20VC: Meta's Muse Hits No. 1. ChatGPT Finally Has a Rival | Menlo Sounds the AI Bubble Alarm | Factory Triples Its Valuation to $5 Billion | Keith Rabois vs Airwallex: Who is Right? | Crusoe's $3.9 Billion Round. Is the Data Centre Trade Overheating?

Boys, another week, and I wanted to start with the news that Anthropic pushed its $2 trillion IPO from October to November. Some people are suggesting that it's the first sign of a crack in the market. It's following the pacing of the frontier that we discussed last week. To what extent do you agree with that? Or to what extent do you think it really is just them wanting to have a great Q3 prospectus and needing more time for those numbers to show through?

1. Anthropic Delays Its IPO

I think that's the answer. I think the whole crack-in-the-market thing, let's leave that to one side for now. We'll talk about it later. Ditto the pacing stuff, which clearly has also been shown to be not really believed, even by the people who said it.

It really boils down to what you said. They had an amazing Q2. They finally became bigger than OpenAI. And remember, we talked about this: OpenAI then responded furiously in July, started tweeting all their "My Q3 numbers are amazing" stories, and really pushed back. Some of the data supports that.

So my guess is the banker said, "Hey, if you want a really clean story, it would be great to incorporate a quarter that reflects this noise. Instead of going out in October, where you won't be able to share your October numbers, it's kind of that weird thing. Go out in November, where it's clean, you print your October numbers, you drop them in, and you go."

From a programming and timing perspective, it just all made sense. It's a cleaner deal. Whether it turns out to be a wise decision or not, we can come back to in a second, but it's always weird to go out after you have your numbers, but before you can share the numbers—after the end of your quarter, but before you've finalized and audited them—especially for what looks like quite a pivotal quarter.

So it totally made sense. If I'm a banker, I'm thinking, "If we do this in October, it's going to be a lot of explaining. If we do this in November, the numbers will talk."

Having said that, that sounds all right to me, right? But isn't there maybe just a hint of worry that getting 30× or 20× oversubscribed, or whatever they want at the most massive IPO of our lifetimes, maybe there's just a hint of stress in the pre-conversations?

I agree, and that's why I made the comment earlier about whether it's a good decision or not. If the world goes to hell in a handbasket in November, you'll look back and go, "Damn, we should have gone when we had the chance."

You'd have a CFO who said, "We're just doing it. I don't care about your messy story. We're going to tell a good story. If the price ends up at $1.5 trillion instead of $2 trillion because it's a messy story, I don't care. I'm willing to live with that." And that's the approach you take if you feel it's hell or high water: I have to get this money.

Clearly, the fact that they didn't take that approach means they decided it's not hell or high water. They're confident. They're willing to take another month of timing risk, probably for some significant valuation pop, which is what you do when you think you have lots of time and you're in a good, commanding position.

I mean, if I was on that board, it's 90% certain that this is a good decision, and a clean story in Q4 is better than a messy story in October. There's always that 10% chance the world goes crazy and you look back and go, "Damn, should have taken the $100 billion."

2. AI Liability Cannot Stop The IPO

I was walking with a friend of mine the other day, and he said, "Harry,"—a very successful, multibillionaire investor—he said, "Harry, how the hell do these frontier model providers go public when no one is willing to provide liability insurance? You have swarms of rogue agents doing whatever you want. How on earth do they go public? Who's liable?"

Bullshit comment, and I'll tell you why. Once you've said publicly that there's a 10% risk that your thing can blow up the world, sweating product liability is in the noise, right? This is a $2 trillion market-cap company. It can self-insure, right? They don't need reinsurance from Munich Re, with a market cap of $200 billion, to reinsure their $2 trillion market cap. I don't think you need product liability insurance to get an S-1 done.

Now, separate comment: the fact that no one will insure them is a data point about the dangers of the product, but that's long since been internalized. To be very fair to the Anthropic management, anyone who buys that stock and doesn't know that senior management think it's the most dangerous thing since the atomic bomb has, rightly or wrongly, not been reading their tweets.

