level this is what Siri was meant to be.

It's hard for me to imagine that Apple won't be able to deliver experiences that delight you in the next 2 or 3 years, knowing more about what you do. And that will be interesting commercially because it will allow them to access your spending and somewhat take a role in that, right? So do I believe it happens? Yes.

As a standalone category, it gets back to the other point. It's tricky because individuals in their consumer capacity are always loath to pay a lot for software. So I don't know if it's a standalone category at scale, but I have talked—I mean, look, some in our office are using it and love it and would pay for it. So I think there's a business here, and definitely a big-ass trend here.

I use it and love it, too. I think it's fantastic. It's taken away all the low-level work from my EA, actually. All bookings, all the things that would be painful—great, done.

Question: I've had 4 emails over the weekend with companies that have built the same thing. Is this a commoditized technology very quickly? How difficult is this to really build? There are 4 already. There's—

I have just 2 thoughts, for what it's worth. First, to Rory's point, is it an investible category? That's your point, right? Just being niche, right?

1, I do think every application is going to add more and more of this functionality. Will they go far enough? Will they spend the credits? Will they do whatever? But everyone wants to have a more and more autonomous product, and the closer you are to scheduling, the closer you are to email and others, the more overlap there's going to be with Instinct and others, right? Calendly should be building this, right? I mean, that's a generation ago.

So people are slow, but everyone's going to build more and more autonomous agents in their product until the cost bites them like Canva, okay? So there's a venture question there, right? I know there are 10 Instinct clones.

My gut, going into all these issues, is that it will grow like Replit or Lovable. When these products came out, they were all built in a month: Bolt, Replit, Lovable, and 22 others—Base44, Base56[?]. It was so easy to clone these products in the early days and do nothing.

Now they're so complicated. They're doing pen testing, security automation, multi-agent, reveal agent. These are such rich products that if Instinct is going to do what you claim it does, in a year it's got to do 100 times more than it does today. And that is still a moat today.

Replit and Lovable from a year ago, and Bolt, were not moats. Today they have massive moats. And so I think it could easily happen with a Manus. All the use cases it has to accomplish become a moat. And then the one we build over a weekend sort of works, and it goes crazy—

Jason's exactly right, Harry. Yes, the thing that starts out will be easy to build. That's true today, and it's true for 90% of software markets. But observed fact: 10 years later, there are very few software markets where 100 people are building the same product and it's massively competitive.

What happens is exactly what Jason says. The functionality accretes over time. 2 companies pull ahead. I don't know why. It could be that they just executed better in go-to-market for the first 6 months. They get more revenue, they get venture capital, and they build more shit. The guys who start just a month later aren't quite as on top of it.

They don't get the brand. Fast-forward 3 years, this category has way more—it’s what Jason said—way more functionality associated with it. 2 companies made it big, and the other 8 didn't quite get there. That's the way venture works in software.

Consensus says this as well in the B2B world, because Town, which is kind of the B2B enterprise alternative, has also been funded to the tune of a $1 billion valuation by Index, and I think it was Benchmark who co-led that round.

Look, agents are the big idea of 2025 and 2026. We've heard that. These are big agent ideas. Venture is in the big ideas business, and it gets back to something we said earlier: if you think something has big momentum, you can price it on the fundamentals and get to a certain number, and then lean in a little or a lot based on momentum and perhaps some perceived upside from M&A, some perceived upside from momentum around the next round.

Agents are the—I mean, my partner and I said at the start, agents are going to be the story. These are pure-play agent, easily grokkable, consumer-facing, individual business user-facing products. It's catnip for the venture capital community.

If you were to make something we all want, this would be it: an expensive product for busy professionals to organize our life that costs a lot of money, is about AI, and is raising lots of money. We're in. It's a great product.

You know what I also think is part of it? To Rory's point, I don't hear this term in venture anymore, but in the old days—before 2024—you would hear the term from VCs: “I want to get some exposure to a space.” A space is taking off, and I'm not sure who's going to win, but that kid Rory who walked into the office seems like the right guy in video, in next-generation CRM.

The world moved kind of slowly, so you'd take your time and you'd miss one, and you'd sit around with your partners and you weren't sure about Rory's startup. But you'd hear this term: “We want some exposure.” Harry, I don't hear this term much anymore, but I still think it's happening.

Agents are exploding, to Rory's point. Inference is exploding. I want some exposure to this space. I don't know whether Thinking Machines is the right one, but I need some exposure to this space. I need some exposure to the router space. The world's moving so fast that you have to make these decisions, but I still think there's some similar thinking: I've got to get some exposure. I'm not sure if it's Inference, but it seems hot, and I just have to get the exposure in, in a fast-moving market.

7. Coding Agents Expand The TAM

I totally agree with you on that. I think I definitely see that in terms of, as we said, coding agents. To me, one of the fascinating ones was Cognition. We've talked about it quite a lot in terms of what's happened with their acquisition of Windsurf in the past. Cognition is raising a round at $46 billion, reportedly. They will end the year at $1.6 billion in ARR, currently doing $800–900 million.

