Maybe the Pentagon is wrong and they need to buy more Anthropic and just point it at the enemy. It'll bring China to its knees, right?
This is 20VC with me, Harry Stebbings. This week we analyze Anthropic's security release, Figma's earnings breakdown, and of course, the Citrini research piece which wiped billions off the stock market.
When you are priced for perfection, anything less than perfection will be a kick in the nuts. If you look at all the publicly traded B2B companies, there's only one that has a competitive agent. It's Palantir. I'm going to say we're going to produce 100,000 decamillionaires out of these AI leaders. Okay? Almost all the B2B software we use today is terrible now because AI software is so good. If the only thing that's impacted here is the B2B software industry, my suspicion is the rest of the world would go, 'Yeah, I'm willing to lose those guys.'
Here's the greatest dislocation. If I look at the public stocks, right, ready to go, boys. It is so good to be back. I love this.
We're going to start this week in Anthropic with some news on Anthropic. Surprisingly, security review feature wipes out 20 billion from cybersecurity stocks. Obviously, Anthropic released their latest security product and it massively hit some of the biggest players, Cloudflare, CrowdStrike to name a few. Is this a dramatic overreaction from public markets or is there underlying truth to this?
Well, I'll tell you the interesting thing to me. This kind of shows where the markets are and where our mass panic is. Although I think our panic is well grounded. I think we should be panicking. But most of this already exists. You can go to Claude Code today and I literally did this on the plane flying back last week. Okay, I did it inside of Replit and you can say run a detailed security audit on my code and it will already do it. It will already run a total static code review and if you're on Replit or Lovable or Vercel or something, it will actually do penetration testing and everything for you. Like literally, it will do everything. It can already do, I think, a better security audit and testing than a mediocre board engineer will ever do. So what I mean is, it's obviously the pace at which things are getting better. I mean, each week we can't keep up, you know, on this show. But it's also interesting that conceptually, just like reviewing Cobol code dropped IBM 10%. It already existed. Like we're panicking about features that have already been in Claude for months in many cases now. Is this a sign that it will go more aggressively into the space? Of course it is, right? Of course it is. But I just want to point out to folks that aren't doing it, so much of the stuff that we are panicking about I think is worth panicking about. But a total nothing burger in the sense that if you're paying attention, this is months old news in some ways. The first comment, the general comment is, you know, Anthropic markets itself as the nice AI company. And for a nice AI company, it sure creates a lot of damage and kills a lot of stock portfolios. Maybe the Pentagon is wrong and they need to buy more Anthropic and just point it at the enemy. It'll bring China to its knees, right? This is an astonishingly destructive...
Incredibly distracting. Incredibly destructive.
Have a press conference, clip 20 billion on something. But let's do security first because in this case, from a technical capabilities perspective, no surprise. Second, from an adoption in the enterprise perspective, I don't see Anthropic eating up CrowdStrike's business. I just don't see that. People are going to have a security layer. And there was a great, I think it was an HSBC report out today after the Citrini madness, which we'll talk about in a second, that basically said software is the means by which AI will diffuse into the enterprise, which I thought was a wonderful quote and I think it'll be true for security also. In other words, these capabilities like code scanning will diffuse into the enterprise probably by means of companies like CrowdStrike and other companies like them bringing it to bear. So at that level, a, no new news, and b, it shouldn't be such a big panic. But, and this is the big but, unlike some of these other sectors, CrowdStrike and the security companies were effectively trading at a price that assumed nothing could ever go wrong, right? Even on Friday after the correction, you know, I went in to look, I was stunned. It's 22% revenue growth projected, 31% cash margins, and it was trading at 16x revenues on Friday after the first hit, right? The thing is this, when you are priced for perfection, anything less than perfection will be a kick in the nuts, right? And that's what's going on here, right? I mean, it's entirely separate when you're dealing with the companies that were priced at six times revenues and they've gone to four. That's a separate discussion we'll have later. But I think what's happening on some of these super high-ric stocks is the occasional reminder when you're priced to perfect, literally any little thing, any increase in tail risk of you not being the winner logically, you know, corrects pretty substantially. So I think that's all that went on here. I don't think CrowdStrike is obviated by this. I think it'll totally be fine. It'll continue to be a business. But when things are priced at extraordinary high prices, it doesn't take a lot to knock the narrative off kilter.
Would you say that CrowdStrike with the repricing is now fairly priced or underpriced?