In the list of risks disclosed—and I'm really looking forward to reading the S-1 risks, by the way—product dangers have been pretty thoroughly discussed for a decade here. So I don't buy that at all. There are 100 reasons why you can worry about valuation and traction and all that, but product liability insurance is in the noise.

Yeah, they're just going to have a big litigation legal team. They're going to fight this stuff forever, just like tech leaders have always had to fight IP trolls with large teams. They're going to fight liability suits.

They’re going to have 200 folks in-house, and all the top law firms fighting this, dragging it out, saying they’re not responsible, trying to get legislation passed. But you don’t stop the IPO. I can’t remember an IPO where this was as explicitly dangerous, to Rory’s point and Harry’s point. I mean, it is novel, but it’s game on, man.

Agree. The most important part of securities laws is not that you sell stock that has no risk. It’s that you sell stock where you disclose the risk. Provided somewhere in the S-1 they say, “At least half of our crazy employee base thinks this thing is going to blow up the world. I don’t, for what it’s worth,” says Dario, “but my people think that. Just letting you people know they’re out there in S-1 land.” Provided he discloses that, he’s covered.

And obviously, stating the obvious, he’s not covered if, in fact, it does end humanity; he’ll die, too. But let’s just ignore that for now. By the way, I want to call out Jensen for his big-ass call. I love it: his probability of doom is 0. Finally, an unequivocal statement from a no-bullshit investor and tech leader. There is a 0 probability AI will destroy all of humanity. I love it.

That’s because his LLMs haven’t caught up yet. His open-source LLMs haven’t caught up yet. That’s why.

That’s why he can say not until 2030, because he knows he needs until then to catch up.

3. OpenAI Faces A Capital Problem

Before we move to Meta and Manus, OpenAI is going to burn $278 billion by 2030 and be out of cash by 2028. Is this just more noise and, of course, to be expected? There are rumors of another round at a $1.5 trillion valuation.

I bet it’s more. It doesn’t have a history of the burn coming in less than planned. They may need $400 billion. We’ve all had a portfolio company like that: top line great, love the team, but burn, no matter what they say, always comes in 30% to 50% higher than the model. It doesn’t matter who we put in as CFO.

Jason is broadly right. Look, big picture on the forecast, there are actually 3 numbers that matter, not 2. They’re forecasting growing from $35 billion in ARR at the end of this year to $350 billion, I think, in 3 or 4 years, so 10X growth. Worth pointing out, by the way, that last year Anthropic grew 10X in 1 year. This is a 10X forecast over 3 or 4 years, so almost modest.

The second thing is the burn. You’re right, they’re forecasting a net burn of $278 billion, and that’s money out the door. They have $122 billion of cash on hand, so they have time to raise the extra capital. The amazing number is the other one, which is the CapEx required to do all this. Not all of it is on their balance sheet; a lot of it is on other people’s balance sheets. It’s around $700 billion.

So it’s just a reminder that this is, unlike software, an extraordinarily capital-intensive business. To bring it back to what you said, Harry, if all that burn—all $700 billion of it—had to appear on the balance sheet, it would be even worse than that. The only reason they’re able to do it, only burning $278 billion, is because other companies like Oracle and NVIDIA, with revenue support and backstop insurance, are able to say, “We’ll do the CapEx and lease it to you.” So this is an extraordinarily capital-intensive company. It is going to consume, directly or indirectly, $700 billion worth of CapEx to get there, to get to $350 billion in revenue. Intelligence is not cheap.

I think that was the most efficient coverage we’ve done of OpenAI and Anthropic.

Yeah, we can move on.

4. Muse Gives Meta An AI Reprieve

So we’re going to move to a story of the week, which is going to be a new feature of the show. Jason’s IC has been a massively popular segment, and we’re going to add a story of the week, which for me is Meta’s Muse taking the number-one spot for days, the stock ripping, Alex Wang and team absolutely crushing with this launch. Thoughts? Bad for OpenAI.