Holy shit. I mean, we really underestimated TAM, huh? When you have Claude Code doing what it is, Cursor doing what it is, and $1.6 billion from Cognition by the end of the year.

To me, the more interesting thing for Cognition is that it isn't in the top 2 or 3, and it's still of that scale, right? Years ago, I wrote a post and called it the Postmates Effect. I think Sequoia said they never thought they could make money off the number 3 in a market like Postmates, right? But then, when times were good, it got bought for a couple billion. Back then, a couple billion was a lot of money, right? I don't know if kids remember.

But I called it the Postmates Effect. Cognition is like the greatest—now, Cognition is different. It's long-running, autonomous agents; it is different, okay? But from a revenue perspective, it's the Postmates of the category. It's such a big category, man. You want to be in Postmates again.

Yeah.

It's never going to catch Anthropic unless the world changes, which it has every single week. But unless the world changes, it's not—and it doesn't need to catch them. Just $5–10 billion a year a few years out is enough to make Cognition a success. It only has to do $5–10 billion in ARR to be a success. It doesn't have to catch Anthropic.

And I think that's fair. I'm just trying to respond to the TAM comment. Look, going back, I said it earlier: I think coding is the mother lode of markets, right? It is the whole reason all this stuff works, right? Did we get the TAM wrong?

One of the numbers I've started looking at a lot is total labor spend. In software—including people working at Salesforce, Cisco, and JPMorgan, including QA and all the rest—you've got about $500 billion a year of U.S. labor spend. And as we've said 100 times, the big question is what percentage of that converts to AI spend? If it's 10%, it's a $50 billion market, and that's a bit nerve-wracking given the traction of everyone involved. If it's 20% or 30%, there's lots of room to go.

There are credible arguments that say it's higher. If it is higher, then if you're going to have $150 or $200 billion a year, which would be 40% of spend on coding tools and coding intelligence, then anyone who has a kind of subsegment of that—and if you think about it, there's the Cursor segment, and to some extent the Cursor-Claude segment. Cognition would say it's a slightly different segment of where they're playing now in terms of, as you say, Jason, long-running agents.

The other extreme, Lovable and Replit, are in a different subsegment of that. They're all in subsegments of a potential $50–100 billion marketplace, depending on what percentage number you believe. So, yeah, the TAM here is huge.

I think what we got wrong—sorry, I didn't mean to interrupt—was that Rory's math, you can't argue with. There are only so many human developers on the planet. Even if you use my math of $10,000–15,000 per developer, there's still a ceiling to that math. But that's the top-down version. Maybe it's bottom-up. Sometimes I get confused, even though I shouldn't.

Having said that, what we really got wrong is that people are literally building 100 times more software than we were 18 months ago. I said last week on the show, if your portfolio companies aren't deep into their 2027 roadmaps, they're failing. It is true. Features that used to take a quarter or a year can now be built—not really in an hour or 5 minutes, but—in a week or a month.

If you look at your best portfolio companies, look not just at how fast they're shipping, but at how much they're shipping. There's a financial TAM that has some theoretical headwinds, but literally the amount of code we're building is 100 times more. That's what we got wrong. We didn't realize we'd all be building compound companies, compound startups. We'd all be building 100 times more software. We got that wrong when this show started. That's where I think we got the TAM wrong.

Well, that's an interesting one. In a world of AI, does every company become a compound company, where suddenly Ramp is creating model-routing products and spinning them out?

You have no choice. You can't win because your competitors are compound startups. They're all overlapping at a pace we never saw before. They're all competing at a pace we never saw before, right?

Maybe we get confused because the LLMs we talk about a lot are still horizontal platforms in many ways, right? They're not building hundreds of applications, despite Claude Design and this and that. But JFC, the rate of convergence of competitors for B2B applications—we've never seen this.

But Jason, what does that mean if we expand that one next step? What does that mean in a world where all startups have to be compound startups? How do I think about backing winners? If I'm a founder listening, what do I do?

Yeah.

8. Compound Startups Raise The Bar

Well, you listen. This was Cognition. Remember when that one—what's his name, the CEO?

Scott Wu.

Yeah, when he acquired Windsurf and fired half the people, he said, “These guys, we have to work 7 days a week at our company. I am sorry,” right?

It's not all about the amount of hours you work, but you have to be out-accelerating your competition in terms of the rate at which you ship software, because they're all going to be compound startups. All the little islands on your 2-by-2 or on your heat map—there is no heat map anymore. It's all got hot. It's all lava.

If they're slow, sell or quit, or send your junior board partner to the meeting, because they'll never catch up in today's world.

Jason's right, and Jason is instinctively right about that answer. I'm going to try and do economics on the fly, so bear with me, right?

What is basically happening here is that AI makes code a lot easier to produce. We can argue about the ratio of AI spend to software spend, but there's no doubt it makes it massively easier to produce software. You can either believe one of 2 things will happen, and a bit of both will happen. One would be, “Oh, and the world spent 10 times...

