Anthropic Buys Compute From Elon & Commits $200BN to Google | Cerebras IPO | Ramp Raises at $40BN

There are categories of software where if they don't have a reason to exist in an agentic world, they will go into a terminal state of decay. If you're not accelerating, you're going to be destroyed, right? And at a minimum, you've got to raise guidance. ZoomInfo's growth was stolen from it from Clay and Friends. And it's a brutal case study. Give me 30% growth. Give me profits. Give me a story that's got some future in it, and I'll get you back to five times. I'll never give you 20 times again. I won't fall for that one this time. There you go. In capitalism. Great.

In one sense, not surprising. Companies like to keep control of their cap table and they probably lost a little bit of control. And stepping back because the SPV word is a little misleading. There's a lot of different ways shares change hand. A lot of different ways SPVs are used. So, let's distinguish. One is sometimes an SPV is when a venture firm has a big allocation, can't take it all up. So it forms a special purpose vehicle and that company invests directly in Anthropic, maybe raises money from the venture firm's LPs and the SPV appears on the cap table. Nothing wrong. Nothing to see here, right? That's not what we're talking about here. What we're talking about here is secondary sales of Anthropic shares, outright sales, we'll come to that in a second, or transfer of economic value. In other words, I own shares of Anthropic. I'm an early employee. For whatever reason, I couldn't access the company structured tender offer and I decide to do something outside of that. So, the first thing I do is I try and sell directly to Jason. He's willing to buy. I'm willing to sell. But Anthropic says, 'We're not going to effect that transfer. We have rights to approve transfers. We're just not approving this.' Right? And I don't know what they've been doing in the past, but that's what they say they're going to do going forward. In other words, they have the right to say yes or no, which makes sense. It's not unusual. But the real thing that's going on, and this is where it gets interesting, is what I could also do is say to Jason, 'Hey, Jason, I can't sell you my Anthropic shares because they won't let me. But I contract with you that whatever I get for my Anthropic shares later, I will transfer that value to you and I will structure a document that says, you know, I owe you contractually all the money there is that ever comes from that share.' Right? So now Jason thinks he doesn't have the shares, but he kind of has the economic right to the shares, so his money's good. And there's two problems with that. The first is probably the Anthropic documents say, Rory, not only can you not sell your shares, you can't transfer beneficial ownership of those shares to Jason. So I may not be able to do this. So Anthropic could say, 'Hey, we don't agree to that.' And that's fine, though. But the funny thing is Jason and Rory might still have their contract because I promised Jason something. Just because Anthropic says they don't agree to it doesn't mean Jason and Rory can't contract. But the tricky thing for poor Jason two years later when we go public is he doesn't have any structure. He can't go back to Anthropic and say my shares are in the cap table. I'm good. I can just sell. All he has is a commitment from Rory and Rory turned out to be a liar and a cheat and he never owned those shares or he owned them but he sold them to someone else as well. And it's just going to get really messy at that level as individuals start trying to enforce contracts not against Anthropic because they're not in the loop, but against other investors who sold. So there's going to be a lot of losses at that level. And I think Anthropic is just trying to distance itself from that and probably also worried that if there's been a lot of transfers, you know, I saw someone suggest that even though they say they don't allow these, you get into all sorts of equitable remedies. In other words, the courts start saying, 'Hey, dude. You knew this was going on forever. You kind of acquiesced to it. So maybe you are involved in this mess' and they don't want to be involved in this mess. So totally makes sense. As you get ready for an IPO, you're like, 'I've been a little sloppy here. Time to tighten up.'

And then the real thing is we'll see the secondary impact. To me, listen, maybe I'm missing something. To me, this story seems like a nothing burger made up on social media. Let me tell you why. First of all, all companies have had this for a long time. Most of our portfolio companies have the same provisions in it. This has gotten more and more locked down over my investing career, right? It used to be the lawyers wouldn't put this stuff in by default and then maybe four or five years ago, every set of charter documents say no transfer without the board permission. Just like Anthropic and OpenAI, okay, so nothing new. Two, Anthropic last year warned Menlo, one of its best investors and others, enough with the SPVs. They said publicly, enough with the damn SPVs. This isn't new. They said, stop the SPVs. If you want to invest, you have to invest directly. Okay, they were clear on this last year. I think this is Anthropic just saying this is still happening. They named a bunch of entities in their statement, which you never see. They named like five hedge funds or something and saying these guys are bad actors. So I think they're just, I think no one was listening to them is the problem. They banned it. They were clear that one of their lead investors, you know, one of the greatest venture investors of all time. They told them to stop SPVs and they told everybody and there's so much greed in AI people wouldn't listen. And so they named actors. And I think the interesting thing to me is when folks don't listen. I don't think they listened to the company last year. So then you have to go public on the megaphone and make sure everyone hears, 'Hey, these are bad actors and don't do it' because people are so greedy they take the risk anyway. The triple layered, quadruple layered 2020 SPVs, just you, this is the hottest share of the century.