That’s a good point.

It’s the first real ChatGPT competitor. Anthropic never really cared about Claude until recently. Manus is many things. It’s one of the best pieces of software I’ve ever used, but it’s also a Trojan horse to fight ChatGPT because the LLM is pretty good.

Every time you’re doing an agent, you’re also asking questions. You’re not just telling it to get your movies. You’re saying, “Hey, what is Plane? How was that latest superhero movie? How was the latest 20VC?” It has opinions. It’s a darn good, normal consumer-grade LLM. I don’t think anyone’s writing wet-lab biotech software on Manus, but at a lay level, if it’s free, it has agents that are truly autonomous, which ChatGPT doesn’t, and it can do all the other questions you have. It writes you cute letters when it makes mistakes.

If it’s free, why would an ordinary person pay? And the tokens are vastly more than you get from ChatGPT. I just think, in addition to everything else, it’s a ChatGPT Trojan horse because it does everything ChatGPT can do, and it has autonomous agents. It doesn’t have to just be agents. It’s doing all of it, so it’s pretty cool.

Could not agree more. I think it’s excellent. It’s funny: as I said 2 weeks ago, we were saying, “They’ve got to ship this,” and literally between our conversation and release, they shipped it. We’ve been skeptical and a little harsh about some of the Meta/Facebook investments in AI with an ill-defined enterprise model. This is the exact opposite. This is spot-on; it plays into their distribution leverage.

It’s intuitively something that you would take. To the extent that you already trust Facebook with everything else, that’s a longer discussion. At this point, if you’re deep in Insta or Facebook, you’re in already; you don’t care. And it’s a good product. Jason, you’re exactly right. It’s a good UI. It works. It gives me recommendations, and they just have the distribution.

So I agree. I thought it was a wow moment. I thought it was a real win. I’m not sure about the economics and if it’s worth $100 billion, but it’s worth pointing out what it meant for Meta. They’re up 7% or 8%. They made $100 billion in market cap this week because of that product, right?

And that just shows, when you’re dealing with these huge end markets, the value of what’s possible. You can see the value of a play in that space. That’s the argument for why OpenAI might look at that and bring it back to venture, right? “Hmm, they made $100 billion. Should we buy Instinct? Do we have to do something in this space, like, now?”

And it’s up 34% this month. Some people say it’s up more because people who were in the beta test were buying, right?

That’s interesting. I saw that, yes, that the beta testers will buy, which I just love. For the longest time I was like, “Facebook—it’s a great business that’s spending a lot of money in a bottomless pit of enterprise AI.” Now it’s a great business that could have a next act in AI. That’s a big change.

No, that’s worth the re-rating, and give them huge credit. I think it was a big week for Meta, to state the obvious.

The reason I say it’s a Trojan horse is that it built me an entire CRM. This isn’t just reservations, and it’s very good. Now, it’s very limited 2 weeks in. It does everything. It built me an entire CRM for every SaaS or sponsor. It tracks every email about them, every item in real time, and it updates it.

So it is sort of a CRM of 1. It’s limited because it can’t really collaborate. You couldn’t use it for Salesforce, but imagine yourself: it is some of the first composable software that I’ve ever actually seen work. This has been a myth since this show started, right?

I really said, “I need a CRM for myself. I need you to track 150 sponsors daily in real time using AI,” and it just built it. For what it is, it’s very good, and it cost 0. You need all the pieces: you need a database, you need an LLM, you need intelligence, and you need access to email. You need all of this for that to work.

You’re obviously not the typical user, and I don’t think their economics are structured around everyone building their own CRM, or it’s going to eat compute. But yes, through that interface, you have a sophisticated LLM at the back end, and you have your own little sandbox. They have a lot more standalone compute and sandbox for you than I think Instinct does. So yeah, that makes sense.