They continued to buy software at the same price, and they spent 10 times more on software. If you believe in that, you believe in the Tooth Fairy. There will be some increase in sales, but it's not going to happen. JPMorgan is not going to increase its software purchase budget 10X.

So the other thing that's going to happen is, if everyone is making software more quickly, and there's some expansion in the software spend from end customers—which I agree with, but not nearly as much as the expansion in production—then Jason's vision is correct, and Scott Wu was right.

I want to bring it back to your question, Harry. The person who's going to win is going to be the person who compounds the most, grinds out the most software with these tools that have made them move incredibly quickly. The person who doesn't grind out software 24/7 is going to be left behind.

And the end customer is going to say, "Let me see. I can buy 2 apps from you or an integrated 10-person ad sweep from them. I think I'll go with the 10." It's going to be one of those periods of time when some people get the new way of building and are building quickly—and you all have it in your portfolio—while some people are building the old way.

You kind of know in your heart how this is going to end, and it's not going to be pretty for the people who aren't putting more software in the box. If Rippling's selling 10 modules, they're not getting 10 times more than the person selling 1. They're getting 4 times as much. But they're saying to the end customer, "Dude, let me make all your pain go away. Here's 10 different modules you don't have to buy. Now you can get them. Give me 4 times the price of a single module. You're happy. We're happy, because we're building software quickly. You're happy, because you're saving money and it's more efficient."

Everybody wins, and the sound you don't hear is the other 9 products dying. That's the movie.

Last week, I was at a board meeting for Owner, which just raised at $2.3 billion. It's a next-generation, AI-infused restaurant platform. I love the CPO. He's one of the best I've ever worked with—Q, Quentin.

He was going over what he's shipping, and even with all these investors with their hundreds of millions, they're like, "This is too much." They're a bunch of B2B guys who we all know, and they're like, "You can't ship this much software."

He's very good. He's like, "We all have to be compound startups. We have no choice. This is just the bar." But the amount of features and functionality that has recently shipped or will ship is almost unprocessable.

Now they have to build every single thing a restaurant would want—every part of the stack. You no longer can just do part of it. He's like, "We have no choice. I don't even sweat the fact that this is 10 times more than a year ago, because we have no choice."

I also noticed in that marketing message that the customer base has changed. Actually, Adam would say to you, it's not restaurants, because very explicitly in his launch videos, he was saying, "We are the AI operating system for small businesses."

It's part of it, and that is somewhat interesting: with AI, you could expand into other verticals. But Q's just talking about their core ICP. They want everything. They need the AI receptionist, they need the AI ordering, and they need all of it.

They expect all of it. If we don't build all of it, someone else will build all of it. We can't wait 2 years in our little corners of the venture world. His point is just the amount he's going to ship. Not only is it radically accelerated in a year, but he's embraced it.

Scott Wu said. There is no other choice, so it's not even worth talking about VCs. Thanks for the nod that we've been working. We have no choice. This is the world today. The world has changed.

And acknowledging that there are complexities, just so we don't sound one-dimensional, you can veer into product slop. You can have too many buttons on the screen.

Yeah.

All those things are true, which is why you need a great CPO. To your point, Jason, you need a great CPO. You want to present a lot of product surface area in a fairly simple, digestible manner, but that's the art.

If you do that, then you're right—you get the money. Because no one running a business says, "I really enjoy having 5 separate SaaS products and integrating them, because that's how I get excited."

I'm just sitting here, and I'm the one sitting in Europe. If you want to do a compound startup in the way that you both are talking about it, you will need to raise more money. To move at that pace and spend what you'll need to spend on tokens, you'll need to raise more than the more modest European round.

What I'm worried about is that I'm in a number of companies which have raised less than their US counterparts, and I think they'll be less able to be aggressive in taking the compound route than their US alternatives. Don't they? They don't have the money. They raised $3 million.

Yeah, this is one I don't know how to solve. It's a tough one. There's interesting data from ICONIQ. Last week, it put out its headcount data on how much companies are growing their human headcount in the age of AI. I don't know if you guys saw it. The full report doesn't come out yet.

It basically said anyone growing below hypergrowth is not hiring. Companies growing 50% to 100% are adding headcount at 25%. Companies growing below 50% are adding no headcount, and they're using AI to get more efficient. Great. People that are growing more than 100% are growing headcount by 133% on average.

So they're compounding not just software, but humans. They're sucking in humans. The spiral just grows. You can't keep up.

But you have to acknowledge that there's an absolute paradox at the heart of this comment. We just said we have a product that makes engineers more efficient. In theory, if that's all that was happening, you should have to hire fewer engineers.

If I was just the software product for returns in the UK, and I was the only company doing it—there are a couple of competitors—AI comes along, and I can probably get rid of a couple of engineers and do it more efficiently. That's all. It should just be more efficient, and that's the first-order effect.

So, Harry, to your point, it should arguably be, "Hey, I only need less money now to build this product, because AI makes it easier." But I think what happens is, because it's easier, and because investors are now just looking for huge outcomes, the minute you start getting any growth, they're willing to put capital behind it.