It just shows, you know, needs must when the devil drives, as they say. You know, I mean, let's be honest, Elon in the past has said, you know, horrible things about Anthropic. They're evil, they're woke, they're anti-white, they're anti-Chinese. It's been, you know, as recently as 3 months ago and suddenly he wakes up one morning with excess capacity and Dario wakes up one morning with a need for capacity and Elon ends up in trial with OpenAI. So kind of the enemy of my enemy is my friend and here we are, right? It's a totally sensible deal. I mean it says a lot about both markets. It says this is a market consolidating and even though they will deny it till the day they die, what this is is xAI/SpaceX/Grok basically saying we're not going to be right now a leading edge model contender. Grok is not growing like OpenAI and Anthropic is and we're going to effectively switch from being a net demander, net buyer of capex because we're trying to build Grok to being a net seller because we're not going to be able to build Grok. I believe the data center was 11% utilized. Right? So this is the market consolidating. The stronger players have the capital to buy more capex and in the context of this, you know, Grok is a weaker player and they very wisely are opting to stop for now at least the dream and instead take probably four or five billion a year which is a big slug of revenue. I mean depending on the estimates 3 to 5 billion a year of revenue which you know for an asset that was losing money and you know for context the total SpaceX revenue run rate is around 20 billion so this is like a 15% revenue lift from something that you know a couple of months back looked like a money pit so the combination of this and Cursor I think they've taken xAI from a $250 billion which is what they paid for it in SpaceX stock which is astonishing from kind of a huh why are you doing this to okay I get it it's at least notionally profitable and has some kind of existence right but it's not a competitor anymore to Anthropic and OpenAI and trying to pretend it is is different so it's you've got your own little CoreWeave congratulations.

You know, we talked about this a couple weeks ago and I said something, you know, I usually don't use these kind of lines but I said something like this wouldn't be hard to imagine because like Samsung does it with Apple, they compete and sell them components. And we talked about this and said, 'No, you're ridiculous.' Yeah, they just, they would sell it to each other, right? Actually, the most obvious candidate now is SpaceX because they've got to have different business units, right? So, you've got, how, I can't even keep track. You've got the rocket guys, you've got the Starlink guys, and now you've got xAI, which is two things mashed together, right? And they each have their own P&Ls, business units, and responsibilities. And if I'm stressed and I'm running this xAI Twitter thing and I can get another four to five billion in revenue, it doesn't just make sense for the IPO, it makes sense for me. It just selfishly makes sense for me. So, I don't think the deal is as surprising. It's a reminder like, you know, keep meeting with your competitors. This is a classic disaster post. Always meet with the CEOs of your competitors. You never know where it's, it's never a bad idea to have lunch once or twice a year with your competitors. It's never a bad idea because this is one of those deals. And two, the more interesting thing to me is my god, you know, this changes so quickly. Just a couple weeks ago it was like, well, Anthropic can't launch Mythos because it doesn't have enough capacity. And now Anthropic, I hate the 5D, 8D chess metaphor, but now it's figuring out a way to have more capacity than OpenAI. Like, it's pretty epic to just hoover up anything that's available on planet Earth and possibly orbiting soon enough. But it will hoover up everything available, right? CoreWeave, xAI, anything, right? It'll probably buy capacity from OpenAI if Sam lets them for some reason.

You know, capitalism works and assets should get reallocated to the person who can create the most value from them and you know, right now Anthropic can turn that capex into the most amount of money the quickest and Grok could not, so I agree.

The person who is in charge of making the P&L work for the data centers is sure glad to get an extra 3 to 4 billion to make his math work and not get fired by Elon. Pretty happy.

It will be fun, by the way, to watch the SpaceX road show because someone's going to have to perform out. Well, this is what we owned a year ago. Then late last year, we bought xAI, so bear with me. We dropped all this stuff in here. That's what it is today. And we have a whole quarter of combined revenue. Oh, and by the way, we then sold all that capacity. And that should hit next quarter. So, you got to perform that in. And then the quarter after that, we're going to drop in Cursor, which hopefully Colossus 2 will be online then, and you got to perform that in. So basically, Mr. IPO investor, you're buying a $15 billion run rate company today, and it'll be a $23 billion run rate company in two quarters with two totally different businesses on top. That's why the bankers are going to earn a couple hundred million, but it won't just be an extend the model and grow 20% Q kind of analysis here.

To some extent at places like Google it's like okay like obviously we want to win. We want Gemini to win. We want to beat everywhere. But we're not strong everywhere. But let the best model win. Like we're going to win, but we're okay with winning with internal competition by selling capacity to Anthropic as well. Like it's okay. Like it's okay at their scale to let the best buyer of these different things, as long as, I mean they don't have infinite capacity, but it's not necessarily a bad way to keep everybody on your toes to have a little bit of the competition inside of you. It's not necessarily a bad thing.