It's actually a very good question and I'll make it more general. Are you saying, do you prefer the stories where the AI disruption is perhaps a little remote but the values are still pretty lofty like CrowdStrike, or do you prefer the ones where there's some of these stocks that have corrected to six, seven and eight times 2026 EBITDA where you go, just on a value basis, that's wrong? I mean, I was thinking about companies like Toast, I was thinking about one of our own, DocuSign, at seven or eight times revenues. These things are absurdly cheap whereas CrowdStrike is still not yet cheap. It's just marginally less. But it does have probably more clarity on the way the world is going. I probably, and I'm doing this on the fly and I reserve the right to change my mind next week, I probably, even though it's a, I probably err to a bit of value. In other words, I just say to myself, buy the things where the cash flow alone makes it easy. Would you rather have your money in CrowdStrike, which is still priced well, or Monday priced at one and a half revenues? I'd like a basket of one and a half times revenues, not an individual stock. Because I think at the level of individual stock, it's hard to say because it's very idiosyncratic. At the level of the basket, if you buy 20 stocks at an average of three times revenues, eight times EBITDA, I think you'll do just fine.
I don't know, man. Just last week, you kind of mocked me and I can take it, it's good. You mocked me for saying Shopify is at partial risk of disruption. And I said it's not at total risk of disruption, but the fact that whoever builds the agentic layer, more and more value is going to accrete to them, you only have to have a partial deceleration to your numbers. You only have to have a partial risk. You only have to see more of the value of HubSpot or DocuSign flow to an agent, right? For these stocks to do worse.
Look, I'll give you an example for e-signature. I know this space, DocuSign, right? It's a great company, great company. We both know and a lot of folks were saying, oh, this is going to be destroyed by Claude because all you're doing is creating an image of a signature. No, this is a very complex enterprise workflow system that also is a partial system of records. Someone has to say these contracts are true and valid and someone has to route it through a hundred steps in transformation and negotiation and all this stuff, right? So that is not going to be destroyed inside of a Claude chat, but some of that can be done by an agent. And an agent that truly auto-contracts for a business, okay, let's say it does all of your commercial transactions. That could take enough of the value away that these companies are maimed. And I think, and I even think it's logical for CrowdStrike. Obviously, Claude Code Review is not doing endpoint security, but can anybody be maimed by Claude? Claude just launched an entire enterprise agent solution today. You think anybody can't be? I think anything, this is how I think about it. Claude is like inverse, sorry, what's the game where you fall out of the sky onto the island and it keeps shrinking? What's that game called? I used to play it. What's the game? Everyone plays the game. The video game where it keeps shrinking. Right. And so...
And the one, okay, but that's like, your territory keeps shrinking. Claude keeps consuming more and more of you and you're stuck in this smaller and smaller island that you got to own more and more market share of. Almost everyone that's public, your surface area is shrinking because of Claude and AI and the question is how much. And so, you know, I don't know that we've seen the bottom of these impacts. I think the impacts are accelerating and that's why Anthropic today had to publicly say, hey guys, their head of AI said, listen guys, on the one hand things are accelerating so it's bad for software. On the other hand, we're best of friends of software on the day they launched enterprise agent. We're best friends. We're enabling them, but the pace of change is so fast. So I want to believe there are these safe islands. But I believe the agents are going to own enough of the value that just owning less of the value in your space can create terminal decline. Terminal decline.
Put me down for a strong... Give us the rest of the year. You're going to come around on me on this one. You're going to come around. Just give me the rest of the year.
Literally on the plane last week, I did a security audit in Replit, which is using Claude Code. It's, I mean, some of that is Replit. We could talk about it. I don't know the details, but it's great. And I literally sent it to the Replit team, the entire technical team, because they're on like, I'm like, 'Do you see how good this is?' And they're like, 'We didn't even know it was this good yet. We didn't even know our security audit had become this good last week. It's so good and so much better than before.' And they didn't even know. And so to think that we're in some sort of static world for the rest of the year, I don't even think, like literally, we could delete some of our podcasts from four weeks ago. They're so dated.
So let's try not to be like that. So a comment here, intelligence will infuse all software over the next decade. That intelligence is generated by foundation models like Claude and like OpenAI. I think that's a given. Agreed. So what we're really saying is how much of that do they do themselves. That intelligence has to get to the enterprise and there's probably four ways it can get there. Lay out one is they buy it from Claude directly, everyone buys all the software from Claude. The second is that they build it themselves, they construct every enterprise constructs its own agents. The third is they buy it from existing incumbents who integrate AI. And the fourth is they buy it from gazillions of new companies, all of whom are leveraging, you know, like the Hobbies logo, all these companies who are building on top of foundation models. You founded post-22, right? And you have to figure out which of those scenarios you believe in. And actually the post I read and I actually agreed very much is I think it's three and four. I don't think enterprises are going to build their own systems on top of Claude directly and I don't think Claude's going to build all these focus systems for everything. So I think there's going to be this software mediation layer between them and that's the opportunity.
Here's the thing. If you look at all the publicly traded B2B companies, there's only one that has a competitive agent. It's Palantir. No one else has seen a single ounce of revenue acceleration due to their AI agents. And yet, the companies we talk about each week have jaw-dropping acceleration. It's not just Anthropic because they have built the agents that matter in their space. It is not sprinkling AI dust on top of their analytics software. There's not enough value there.