5. Agents Challenge Amazon Marketplace

For those that don’t know or haven’t used it, a lot of people use it to buy things from different providers. Amazon has, in response, blocked it. Shopify has decided to partner with it. How do we think about those 2 decisions from Amazon and Shopify, and who do we ultimately think is right?

I think both could be right for them. From Amazon’s perspective, what they’ve discovered in other agentic-type commerce is that 2 things happen when you use this agentic commerce. One, you don’t get any revenue from your ad business, and Amazon’s ad business is now larger than its e-commerce profits. In other words, it is the entirety of the profit stream for e-commerce, as distinct from their AWS business.

The second thing, which I didn’t know, is that Walmart did something recently where the basket size gets reduced because you don’t get the chance to shop. When you go on Amazon directly, they’re like, “People also bought…” Then you have, “Oh yeah, I need to order that.” Whereas if I do this, I just order the thing.

So from Amazon’s perspective, because they’re such a big player, they’re like, “If I let these people do it, then I don’t get my ad revenue, I get a smaller basket.” And the next sentence is key: “If I block them, they’ll probably come to me anyway because I’m Amazon.”

So I have leverage. This is tech behemoths bumping into each other. Amazon is now saying, “Screw you, Muse, Meta. We’re going to have to talk about this before I roll over.” At some point, there probably will be a more aligned exchange of value.

Shopify, on the other hand, represents lots of long-tail merchants. They are very glad of the extra business. They probably don’t see that much compression in order value because if you go directly—if you’re buying some obscure thing from a mid-tier Shopify merchant—you go buy the thing directly or you buy it via Amazon. No matter what, Shopify’s broadly happy. They don’t have a big ad business.

So from Shopify’s perspective, it should be, “Have at it.” What Shopify likes is they have the common payment app, so they like to have that go through that. In each case, capitalism works. These executives are very logical. They’re like, “It’s great that you have this new source of demand. What matters to me, in Amazon’s case, is that I think I have a lot of leverage. In Shopify’s case, provided you let me charge for my payments rail, I’m good with this access.”

So it all made sense. It’s the early days. There will be much negotiation. But the bigger lesson here is that there was a whole bunch of Google and OpenAI blathering on 6 months ago about various different agentic payment mechanisms, right? It was all your smart people getting in a room and thinking things through. The truth is, in tech, none of that shit ever matters.

What really matters is someone aggregates consumer demand, like Instinct and Meta have done. They start pounding on the API, and then everyone has to focus. I’m willing to bet someone at Resy, even as we speak, and at OpenTable, is formulating their agent API policy. I’m willing to bet that at Amazon they’re talking to Meta this week. Demand creates urgency to sort all this shit out.

They’re all going to lose. I think it’s the last stand of the unnecessary system of record. Resy’s going to lose. Amazon is the least lossy in the short term because they’re at the edge of a monopolist for what they sell as a merchant, but the agents will bypass them ultimately.

You’re going to have to decide. The agents have very clear positions. And you’re right, this benefits Shopify. Of course Shopify leans into this, right? Of course Stripe and PayPal lean into this. It benefits them. There’s no question. Amazon losing ads, losing upsell, losing shopping cart size—all negative.

To the extent an agent can route around Amazon, to the extent you can, the agents not only will, they will gleefully do so. They will gleefully do so. “What would you like me to do? Would you like me to call the restaurant directly, Jason? Oh, I found a back-end API where OpenTable still works, even though they’re exclusive on Resy. Oh, I found another way to do this on DoorDash that’s left open.”

The agents are wonderful at finding broken APIs that have other surfaces that shouldn’t be exposed. This may be their only choice, but I believe all the systems of record, the places of record, are just battening down the hatches, and they’re fighting the agents. Net-net, it’s not a positive for any of them. This is not a positive for Amazon.

I hear you. The agent, as Jason points out, is tireless. The agent does have the ability to call and hit all 17 websites. It’s only compute. They don’t care. So you’re right: a lot of these digital-only aggregation businesses, I don’t think they’ll fall as brutally as you say, Jason, but there definitely will be some end-run pressure to go around demand systems like Resy or something like that. I’m not sure it’s as fatal as you think.