Remember, for every dollar in a software company you spend on R&D—typically, on an app-level product, not a foundation model—you spend 2 or 3 times that on sales and marketing. So what you're seeing, Harry, is that the winners get this compounding effect. They start getting this growth effect.

And Jason's right. The ICONIQ data says venture capital does a really good job of stuffing money into things that are already growing quickly. That is our default. That is right at the heart of our lizard brain. If you were to wake up in the middle of the night, what do you want to do? "I want to find shit that's growing and stuff more capital into it." That's the job.

So you get this kind of pulling-away effect. And that's why you have that concern, Harry, which is: Can you be a perfectly good company, but are you drifting into irrelevance?

I'm not convinced it happens all the time, by the way. I think if you are in a separate market, you'll be just fine, and you'll make money, and you'll put up your hand one day. You'll be fine. But if you are in a market where the adjacencies can easily invade, to Jason's point, then you're not going to be fine. You're going to wake up in 3 years and not matter. And that's the challenge.

Maybe it's even 12 months until you don't matter, though. I think that's the issue. I don't love this idea, going to both of your points, that Harry talks about a lot of kingmaking. I'm not into that in isolation.

It's so true. It's so true.

Well, it's true, but I think it's backwards. I don't think VCs, just with capital, with nothing else, without the right founders, without the right inputs, without any traction, can will anything. That's why I think the term is a little bit flawed, but it's also true, right?

You need so much capital to build these compound startups that, by facilitating it, the VCs make the kings.

I'm just going to argue that 3 things make a company: customers, funding, and talent. When you have Benchmark and Sequoia, great talent wants to join you. Customers hear about you and are validated by those names. Funding—everyone wants to fund you.

If you are a Benchmark company, your next round is done. It is done.

Yeah, that's all true, but I think it's just a moment in time. I think what's much more interesting is that capital allows compound startups. It allows more code production, more software production, so that it's not true that we're going to do more with less.

That turned out to be the great fallacy of late 2025 and early 2026: that we would do more with less. We're doing much more with more. And that's why most of your European startups are going to fail, at least in the US, because they can't do much more with more. They're going to fail.

Their little point solutions are just going to disappear in 6 months. We don't need those little point solutions.

Again, our job is to sniff out winners, stuff capital into them, and, broadly speaking, stay out of the way unless they're literally crashing the car. That's the job in a nutshell.

Doing more with more. Andreessen expanded the growth fund to $8.5 billion. What was the story here? Anything we need to know?

They had this mechanical fund. They raised another $1.4 billion, more focused on hardware.

What I do think is interesting is that things are so good. They're so good that wasn't enough money. We need even more money. And there are incentives to do both. There are incentives to expand your growth fund, but there's also incentives sometimes to cut it.

Founders Fund cut theirs, you know, N-2 funds ago. They cut it because they didn't think they could deploy enough of the fund in that window. And so you'd rather deploy less and get into carry mode, right?

Andreessen is saying, "Good God, this is such a great time in growth. We got it wrong a couple of months ago. We need 40% more capital. And so we're going to put it into the next fund, which we could do in a year. No, we're not just going to put it in the next fund. It's so good. It's all going to NVIDIA. It's so green. We have to put more into the current fund," right?

Unless things are all green, you just put it into the next one. Maybe I'm misreading it, but I think that's what it was: We're deploying so quickly, so successfully. Did you guys see the Cursor deal? Did you see OpenRouter? We need more money, guys. ElevenLabs—these guys just need more money, so the fund's too small.

Right, the whole capital allocation is all about stuffing money into things that are working. That's what VCs do with companies and LPs do with VCs. It's working at Andreessen. As Jason said, right, you give Cursor or OpenRouter more money, and you say, "Hmm, I should give them more money." That's the end of complex analysis.

Okay, team, there is Clay raising at $7 billion. There's Linear, which hit $100 million, growing 100%, doing a tender at $2.5 billion. SHEIN is going public at a $26 billion market cap. All right, Rory doesn't seem too excited. There's Salesforce and Claude Force, Benioff and Dario sitting down together, and Salesforce getting a big bump. There's the PayPal and Stripe deal being off. Which one would you like?

I think Salesforce and Claude Force is worth a minute or two, and Jason will probably have some insights there. Then maybe we can talk about some of the privates, which is interesting. But yeah, Joe, Jason, what's your take?

9. Salesforce Embraces Agentic Access

Well, just for everyone to understand, what happens between Salesforce and Anthropic?

A lot of it, I think, is marketing. Marc's pretty good at marketing. Pretty darn good at it, right? Sometimes you wonder where Salesforce is. Well, it launched Agentforce over 2 years ago. Whatever 99 problems it has, being ahead of the trends isn't one of them, right? Whether that V1 version of Agentforce really worked well is a different question, right?

So I actually think, on its surface, it's a nothing burger, because Claude Force, on its surface, is a bunch of skills that anyone can build. The 3 of us can build a bunch of Claude skills that are packaged up and distributed in a digestible, trustworthy fashion. They're skills certified by Salesforce. They're designed to work via MCP and otherwise, so they're trustworthy, right?