But let's talk about that kind of fourth category. You would also be that there are many privately held recently founded companies exploding in revenue also. Let's talk about, like, do you think it all goes to the model companies or do you believe any of these companies like, you know, that are raised, I mean, you mentioned, you guys talk about Replit and Lovable a lot. Do you believe that's defendable? Talk about, you know, they even might not be defendable in the rate of because Claude Code just last week, just since we did the last show, Claude Code launched the ability to see apps inside of Claude Code. So for a lot of product people, you don't need Replit and Lovable anymore as of last week. Now you can change Claude Code and visualize your app inside of Claude Desktop and inside of Claude Code. You don't need to do that anymore in a third-party app. And so they're all being... but my real point is, sorry, and Harry, you're the boss. I've talked to three founders over the weekend of like public and near-public companies and they're all, and this is the advice I gave them. Your agent is not great. You're being disrupted by the agentic layer. Now I hope that ServiceNow builds these great agents and I believe Agentforce has a shot and I believe others, but they are all being disrupted in real time and that's why the folks Harry interviews on 20VC are stressed as f. They are stressed as f because no matter what they say, they know they do not have the dominant agent in the space. No matter how many LLMs they stick inside of a feature. When I see what many leaders of top public and private and our hundred, you know, nine-figure unicorns are saying, it's lip service. I see it on LinkedIn all weekend long. I see we're adding AI to our email feature. We're adding the ability to process emails more efficiently. You're going to go out of business and you're not going to fail because your customers are going to renew, but your growth is going to fall so far that you become irrelevant in two years. And that's why I'm going to buy my four stocks, don't get me wrong, but I'm already changing my mind since last week. Because things continue to, the panic is overdone and real at the same time.
So you made a very clear statement that we have not seen any of the public providers make great agentic use cases work and have a meaningful impact on revenue. I'm very naive. Why? Toby's a brilliant CEO of Shopify. You know, Mark is trying with Agentforce. Why have they all failed so far?
Well, I'll give you two reasons if you want. There's a long list of reasons, right? I'll give you two practical reasons. One is it's a lot of work, man. And no one wants to do this at these companies. No one wants to. You don't. Here, let me finish. Every agent, here's the problem today, and this will change in the next two years. It is not true today. Every agent is essentially custom, okay? Every agent needs to be trained. Every agent needs to be onboarded if you want it to be great, okay? Every agent needs its data cleansed. And so this is a vast amount of work for organizations that already think they're overworked and working too hard. There is huge institutional momentum to overcome. The second is what I just said is true. You need a massive amount of forward-deployed engineers and trained workers that are technical enough and smart enough to train and deploy an agent. One, the workers don't exist in most companies. The average customer success person that shows up with a green, yellow, red light dashboard cannot train and tune an agent. And two, then there's a meta challenge for the Shopifys, Mondays, HubSpots, and Toasts and others, which is you can't afford the human to do it. And that would be fine if you have... So what we're all at a niche level, we're seeing it with startups, we're not seeing with publics. Hyper-niche agents work really well because they have a small set of things to do, okay? As soon as you get to spaces like Shopify or even worse, Monday where you have a hundred verticals, right? It's very, very hard to build a very specific agent automatically that does everything that churches need and basketball courts need and refrigerator businesses need. They're not all the same needs. And so the agents aren't good enough. And in fact, a lot of these leaders that I describe, if you, they're in beta. They have six people using them, sixty people, because it's too hard for them. So they're going to get killed by the startup that does it. They're going to get killed.
Killed. But that's the key, that a startup that does it.
What it's not going to be is the foundation model directly selling to the church, directly selling to the thing. They will provide the raw intelligence, but there will be software opportunities to build compelling software companies in most of these verticals just as there was in SaaS, right? And that's so maybe we're more in sync than we think. Anyway, I think, let me just say clearly, I think intelligence-led applications are the only applications that are going to sell and grow quickly over the next ten years. Non-intelligent applications will at best be flat to mild growth if they're not obviously disrupted by intelligence and at worst be down. Right. Yeah. And I would argue, for example, I'm picking on Toast because I'm not an investor, so I have no emotional connection. I think that's a good one because I don't think there's a ton of agent work to be done and it's a lot of payments and restaurant kind of organization, I think is fairly durable. We can argue, but yeah, whereas something like Monday, it's very knowledge workery. I can see much more disruption story, right? But that's the incumbent. But your point, what you, I think we're now in sync and saying is that those opportunities can be grabbed by standalone companies perhaps built, you know, started, founded 21, 22, leveraging directly on top of the foundation models. What you're not saying is all that revenue just accrues to the foundation models, correct?
No, I think just people are going to be maimed even more than they think if they don't own the agents in your category. You got to own the agents in your category. And whether those agents are owned by a startup or whether some version of that agent can be done inside of Claude, they're going to maim you. And it's accelerating.
