Anthropic 对决五角大楼：谁赢？｜数据中心军备竞赛｜终极选股

1. Anthropic vs. the Pentagon: The Billion Dollar Supply Chain War

Boys, we have a big week of news. As always, we're going to start with This Week in Anthropic. We have gone from a couple of hundred million dollars in lost contracts with the Pentagon, with the DoD, to potentially billions of dollars at risk.

If we start here, with the billions of dollars at risk, how did you analyze that? And maybe, Rory, the wonderful context giver, if you want to provide some context.

Rory O’Driscoll

Yeah, I will. I think that Anthropic, in my view, correctly sued the federal government in both California and subsequently in D.C., just for procedural reasons, basically saying that the government's designation of them as a supply-chain risk was incorrect, badly done, and should not legally stand up.

The reason they're suing immediately versus talking is that they're stating that quite a lot of revenue could be at risk for them because the Department of War and the government—I think, actually, I should say the administration—has articulated a very aggressive definition of supply-chain risk.

If you recollect last week, I was not entirely pro the Anthropic position, and I stand by that. But I think the government is now overreaching as well. So, like all good disasters, both sides are doing things wrong.

It's totally rational for the government to say, “You want $200 million of revenue from us, and you want to get in our shorts and tell us what to do. We're the Department of War, and we ain't going to do that. So you're not getting your $200 million. That's totally fair.”

But designating them a supply-chain risk has a series of escalating consequences. At a minimum, it implies that no one selling to the Department of War can use them for those contracts. No other company selling to the Department of War can use them for those contracts, which is one step more. Again, I see how the Department of Defense might get to that place.

But then you're seeing the government articulate this really expansive definition, which is: throw them out entirely from every government contract. That seems an overreach. Then, even more overreach—and it's actually not happening because the various cloud providers have pushed back on it—there was one version that said, “If you use Anthropic at all, Microsoft or Amazon, then we won't use you at all in the U.S. government.” That's an element of, frankly, almost ludicrous overreach.

What's happening here is that Anthropic and the Department of War are trying to figure out, in this spat, how big the blast radius of the consequences from this $200 million contract is. In my perspective, there should be consequences, but having it be this big a blast radius doesn't really stand up. I think the consensus is that, in law, Anthropic would probably win a good slug of this case, which is different from saying they're going to win the fight.

2. The Era of Gentle Deceleration Is Dead: Public Markets Turn Brutal

Guest 2

Well, look, there's something to me that's minorly interesting, and then something more interesting. The minorly interesting thing is how hard the hammer came down. I mean, they're going to lose. It's like tariffs: even if, for some reason, Anthropic wins their case, the DoD—the Department of War—will just come back with another issue. Just like tariffs, they'll find another justification to block Anthropic.

3. B2B Panic: Why Leading Companies Are Losing Deals to OpenAI

At a minimum, I think the real panic—and then I'll talk about what's more interesting to me—is that they didn't realize what a blocker this would be with their customers. This is a classic B2B sales issue. You walk into these big deals, and they said it in the complaint: deals are being cut in half, and they're struggling to close deals because prospects are worried that some of their business has exposure to the federal government and the Department of War.

Then their competition, OpenAI and xAI, says, “We don't have these problems,” and they may steal the deal. They literally said in the complaint that this was their issue, and this is classic B2B stress. The risk only has to be ambiguous in a B2B deal for you to lose the deal and go with another leader. It just has to be a little bit less secure, or they went down for 2 days, or the database was left open.

There was a classic B2B panic. I don't want to get into politics, but I honestly don't. The government is saying, “Listen, we need you for war.” In fact, it's more than that. The government is saying, “We used Palantir without Anthropic's knowledge to conduct a war in Venezuela and depose a government. We did that without Anthropic's knowledge. So it happened, but the government's going to say, ‘You're either going to agree to this, or we are going to freeze you out of the U.S. economy.’”

It doesn't matter what a court says because they'll come up with another justification. There'll be another one. There'll be a different type of supply-chain risk. It'll be endless until they bend to the will, I think.

Rory O’Driscoll

Yeah, but you may be right in terms of what the government will do. Having been critical last week, I want to say it's wrong to try and drive a wildly successful American tech company out of a good slug of the economy simply because they don't want to serve this one particular need.

It's interesting if you read the complaint. A lot of what Anthropic is talking about is literally their First Amendment rights. This is not a contract. What they're basically saying is, “We, Anthropic, articulated our less-than-love for some of the things that the Department of War might use this technology for, and you're trying to ruin our business simply because we used those words. That's our First Amendment right, and that's un-American.”

While I think you are entirely correct, it's a really tough position for Anthropic to be in. As I said, maybe—I don't know if they should have picked this fight. I also think it's particularly unfair of the U.S. government to try and be this sweeping in its actions against what is one of the great American success stories.

If AI really is all about the leading edge of the economy, it's probably a mistake to pound the crap out of one of the 2 leading companies in that space simply because, at the margin, you find them slightly sanctimonious, which many people do, and you don't like their politics. You don't have to buy from them. Put your $200 million in another place.

Maybe don't let them into the Department of War at all, but stop trying to cut them off at the knees across the entire B2B infrastructure. I think it's a little bit of an overreach. But I think you're right, Jason. Me thinking that is very different from them changing their mind.

There's a style in the administration that's prone to overreach, and I'm sure Anthropic deeply regrets getting down in the mud pit wrestling with this particular opponent because it sucks.

4. The Anthropic Endgame: Will Claude Eclipse ChatGPT?

What actually happens here? Realistically, in a year's time, does one cave and give up? Play out the realities for me.

Rory O’Driscoll

I think the interesting thing is that, on the merits of the lawsuit, Anthropic will probably prevail, and maybe even prevail quickly on some of the issues. I think they've actually moved for immediate redress, basically to get a reversal.

If I were to be logical, just based on prior cases, they might prevail on some of the legal issues. They clearly won't get a $200 million revenue deal with the U.S. Department of War.

But in a perfect world, if they bend the knee and prevail on the law, what we've seen in multiple other situations is that you end up with some kind of acquiescence to the administration in return for, “Leave us alone to pursue the rest of our daily business, our B2B business.”

The administration's not going to be saying to Microsoft, “Hey, if you use Anthropic anywhere, you don't get any government business,” which is the most overreaching version of this. In return for that, Anthropic will probably have to bend the knee and be slightly supplicant.

That's my gut. They'll win on the law, just like many of the other institutions have won some of the cases, but will find that they've picked the fight they regret and just want to settle.

Does this impact their going public later this year? Polymarket predicts Anthropic no longer does. And Polymarket and Kalshi know.

Guest 2

I mean, Polymarket and Kalshi were able to tell you the date at which we were going to attack Iran. The amount Polymarket knows is a lot. The only thing is, they're tied, right?

Rory O’Driscoll

The only thing I would say—and maybe there's a related point that's broader than this that I'd like to make—is that I do think we overstate IPO risks in general.

Rory’s been through this, right? Listen, it totally makes sense that people would react: “Oh my God, Anthropic has to work to perfection to pull off the IPO they want, given the amount of capital they want and a valuation of $1 trillion. Maybe the supply-chain risk would be risk factor number 1 and freak out the public markets.” I think we’ve all seen every public startup we’ve ever worked with IPO, with several things that had to be highlighted in their prospectus, and it never seems to be the end of the world. (Laughter.)

Rory O’Driscoll

Several things had to be highlighted in their prospectus, and it never seems to be the end of the world. IPOs are very binary, right? Either folks get super greedy and they’re genuinely massively oversubscribed, or they barely get done at all. I just think that the media overstated this existential IPO risk. Until it shows up in the numbers, I don’t think fear overcomes greed.

And so we didn’t actually get to that, which I’m going to push you on. Does it show up in the numbers, yes or no?