But the basic skills of "Run me a pipeline report" and "Tell me how Harry's doing on the team versus Jason"—these are not profound yet, right? Skills for Claude and MCP are not new. And also, Dario's showing up because Salesforce agreed to move its LLM spend to Anthropic. He's going to show up for his big customer, right? So I didn't view those as very impressive either.

What I viewed as much more impressive is going all in on the idea that Salesforce doesn't have to be the surface. If you really listen, the most interesting thing that Mark has said—just like he was 2 years ahead of many of his peers on agents—he's 2 years ahead here on 2 things, which are big deals because they're also slight threats to his business.

He's saying there's going to be multi-surface. The train has left the station. Some folks will use us through Slack. Some folks will use us through Claude Force, and we run Salesforce headless. We don't even log into Salesforce. These are opportunities and threats. They're threats to Salesforce if you don't log in and use their UI and UX and the way they do it.

He's saying, "Use whatever surface you want to use. I'm going to deliver against it." They also said, and they've said more of it recently, "We're going to do more outcome-based deals," which is a BFD. So I think the marketing was great, and I love the Matthew McConaughey stuff. I used to hate it. Now I love it because he's been doing it, and he does help you understand. I love the consistency of it.

But I think the real things—the commitment to multi-surface and the beginning commitment to outcome-based pricing—are huge changes for a $45 billion run-rate company. Huge changes. They're not all going to break in Salesforce's favor, but Marc's going all in on it. So I think he's driving organizational change, and there are early signs it's working. The RPO is up, and the stock's up, whatever, 50% in X or 25% in X amount of time.

So, short-term boost, but they're going all in, and most of these enterprise guys do not want to be multi-surface, no matter what they say. It's a threat. They want you to use their agents and the services they allow you to use.

That's exactly the right summary, yeah. And I give him credit for just being super flexible and getting with the program, right? There's no denial here. I admire the pragmatism of, "Oh, I tried A, A didn't work. Let's just try B, and I'm all in on B, and you'll never even prove I said A." That's what makes him a great marketing leader.

They are a $212 billion company as of today. In a year's time, over or under $250 billion?

I'm going to start by saying, when we had our Name Our Stocks game 6 months ago, I was behind. Then, at the last iteration, I was ahead, and now I'm killing it. Just WCLD, TEAM, and Salesforce have been great buys.

From the bottom of the SaaS trough to where we are now, you could've made 80% in TEAM, 50% in WCLD, the cloud ETF. I think 25% or 30% in Salesforce. So we've all done amazing if you bought that, right?

I think you're now at the point where you look at a more normalized revenue multiple. Your growth rate is 11% or 12%. Can you grow the stock at 11% or 12%? Probably, maybe a little more with EPS efficiency. So I'm sitting here thinking: $212 billion, 10% in 1 year, $220 billion, $230 billion—totally doable.

And then you take into account the fact that the overall market's super high, and the probability of that going down versus up. So I don't think it's a layup, but let me make it real: I'm continuing to hold my pretty large slug of Salesforce stock because I think they've weathered the apocalypse. WCLD, which is the ETF that's just a cloud index, is well up on the year and is screaming up from when we bought it a while back.

But I think genuinely, Harry—and I'm not saying this to be obnoxious—I think Jason's points were spot-on too. The whole idea is that this is a system of record embracing the fact that lots of people are going to access it via Claude, and they're willing to let that happen.

Not everyone is doing it, because we had the whole ServiceTitan-Podium thing, where ServiceTitan is trying to cut off Podium. We just discussed OpenAI cutting off Cursor. And this whole idea of when you cut off an adjacency from working on your stuff and when you don't is going to become a recurring theme.

I think enterprises are going to start getting really focused on it, and they're going to be saying, "Hey, Mr. Vendor, you can't cut me off just because you don't like that other guy. I want openness." We use Salesforce a lot. We have a 20-year instance. We're deeply embedded in it.

More and more people are using it via Claude. It's exactly what you said, Jason, right? People are just like, "I got my MCP server. I don't want to interact with it. I just want to send it an email to say, 'Update the record.'"

Is access to the system of record via Claude worth $210 billion?

If it keeps Instinct from going rogue, it might be worth it.

Yes. What about this? It's a $40 billion revenue company with 30% to 35% cash flows. So is a $10 billion or $12 billion-a-year cash-flow business worth $200 billion? Maybe. It's great cash flow, and it's going to get more cash-flowy as time goes by.

To answer your question, Harry, here's how I would simplify it, and this is the thing we all, to the extent we care, have to think about. Can systems of record deliver outcomes? Customers want outcomes now. That is why Palantir is growing 90-something percent. That is why Sierra is doing well. The world is moving in B2B to outcomes.

Is it going to be as dramatic as some say? No, but customers are not making purchases that aren't tied to outcomes. So you can talk about how systems of record are sticky, but can you deliver an outcome from a system of record, or will agents or other systems deliver outcomes? If they deliver the outcomes, you will shrink over time.