Jason, I want to be very direct with you and you'll give me a direct answer. You're better at direct than Rory's sometimes nuanced answers, which people love also. But will Claude Code make Replit and Lovable weaker in a twelve-month period? Like, do you think they will meaningfully enter their space and take market share?
I think it will do everything that it can do. Going to Rory's point, anything that can be done either inside the browser or inside the desktop, Claude will do. That's what we've learned this year. Anything it can do. So right now, if you go to design something in Claude Code, you'll laugh. All the crappy Claude Code websites look the same. They have the same artifacts, the same icons, the same purple color scheme. You can laugh. But can all of the design, all of the parts of Figma that are design be done within the code? Of course it can. So I believe it will aggressively attack parts of Figma this year even though they're key partners. Replit and Lovable, it continues to do more of them. Now, will Claude Code want to host entire websites? Will it want to go into prod? Will Claude... the only thing that ultimately will protect them is this is all their teams do, right? And I don't think they want to build databases and build production websites and host domain names. But if they change their mind, there's no reason an agent, there's no reason they can't license Supabase or Neon or fork their own Postgres. They're pretty good there at that company. They can build their own database. They've already got plenty of servers. They can spool up a few more to host websites if they want to. They can. And I just think that, oh, Fortnite's the game, you know, at the end of Fortnite, the circle gets... and so if I'm at Replit or Lovable or even Figma, like, don't, I would be worried if Figma, like, the circle's just starting to shrink. But I think at many companies that circle's shrinking. And I have two investments I made last year, I love them, those products no longer have a reason to exist today because of Claude. I made two investments, now it's not going to bankrupt the fund and they've tilted, they've evolved, but these were standalone investments last year, I'm not going to go into it, that were great, that blew up in the early days and they just have no reason to exist today in that prior form. Just no reason. And when we turn around and all of a sudden you can preview your entire app inside of Claude, which you couldn't do last week, if I'm any of these, Figma, Replit, Lovable, Vercel, all of them, all of them, I love that, just the Fortnite circle shrinking. So you know, you got to do something about it, right? And will it stop? Like at the end of the game, it does get pretty small though, right? It's stressful. But the flip side is if you nail the agent, look how much revenue these guys did building essentially an agent on top of Claude Code. They built a billion dollars of revenue building the agent that didn't exist. So that's the flip side. That's our job is to build these billion-dollar agents. If you can do something that is extremely high value that could not be done before, you can close millions of revenue your first week. It's never happened before in the history of software. Right.
Before we move on, I do just want to stay on Anthropic and I wanted to discuss they've lined up five to six billion dollars for an employee share sale at a $350 billion valuation. Obviously, there's people queuing up out the door for this. This was following obviously OpenAI doing the same a couple of months ago. Have we ever seen liquidity at this scale when we look at the number of millionaires minted from OpenAI, soon to be Anthropic? When we look at the Nvidia millionaires that exist already, is anyone going to be able to buy a house in the valley?
There's a lot of things that seem silly in AI twelve or eighteen months ago. But if Nvidia has twenty thousand decamillionaires, right? I'm going to now do a WAG, but based on the math, I'm going to say we're going to produce a hundred thousand decamillionaires out of these AI leaders. Okay? And Rory's better at the math than me, but Nvidia already has twenty thousand. So if that's the case, that would be great if everyone was at full employment and we were all full of companies with six-figure employees making money. The worry is, I don't want to spend too much. The worry is that we do see this concentration of wealth at the same time as everyone else gets leaner. And will Jevons paradox create more employment in tech or not? I believe it will create, I don't believe, I believe we will need more engineers than ever, like when we talked with Michael Kenon Brooks. I don't know if it will create more employment. So.
Listen, the knock-on effect of that is bluntly what has shaken markets so much. When you look at Citrini, or however you want to pronounce it, 2028 Global Intelligence Crisis, which was a piece written kind of forward-looking or predicting kind of the state at 2028. And I broke down kind of some of the core elements there. And I think we can start with actually one that you kind of mentioned there, Jason, which is kind of ghost GDP disconnects market from the real economy. So you have AI boosting productivity and corporate profits and market caps of companies inflating headline economic figures, but actually consumer income doesn't scale with market cap and enterprise value. How do we analyze and assess that?
I'm going to call [__] here.