Rory O’Driscoll

At the level of growth we’re seeing, it can’t show up in the numbers yet. The beauty of numbers—and this is why I think 80% of public B2B companies are in much worse shape than they look—is that they’re backward-looking. Every set of financial statements is backward-looking. It’s even worse: unless you do a SPAC or something, you can only project so much. So you’re stuck in backward-looking land. The closer you get to the IPO, the less you can say about the future, and the more you have the benefit of hiding in last quarter.

I’ll be clear: I don’t think it shows up in the numbers, provided what appears to be the consensus legal outcome happens relatively quickly. First of all, the process of designating a supply-chain risk has more steps than, if you read the actual legislation, just deciding at 5 minutes’ notice that there’s a review. Like all government actions, there’s a review process, there’s a comment process, and all that kind of stuff.

The interpretation from the hyperscalers—who, remember, do a huge amount of distribution for these guys—and the developers, who are clearly flocking to the app right now, clearly is: “I can use this for definitely non-Department of War and possibly non-government use cases without any supply-chain risk whatsoever. That’s a rounding error, right?” If that version of the world, which appears to be the correct legal interpretation, prevails, then yes. If they went to court and they started to lose those cases, then they would be on the back foot, and they would probably have to settle pretty quickly. But the consensus appears to be that, from a legal perspective in this case, the administration is overreaching.

So right now, no, it’s not showing up, because, yes, you’re clearly down $200 million in license revenue. If you’re the sales rep for the Pentagon and Anthropic, it’s going to be a slow week, right? Conversely, if you’re the sales rep or the guy running PLG growth for the Anthropic Claude app or for Anthropic’s developer products, you’re exploding.

Adjacent to that—against that wall of growth—there’s nothing showing up if those numbers are correct. Let’s do the math: $200 million, let’s call it $17 million or $18 million a month. So they’re doing $1.5 billion. $17 million a month is 1%. This is a company that’s 10x-ing, so it’s going to drop 1%. It’s lost in the noise.

I’m sure they deeply regret ever taking on that $200 million contract because it’s in the noise compared to the business, and now it’s produced this contingent risk. One of the things I reflected on a little bit was that, on the one hand, it can seem very localized to Anthropic, right? It’s its principles and a Department of War deal, but I think a lot of us are going to have to wrestle with these issues.

I’ll give you a very small example. I basically worked with a ton of next-generation CRM products that everyone’s building. We talked about Dad VC, right? Everyone’s building. The approaches to next-generation CRM are very different. Some are very agentic-focused, some are redoing the classic core—leads, contacts, opportunities—and some are building together with existing folks. There are a lot of approaches, but there’s 1 thing that all of them seem to have in common: every single thing that a human does is fully tracked and logged. Everything.

Now, if we go back in time, when things like Gong blew up, at first we paused. We were like, “Oh my God, my business is tracking not just a random call, but every single call?” And then we had to pause for a moment: “Is that okay?” This became the way we worked, right? Gong and all.

Then, even during this pod, we kind of Granola’d people. Is that okay? Is Granola-ing people in meetings okay? Now every app—this is becoming necessary core CRM functionality—records and manages every keystroke, every interaction, everything you type, everything you do, because otherwise you can’t automate a CRM. The next-generation CRM just does not work if you have any privacy in the workplace at all.

It’s just an example: you wouldn’t think that a next-generation CRM would have to deal with the same issues as Anthropic and the Department of War, but I think many of us are going to have these issues. Are we going to change our morals? Are we going to say a 40% or 50% layoff is no big deal at Block because it’s all part of it? I think, as we chase these faster- and faster-growth companies, we’re going to discard more and more of our previous moral standards. Maybe that’s the AI life, but I have a little bit of anxiety here, just a little bit.

5. The Data Center Arms Race: Is the AI Hype Cycle Finally Dead?

A technology that’s so important that it’s consuming 50% of the capital investment of the entire United States probably should cause a significant number of pretty significant societal changes. Two signs that maybe we’re facing or entering an age of more realism: one is Oracle and OpenAI’s plans to expand their flagship data center. Stargate’s Texas data-center expansion to 5 gigawatts is potentially being axed, capping it at 1.2 gigawatts. Is this an early sign of the end of the capex cycle and hype cycle, or not?

Rory O’Driscoll

More no than yes. I’m sitting here going, “I think we are overinvesting, and I think at some point there’s going to be a reckoning and a bunch of companies are going to realize they’ve massively overinvested.” But I don’t want to be that kind of guy who’s trying to call the turn 3 years before it hits.

Look, it speaks to—Rory, how can we be overinvesting when Jason tells us that this is the year of inference and we’ll have inference running? Can we come back to that? I will come back to that, but it can’t possibly be overinvesting.

Rory O’Driscoll

I mean, the point is this: on the question you’ve asked, no, because instantly Meta said, “No, if you guys don’t want this data center because you’re having some issues, we’ll take it.” I think if you just look at all the comments from the hyperscalers, if you look at Jensen’s comments, right now, at least to Jason’s point, the man is insatiable.

So demand right now is insatiable. If demand is insatiable, then you can’t use this 1-off set of dynamics around Oracle as a sign of, “Oh, the capex cycle’s turning.” I think that would be overextrapolation.

6. 24/7 Persistent AI: Why You'll Soon Need Data Centers in Space

I think Meta is playing a game, Google’s playing some of it, and Amazon’s playing some of it that’s hard for Oracle and anybody else to play. Ultimately—and I think Meta is the only one that I know that’s explicit—they’re betting on a world where your AI is 24/7, persistent, and infinite.

Right now, most of us are using a little bit of ChatGPT, a little bit of Claude. Maybe we’re using a couple of hours of coding compute, which is quite expensive if it’s not bundled, and that’s it. None of us are talking—I mean, we made fun earlier of the OpenAI pin with Jony Ive, right? We made fun of it. But I think we missed the point, and at least I did. I missed the point when we talked about it, which is I didn’t fully understand what the world would look like when our AI agents run 24/7.

The amount of compute that we will need to do that is obviously orders of magnitude greater. Then take multiple agents running in parallel, which is everything Cursor’s about, everything Claude Code is about. Replit V4 is all about running 5, 10, 20, 50 agents in parallel, 24 hours a day. I’m not even smart enough to do the math, but we’re talking about multiple orders of magnitude of compute that we need today.

Does it get more efficient? Maybe. On the RAM side, it’s not. We already ran out. (Laughter.) My point is, one thing that may be happening—it may be because Meta is so consumer-focused and also, in some way, so far behind—is that they’re making this bet on 24/7 persistent AI in your life as a consumer.

They will buy all the compute that is available, that they can afford, to make this dream happen, because they’re well positioned for it. As silly as it sounds, they’re probably the most well positioned—unless Claude, which is playing its own endgame, or ChatGPT can figure out its social side—to be the AI that lives with you 24/7. Them and TikTok. (Snorts.) They’re pretty well positioned.

I have a genuine question, Jason. What is that thing that they’re shipping that you’re talking about? What is my AI that’s 24/7 that I can get on Facebook or Instagram? I don’t see it.

Guest 2

It listens. It actually can see.

It can see everything you do. It sees everything you do. What tense are you using? Are you using the present tense or the future tense? What should happen today? What is entirely possible today? It’s just that we can’t afford it.

No, what will be true is that as soon as this call ends, as soon as this Riverside ends, my AI talks to me: “Hey, Jason, that was something pretty dumb you just said with Rory on the call. Let me explain where you got it wrong.” Let me explain this again: I don’t even do anything. My AI is already ahead of me the second this Riverside ends.

That could be true today in everything you do in life. Every partner meeting, every pitch you make—it’s just that we cannot afford, for the moment, the level of compute it takes to have that.

Rory O’Driscoll

I’m willing to stipulate, partly in humor, that the amount of compute it would take to correct all the stupidities that I say, you say, and Harry says might well require data centers in space. You might well be right, but I’m not sure we can afford that.