It is interesting though. I mean I think there, so there's a couple of things with it. One is you know I think now the Anthropic revenue commitment is about 40% of Google's total future backlog right so it underlines quite how heavily dependent the hyperscalers are on these two privately held companies which are effectively you know providing all those revenues and you're right Jason yeah I mean reminder everyone Google has Gemini which in theory is a direct competitor of OpenAI and Anthropic and this is Google giving, a separate part of Google giving Anthropic the compute they need to grow and I think you're right in one sense you're right win both ways I'm sure at the margin you'd prefer at Google to be the winner of the model company because I think in, and I could be wrong, I think in the end the value will accrete mainly to the model providers and everyone down the stack that's selling to them even though they're all making out like bandits today starting with the memory guys all the way up to the hyperscalers over time that's not the obviously differentiated place and I could be wrong on that. Maybe capex and the ability to invest hundreds of billions of dollars in a data center is in fact the moat itself. But over the long term, if I'm Google, I'm like, I'm happy that I'm doing 200 billion dollars of revenue with Anthropic, no more than Microsoft is happy they're doing 200 billion dollars of revenue with OpenAI. But deep in your heart, you should be saying to yourself, 'God, I really wish Gemini was so busy that they needed 200 billion dollars of compute.' And in Microsoft's case, I really wish I even had a model that was worth a damn, which I don't. Right? Because all you're doing is you're enabling with your balance sheet the two most exciting next generation tech companies who, you know, are going to draft on your balance sheet air cover and become huge.

It's for sure. I mean, there's a trade-off. You're enabling your competitor, right? It's just, it was interesting. The Wall Street Journal today published the market shares in the enterprise for OpenAI, Claude, Gemini, and Grok. Grok is a rounding error going to the prior conversation, right? It's not making any progress. But everything is so multimodal that it said Gemini went from 27 to 40% market share, I think, in the last 9 months and Claude went from 21 to 48. Okay. And obviously, OpenAI actually only went down a little bit. They don't sum to 100 because you're multimodal, right? But if Gemini's gone from 27 to 40 and Claude as we know has gone from 21 to 48. Now Google's got both pieces. It's got its own winner going to 40 and 50 and 6% of the enterprise and it's got a large share of this other leader, 48%. There's worse ways to solve revenue growth than having both the fastest growing players yourself and your competitor. You get a piece of each.

Well, I wish I had the fluency in numbers that Rory has, but 24x sounds, I know I highlighted this, it sounds just way too low.

The theme of a lot of the rest of the year is parallel agents. Now, we don't need parallel agents in everything. We don't need 100 SDRs hitting up our 1,000 potential customers every minute. Like, you run out. There are definitely plenty of workflows that do not need 10, 20, 100 parallel agents, but you don't need a thousand flights going to the Bahamas for your vacation, but workflows that can benefit from parallel agents, it's just kicking off inside of these LLMs, their models. And so, not only does 24x sound low, what if we have 10 agents? That's 250x, right? And more. And then we also forget how underpenetrated the enterprises are. It's so early outside of tech, right? So, I don't see why it's not 250x, but I got to put it on a better spreadsheet. But I think we're underestimating the potential impact of parallel agents. It's still, most of us live in a sequential world. We fire up something, Claude, Cloud Code, ChatGPT, it doesn't matter. And we kind of have a human interface where we're doing things sequentially because that's how our brains work. But it's also how the LLMs have worked. But now that they can natively run parallel agents, we just get these superpowers we didn't have a couple months ago. Like, and what these parallel agents are doing, if you haven't seen even in action, is they'll be even better because like for things like coding, they'll go out and do 10 different versions of the same feature or of the same iteration. And then they'll just, then we'll decide which is the best of the 10. It can present you the two or three best options and you can approve it. So why build a feature once if you can build it 10 times, have the LLM decide which of the two to three expressions of the feature is the best and then you pick the best combination of the expression, right? And so you can see hints of it like in image generation, you know, you use some image generation, sometimes you get four images, right? But what if you were in real time blending the best of all of these with much more complicated workflows? So 24x just sounds, it sounds conservative. I think that's why Anthropic's right to buy every TPU, GPU, every Cerebras chip, 12-inch chip they can buy.

Yeah. And I think the token count is always a mis, because I mean I've tried to do these numbers, Jason, and they're just so hard because, and you, I saw the summary of the Goldman report. I haven't read the detail and I really want to because I was trying to think through the same stuff myself, but like what you, what do you know? Every 18 months, you know, the raw performance of the chip gets, you know, roughly 3x faster. Then on top of that, other optimizations, how they run LLMs, how they do all the quantization, all the other clever stuff that you know, you can read about and try and understand gets you another 3x. So you're probably roughly 10x-ing the number of tokens per unit of money every couple of years, right? So you have that, right? So yeah, so if you were using more tokens, right, ironically, the total revenue would be going way down because, you know, it used to, you know, if you only needed a million tokens and it cost you, you know, five, you know, I mean, you see it in the prices, they're down 10x. If it cost you five bucks now, for the same number of tokens at the same efficiency, it would be way, way, way cheaper, right? But then what you have on the other side of things just getting better, right? So you use more of them to get to a better result. But what you see as Jason said is the, to what I looked at the token count, to maybe the way to say is this, the token count to support a chat two or three years ago when you're interacting with, you know, an LLM is whatever it is, is X, it's 10x that now to do some kind of simple co-work analysis and it's 10x that again if you're doing coding and it's 10x that again if you're doing kind of parallel work and kind of where these agents are going. So you're just going to see token utilization, the cost per token is going to go way down and the tokenization, that the value, the use of those tokens is going to way up and then the question is, you know, the interesting thing is you're dealing with two numbers, one on each side that's moving in order of magnitude every 18 months and trying to forecast where the net multiplicative effect of that comes out is hard. I mean, you know, that's what you're trying to do and I wouldn't fool myself into saying that you can be wildly accurate on that. You can get the rough direction, but it's hard to say, you know, I know exactly how this is coming out.