Right. I love the Noah Smith descriptions Substack who I follow in economics a lot. Yeah. He called it basically scary bedtime reading. Right. And you know, but I think if you want to have a conversation about, I actually think the way you need to break it up is to, I'm trying to say how to have, because I knew we're going to have this one, right? And I saw your seven points and it's too much detail, Harry. Big picture, I would suggest we approach this in the following fashion. The first thing you have to figure out is actually know micro macro at a micro level for each of the things he says are going to happen. Do we believe they're going to happen? In other words, is AI going to replace coding? Is AI going to replace DoorDash? Is AI going to replace Amex? Right? In other words, and the wonderful thing is I think that's something this group is well equipped to do because we're all investing in venture companies who are the tippy point of the spear in terms of adoption. I'm allowed to have an informed opinion on question one, which is micro level. Are these changes going to happen? Then the second big picture question lumping all the other things in together is this. What are the macro consequences of this? Right? In other words, if you assume that there's a high level of AI adoption over a short period of time, then everyone, and that's as I say what we can talk about because we understand, then everyone gets to pontificate on global macro, which is what you were starting to do there, right? And I'm fine coming back to that, right? I think he was, and on the show in other words, and revisiting this in a few minutes, but I think jumping straight to the, oh my god, the GDP is ending with no thought process to, I even believe that basically he was hypothesizing a two-year, two-year adoption cycle of almost everything such that everything from ServiceNow to DoorDash to Amex gets rolled over in two years. And I think if you believe that's happening, going to happen very quickly, then you do graduate to the global macro question. And I still think he's wrong about that. But if you don't think the adoption is going to be that quick, you can literally ignore the rest of the piece. Does that make sense as a framing first of all?
It totally does. Why don't we start with the micro then, where you think we are seasoned and responsible enough to have an informed opinion?
Well, yeah, let's break it up. I mean, there was, I mean, he points about software development, about SaaS apps, about companies like DoorDash, and then companies like Amex, right? In other words, interchange, right? Let's do the stupid ones first. DoorDash, right? The idea was you're going to want to delegate to your agent purchasing of an opt, you know, optimizing for, there's going to be six different versions of DoorDash and your agent will do between them. And the only reason that you don't do this today is because quote unquote friction. And if it was automated, you'd let the agents order your pizza. And I just call [__] on that. I mean, at the consumer level, I mean, I don't know how you felt when you're organizing, when you're buying pizza on a Friday night, you're talking to your wife and you said, 'Honey, you know, we had the freaking kimono last time. I say we go with the pepperoni.' She was like, 'No, I don't like it. I told you we want to stick with the two salads on the pizza. It's not something we want.' No one wants to delegate to an agent how to decide what food they get and then have the thing come up. 'Good news, I saved you two bucks. Bad news, you like the high-end pizza, but I got you the crappy little pizza.'
Rory, can I add just one thought? Here's Andy Fang, CTO of DoorDash.
'We strongly believe agent commerce will be transformative to our industry.' He believes this. I believe it has. I have a large investment exposed in the space and I can see agents and AI ripping through it. This is what I don't mean to interrupt, but like I think the examples we think are safe. This is CTO saying we need to earn the right to service customers agents end to end, discovery, ordering, delivery and support. We need to earn the right in the new world. So to think that these spaces aren't just, it only has to be maimed. It does not have to be the idea that we're going to vibe code our own DoorDash is stupid, right? It is stupid.
And he's trying to get people millions of views, right? So the core idea he started with is stupid, but keep on extra, right? Now that we've agreed to...
that an agent can decide for you, here is the CTO says it's real. That's the threat.
But he didn't say, 'Okay, you caught me. I'm going home.' If you want to make more automations around a recommendation, say, 'Hey, there's three different pizzas.' But do you really think it's going to be rep... I mean, I just think the level of customer inertia on the consumer to move to have this vision of five different DoorDash competitive companies being enabled in this world. I just don't see it. DoorDash is a combination of a huge amount of logistics, a huge amount of customer aftermarket service, a huge amount of signing up restaurants, right? And the only thing that's changed, so what's, and you know, they beat off four or five other big competitors to now have some kind of stable oligopoly with Uber Eats and there was one other smaller player in the US, right? What in software is going to allow a new, talk me through the new competitor emerging and taking market share in a high fixed cost business like this?
Well, I mean, the most simple one is that a new competitor can decide whether Uber Eats or DoorDash is the right thing for Uber Eats, DoorDash or Direct. You have three options in the US. Okay, there's nothing else really exists, right? The agent may make that decision. In fact, I would prefer that because I don't want to figure out which one to use, DoorDash, Uber Eats or Direct. I would prefer that Uber Eats knows my favor, and I'm just picking one example, but this is a real threat today. This is what Andy is saying, it's a real threat that the, you don't, it's no longer, we don't need to decide which is the best place for us. The agent decides which is the best deal between these options, which is the best source for me, which is the best for my family, and it makes the decision. It only, if the agent is who we go to, as long as the agent is who we go to rather than the first party, it just risks disruption. It doesn't destroy the company. It doesn't destroy it.
Yeah, it doesn't destroy the company. Okay. You see, now we're... Okay, so tracing through the unrealistic statements there. Let's go an existing example today. Netflix recommendation.
But I'm right about DoorDash because Andy Fang said the same thing. Why do you, why are you dodging the one that everyone thinks is free from disruption when the CTO says the ground is shifting underneath his feet? He literally said it this week. The CTO.