But having your AI with you every minute of your life—at least in your personal life and your work life—helping you be better on every single thing you’ve said or thought. Every single decision you make. How many micro-decisions do you make all day long? Not even the big ones of $100 million—how many checks do you make every day, every week?

It’s impossible. If your AI has full context and full history, every deal that’s ever been done at Scale, every deal in the last year, every deal any VC has done, every article, every interview, is it hard to imagine that AI, immediately after every conversation you have, won’t enrich your work life? Of course it will. Of course it will. It will be instantly better, right?

I don’t know that the problem, in the short term, is—I don’t think Oracle has the cash to keep up. Going to Harry’s point, it just doesn’t have the cash.

Rory O’Driscoll

One of my big ahas on this whole “Is it all going to go wrong?” question is the zoom-out comment: the people who say it’s not talk technology, and the people who say it is talk economics, and we don’t quite overlap. That’s the core of the challenge in this discussion.

There are a million things you could envisage doing with AI. The question is: are they worth doing, and at what price? I was just doing the math in my head. There’s $600 billion being spent on capex, and there are 150 million people working in the United States of America. $600 billion divided by 150 million is roughly $4,000 per head, right?

I don’t know if an employer wants to spend $4,000 per head on automating with AI. If you run through all the people, there are a bunch of people serving coffee. They probably don’t need $4,000 worth of AI. You start doing that math and you say to yourself, “Is the return there?” So, I stand by my comment. I think we probably are overinvesting, and at some point those chickens come home to roost.

I don’t think it’s a controversial statement, Harry, pushing back. Zuckerberg, who I think is one of the smartest, has even said it. Everyone’s said it. Business has said it. This is probably going to go wrong, but I’d prefer to ante up and be in the game and have a chance of winning than not ante up like Apple and know I’m going to lose, right?

By definition, if 6 or 7 people articulate that that’s the game they’re playing—and only 6 or 7 people matter in this discussion—then it is almost tautological to say we’re going to overinvest, because that’s what game theory says. Until such time as someone realizes, “Oh, I’m in the poker game, but I don’t have the right hand. I’ve got to fold,” we’re going to overinvest. And I don’t think that’s happening in 2026.

The only person that’s even struggling slightly to cover their nut is Oracle, because they had the weakest balance sheet and the least compelling use case. At some point, they’re going to throw 20,000 or 30,000 people over the side to make the P&L work, right? Meta can borrow. Google actually has a business, and Microsoft is quietly stepping back from the table.

This game goes on for another year or 2, but I do stand by my comment: there’s a level of overinvestment going on. I thought it was interesting—it was rumored, and this is rumor, so Rory does not like rumors—that Alexandr Wang is being sidelined, and that his position is no longer as secure as it once was, with the creation of a new lab and him not being in charge of it. That lab director reports directly to Boz, and it appears that the Scale acquisition was, bluntly, hastily done and a mistake.

To your point, Jason, on Meta being behind, that is the kind of consensus that’s being shared. That would be overinvestment. I paid $15 billion for an asset that, 12 months later, I put on the bench. That is the very definition of overinvestment.

So, I actually think you made my case for me, right? The fact that Meta’s response would be, “Oh well, we’ll do something else,” shows that we’re not yet at the stage where overinvestment is terrifying. We’re simply at, “I’ve got to win. Try something. If it doesn’t work, try something else.”

Can I come in? I’ll add that I think this multibillion-dollar deal is an interesting micro-story of Meta’s very interesting AI M&A strategy. Before I get there, on this overinvestment thing, I just want to bring up one thing that happened this week, which I think people misunderstood and is so important to my point: we’re not even at the amount of AI we want to use when we’re running 24/7. We need massively more compute than we have.

Anthropic launched a true Claude Code review to find bugs and issues. They said it’s $15 to $25, and the internet blew up: “Oh my God, this is so expensive. I only spend $200 a month for Max. I spend $20, and now you want $20 to do a code review to check for bugs and issues?”

There are certainly folks who are trying to manage their Claude Code costs to the bare minimum. They’re vibe-coding from a café in Thailand, and they’ve got to keep their costs to the bare minimum. But the response back from the head of Claude Code—or at least whoever built the feature—was, “My God, we’re spooling up 10-plus agents in parallel to run for 20 minutes to find every single bug in your product.”

You could do this manually—I’ve been doing it manually in Replit for months—but to do this in 1 click is profound. Humans can’t even do this: find every single bug in 20 minutes. The amount of compute Claude Code needs to do this is an extra $20.

What you’d really like to do is run this after every commit—not episodically, not when you’re getting up the nerve to spend $20 worth of tokens, which is a lot, because this isn’t subsidized. This is Anthropic’s. You’d love to run this 10 times a day, after every commit you make. This is orders of magnitude more compute being used for code review than we used before.

Between that and everything running 24/7, we’ve just scratched the surface. If you’ve built anything in Claude Code, Replit, Lovable, or whatever, what you realize is that it’s just the beginning. Who’s doing all the QA? Who’s doing all the code review? We’ve only scratched the surface. We do need data centers in space so that I can keep coding.

But I think the point is interesting from the news. People missed the point of this code-review feature, thinking, “Is $20 a lot to review your entire code base and find all your top bugs?” Is that a lot of money? It’s $20.

Rory O’Driscoll

Not. You know, we’re happy investors in CodeRabbit, a standalone company doing it across multiple tools. Now we have exciting and interesting competition from Claude. That’s Claude. That’s bad. On the other hand, it validates the space and says you need to do this.

I think your description of the product and the need is exactly right. If you’re generating infinite amounts of code using all these agents, the idea that you’re not going to do code review is absurd. You’re going to have automated code review. You’re exactly right.

Your comment about spinning up all those agents illustrated my point. From a technology perspective, there’s infinite need for large amounts of agents to do large amounts of code review. From an economics perspective, you have people complaining about the fact that they’re being asked to pay $20. And you’re exactly right, Jason.

The correct answer, which Claude was too nice to give because we know they’re nice people, despite what Peter Thiel thinks, was, “For Jesus’ sake, I’m giving you—you’re automating an entire developer, and we’re letting you generate infinite code for $200 a month, and you won’t even pay $20 to check it before you put it into production? For God’s sake, man, pony up.”

That’s the correct response, and that’s where technology’s meeting economics. The idea that you can just have all this shit for free is, at some point, going to stop, because someone’s going to have to cover their nut, right?

Then we’ll discover the really interesting question: how many of those users are prepared to pay $20 per code review, or $200 for a big code review, or whatever? I do agree with you. The demand is infinite. The number of things you can do is just freaking amazing. The question is, how much? What are the things that people are prepared to pay full boat for?

It was interesting, by the way, and it validated a little something we were thinking.

Claude. I would love to be the pricing person for Claude, and you've done pricing, Jason. How do you decide what things to price? They've got some really interesting things going on. The API is expensive and metered. On the other hand, Claude Code is a subscription. As you point out, you can get a whole ton more than you should be able to for $200. So, they're not metering that.

Then they decided—and, in my view, wisely—maybe code review is something that feels managerial, and maybe there is budget for that. You and I both did deals where they were in low-end SaaS, which were PLG for individual users, and people were like, “How are you ever going to make money in small, low-end segments?” But the more you go to management of seats in the enterprise, the more there is a propensity for you to pay. I think this is the beginning of the same thing.

Mr. Hacker, you can have all the free code you like for $200. But if you're going to be rolling this out into Bank of America systems, you're going to want the deluxe, not the $25 code review. You're going to want the $250 security-approved code review, and you're going to see that kind of push to make money here. I don't know if that makes sense or not, but I think that's the whole dynamic.

I think related to that, there are certain things we've talked about in the past that were true, but we're also willing into existence. The enterprise is willing its desire for AI to replace humans into existence, so it's going to accelerate because they want it to accelerate. My God, the one that is being accelerated is, “Hire no junior developers. Hire no juniors.” This is being willed into existence. No one wants to train anybody.