There's a little bit of a growing counterargument that's growing here when you talk to some of the best CTOs and engineering leaders that we don't need as many tokens as we think. And if you talk to folks that didn't just get their team the last quarter or so to get going, right, on Claude 4.7, but have been deep in this for a while, right, that are multiple releases into their whole team being AI-pilled. You know there is a theme that like we actually don't need this much code. We don't need this many lines of code. It is too much. We cannot process all of it. And this is just a lot of token maxing that is just unnecessary to deliver what the end customer needs. And we're all learning. We're all excited about these tools. They work and people are running, they're running Claude Code and Codex 8, 10 hours a day. But it's not necessary guys. You're producing too much code that's never going to be committed to production. And you're wasting your energy. And there is this, there's the token, there's like the whole from the press, the whole token max at Amazon where people are pretending to work because they have quotas. But this is different. This is some of the smartest people saying the folks that need this many cloud coding, these are the mediocre webheads, web developers that don't know what they're doing. Okay, these are like a little bit better than Lmin.

Yeah. But the question is can the average Atlassian engineer really effectively consume $10,000, $20,000 a month of tokens and be productive? Okay. And there is a growing micro backlash that the best developers and engineers now don't need $20,000 a month necessarily and that this is going to fade and we're not going to go back to handcrafting code, right? But that this is, we don't need to be running Claude Code 10 hours a day. This is a bad way to ship enterprise-grade software. This is a great way to ship hacks and proof of concepts and impress my boss. But is this really, do we really need this many lines of code a day? Do we really need it?

Yeah, there's a lot to unpack. I'm a little tentative because it's more my info colleagues doing this, but I've been talking to him about wrestling with this question. So, I just want to pick this apart, Jason, because there's a lot to do. One is this whole idea of, you know, monitoring people's token consumption and what impact that has. There's an economic concept. There's actually an LSE professor, Goodhart's law, which basically says whenever you monitor a variable you actually change the causal relationship of that variable. So once, which is a way of saying if you say to people I'm going to monitor your token production and how much you use you will in fact distort the result. That's what Jason's hinting at, and I think we saw it for either Amazon or Meta or one of those where employees are internally just burning tokens on stupid tasks, which is easy to do, just to make sure they quote unquote make their quota. To assess the market size for these companies at a highest level on Anthropic, you do have to have some mental model of how much LLM spend, for which token is a tricky but meaningful proxy, it is going to be relative to salary. And that's one of those macro numbers I'm kind of trying to figure out all the time. Does the average developer spend 2% of their salary dollars on tokens? Is it 5%? Is it 10%? It's a huge number and it's a very important number. But the odd thing is the more these big companies target their employees based on it, the more likely the employees are to distort the outcome. So that's the first big picture comment is that and I agree. I think there's a ton of that going on. And to some extent when people find that out and I think there will, I think there is going to be a push to get more grip on cost because there has to be because if you think back to December when Anthropic was at $9 billion run rate by definition very few of those CIOs had in their budget oh by the way you're going to spend 10x that next year right and it's a big sum of money someone's going to have to find 50 or 60 billion of budget across you know US corporates and that's real money and what that means is it's going to start being some kind of pressure on where is this money being spent even if they still want to kind of AI max conceptually so I agree you Jason so on that I agreed right the thing that I don't know enough about and I'm just going to ask you because you said it is there's this backlash do the quote best engineers not need as much tokens I hear you but I have two questions on that and they're questions that I ask my info guys where I don't have the answer the first is I've heard them articulate a perspective that actually says it's the opposite the very best engineers can in fact manage you because they have the conceptual vision of what they're trying to build, they can actually, they can be more productive with these tools and the less productive you are, the more you're getting yield loss in the sense of you're using tokens that aren't turning into effective code.

Yeah, but I think both are right. I think they really are becoming 100x engineers, right? Let them use whatever they want, right? Give them all the tools in the world. The concern is that you know there's kind of a snarky term web devs. There's others, folks one or two steps above me or the mediocre folks on your team that just aren't that good. Okay, they're web devs. They're consuming massive amount of tokens for relatively low amount of productivity gains, right? And it's not all performative like the Amazon thing. Some of it is attempting to keep up, but they need so many tokens to contribute so little value, right? It's the 100,000 line of code where some folks are made up on Twitter. Do you really need 100,000 lines of code to run a blog? Well, maybe you don't, right? And maybe you don't.