Okay. No one's going to say, no one's going to say when the CEO of a public company, 'I don't believe that stuff.' They're going to say we're on it because that's the message you got to say, right? But I don't think they're saying, 'Oh my god, three more companies are going to displace me.' Right? Our job as investors is to be kind of analyze the facts and try and come to our independent conclusions. So I'm just going to take the idea. I think two of the most personal things at the consumer level are the food you eat and the TV content you watch. Right now the good thing about net, about the content you watch is we've had ten years of AI already. The Netflix recommendation engines, let's be clear, it is a massively useful tool to them because at the margin it helps them predict what people want, right? So I do agree there is core value in knowing people's preferences on the aggregate, right? But how often, what percentage of your recommend, what percentage of your content viewing do you base entirely blindly on the recommendation engine when you sit down at Netflix? Five percent, ten percent. I think it's light.
Sorry, but what DoorDash is barely in B2B? What the, what the Netflix point is? I'll answer your question if you want the answer, but I'm missing the point. Do I think AI can disrupt Netflix? Netflix thinks AI can disrupt Netflix because we're all watching short-form content. And as of the last forty-five days, you can watch an incredible short on YouTube that was entirely AI generated where you can watch Star Wars stories that are better than the crappy last three movies that are AI generated. That's utterly disruptive to Netflix. Utterly, utterly, they're so panicked they have to buy a studio.
I'm breaking it apart into two separate things, right? If you can generate, and I think that's actually useful. If you can generate content using AI, that's very disruptive. I was trying to focus on recommendations because the idea was, I'm just looking at the idea of the idea is for consumer preference. You will entrust your decision-making to an agent who will quote unquote know what you want. That's what you're saying about DoorDash. And we're trying to prove something that hasn't happened yet. So I was making the point content is another thing that's quite personal. And Netflix has had this agent running for the last ten or fifteen years, the recommendation engine. And at the margin, it does a good job of predicting on aggregate what people want. But you, if you had two choices, one program that gave you exactly what they recommended, and then the other program that allowed you to pick, I'm going to tell you, you're not going to go with the recommendation engine. You're going to say, 'I didn't like that recommendation last week. I'm done.' In the same way, I don't believe I want to entrust my eating decision entirely to an agent. I think we're just caught up in this. Anything could happen, right? But dude, YouTube is the number one way we consume video and it is entirely based on a recommendation engine. And it is the best recommendation company on planet Earth. There are no channels on YouTube that matter anymore. Followers don't even matter anymore. Nothing matters. Every day I log into YouTube and it gets better and better at knowing what I want to watch every day. It is epically, it is epically better than anything else on the planet and utterly disruptive to how we view few things.
It's a good argument. I'll give you that.
Everything is, everyone is stressed today. It's a good thing. Everyone is stressed because every, the further you go, again, I don't mean to, every, the further you go, the more folks are at risk of being maimed by AI. Just maimed, except Georgia CrowdStrike. He's fine, but everyone else is at risk of being maimed by AI. It doesn't, even if your growth goes from thirty-one percent to twenty percent. Like that's a big deal. That's a big deal.
Agreed. But do you really believe that you want an agent to recommend you food?
Well, listen. I don't, I have, I have the data. I have an investment in this space. I already know the answer is yes. It's not my opinion. I have, I have data from over ten thousand restaurants. I know the answer is yes.
Right. If your agent all the historical context on every pizza order you and your wife have made and you know the price point, the location, so you know the delivery time that you're estimating and it can also analyze every TikTok, Instagram review as latest food trends and tell you about the latest within that price band with that crisp topping that your wife likes. Because it already knows that because you left a review or because you said it in a WhatsApp, say it's got open Claude and was able to deliver that to you. I think most people would, I'd love it. I think that, but Harry, I think that at the margin, yes, I'd be sitting there on DoorDash going, I get a five percent extra satisfaction rate and selection rate from this. But do you think AI is so disruptive that it can warrant the creation of an entirely new company that says basically what you're saying is all the investment you made in restaurant relationships, logistics, the app, consumer, right, are as nothing because this new thing is disruptive enough just... No, I don't think anyone said nothing.
So it's not. So do you really think? Let me...
all these SaaS apps become dumb databases and Toast just becomes a POS system. They're not going away, but they become more and more commodified. And more importantly, they don't, the real issue is they don't capture enough of the incremental value. It's the incremental value we're investing in.
Jason, I wanted to deal with the easy ones first, right? Do you really believe that there will be a direct competitor to DoorDash enabled by AI?
No, listen. We all, we all agree that that is, that is clickbaity. Okay. But, and so is all of it. But listen, even if, let's talk about Ghost GDP for a minute, okay? I admit I'm living on the bleeding edge. Okay? Let's concede that I'm a laboratory. But we have gone from twelve people to two people on my little team. Okay? That, that not only does investment, it generates eight figures a year in revenue. Okay? And that is Ghost GDP. Those folks that are gone, that value, the profits that are left go to two people. Well, three people.