It ties to this code review because, listen, if you're going to hire no juniors, of course you can spend $200 or an extra $500 a month to do code review for your senior developers. They're already insanely productive, right? I think the death of the junior worries me. I had dinner last night with my son, who's off-the-charts smart, but he's at Penn State, which is not a top-10 Ivy League school, and he said there are just zero jobs for anyone in CS or math. There's just zero. No one is even coming by to hire them for C-tier positions.

He has job offers because he's publishing things on Jensen's fine-structure theory that I don't understand. So, great, 6 people in his class at his large state college have tech offers. But it has become—not only in our world, but in his world—it has become, “We just don't want juniors.” We don't want to train them for 3 months or 6 months. They don't know the tools cold.

That's where the budget comes from when you hire no juniors. When we're just a world of the middle of the bell curve in terms of experience, we don't want folks that don't want to pick up the tools, and we don't want folks that aren't experts. It's going to ripple. This is my big worry for 2027 as it happens.

7. The Death of the Junior: Why Entry-Level Jobs are Vanishing

It's going to be true in sales. It's already true in support. You want humans in support, but you don't want juniors. My God, I want a human that knows CodeRabbit cold, who can handle the escalations. I don't need junior marketers. I don't need junior anything. Harry, not to get us off track, but I think it ties to this. I think getting rid of juniors is where we get budget for these data centers, in part. They're just dying faster than we ever thought. It's happening in front of us: the death of the junior. Bad for society, but it's a reality.

I think it ties to 2 fundraises that we saw. Rory, you may disagree as an investor in one of them. I would agree as an investor in another, but we saw Intercom raising $250 million, and we saw Algolia raising $500 million. I think both tie to your point there about the replacement of juniors and lower-rank employees.

Rory O’Driscoll

The death of the juniors. Why would I want to wait for an associate to scale up for 2 years? That was the classic legal hire. You hire someone real from a top school, ideally someone who was in the top 10% of their class. They have the IQ, and you basically subsidize them for 2 years until they cross the line and become senior enough that you can give them a top client or let them take a case. Who wants to do those 2 years now? Unless you have to. Unless you have to.

This has been one of our long-running discussions, and it's good that I'm evolving my position, Jason. I said this before, but I've come more to your perspective that there clearly is something going on here. Even though I don't believe in the catastrophic mass-unemployment scenario, I think you're exactly right for a couple of very targeted demographics, one of which is entry-level computer-science jobs. The other would be actual customer-support jobs and actual legal-associate jobs. There is a meaningful impact on unemployment right now, so I just want to acknowledge that you're right on that.

We're very much right. It is interesting because I think that's true. On the other hand, Paul Graham says that, typically, when there's new technology, the best adopters are young people because they come in with fewer priors. There's a part of me that says, are these schools doing their job, especially now that it's 2026? If you're graduating computer-science graduates who aren't AI-first and totally awash in using these tools, you're just not doing your job.

Rory, are you unable to alter the curriculum fast enough to update it for Claude? Curriculums take time to change.

Rory O’Driscoll

They need to change that because if you're looking at investing $100,000 a year for 4 years in a 4-year college education that turns into unemployment and a barista job, you should be pissed, and you will be pissed.

But I'm with you. When you look at Andrej Karpathy's journey in 6 months, from blindly using it for 20%—and this is Andrej Karpathy—I think it's unreasonable to expect an educational institution like a university to alter its entire curriculum based on Claude's ability to progress throughout that skill journey. It certainly doesn't appear to be happening.

Well, Rory's point is, should we be expecting it from these institutions?

Yes. They're becoming just babysitting—expensive babysitting—for 4 years. They shouldn't be. Give Peter Thiel credit: 15 years ago, he articulated the perspective that the ROI on higher education is going down, and at the time it was still high.

It's worth pointing out that, even to this day, unemployment among college graduates is lower than unemployment among high-school graduates, which is lower than unemployment among high-school dropouts. You're still better off going to college in terms of employment. But the point is, now you take into account the net present value of that $400,000. You're not getting your money back on anything other than the best degrees.

These are not the most deserving people on the planet. They still have lower unemployment—3.something% versus 4.4% on average—but there is definitely a trend here that would be troubling if you're a recent graduate. So, I'm kind of with you on the trend. Rory, when you say you don't see the mass-unemployment case, or that you think it's an overreach, when you look at legal, computer science, customer support, and bookkeeping, those alone, if they had no impact anywhere else, would be mass unemployment.

Rory O’Driscoll

I don't know if it would be. Again, look at the numbers. I was actually looking at them today. The entire tech industry, excluding tech people employed in non-tech companies—technology NAICS classification 51, tech companies—is 3.1 to 3.2 million people in the United States, of whom roughly 800,000 are software developers, right?

If you go down by half, that's 400,000 software developers. It's a lot of people, but it's 0.2% unemployment. How many are in customer support? A much bigger number, right? You're exactly right, but it's also way more fungible, in the sense that a lot of the low-end customer-support workers can do lots of other service-type jobs.

I'm a believer that, over the medium term, the technology takes longer to diffuse than we in Silicon Valley allow, and people are more adaptable. Over the medium term, I don't see any kind of 10%-plus tech-driven unemployment, which is— I know you're going to disagree. The point is, we don't have to argue about this because, in the end, we'll know. I just think I have 200 years of facts on my side, and you have bullshit, but maybe you're right.

That said, I do believe you are right in the short term. In isolated pockets, very vocal people are really struggling. I have 3 recent college graduates over the last 5 or 6 years, so I'm totally aware of how hard it is to get a job as a college graduate. I think it's a real issue, and we better get on top of it because one of the other things the French Revolution teaches you—and, actually, history in general says this is going to sound really cold.

Guest 3

You can have dispossessed urban poor forever, and nothing happens. But if you piss off the 20-something-year-old middle class—the overeducated elites—they tend to cause trouble, right? And I actually think the continued production of masses of college graduates with sociology and English degrees, finance and computer science degrees, who have played it by the rules for 20 years and then been given mass unemployment, is going to be problematic for that cohort.

I want to be clear: just because I think the macro trend is not there doesn't mean I don't think the micro trend is a real issue. And that's to Jason's point. I don't know if it's a moral issue or an economic issue, and I predict in 2026 it's going to be a political issue.

I mean, it's so funny to watch the NPS survey, right? The NPS for AI polls really negative, right? As someone tweeted, “You've been telling us you're going to blow up the world and make us all unemployed, and you're surprised to discover we don't like you.” I mean, it's almost funny.

At the same time, if you look—I'm sure Intercom's data says this, as does everyone's—everyone who works in a heterogeneous human-AI-agent environment, the CSAT is always in the top 10% for the agents. It's never number 1. They never say that the number 1 most popular person is always a human, but the agent is always in the top.

That is more telling than most people realize, because I think the theme—I've said it's early, but it's accelerating—is that we'd just rather work with agents. We'd rather hire agents, we'd rather buy an agent. We don't want to—I mean, Harry's got 3,000 applicants for his EA position because he needs a human, but I tell you he'd rather have an agent if it worked.

You don't want to interview 3,000 people. If he could have an agent, he'd pick an agent, and we're all going through that. It doesn't even matter if it's cost-effective. I do agree: we'd rather have an agent.

I've been thinking a lot. I know this is Captain Obvious stuff, okay? But my God, why do some of the AI B2B startups grow so quickly? The simple answer is product-market fit, product-led growth, agent-led growth. But as you dig deeper, as the year goes on, and as we roll into late 2026 and 2027, I really think people just want to buy an agent. They don't want a human.

If you can deliver perceived massive ROI, that is so appealing. It is not layoffs; it is not this. It is, “Oh my God, I can get customers, I can run campaigns, I could do support without humans. My God, I want that.”