Which is why Jason, you're right. Which is why the question I'm always asking the rest of my team and I don't know the answer to is what is the objective? How do you think about measuring this? I mean, you're right. It's not lines of code is a dumb measure because this stuff grinds out lines of code. You know, you need some kind of conceptual effect of lines of code. And I haven't found anyone. We in fact we just surveyed I think 30 of our VPs at Anthropic to try and understand you know what's going on in terms of spend and you know they all spending a lot they all think they're going to spend more but you're right I didn't get clarity on the what is the heuristic for success and there has to be one over time.

Right, well but it's just hard to know what the ratio of like web devs and token trashers is to 10x or 100x engineers, right? And that this, it just goes to the, like I'm not smart enough to answer the question, goes to the question we asked before. Are there really enough developers in all of the solar system to keep Anthropic on the unprecedented growth path we have this year? Probably. But a counterargument is these web devs and trash tokens and we like, we're going to, this is going to play itself out. We're not going to, in a year from now we're not going to be wasting tokens on mediocre web developers playing with stuff and we're going to clamp down on it because it's a huge waste. We're going to give the S-tier guys all they want, the 100x, but that's always been true.

It might be. And look, I don't want to go too far. I just, we're still trying to see. We talked on the show about Claude design and whether it was a killer, right? It isn't a killer yet, right? I just don't, it's hard to predict some of the stuff. Listen, there have been many startups and other startups that have been destroyed by a Claude feature, right? I've invested in one or two. Being able to innovate faster than Anthropic is tough, okay? It's pretty effing. This is not classic slow company, slow pace, right? Having said that, who, I mean, boy, if I'm a lawyer billing $2,000 an hour and doing this, I don't want to take a risk. Like, it's not that I'm not going to ask Anthropic and ChatGPT my questions. I'm going to ask questions too in addition to Harvey, Legora or other tools, right? But man, I don't know. I don't want to take any risk that this is not have the level of domain investment in anything that is close to regulated or has other, I mean we, you know, you can't, it's people already got the memo the Wall Street Journal, you can't submit briefs to a court with hallucinations in it. Everyone gets that's a problem. Now,

the question of are these guys going to take away everyone's vertical business? And I don't know. I mean, I think it's what Jason said is that I think the model companies have a lot to do building their models, building broad horizontal harnesses, building products like co-work, right? It's not clear to me that they'll be able to have the time and the focus to do all these specific verticals. Now, should they?

I don't think there's any chance that Anthropic is going to do applications in CX and actually, and I could be wrong. I will bet you a lot of money they're not going to go all the way in legal at the app like that. That's why design was interesting. It was an application. Okay. Now, what does Harvey cost on average? $150,000 a year per law firm. Okay? Like, I don't want to roll that into my $200 a month I'm paying for Claude if there's any risk. It's just not worth it, right? It is just not worth it, right? I need a solution. I needed to do everything. I needed to integrate with DocuSign. I needed to prepare the brief properly. I needed to review it a different way. It's just not worth, this is not a career where saving a few pennies is worth it outside of the low end of the market. You, the low end maybe the ambulance chasers do use Claude, right, that's fine but I don't and I, so I may be wrong next week but I'm not confident that these verticals are going to put the resources to build an application. That's the thing, they're not going to build a CX application, they're not going to rebuild Decagon or Sierra or Fin or Gorgias or any of these other, it's not that they won't build chunks of it, it's not that they won't take away, it's not that like OpenAI hasn't taken away pieces of ElevenLabs or other pieces of other folks, but the app, but we haven't seen them commit to building applications. That's different. If the LLM can express what an application does, that's where the YC companies get killed because all of a sudden you don't need an application. It just works in the prompt, right? That will kill a thousand. It already has killed one or two of mine that are on like, one is on its third version because it was super innovative a year ago and then just, it is built into Claude today.

Step back here. This is not the first time you've faced this question. If you look at every single platform, there is a dominant compute level provider. And the million-dollar question is how much of the app layer they take over, right? I mean, let's just do two obvious ones. Microsoft in the '90s, they dominated. Yeah, they were the operating system. So, they dominated broad horizontal application software for the consumer and the individual knowledge worker. The office suite, I think co-work could be like the office suite, right? And they dominated networking at the kind of infrastructure layer. No, you know, right, just connecting things. But there was hundreds of application software companies on top of that. Siebel, Vantive, Scopus, Clarify, Baan, SAP that said, we run on top of Microsoft. We build specifically for this vertical or this horizontal use case. And Microsoft tried. They bought Great Plains. They never really made it happen. Fast forward a decade later, hang on, Amazon, right? Same kind of question. Does AWS eat everything? Right. And I know lots of brilliant investors who passed on Snowflake, which is even not even an app. It was an infrastructure level player because they said, 'Oh my god, Microsoft Redshift, Amazon Redshift is just going to eat the lunch. It's not going to be a thing.' And it turns out it was a $50 billion thing. Right? So my point is you have to have some approach to thinking about this, but you have to face this question every compute revolution. And this is just the latest turn of the crank, right? And to me kind of the default is the Microsoft outcome which is broad horizontal, the equivalent of broad horizontal compute is now broad horizontal intelligence and that's going to be provided maybe not by the monopoly like Microsoft but the oligopoly of Anthropic, OpenAI and maybe Gemini, probably the kind of knowledge tool Microsoft will regret to their dying day why they let, you know, slipped on this and let, you know, Claude co-work be there take their lunch but that's their problem not mine. Right. But at the app layer, I'm kind of with Jason. I think all these individual apps, especially, and this is a key point, especially to take legal, if all you're going to do is mark up a document for an individual user, I think there is an argument that an individual user might get a skill from Claude and they'll be fine, right? But most of these companies are selling coordinated workflows across an enterprise. And I think once you get to that point, I'm with Jason. I think the, you know, Legora, the Harvey rep, if you sell into AmLaw, GCI, which is ours, if you sell into corporates, you're going to want both the relationship. You're going to want the ability to customize it to what you want. There's just going to be a whole bunch of work that's just better done by a focused firm. I mean, I think history's on our side. That's the way it shapes out.