Ghost GDP is that this productivity is not going to human workers that then spend it. That's the fear that we are creating this productivity, but it is not going to any, there's no, there's no humans to spend the money. It's great that I can spend more money, but that doesn't, that I don't think that's great for the economy if I get a little bit richer. Like we lost eight people on our team, right?
Okay. Stop. When, first of all, I agree we're now doing macro because, and in this case we can do macro because we agree we can only do macro once we've conquered micro. And in this case you've conquered micro. You've said it has happened. I have two, I had twelve people. Now I only have two. Ten people no longer have jobs. Is what does that mean? Well, at least, at least whatever they have other jobs, but the value that we're creating, right? This eight figures of value, it's, it's accruing to fewer and fewer people and there aren't as many people to buy, to buy handbags and to buy shoes and to buy t-shirts and to buy Netflix. Even just to buy Netflix, there's less people, right? That was the point of the Ghost GDP in this inflammatory annoying article. But I don't, I don't think that these things are wrong. I think he's just trying to claim everything's going to happen in eighteen months, that, that, that, that it's not going to happen in eighteen months.
What we're basically, but implicit in that, and you have to be logical, implicit in that is productivity gains which have been the engine of growth for the last two hundred years are bad. There's some embarrassed statement here is that it is bad that Jason is now able to do something with two people that he was hitherto for only able to do with twelve. Right? And I'm going to say something. Across the arc of the last two hundred years since the industrial revolution, productivity gains have been good, right? Because the other ten people who used to be wasting time writing slop for Jason can now do other things and the sum total of human achievement will contain the extra work that those people do. In the long term, I don't think you can argue but that productivity gains are good. Right? We used to have eighty percent of people working on farms. We now have ten percent of people working on farms. Actually, four percent of people working on farms. And we sell so much food that we're all fat, right? And those other eighty-five percent of people are doing other [__]. Across the scope of history, productivity is freaking awesome. It's the only thing that's made us rich. So I want to say that so clearly because then all these arguments, these macro arguments that I was hoping to ignore are basically some version of even though productivity is amazing in the long term because you can't disagree to that, something bad is happening in the short term. And then the onus is on you to say what that is. What is it? What's so bad about constant change and the fact that jobs go away and new jobs emerge? What's going to happen that's so bad?
A softening of consumer spend with a concentration of wealth to fewer people. There are less people to spend money across different parts of the economy and that impacts a large amount of people. And to your point on how does productivity actually lead to a worsening in the economy, you only need to look at Japan in the nineteen nineties and the need for economics. And actually you saw massive productivity increases in the nineteen nineties with massively improving mechanical infrastructure that they brought in and actually it didn't disperse to a huge amount of the Japanese population. So that there's very recent precedent actually.
If you want to see a sat, a dystopian version of this, go to Japan and meet with B2B founders. I was at a dinner last November with all founders from IPO twenty million and up only. It was a VIP dinner put together. It was great. It was, it was the best of the best. Okay. And they're all talking about how inherently their seat base shrinks. This is not just the AI topic we're talking about today about seat-based risk. This was last year all talking about how just each year their seats shrink because their economy shrinking, right? It is, uh, just to Harry's point, it is a, there is, I think in the short term this is all great for us as investors. It's terrific for us to get more productivity. Well, we're going to get, we're going to make money out of it and we should put it in the bank and flee to Miami or Monaco because I don't, but I'm not sure it's good for everybody.
I'm going to call [__] on that. There's two or three different things with disaggregating and seat-based comment, right? Um, there is, are you making a comment on Japanese depopulation, which I don't think we can blame on AI? It's been a trend for thirty years. I just think there's a loose parallel to this Ghost GDP idea to Harry brought up of the depopulation. Well, there's different ways you can, you can depopulate a worker force even if the humans are still there, right? But it is, it's a similar, it's, it's, it's a structural headwind, right, to folks buying stuff. Our ten agents at SaaStr generate millions of revenue, but they buy nothing. Our agents buy nothing. They, they work all weekend long. Repley, Art, Qualie, Monty, they're good kids, okay? They, they, they create a lot of noise. Like they're a lot of work, but they buy nothing.
Nothing except, except, except tokens. That's the only thing they buy are tokens and they, they buy millions and...
Millions of tokens for real, right? For real, that is a little different than the past point is.
Productivity increases are only good if the consumer wallet is dispersed and they are able to spend money. If that shrinks, that is not a good thing.
Yes, but so what do you do? Ban productivity increases? Good news, we're all doing fine. Bad news, we're all in 1790 and we're all, you know, one bad harvest away from starvation. But yay, we're all fine as long as nothing goes wrong, right? It's not a credible argument.
Look, I do agree at a very micro level you could argue if a... and this is why we actually have to go back to the micro. If the disruption happens extraordinarily quickly and people don't have time to adjust, then in the short term you will have some element of structural dislocation that will result in some form of, definitely some form of recession, GDP slowdown, if those folks can't be digested in new jobs quickly. So I do agree.