When I look at so many of the mediocre products the public B2B companies have launched, they do not service that demand. I don't want humans at Atlassian or HubSpot or other companies. I don't want them. I don't want to make my humans 8% more efficient with your AI. I don't want these humans; I want an agent to do this work.

And whether it's Legora, Harvey, or whatever we've had on the show, whether they fully can realize that potential, people will line up at your door when you can give them an AI instead of a human—an agent. That's what they want. These are the companies we want to build and run. They will keep shrinking in size as revenue grows for this reason.

Guest 3

As is so often the case, I agree with the broad direction and might disagree with the sequencing—not so much the sequencing as the narrative. I think a lot of the B2B adoption to date, probably up until now, with the exception of customer support, and in the last year with the exception of coding, has not actually been about human replacement, except very much at the margin.

I think folks like Harvey, Legora, and GC AI, they're all about—they're not actually about human replacement, except very much at the margin. At some vague level, maybe I don't need the 17th associate at Linklaters, but so far it hasn't been agentic. So far it's been primarily, “Yeah, let me help you, Mr. Lawyer, be smarter, work more efficiently.”

At some level, you're not asking an agent to do something; it hasn't been a wholesale replacement story. I think you're right, Jason. It's been at the task level. I've been able to increase the number of tasks that I can do and therefore decrease the number of tasks I have to pull in an associate to do, right?

8. Agent-Led Growth: The Secret Reason Startups are Exploding in 2026

Now, with agents, and I think we've seen it first in software, it's stunning what's happened in the last 6–9 months in terms of the ability of these things to genuinely just code start to finish. And therefore, this is where I think you're all right: we might be at the start of some of this agentification, and the next 12 months will be B2B companies saying, “Now I finally can automate this task entirely.” And then we'll find out what the demand is.

I like to think through a frame of what my children will look at me and how I used to live and say, “God, I can't believe you used to do that.” And I think one of the things they will say that to is, “I can't believe you spent years training people, and then they left.” Like, all those years wasted, and then they left. That's so nuts. It's crazy. I think that will seem incredibly archaic.

I look at my EA today—we've mentioned it. I've spent years training her. She's amazing, and then she leaves. It's fucking nuts. An agentic world which compounds education and learning and they never leave you, Jason—would I prefer it? Fuck yes. All of that time was wasted. I've got to start again.

I just think if you're founders in B2B, this is what you should be building: something where folks will line up at the door because they'd rather work with your agent than a human. We're lining the hell up. Most of the public companies are not even thinking this way, but the startups that do blow up, well, they line out the door.

Of the agentic ones, please? As distinct from the automated-task ones, what level of agent are you talking about?

I mean, literally—listen, there's a bunch, but I've done so much with these GTM agents, these go-to-market agents. This last one that we deployed, this Monaco one that Sam Blond did, they closed 7 figures in their first 5 days, and they have 60 days of people lined up.

And why are they lined up out the door? Look, I'm not hyping it. They're all good, but why are they lined up out the door? This is the more interesting question: why are they lined up out the door? Because it goes—whether they're fully successful on their whole journey or not—it's worked for us as a tool. It talks to the prospect, it texts them, it pitches your product, and it sets up the meeting, and the meeting is fully ready to go and close.

If you can deliver that, and it's a high bar, and I don't even think it was possible 60 days ago, and all the other tools we use—just before you critique it, we can all poke holes at it. It's early, but people will line up the fucking door for this stuff. They'll line up the fucking door for this stuff.

Whatever category you're in, you want to build this agent that is sufficiently better than 90% of humans that everyone will pick it, and the budget magically comes out of nowhere. It literally comes from nowhere. That's why we're just shocked at the TAMs for some of these products. They don't even make sense to us. They're not consistent with the past, but people will line up if you can do this without humans.

I mean, we built our own VP of marketing literally yesterday. This week was the first time it led our human staff meeting. Our AI VP of marketing led the team meeting. It led the meeting. It summarized every single metric, every single thing, every single thing the humans on our team had to do—10K rows of it—and led the meeting.

As great as Clay is for enrichment, it's great. If it could do that, I'd give you $50,000 tomorrow. An extra $50,000 tomorrow.

You were doing the Intercom deal. Are you betting that they beat Sierra?

You know, again, I prefer to let Eoghan speak for himself on what he wants because he's a very ambitious man. I'll tell you what I am betting on: I'm betting that they can carve out a significant space in that market for that product and that they've had explosive growth in it.

And coming from SaaS, I'm betting that that team, coming from a SaaS background but being grounded in what it takes in AI, can take a significant portion of market share.

Guest 3

They tend, on average, to be slightly smaller in terms of customer size than the high end of Sierra, and they compete more with Decagon. But one thing that I think a lot of these categories that have traction—if you want to say I'm only a partial expert, but if you want to say Intercom versus Decagon versus Sierra versus others, Zendesk, Salesforce—one thing that people miss a little bit in terms of the winner, which you kind of asked Rory, right, is FDEs: forward-deployed engineers.

For the moment, they're a limiting factor. Let me explain what I mean. What it means is no one has enough fully trained FDEs that can get a customer up and running fast enough. No one has enough of these resources, okay?

And what it means is that almost every vendor has to pick lanes. You can't do—as much as you'd like to do it if it was all software—you've got to pick where you're strong, right? And if you're best at digital transformation for enterprises, maybe that's what Bret Taylor is. You've got to pick that, right?

Not that these folks—I know the sales teams and other people think they're direct competitors—but everyone has more demand than S-tier FDEs to service that, so they're all picking lanes today.

And so, in that sense, they’re all services businesses. In a sense, this VC meme is true. They’re not long-term services businesses, but if you need humans to train it, if you need humans to manage it, and if you need humans to iterate it, you have the same constraints that services businesses have.

This FDE thing is a limiter. It’s exciting because I think a lot of the losers out there can’t even get good FDEs. If you look at the folks that are struggling in AI B2B, if you peel the layer back, sometimes their agent isn’t competitive, sometimes it’s too slow, right? A lot of issues. But if you really dig down, how many great FDEs do you have? How many folks do you have that can spool up a customer in 30 days? “Well, I’ve got 3.” Well, you’re going to lose.

Just staying on venture from private markets and covering the news items that people are up to. As I said, Legora raised $500 million at $5.5 billion. Accel led that round. Obviously, Harvey raised at, I think, $11 billion in the latest round. Databricks did a tender at $4.5 billion. Intercom raised $250 million at—I don’t know the price for it. You might know it.

No, in debt. Yeah, they’re a large company.

Yeah, they’re—yes, it’s a debt round. Great. Founders Fund is closing in on $6 billion of new funds. They invested the last $3.3 billion in 11 months in amazing businesses like SpaceX, Stripe, and Anduril. I’d be very happy if I were an LP.

Base44 hit $100 million in ARR. I don’t know if you have anything to say about that, Jason, given Replit.

Does it save Wix? Well, you know, I was thinking we had on the agenda: CrowdStrike crushes the quarter, trades down, right?

Yeah, I was going to move on to that with our prediction.

No, but it’s related to the Wix thing. Then I was looking at Cloudflare, which is on my top 4, right? Cloudflare: 27% growth a year ago, 34% growth last quarter. I mean, that’s pretty good acceleration, okay? That’s pretty damn good acceleration. A 40% year-over-year increase in net new customers.

I think the public markets—we were wondering what’s happened this year, and you had Eran from monday.com on a week ago on 20VC, and he said it. We’ve got to accelerate to get back credibility with the public markets. Accelerate. Not manage a genteel deceleration, which was 2022 to late 2025. Not manage a gentle deceleration with higher net margins, which everyone thought was the job at scale.

Now, the truth is you’ve got to be Cloudflare. It’s not just Palantir. You’ve got to be Cloudflare or better. And so—

I wouldn’t want to be running Wix, because $100 million, right? It shows a lot of new things. But their larger customer count is flat to down. So they’ve got the gravity of the decline of the core business. They’ve got to get back to so much growth that $100 million just—it’s not enough.