Well, I would just offer two thoughts, maybe. One, this is more OpenAI than Anthropic. You know, Sam did hire Fidji to be the CEO of applications and there no longer is a CEO of applications. Now she's CEO of AGI. So that is really walking back. Now you could say it's a little bit different but just think about hiring a CEO of application and say, you know, that's not a business we want to be in. Claude not seeming, Anthropic not seeming to invest 100, 500 people in design, right, which they could and so that could change next week but right now neither of them seem to see applications. Now paradigms do shift. They're also going to Rory's point of the low-end redlining or whatever that could get more powerful. I'll give you a different example. It's not the same. But I don't think in a year we're going to need any traditional marketing automation software. It's too dated. It doesn't work for agents marketing. There are categories of software where if they don't have a reason to exist in an agentic world, they will go into a terminal state of decay. Okay. Agents do not need HubSpot or Marketo or Salesforce or any of these marketing automation tools because they have no need to hand compose an email in a third party template. And so I'm not saying that would happen in legal but you could see if they don't keep up ahead of it and the paradigm changes it, you can become obsolete over time.

There's two separate dynamics. One is the does new software get eaten by new models. In other words, Jav, Legora eaten by Anthropic. You're referencing something different, which is does old software get eaten by agents either Claude agents or new agents,

but new software can get old, too.

Yeah. Well, yeah. So, you're conflating the two, but yes, agreed. It can, but I just think they're different questions, right? And I'm just trying to,

I just think the rate of decay might accelerate in the agentic era. That's what I was, that's the connection I didn't make. It used to take a decade. Now, it could be 18 months. And if you talk to folks, senior folks at Lovable and Replit, you will hear they are well aware of this. They do not want to be clotted out of existence. They think about it every day that they have to stay ahead of it because they're closer to the core, right? And so I just think stuff gets old much faster than it used to. The evolutionary pressure from a model that's underneath you, an intelligent model that's underneath you, a company building that intelligent model through, not Anthropic, for it is, you're right Jason, more powerful than the pressure of a, you know, an operating system or a, you know, compute system like AWS. So I do agree with you every day, every day you wake up as Lovable, a Replit, where the model provider underneath you is doing more of your, you can't, if you could, if you could miss, if you could be behind by a year competing with Redshift, you can't be behind by a week competing with these guys.

Maybe it's not as exciting as the Anthropic wars. I just thought that the slight contrast that was interesting was Monday versus HubSpot. Okay, so they're both decelerating. They come out with a quarter. And Monday trades up after being maybe the most beaten down stock out there, right? Trades up I think 20% and HubSpot down 18%. Right? And what's the difference? They're actually not much further than each other on their agentic journey which is early, which is, you know, Monday's in production and HubSpot sort of in but I mean they're pretty early. All Monday did that I can tell was for the next quarter they're still decelerating but at least they raised the guidance for real, right? HubSpot couldn't do it. HubSpot lowered their guidance. So if you're not accelerating, you're going to be destroyed, right? And at a minimum, you've got to raise guidance, right? Even if you're not accelerating like Monday, you've got to at least raise your guidance to get some breath because at least it says, I think it says since when Monday says, 'Hey, we're raising our guidance next quarter.' At least it says we're not going to zero, right? We're not being destroyed by AI if you're raising your guidance. But I think even though the SaaS apocalypse is behind us, I think for many there is still a worry the terminal value is zero. I don't think that fear has gone away if you keep decelerating while budget is accelerating everywhere around you. It's not that everyone's decelerating. The budget is accelerating and you are decelerating. It's not, those aren't two good lines to cross over. They're not crossing over each other. So, but at least it's nice we see some bounces off the hard deck.