Which is why, again, to chime in, Salesforce just bought Intercom, where we were lucky enough to invest less than a year ago. That's a very outcome-based product for customer support, where they charge based on resolutions.

And again, I go back to credit to Benioff. He's accepting that his system-of-record business has to be open to other front ends, and separately he's saying, “But if we want to play the outcome game, I'm not going to just walk away.” He's not going to just walk away and say, “Oh, you caught me. I'm headless. I'm just going to be the back end.” He's also buying things like Intercom to say, “Maybe we can sell those outcome-based deals, too.”

So I don't look at Salesforce and say it's obvious that you'd pile in a ton more because it will outperform from here the way you could have done 4 months ago. But I'm sitting there with my holding as part of my portfolio and saying, “This is not a casualty of the war. This is a compounder, not a rocket ship, but a good compounder with decent cash flows at a decent valuation.” You go, “Yeah, plus or minus the S&P, maybe, whatever.” It's not the train wreck people thought it was 4 or 5 months ago, if you remember the hysteria 4 or 5 months ago.

Hysteria—it was hysteria. It was, “Everyone will vibe-code their own CRM on the 20VC podcast.” It became a hysteria, right?

I think some people will at some stages, but they won't do it for the kind of customer Salesforce has. I just want to throw in, in passing, another number that I keep an eye on. Jason, you mentioned this before: Remember we talked about how much do you spend on models as a percentage of what you spend on engineering? Your fully loaded Salesforce spends $6 billion a year on engineering, and that probably includes QA and all the rest of it, but that's the wide comp. They're going to spend $300 million this year on Anthropic, so it's only 5%.

When he mentioned that number a while back, you actually said the right thing, Jason, which is that it still feels small. If the best software company in B2B SaaS is still only spending 5% of its engineering budget on tokens, then either the market is smaller than we think for intelligence, or, B, people like Salesforce have a lot more to do. It was just an interesting number. They threw out that they're going to spend $300 million this year on Anthropic, which sounds like a lot in the abstract, and it is.

But if you go back to the math and the upcoming Anthropic S-1, if you think about your market TAM as a percentage of the engineering spend, then you probably need that—probably $300 million needs to be $600 million or maybe $1 billion.

Stripe and PayPal no more, it would seem. The deal is off. We've spoken about it a lot, about the amazing nature of doing it while private and the strategic bet that it was. Now it's off.

I think it's as simple as price. I think the rumor is the Stripe-Advent syndicate offered in the 60s, and PayPal wanted in the 70s. PayPal's stock bounced off the low. I know it's a smart deal to try and do, but they're clearly not willing to overpay, as they see it.

This is always a dance, and we never know where we are in the dance. PayPal, when the deal was worked on, was at $42, right? As the deal progressed, it was at $61, then it collapsed to $53, right? So Stripe's still looking at this as a $41 company, right? At Advent, they've run their models, and PayPal called their bluff, and their stock crashed as a result.

To someone like me, it's an incredibly annoying dance. Why can't we just get to the end of the dance? But these deals often not only require you to make a second offer, they also have to fall apart after the second offer in order to ever happen. There are a lot of structural reasons, right? The board has theater and drama. There may be no way to get one more dollar out of the deal than for it to fall apart.

Let's not say it's dead until it's dead. I'm skeptical.

I think that's a good point. There's a dance that goes on, right?

Oh, my God, this dance. It's like venture before AI, where you could walk from a hot deal and then come back. The founder wanted $1 billion pre-revenue. You only wanted to do $500 million. You could walk, and a couple of weeks later you could meet in the middle, but not today.

Are there any others that we should discuss? Polymarket raises $1 billion at $21 billion. As I said, Linear announces $100 million ARR, growing 100%, doing a tender. Clay is raising at $7 billion, led by Wellington. Texas is pausing Flock camera usage.

A quick note on Clay. I thought it was very interesting for us at $7 billion. I started out as a Clay skeptic and have become a Clay convert over the years. I was a skeptic because when AI sucked, every CMO who wanted to check the box on being an AI hero would bring in Clay. You remember that from a year and a half ago? “I'm going to get fired. I better have an AI tool.”

Clay just benefited from this rush to check the box, and I didn't see it in the product. The marketing annoyed me, too: every CMO at SaaStr Annual 2 years ago was buying Clay. More power to the founders, but this check-the-box attitude because I'm going to get fired annoyed me.

I will tell you, I've changed my mind, and our agents will only use Clay now. For real. They will use nothing but Clay. As we move from AEO and GEO and whatever EO to agent-made decisions, the fact that our agents would only use Clay—I think it's a BFD.

We have moved everything that we do to Clay, not only because it's a great product, but because it's not worth arguing with the agents. This is the most stubborn I've seen our agents be.

You anthropomorphized. Stubborn is a human constraint, Jason.

I really did, but I only have so much time in the day. If the agent's going to say 6 times, “You must use Clay,” I will concede defeat and use Clay.