It’s not enough, and Rory could help trail out the math, but I think Base44—and there’s some cannibalization there—it’s got to be doing $500–600 million to move the needle for Wix with this core growth decelerating. It’s just so hard, and for startups, you’ve got to start wondering this year when you should just give up on your portfolio if they’re not accelerating. At what point do you just give up? Because if the public markets won’t tolerate gentle deceleration, how will the private markets tolerate it?

It’s just the era of gentle deceleration has ended. It’s dead. We didn’t think it was such a great period, 2022 to 2025, but it was pretty nice. The gentle deceleration was tolerable.

I mean, again, so much to unpack. First of all, to be clear, gentle deceleration is like entropy. It is the end state of the universe because everything decelerates to GDP growth, right? I mean, Anthropic is, as we speak—

The universe will be dead before some of it happens.

[Speaker?]

Agreed. So, I just want to be precise, because you know me, I’m that guy, right? What you’re really saying is that at 10x growth year-on-year, Anthropic can afford to decelerate. At 40% growth, like Figma had a couple of quarters ago, yeah, you decelerate to 30%, you’re going to take a little hit. But you’re exactly right, Jason.

What you’re really saying is that at 10% growth, you’re now so close to GDP that you’ve got to reaccelerate. So that’s really what it was. Just being precise, right? I don’t even know if the public markets think it’s okay that Figma might decelerate, right? I got pushed on that, because just take CrowdStrike, right? Slight deceleration: 27% to guidance of 23%. Twenty-seven percent revenue growth, guidance of 23%. They’ll probably beat that by 2.

You know, it’s okay to go 25% when you have scale. I mean, Microsoft and Google are all going at 13%, right? So, yeah, you just have to be objective. When you have enough scale, 25% is enough to matter, right? Not everything’s going to accelerate all the time, Jason, because it turns out—

The markets have been brutal this year, and I think it’s not that you’re wrong analytically or empirically, but the markets have given up on folks not accelerating. They’ve just entirely abandoned it.

Well, yes. But, again, I’m being precise. What they’ve done is they’ve said, “For 3 years I thought it was temporary, and now I recognize it’s permanent, and now I’m going to value based on this growth rate.”

Take a hit off you for technical obsolescence in the terminal value. Take another hit off you for free cash flow, and I’m going to give you 8 or 9 times revenue, EBITDA, revenues. That was sweet that I said that, wasn’t it? That was a good one.

No, you’re exactly right. I mean, so they’ve all corrected to fundamentally what they’re worth with no pixie dust. And that’s just a tough place to be long term.

I agree, those guys—the negative spin is they have to reaccelerate. The positive spin is even a small amount of reacceleration will get you some lift from here. But you’ve got to do it. Anyway, Wix is a good example of that. It’s funny. Let’s try and take apart Wix. I hadn’t prepared on this one, but I can’t remember. The overall revenue is what? They’re $2 billion, growing 13%.

Yeah. If that $100 million business grew at the same trajectory as Lovable or Replit, I’m going to do some math, and you guys know the numbers better. Let’s just play it out. They went from $100 million to $300 million, plus or minus, in a year, fair? And $300 million to $600 million, plus or minus, in another year, at least.

If Base44, which is inside Wix at $2 billion, went from $100 million to $300 million this year, that would mean it’s small today. From $10 million to $100 million, it gets lost in the noise. From $100 million to $300 million, it’s a 10% lift to Wix’s growth rate. So if it’s 10% before, now it’s 20%. And from $300 million to, say, $600 million, it’s a 15% lift, right?

So you can’t just say it’s not big enough. Wix is—because the nature of a small thing is it’s small and then it exponentially compounds.

Agreed. That’s where I was going to go. It’s not that 0 to $100 million is bad, Harry. But where you are correct is this: if you’re a stale SaaS company, if you’re an old-school SaaS company, and you have this bright, shiny thing inside you that’s growing from 0 to $100 million in 1 year, you have to run it in such a way that it can keep up with the Lovables and the Replits.

And it has to go from $100 million to $300 million, and $300 million to maybe $1 billion, maybe $600 million. What you’re saying, and it is correct, is if it ends up hobbled by the wider company, the bureaucracy, or the SaaS—the blah, blah—then you’re not going to get the needle moved. Because that’s the—

Their core paid customers declined 1.2% last year. So they’re at a terminal state where a lot of B2B companies are, where they have reached a terminal state for their core customer base, right? And what they’ve done better than most public companies is, hell, with a great acquisition they got $100 million of AI revenue on top of it. That’s better than most, but it is not enough yet.

Not only because of your math, but the core is declining 1% to 2% a year, which doesn’t sound like a lot, but, man, that’s rough, right?

It doesn’t matter. Stop. It doesn’t matter. If Base44 could grow like Lovable or Replit—

Base44, but yes. Base4—sorry. Yeah, Base10 is a venture firm. Yeah, cool.

Base44 could grow like Lovable or Replit, then within 2 years they’d be out of the woods and they’d be a 30% growth company, right? So you’re exactly right. They made the right move, but what you’re effectively saying is that inside that other company they might not be able to, right?

As an investor, you should be saying to yourself, if we think Base44 has the same trajectory, then great. Then you have a diamond in the rough here. You get the existing business for free, and you get the new business. But I think implicitly your board’s saying that it’s not going to grow at that trajectory.

It’s probably going to grow from $100 million to $200 million. So it’s probably going to get the growth rate back up to the mid-teens, but it’s not going to be enough to dramatically change the trajectory. That’s effectively—

I think time will tell. If we’re using Wix as a case study, here’s why every B2B company at scale should hope that Base44 works. Look, I love Replit and Lovable, they’re great, okay? But Wix has 6.11 million customers.

The simple reason—and I’m sure the founder who’s been on Harry’s show will challenge me—but I know half the reason it’s working. It’s a good product, right? They have 6.11 million customers to sell it to.

Yeah, and that’s okay.

No, it is okay. They should be able to pull it off. But $100 million ain’t enough. All the tweets about how great it is—the scale is so intimidating that it’s tough. You’ve got to do what Mark did and buy Informatica, too, to bridge the gap.

If you’re in a game as an existing SaaS company where what Jason’s effectively saying is, even when you make the right strategic moves, it’s “not enough” because you can’t pull it off for whatever institutional reasons, then that points to a grim conclusion for most of these companies.

The case study will be Base44. It’s a good product. I will stipulate that; I’m not going to argue where it stands on the hierarchy of vibe coding. I have used it. It is a good product. It’s a lean team, but it’s a good product.

If Wix can’t freaking cross-sell out to 6.1 million people, what hope is there for you? This is a rhetorical question to many founders. It’s a bleak question, because I can tell you anyone would rather vibe-code a website than deal with the crappy templates it comes with. So it’s a great use case, right? It’s got 6.11 million customers. If it doesn’t work—

Jason, I see your comment. I think you’re exactly right. I think Wix is a perfect specimen use case. If you can’t cross-sell the new AI product to existing customers who are trying to do the same thing, just with an older technology, then the cross-sell acquisition and cross-sell story doesn’t work.

This is as good as it’s going to get if you’re an old-school SaaS company, other than managing the decline. So you’re right: it’s a total lab experiment. As you’re watching Salesforce, in my view wisely, buying a few next-generation, AI-first companies, they’ve got to do the same dance just with an X of zero everywhere, right? They’ve got to take that $40 billion behemoth and find a way to upsell another $20 billion worth of stuff.

9. Figma Make Is Terrible? The Failure of Quarterly Software Releases

At least they have something in this market to play with. I look at your Webflows of the world and your Squarespaces, and I go, “Ouch.”

Yeah, it’s a mystery why they’re not trying. I’ll tell you one thing that happened to me this week: I tried Figma Make for real for the first time. I know Harry saw it on Twitter. For my use case, it was terrible—much worse than Base44, much worse than any vibe-coding product I’ve ever used. Figma Make was terrible. It was undesigned, and, even worse, all these products have advanced.