And I agree. I think it was a super interesting quarter. Lots of different people reported did different things. Some of them cut expenses. Some of them had decent quarters and the stocks, you know, we had all sorts of weird movements. I mean, Cloudflare, you know, had a good, really good quarter. I think mid-30s growth. I meant, yes, laid off a bunch of its base, stock went down, you know. You're right. HubSpot, decent quarter for the record here. Yeah. Decent growth, stock went down. You're right. Monday, earlier on, a little bit better. Bill.com, one where I used to be on the board of, you know, unfortunately also had to lay off a bunch of people. Stock bounced up and partly because of a buyback. So, you know, you kind of look at all this and go, what's going on? And I think there's really two things and Jason's done a really good job of articulating the important one, which is where are you, where's your business structurally? Are

Right. I think one of the things that happened, interestingly enough, to definitely Cloudflare and possibly to HubSpot, is a little bit of the 'oh my god, you're cutting 20% of your costs,' which had been perceived as a net positive. Now there's a little bit of a 'huh, Cloudflare, you're a good company, WTF, what's going on?' Right, a little bit of uncertainty, right? You know, what's going on with that. But if the first big bucket is that whole kind of 'what's going on in your business,' then separately, it's going to sound really Captain Obvious, as Jason would say, the second half of this is you just got to look at price, right? In other words, if you have a muddled quarter and your stock's already at the bottom, you might get a bounce, right? If you have a tricky story and your stock is like you, Cloudflare, or AppLovin. I think AppLovin was, and the two of those were among the two highest valued companies, right? You know, if you're going 30, 40%, 30% in the case of Cloudflare, but you have a story that's even a little bit messy, people are like, you know, I might be okay with this at four times. I ain't okay with this at 15 times. So sometimes you forget that there's both the strategic journey and then there's the how is price dealing with that? And I think what you saw in the case of AppLovin and Cloudflare is it turns out high-ric stocks at the start of a paradigm shift, no matter how amazing the quarter is, are just vulnerable to disruption. Whereas when you're trading like Monday, poor Monday was at like under two times revenues with a bunch of cash, you're like pretty much anything you do other than burn the company down, the stock goes up. Excuse my language. It was basically priced at nothing because at one point it was based at like almost one and a half times cash.

Yeah. And it's basically just saying, you know, just any positive momentum you will get rewarded. And it's like, as I say, it is very much Captain Obvious, but price is the vector, and we forget about this often in the private side because you're not really dealing with price on a day-to-day basis, but on the public side, that's how it works.

I think the most brutal one, and I love Henry, but the most brutal one that maybe was under-discussed was ZoomInfo, right? And the reason I bring it up is ZoomInfo also tanked when, you know, they're growing 1% now, and I think they guided to some negative revenue growth going forward. The reason it's just a tough one is ZoomInfo just won the pre-AI game of sales intelligence data on your customers, even just basic stuff, emails, phone numbers. It just won that stuff. And then agents, like really Clay and others, you know, they're there in some ways, they're pre-45, right? They're not epic products in some ways, right? We use Clay, we love it, but I wouldn't say it's epic. I would just say it's LLM-infused, it's AI-infused. But that's just a one-to-one loss of dollars. Okay, whatever. What's Clay doing? 200 million, 300 million, right? Something like that. I mean, it's hard to say. So ZoomInfo is doing a billion something, right? That's AI taking all of ZoomInfo's growth away from it. All of it. And so I only bring it up as a tough case study of how even if you're treading ground, you can turn around and these AI, they're not all Harvey and Loras, just these kids can just take all your growth from you. I think ZoomInfo's growth was stolen from it from Clay and Friends, and it's a brutal case study. It's a brutal case study.

Agreed. And for the record, it wasn't, I don't think, I mean, I think the Clay story is amazing, and it's much less to me a kind of AI-first story. I mean, it's even simpler than that. They built a waterfall product that allowed them to optimize among multiple different data providers and allowed RevOps to, you know, kind of waterfall a bunch of different data providers and pick the best. And what that means is instead of being ZoomInfo being the only game in town, you can compare and contrast five or six data providers. Data then becomes a commodity. And you're right, your business is commodified. Then on top of that, they've done, I agree, a much better job than any because they're really to some extent a pre-LLM company. Much better job of building Clay agents and having an AI story. So the funny thing is it's a pre-AI company at its core initial value proposition that's morphed brilliantly. And you're right, it's sucked all the value out of the data providers. So yeah, those guys need to figure out how to become relevant in an agent-first world pretty damn quick.

Nothing stays growing at 1% a year in the public markets for a long period of time.

You know, bought by PE. I almost, I resist that thing because it's such a lazy man's approach because the implicit assumption is, you know, no matter what happens, you can get bought by PE. And yeah, that might be true going forward. Though I, again, contradicting myself, price, a friend of mine used to say, a price clears all markets. There's a price at which PE will buy something like that and say, we'll do the hard things for two years to fix it. Just might be a particularly compelling price. What's it? I don't know what it's trading at. One times revenue with 35% adjusted operating income. It's trading at one times revenue.

Yes. Yes. One times revenue with 35% adjusted. That is a classic take private if you can put the, I mean, you'd have a hard time finding someone, I think, better than Henry. But that's the counterargument, right? But if you, on paper, that's the classic take private, right? Trading at one times revenue, 34% adjusted operating income, right? Reasonably stable, not adding customers, but reasonably, you know, it's net neutral on customer growth.