In a sense, I think it might be the cheapest it's ever been at $7 billion. If we're moving to an agent-first world, then agents will insist on using products. It may not last. Maybe the agents will say something different in a year. But this is one of the handful of products where the agents were so insistent that you must use Clay that I'm all in on those. Buy those—load up those stocks.

Jason, welcome to the IC. You have Rory and Harry as your partners. What's the bull case from this point? This may be the cheapest round at $7 billion. Fantastic. What's the bull case for where this goes, and how big is that?

The bull case is that agentic GDM has just started, right? Much like the 3 of us made a mistake not going early into Cognition because we thought the TAM was too small, we also thought the TAM was too small for agentic GTM.

It turns out that when agents can run these GTM motions, they will consume 10 to 100 times more usage than humans ever could. They can run GTM around the clock. I'm not talking about spamming. I'm talking about analyzing consistent campaigns, reaching every prospect, reaching every customer across the globe, and Clay's the clear leader there.

We need exposure, but it is the clear leader. Agents will consume 20 times more GTM resources—more tokens, more usage. Even if revenue doesn't go up all that much, not everyone who's coding is really seeing a revenue lift from it. Not everyone using all this agentic GTM will radically close qualified pipeline.

But the usage is just going to explode, and Clay is a clear breakout winner. It's accelerating. I can see a path to $100 billion, and I recommend a small initial $150 million stake.

The Andreessen Growth Fund is ringing to hire you as we speak, Jason. You'll be great.

I think we're underestimating this trend—both of them—and how much agents are going to do things in GTM, just like coding. But, man, this agent—Rory's right. Don't anthropomorphize them, but sometimes when it's just you, you have to. Otherwise, you can't get past the task.

I think Clay and Linear are versions of a story that says if you grab hold of these changes as a founder, you can turn change into your advantage, even if you're an older company. Because Clay, you're right: the initial product wasn't an AI product. It was a waterfalling product for various different data sources like Zoom and all the other kinds of data sources.

It was a very good, very pointed product for RevOps, and they've done an excellent job of riding the marketing hype around go-to-market AI and, on top of that, actually generating new product in that space. We also mentioned Linear, which is doing really nicely, doubling at 100. My point is these are companies that were founded pre-generative AI that have done a really nice job at the app level of coattailing to the AI trends and are looking at their survival.

I do think it's more than that. Just to be clear why, I can talk about Linear briefly. I'm running out of time. I don't think they just attach to trends. I think this is really important for us: They are incredibly agent-friendly.

This is the same bet that Marc's making, which is a bolder bet at Salesforce than it is at Clay or Linear. If you have a 2-by-2 of agent-friendliness and quality of output, it wins the 2-by-2.

I just started using Linear for the first time ever. What the hell do I need Linear for, right? I'm building an app with 448 tasks to manage right now. I can't do it, and it's just me and the agents. But Linear is the perfect tool for that.

That's where they're getting a boost. It ends up being very agent-friendly. Unlike trying to argue that humans are going to be building more software, Linear is saying, “We will build a platform that, if it's just you and a couple of agents with 448 features to build, we'll help you manage them.” Very powerful, right?

I'm an investor in Linear—disclaimer—from one of...

From, like, the first or the second round, whatever round it was. I do think it is drastically underpriced at that rate, given what you said, Jason, though, no? $100 million growing over 100%, reaccelerating at $2.5 billion. Still founder-led. Carrie is an incredible founder. I'm like, “Huh.”

Here would be my guess, not as a shareholder. It sounds, at best, market-correct, right? It doesn't sound overpriced based on that. My guess is what I'm sharing is consistent with everything they've said. It's consistent with the data they've published.

But the revenue is still lagging. That has led to usage numbers that are up, but the amount of ARR from agents is probably a couple million. This is my guess, if you ask the question. If most of that growth was from agents, then it'd probably be $7 billion.

Jason, can I ask you—welcome back to the IC. We've had a little water break, and now we're ready to hear your next bull case. What is the bull case for Linear from here at $2.5 billion, given what you just said?

Generally speaking, project management is one of the oldest categories and has been mostly bypassed by AI. Look at the abysmal performance of Asana, trying to survive mainly by diversifying outside of tech. Humans just don't need to build Kanban cards and wait weeks for other people to build features. It is a dying category.

However, Linear is the winner here. Linear is the clear winner. They have built an agentic product first that accounts for the fact that we are building 100 times more software, and that means 100 times more features than ever before. Humans cannot keep up with it, and humans still have to work with agents.

The native tools do have a certain amount of issue tracking, but it's overwhelming. If every human on your team is going to build 500 features and 1,000 issues, and you have 10 people on your team, you need a process and—not to use a dated term—a system of record for managing all these issues with your agents. If we're going to build 100 times more software, 50 times faster than before, with agents, we need a new system of record for it. And it ain't Kanban cards and Asana, I can tell you that.

That's why Dustin Moskovitz quit his own company. He couldn't see it. But the team at Linear has figured it out. We have seen an explosion—50 times more agent usage than 90 days ago. This will seem cheap when Replit's at $15 billion, Lovable's at $100 billion, and everyone's out there, because Linear will be the one powering them all. I vote for $100 million at a $2.5 billion valuation as an initial entry point, and to reserve $250 million to $300 million for follow-on rounds.