I asked—I do this test, and a lot of people make fun of me. They’re like, “Oh, you’re an idiot. You don’t know how to do a prompt.” I’ve done a few prompts, okay? My AI apps have been used a million times. But I also know how these tools work.

When I want to test a new vibe-coding site, I do something very simple that didn’t work 6 months ago and works well today: go to saster.ai and make me a better version. People make fun of you. “There’s not enough context. There’s not enough data.” Of course there is. I have massive amounts of context and data on my website. Scrape it, use your AI, and come up with a better idea.

I’ll tell you, Replit and Lovable, and even v0, can do a pretty good job of it. They come up with ideas. This is what the AIs are supposed to do. Make didn’t even know what was on my website. It didn’t even try. It was the worst thing I’d ever seen, nor was it designed.

Because they would say that the use case here is to take a design and implement it as code, not take an existing website.

Yeah. That—and res ipsa loquitur—is fine in 2026. Your website doesn’t have to say anything coherent because it doesn’t matter for design.

That was the other feedback, which I thought was rather dated, because your comment was that you were criticizing the work in terms of its written content, not its editorial look.

Well, it was all terrible. First of all, it didn’t design the website. It just used dated cloud artifacts from 6 months ago. Giving me little sparkly icons that are on every tier-three demo day is not impressive for Figma. There’s no design, but it didn’t pick up—

They chose not to pick up any context from the asset.

My point is, they chose not to pick up any context from the asset. If Figma Make can’t do that—and I can tell you how folks criticize me: “That’s the wrong use case. It’s not supposed to do that”—but if Figma Make can’t even pick up the context from a website to redesign it, what hope is there for Squarespace and all the other guys, right? They can’t even do it. Even Figma can’t do it.

The worst vibe-coding experience I’ve had in 6 months is Figma Make. And I’m not a designer.

Jason, did this change your opinion on your optimism as a shareholder or potential shareholder of Figma moving forward?

Yeah, it shows that the team’s not there. I think it’s worse than Base44. It’s rough. You’re not going to win today if you’re doing quarterly, best-effort releases. I would say most public software companies are also, and many struggling unicorns—I would say almost all struggling unicorns—are doing best-effort quarterly releases. That’s death today. That world doesn’t exist any longer.

“Best-effort quarterly. Let’s get around the table and decide what we’re going to ship this quarter, guys.” Okay, that’s the way I built software. It doesn’t work today.

I think what’s interesting about this discussion is how generalizable it is to the challenges that a huge amount of the portfolio faces. This challenge of taking an existing business—because one thing I think you can stipulate is that the people at Figma are extraordinarily talented. It’s not a year or 2 ago since we said, “Oh my God, these are the best people out there.” They built something enormous, and they clearly have, and they clearly did.

I think this speaks to the challenge anyone faces in a larger company with an existing product architecture: finding a way to adopt the new technology, the new architecture, and the new way of both building and determining what you actually have to build, in such a way that you can make a big enough impact to change the trajectory of your existing company.

We’ve talked about 2 or 3 examples. We’ve talked about Figma, we’ve talked about Wix, and we’ve talked about Intercom, who seems to have been doing it. It’s just really damn hard. The funny thing is, there are literally $1.something trillion worth of public companies and another trillion dollars’ worth of private companies for whom this is the existential crisis.

It’s really interesting here. We’re seeing it across our portfolio. How do you make sure you matter? How do you cut through the internal noise? How do you staff your teams? How do you create the urgency? How do you create a sense of what’s possible?

The odd thing is, this is precisely when you should be hiring young people who are coming in replete with knowledge and unencumbered by priors. So maybe all those unemployed computer-science graduates—maybe in some of these SaaS companies, genuine comment here, you do need some young talent to see what you can do and get you to do some Y Combinator acquisition of failed teams.

How do you turn around your R&D development, which takes, as Jason says, 3, 6, 9 months to deliver something, when the team down the road is delivering a new version every week? Because if you don’t, really, the challenge is that you’re going to trade at 8 times EBITDA, which at best is 2 or 3 times revenues, and it’s going to be pretty freaking unsatisfactory for the people who paid 30 times.

I don’t think there’s a lack of desire for young people, by the way. I think there’s a lack of desire for people 1 or 2 years into a role who’ve been indoctrinated enough to think that the world is a certain way with a certain set of tools. What you actually want is the 18-year-old who’s on every subreddit and knows every intro to a TikTok.

10. The Ultimate Stock Picks: What to Buy and Sell Right Now

Totally. We have to make some public-market bets, and I wanted to save some time for this. We have 4 companies that we need to choose. Jason, would you like to start with your 4? You can give a minute as an explanation, or you can just say them. Up to you.

Okay, hold on. I got a little confused here. Mine haven’t really changed. Mine were Palantir, Cloudflare, Shopify, and CrowdStrike.

I’m sticking with my 3. I’m going with the same. I’m betting on momentum being the only thing that’s going to save us in the age of AI. So I’m betting on overpriced Palantir, Cloudflare, which is reaccelerating and benefiting from AI, and Shopify, which is gaining market share.

I actually can’t find a fourth from my cohort, but because I made up the bet, I’m going to go with CrowdStrike. I don’t think AI is going to hurt them. I would almost go IGV instead of CrowdStrike, but I’m going to go CrowdStrike.

I can’t, even though I want to do Atlassian, because I think it’s above the fold and a good one, but it doesn’t fit the thesis. So I’ve got to pass. After using Make, I cannot do Figma. I cannot do Figma after Make. This is not a world-class product, so I’ve got to stick with CrowdStrike.

I’m with you on CrowdStrike, so we have that in common. I’m adding Nubank. I think Nubank is actually dramatically underappreciated. They’re going to enter the U.S. It’s a good bet. They’ve still got David at the helm, they’re growing 28% year on year, and I think Nubank are incredibly solid.

I’m going for the incredibly boring Nvidia. I think that’s a good one. Honestly, if we’re going to see inference skyrocket, Nvidia will win. With the acquisition of Groq and its movement into the inference layer, I don’t bet against Jensen.

Then I’m actually going to go for an out-there one: Reddit. It’s down 40%. I think it’s a data layer for a lot of the LLMs, and given its pricing, it’s a good buy.

Well, you had a broader circle than I did.

Yeah.

All good, that’s all good. That’s my gripe, Harry, because that wasn’t the assigned homework.

What was the assigned homework?

He’s the schoolmaster. The homework, as I understood it, was a whole bunch of SaaS stocks: software is dead, so what stocks won’t be impacted by that? That was the exercise in the universe of stocks, presumably tech stocks. I don’t know. Are we picking anything? Can anyone comment here?

What’s the barrier between what qualifies and what doesn’t? Does it have to be a SaaS stock?

Totally. I mean, look, if you take any tech stock, you can probably pile into a bunch of them. Again, I stuck with an error in the homework assignment, but I support Harry’s version as well. I don’t think it’s so far afield.

Jason, what’s the narrow homework assignment? I thought we were trying to find the gems in beaten-down software companies.

I think CrowdStrike counts. I think Nubank doesn’t really count, but I’ll give you a pass on it, right? I think Nvidia doesn’t count, but it’s fun, and I think Reddit probably doesn’t count, but it’s all good by me.

But Palantir counts? You don’t think it’s a software company? You’re still debating this pre-IPO company that’s a services company?

Well, I just don’t see why Palantir would count. I’m not saying your choices don’t count; I’m just saying most people would put Palantir in the bucket of software stocks, and they wouldn’t put Nvidia in it. That’s all.

There’s a reason for this. I’m pushing because I actually think, look, if the question is, “What should you invest in overall?” that’s a totally legitimate question, and we can have that discussion sometime. I genuinely thought the intellectual exercise here was triggered by the meme that software is dead: everything went down, and then the question you had to ask yourself is, where do you want to play?