I even, we just said it two minutes ago. Vis-a-vis Monday, get your, I mean, it's to your point, get your act together for two quarters, get a small amount of kind of AI-enabled growth, and you won't ever go back to 30 times or whatever absurd number you were trading at in '21, but it doesn't take a lot to get you to two times, and that doubles your stock, to state the banal. You know, they're not, I mean, the big picture comment is this: the market is not asking these old-school $500 million to $1 billion revenue SaaS companies, you know, to become the next Anthropic and double, treble, and 10x every quarter. Just said, give me 30% growth. Give me profits. Give me a story that's got some future in it, and I'll get you back to five times. I'll never give you 20 times again. I won't fall for that one this time. But I'll give you five or six times run rate, right? Provided you have the growth, provided you have the rule of the profits, that's what you can do.

Look, it's going to go great. The fact that they've raised the range, which you only do when you're highly confident, especially that much, that you've got a killer IPO on your hands. There's no more new information. It's going to go out and it's going to trade amazingly right now. Entirely separate question: how's it going to do two years from now? That's a business question. But the technicals, look, I'm not a banker, but...

You know, let's put it this way. I think what the bankers would say in the boardroom when the pricing committee starts giving them, are we leaving money on the table? What they'll say is something like this: we believe at this price we'll have that nice 20% pop. Everyone will be happy. I know we should hate the pop. Thank you, Bill. But a perfect, they're probably trying to get that perfect IPO. Can it run away from them totally? Can you have a Figma phenomenon where retail piles in? Entirely possible. Nothing excites the mind like some of this AI stuff. It fires the public imagination. There's a dearth of opportunities to play. It's entirely possible that you have a whole host of kind of retail demand that you can't forecast, and in the short term it runs away from you. But reminder, when Figma popped to 100 when it priced at 35, I think I said I think it's, you know, worth 35, 40, and now it's actually significantly below that. So you can't control the weirdness of retail, and as we've discussed, it's impossible to try and, frankly, it's also impossible to let it drive the narrative, right? I think I really feel for Figma in that their narrative is this: oh, you had 100, now you're down 80%. No, you're not. You were priced at 35 and now you're at 20. It still sucks, but it's not 80%, right? Same thing here. Who the hell knows, will it do the normal pop or will it do something crazy? But fundamentally, it's going to go out, price well, trade to the upside, because otherwise these bankers would have been manifestly incompetent to do that raise. And they're not manifestly incompetent. They're smart dudes.

Well, look, of course I agree with Rory. When they raise this much, there's no question it is so people are going to buy, they're going to buy into the first day, so you're going to get, you should get a pop. Maybe there's examples where the range is raised this high and all that where it doesn't happen historically, but in my limited experience you're always, I mean, it's just, it's almost built into the system, right? But I mean, we just have to see. You listen, it's a fun one to watch because on the one hand, demand for inference is infinite, right? It's great. On the one hand, they've got support from OpenAI and Amazon and everyone, which they didn't have when they tried to IPO last time. Right timing, right partners, right seeming backlog, right commitments, but it's competitive, right? And everyone's going to buy every solution. And if Nvidia Groq is better and they can get the chips, they'll use it. So it's hard to predict when there's this much explosion of, and when there's also hedging happening. There's also hedging. There's hedging for capacity. There's hedging for performance, right? And so I just think this one is, it's impossible to predict, but it's a great derivative. It's great to IPO before Anthropic and OpenAI IPO too. It's a great time because I get into that zeitgeist. It's more interesting than CoreWeave, which is a data center, right? This is real technology that is fueling inference, but it's so damn early to really know how it's going to go. Like this is not two years of massive ASIC chips used in production in data centers proving a massive competitive advantage. It's early. So how the hell do we know where it's going to be in two years? How the hell do we know? I might take my profits. I just might take my profits.

Hard to argue with that. And you know, because you know, when they went and go out, and even now, the historical revenue is very much concentrated on a couple of customers from the UAE, United Arab Emirates, right? What they have, what they're leaning into here going forward is the contract from OpenAI and a less kind of fully fleshed out contract from Amazon. So you're right, Jason, you are leaning into a future that's not like the past. So to that extent, you got a lot of risk going on here. At the same time, you know, their story, and you read the CEO letter, founder's letter, it's great. I mean, what they're selling is speed, right? What they're selling is their inference can be faster than anyone else. And I love the tagline in the thing, you know, like, how much would you have to be paid to have a slower internet? And you just don't, once you see speed, you don't like to go back. And you know, one of, actually one of my companies, Tavis, was mentioned deep in the IPO because we use them for our inference because we have these, you know, kind of AI humans and you need real-time responsiveness. And I think there is a focused market for that kind of real-time blazingly fast inference that, you know, they can maybe have that market over the medium. And that is the medium-term bet if they are the first really reasonable competitive alternative to Nvidia solutions. And Nvidia is a $5.5 trillion company. Just roll with me.

Being priced at 48, if you take a 5 to 10 year potential future.