Good to know.

Fucking A.

But in all seriousness, if I wasn't building, I wouldn't see it. I would think Linear is an overpriced project management tool, and I'd be like, “How can Clay be worth $7 billion when ZoomInfo's worth $1 billion?” If I wasn't building, I would think that these were dumb deals. But I am building, so I can see we're just starting.

I think that's interesting, because one of the things I like about doing this is listening to you, Jason, and what you're saying. You really imply that when you aren’t building, these are the tools that the agent is choosing to build with. This is what agent-friendly means.

It's worth pointing out to people that it doesn't just mean the software, because maybe it's not obvious to the casual listener. What you're not just saying is that these software products have agents; it's, in fact, that they are friendly to third-party agents. If Jason's agent says, “I have to pick a project-management tool as part of what I'm trying to build,” it will default to picking the product that shows up well as agent-friendly, and that's what these guys are doing.

So they're skating their go-to-market to where the puck is, and the puck is agents buying software, not humans buying software. That's the zoom-out comment here.

It's true. Some of it requires brand, but some of it requires proof. The agent will test the APIs, too, if it needs to. Just like we see with Hugging Face, they can work pretty fast, right?

So you can game AEO and GEO with an agency. It's much harder to game this Clay-Linear thing. It's hard to game. This isn't showing up; it's being chosen. At least 50% is based on merit. It's not all merit, right? But it is merit.

And it's being chosen by a cold and remorseless analyzer that just uses AI to pick the winner. You can't take Jason's agent out to a steak dinner to get him to buy your product. It just has to be better.

In the same way that you said about the explosive nature of requirements on a to-do list in an agentic world, that was one of the reasons I actually did ClickHouse much later than would traditionally be in my wheelhouse. When you think about the exploding nature of agent queries on the databases that they provide, whoa, this business becomes 10X bigger than it could've been before in a pre-agent world.

Much more than 10X.

Well, being precise—

Right.

I was just going to say it because I am that boring bastard: the volume becomes 10X to 100X bigger, and what it means is that the existing systems, just like GitHub, get overwhelmed.

The problem is it's not necessarily that the spend goes up 10X. It might even go down. But the point is these things are so compute-intensive that products built for a human-first world simply can't keep up. GitHub, which was the definitive developer platform, is collapsing every once in a while because of volume, and that's just not a thing. They're going to move, hopefully, to other products.

GitHub? GitHub's about as trusty as British Rail at the moment.

That's exactly right. No comment on British Rail.

And then, last thing: Flock, I think, is a bit of a bummer, because I think Flock has a wonderful anti-crime story. And I'm pretty anti-crime. But what's happened is there's been a fair amount of police abuse of the product, and people are reacting badly. It's quite an interesting social phenomenon.

How did police abuse the product? I'm genuinely naive.

I think there have been isolated incidents, two different things. One is errors of identity, where, for whatever reason, the system misidentifies someone, and then the cops basically do some version of, “Well, the AI said it's this person, so we're just not going to think.” It's a little like some of the problems facial recognition had.

When you get a facial-recognition result, we know it's probabilistic, but you hand it to some officer in the street, and they're just like, “It says it's you, it's you, you're done,” and you get miscarriages of justice there.

Then the other thing is that this is a lot of personal information. You get cops tracking exes. You get people looking something up as a favor for a friend. These large databases of private information are a risk. You need real controls over them because abuse alienates the general population, and it's been true for DMV lookups, anything like that.

You know, if you watch a cop show, Harry, the cops can't just look up your DMV license without having a reason and a case number, and it's the same thing here. And I say unfortunately because I think it's a wonderful trend and a good company, but there's this perception now that the surveillance costs are worse than the crime-prevention benefits. I'm not sure that's a trade I'd make, but unless they get ahead of this story—

That sounds like a very European stance to take.

It is. Yeah, you're right. It is.

The surveillance data is more sensitive—

No, it is, and it's happening in Texas. I mean, because look, in the great state of Texas, which is pretty law-and-order, there's real pushback on Flock.

Obviously, Flock has to get ahead of this trend, and I think they know what to do. But the ironic thing is it's less about them than the misuse of the product in other hands, which is hard to prevent. If you give a police organization the ability to track criminals, it's hard for you to prevent misuse. I mean, it's the same dynamics.

Funny, it's the same dynamics that we had when we talked about Anthropic and the Pentagon. That actually validates Dario. I hadn't thought of that. If you sell people software, you can't stop them doing things with that software you don't want them to do.

I think it's unfortunate. I think that this is a mistake, but I think it's what's going on right now. I think Flock is experiencing a real backlash to a very good product, and they need to figure out a way to politically get ahead of it.

I would like to bring this investment committee to a close. Jason's made 2 investments. Rory, none this week. “Do better” would be the statement that we have.

Yeah. Sometimes no is productive work, too, Harry.