So my point is this: in the short term you can articulate a thesis like you're just saying, right? If all the, you know, if all the 45-year-old programmers are let go at the same time and there are 6 million programmers on the street and there's no other work for them and it happens in a month, then in the short term there would be this GDP hit. While I'm still correct over the medium and long term, GDP growth bails us all out.
So that's why it does go back to the micro: is do you think it's all going to happen so quickly? Do we think that all these things are going to be displaced extraordinarily quickly?
I don't know. When I look at Karpathy talk about the evolution of how much of his work has gone to AI in the last six months, I do question it.
Okay. What do you mean by that?
I think more and more of labor will be replaced by AI. We will see the concentration of value to fewer people, and fewer consumers will have money to spend in the economy, which will lead to problems and a shrinkage of that economy.
There are 150-odd million people working in the US, right? So what's your estimate for displacement? I'm just trying to get a sense of it, like, you know, you have...
It depends on the time horizon, but I'm not feeling that great about it. I mean, if we look at the most obvious, which is, you know, customer support, customer support, legal, bookkeeping, um, that doesn't look great. If we want to add in Waymo and what it'll do for self-driving within a four-year period, gosh, I think you could see 30 to 40 million.
I think again, Waymo was a good example. I mean, you look at the projection. I mean, again, you look at the projections 10 years ago when people talked about self-driving, and you can find all the, you know, the Catrini articles of then saying it's all going to happen in four years. It's over. Here we are today, and even though I think Waymo is amazing, they're doing $350 million. I think they have single-digit thousands numbers of automobiles, you know, in a few cities, right? Steady rollout, started to increase probably, you know, I don't know how many years you're talking about before it gets to kind of further mass scale.
So it's a very long... almost every, all diffusion takes longer than you think, right? I think we massively overestimate the pace of adoption here, and you massively underestimate human.
Does it ever take shorter? Because everyone always says about the pace of diffusion, and they use the, you know, industrial revolution where you had to buy machinery, transport it, train it, when it's, uh, nano banana pro and it removes an entire industry.
Well, I think it's shorter, much quicker in my life. It's my life. It's shorter. And the thing that is stressful on this, and it's a way to make money, is the stressful thing is that if you're... and again, I don't want to endlessly talk about vibe coding apps, but I have so much experience, right? I'm shocked everything is faster than I would have ever and better than I would have ever expected.
We started this about security, right? If you told me when we started this podcast that today I could just talk to a vibe coding platform and it would do like an A-tier security audit while I was on an airplane and I didn't have to do anything, I would have said, 'Guys, it deleted my whole database. There's no way it's going to do a freaking enterprise-grade security audit.' Yet, here we are.
And then we turn around and everything and everyone's everything Anthropic can do this year is faster and better and bigger than we thought. I'm not saying at some meta level everything isn't slower than we think, right? Certainly that dumb article, right? DoorDash being disrupted by Base44 next week is dumb, but god, this acceleration, it's just, it's hard to... it's so fast.
I'll give you the practical ramification. Almost all the B2B software we use today is terrible now. It's terrible. I can't talk to it. I can't, like, I can barely bring myself to use WordPress. I can't change anything in WordPress. All these products are terrible. A lot of the ones that we, the founders we love that Harry talks to, the products are terrible now because the software is like, AI software is so good. Blow your brains out to input data for two hours into your system, right? It's terrible.
And if nothing else, that is accelerating so quickly that the leaders cannot keep up with the fact that their products are so dated. They're so dated.
I think that's broadly true, and I think when you narrow it, that's why I said when you narrow it down to the micro of the impact of AI on the million to a million and a half workers in the US in the software and tech industry broadly defined could be way more disruptive, right? And I think that's actually a useful conversation.
But it's worth pointing out that, you know, plus or minus 5-7% of all jobs in the US are disrupted every year, right? And, you know, this is 1%, less than 1% of all, maybe one and a half percent of all jobs. If the entire software industry got nuked, right, it still wouldn't be the same as losing the car industry 10 or 15 years ago.
So my point is again, differentiate the micro discussion of are B2B software companies in trouble and how much are they in trouble is a really good discussion. And if you make your living investing in B2B software companies, it's the only thing that matters, right? But jumping from there to saying civilization as we know it ended is just, as you say, Jason, clickbait, and we should just ignore it, right?
The truth, I mean, I'm going to even more directly. We didn't bleed in Silicon Valley when the car industry went down the toilet. Don't hold your breath thinking they're going to come for us. And say if the only thing that's impacted here is the B2B software industry, my suspicion is the rest of the world will go, 'Yeah, I'm willing to lose those guys.' Right?
So I think we discard the macro, and but then I think you're right. The question is who wins, who loses in the, you know, 2% of GDP software business with AI. And I think that's a valid, that is the question. You're right. Is all that software... So do you... all that software is crap and looks outdated today. That's what you're...