That’s why Jason’s comment was so interesting. You had the first question, which was, in the abstract, without price, is software dead? What’s the continuum of software—from “I don’t think it’s dead, but it won’t grow” all the way to “I don’t think it has a future” to “I think it’s actually going to grow just fine”? Then the second question is, you can have your opinions on that, but you’ve got to take price into account.

God, I can’t imagine going to the Scale partner meetings. We’d be debating this stuff forever. Just get to the thumbs-up or thumbs-down. Rory, what would yours be?

Rory O’Driscoll

I’ll tell you. Of the boring ones that have been savaged and are below 10x EBITDA, I think Salesforce at 8 or 9x EBITDA and even team at 9x EBITDA with decent growth are okay, right? I think those are good stocks.

For the record, I also bought Workday. I put the full $250,000 we said we’d bet into Workday the day we made the bet, and we’re up about 4% so far. Nothing major, but nice, right? In a week.

Those companies are squarely in the strike zone of “could be replaced by AI,” but I don’t think they will be. The next bucket is that I actually think the AI fear is overdone, and they’re in that mid-category of GARP. They’re not stupidly priced, they’ve got decent growth, and I put Toast and Intuit in there because they’re software companies with a huge slug of transactions that aren’t going to be replaced by AI.

At the very top, I actually thought Jason raised the most interesting category: the high-growth companies where the story that AI is going to kill them was totally overdone. The best example of that was CrowdStrike. The proof that it was totally overdone was that, by the time we were talking about it 2 weeks later, it had gotten back almost all of the hit it took on the day. It’s been one of those up-and-then-down things, right?

However, to continue and finalize your point, I think, with more fear—because I’m more afraid of current high prices overall than I am of the SaaS apocalypse—I would, with fear, pick CrowdStrike. You’re paying in the mid-to-high teens of NTM revenue. I think the EBITDA multiple is 50, and the growth rate is 23%. That’s scary, but it’s an enduring company.

I wanted to pick 2 expensive ones where you’re not dealing with my first 2, which are 8 or 9x EBITDA, boring-ass value. The other 2 are early-teens EBITDA, which is roughly the same as Microsoft and Google—pretty boring, right? Then you’ve got the high-priced ones at 30 and 40 times EBITDA. In other words, they’re not cash-flow justified, so you’re still story-based.

I went with CrowdStrike, and then I get scared, and then I get drawn to Palantir. I never thought I’d say that, but the growth rate means they’re 1 more year of growth away from being normalized—maybe a year or a year and a half. In other words, you’re forward-paying 2 years, and I think this administration is going to spend for 2 years. So I reluctantly, reluctantly put Palantir into that bucket.

That’s kind of where I come out. Was that 7 or 4? What were the 4?

Two of each: 2 cheap value, 2 GARP, and 2 “oh my God.”

Rory O’Driscoll

I mean, it’s prices where you kind of go, “You have to get to your points.” That’s why I thought the last category that you put all of them into is the venture category.

Effectively, the only category that I listed that’s even vaguely venture-relevant is the 30%-plus-growth people, and the other CrowdStrikes and Palantirs of this world. The astonishing thing is that 95% of the assets in the public market are utterly venture-irrelevant in terms of growth rate and everything like that, because you just haven’t had any public IPOs.

By definition, the return profile from this bucket play is probably significantly below what you should be able to get in venture if you have the right price, which is, of course, the caveat.

As long as you’ve tracked those 4 areas, it’s good by me. Yeah. Benioff’s going to be unhappy with you, Jason.

No, no, it’s not that I would pick value stocks. I’m a deep, deep believer in Salesforce. I guess I can swap it.

If you had to pick 1 value stock today—not asking you to actually put your money in this one—but where you’re having to reach for a gem in value, what would it be?

I only believe in reacceleration. This is my only thesis, right? I have no other thesis. You have to pick Atlassian because I can’t find another beaten-down one that’s reaccelerating. It’s that simple.

My 4 are Palantir, Cloudflare, and Shopify—it’s just the math. For the fourth candidate, I was comparing CrowdStrike, Snowflake, and Atlassian. I had Atlassian. I would pick Atlassian.

Probably in a year I’d pick Salesforce, because I think it’s all going to work. Don’t get me wrong, but we need to see it. The only reason I dinged Snowflake—and this isn’t because I couldn’t do better—is that it’s not founder-led. That’s my thesis today. I have a bright line: I can’t pick [?], or [?], or Snowflake simply because they’re not founder-led.

You could argue, if you look at the numbers, that Snowflake is a better bet than CrowdStrike, right? But Atlassian hasn’t gotten any credit for reaccelerating. Obviously, the answer is that they don’t believe it will sustain. They don’t believe it will last.

Rory O’Driscoll

I think the argument there—and that’s why I hesitated—is a scary one. I put it in because I like it. The software development life cycle itself is changing so dramatically that you’re effectively a software product that coordinates the software development life cycle. Seats are going to change, workflows are going to change, and if you don’t stay on top of it, how people run engineering departments will change.

You have software that coordinates how engineering departments ran in 2022, and the risk with Atlassian is that in 2027 it’s going to run so differently that you’re not the thing. Now, I’m willing to bet he can figure it out, but there’s a legitimate risk in that one, right?

In a way, I don’t think, for example, that Salesforce has the same risk. On Salesforce, your destiny is a little like SAP. Worst-case, if you can’t reaccelerate, you just optimize. Boring as be damned: reduce costs and increase free cash flow.

It’s worth pointing out that all the big pre-SaaS companies that had systems of record—Oracle, PeopleSoft, which rolled up into Oracle, and SAP—survived and financially thrived, even if they weren’t exciting in the age of cloud. My guess is Salesforce does the same in a way that the mid-tier point products don’t.

In a consolidating world and in the cloud, they can probably be mean and pick up other people. God, that’s uninspiring, but there you go.

Well, listen, Harry, you can bring it to an end. What if you wanted to bet on Salesforce? I’ll tell you the simple reason: the numbers don’t show reacceleration yet, right? At least the organic ones.

I’m long on Salesforce. The simple reason—and I know this from being super deep on Agentforce and everything—is that, unlike a lot of folks on this list, they have more demand than they can serve. And it’s a complicated issue. Not every customer can get Agentforce to do what they want, train it, and get an FTE.

But what you really want to bet on, if you want to bet on turnarounds, is who has more AI demand than they can service. And there may be more than Salesforce, but there probably aren’t too many more folks who are literally turning away customers because they can’t service them.

Going earlier in the conversation, those are the startups I want to bet on too, right? How are you going to build something where the demand for your agent is so strong you can’t even service it? Salesforce does have that.

One of the reminders for me, just to put it, is that I have to revisit in my own head the eternal venture question of how much for how much. In other words, how much extra on a revenue or EBITDA multiple do you pay for 30% growth versus 20% versus 40%?

Jason, to your point, because I like your sort algorithm, right? In a venture business, they’re the only things you invest in. But I will say, sometimes in the public markets, you just get a little scared at the kind of 30- or 40-times EBITDA multiples, 12- or 14-times revenue multiples, in conjunction with the 20% growth rate, when you think back to the fact that the average SaaS company used to trade at 6 times with a growth rate of 30%, right?

There’s a little part of me that says, if everyone woke up—I mean, what happened at the low end was everyone woke up and said, “These things aren’t going at all. I’m giving you 8 times EBITDA. That’s it.” The question is, what happens if at the high end people go, “20% or 30% growth looks great relative to Salesforce, but it looks shit relative to Anthropic”?

Maybe I should go back to my, “Oh, you know, 6 times revenue, 7 times revenue, and you have 50% capital risk.” I don’t know. I’m not saying it’s going to happen. I’m actually reminding myself that just good growth at a very pricey price is always tricky.

The great thing with 100% growth is it covers everything. Great. It solves all. Time for a wrap. Game on.