It’s not that I think they’re going to be killed. I think they’re going to be maimed. It doesn’t matter. If this means that Resy’s growth falls this much, or if this means it impacts 10% of Amazon’s advertising revenue, that’s a big deal. Just like Meta’s stock is up 20-some-odd percent, if this drives Amazon’s stock down 20% because it impacts margins, it’s a big deal.

I can tell you, talking to our agents all day, they don’t put up with this bullshit. They don’t friggin’ put up with it. Listen, I don’t want to name the vendor, please don’t make me, but we got a raise this morning from one of our core vendors—a massive price increase. Immediately, our agent said, “I want to work around it. Here are my ideas. Dump them.” It actually laid out a 12-month plan to migrate off this vendor on its own. They will not tolerate this crap.

There are so many products you can only buy on Amazon. It has so many advantages. It has a warehouse. It has fulfillment. But some of that stuff can be bought on a Shopify store. It could just be enough. It could just be enough. This maiming is a big deal for AI. It’s going to maim the existing folks unless they embrace things they don’t want to embrace.

Two comments, one each way. One is, yes, I do think Amazon’s biggest defense will be its physical infrastructure and ability to deliver. But your point is the right one. I remember someone said 25 years ago, when the internet first came out, that the internet abhors inefficiency. In other words, middlemen get pounded down, right?

If you think about things like the travel sites—Expedia, Airbnb—all of that is getting demand and supply closer together. And I think what you’re saying, Jason, is correct. AI is the same. It’s going to pound down people who are in the middle, who just have information and are using that to make offers.

If that’s all you’re doing, then there will be pressure at the margin. Because if you’re reducing search costs, that’s valuable to me as a human, but the agent may be able to just do it itself. So I’ll go with the maiming comment. If it’s even a 5% or 10% impact, it’s meaningful.

That’s the thing, and it’s easy for Amazon to shut off Perplexity. It’s just a gnat. Amazon will have interesting choices when it’s everybody. You can shut off everybody—it is technically possible—but at some point it’s going to be complicated. Are we sure we want to shut off everybody?

Moving swiftly on, is there anything else that I haven’t touched on with Muse?

No, because it’s not really the story of the week. I mean, it’s the story of last week. We’re going to do the story of the week, and Jason’s going to be ready as a hands-on user to tell us about Jev, right? That’s the story of the week, so how are we going?

6. Jev Unbundles LLM Intelligence

It’s the next one up, and fine, it’s Rory’s story of the week. For those who don’t know, ChatGPT’s inventor ships Jev. It’s a model that decides instead of chats, and it became Vercel’s fastest-ever launch. Jason, I’m sure you’ve played with it. What did you think? Jason’s review: new segment.

Well, first of all—listen, I might not be smart enough to get into semantics. I don’t even think it’s a model. I think it’s a classifier backed by an LLM. I’ve used it. It’s awesome. But I don’t know that it’ll be as disruptive as everyone on X, who probably never used it, says it is. But it is awesome.

Jason, what do you use it for?

What I use it for is deciding who in the Sastre community should meet each other. It’s quite good at that. Should Rory, the CRO at GCI, meet Harry, the CRO at Perplexity? It’s actually not that simple a question. They’re at pretty different stages. They live in different places. Should they meet?

I spent all of Saturday working on Jev because I had to redo the way I did prompts. So I thought it was a failure, right? Then what I learned is that if you invest the time, it can answer that question in milliseconds for 1/100th the price of Anthropic or ChatGPT—actually, even less.

It can just answer a question, and it can answer a subset of questions. But I think, as we’ve learned from data labeling and others, there’s a lot of classification that needs to be done, a lot of data labeling that needs to be done. It doesn’t output text, and it can’t do anything particularly complicated or involving reasoning.

So it’s a big deal, and for this particular use case—I’m talking about who should meet whom—it is disruptive. But it’s maybe 20% of all the LLM calls associated with this application. So I love it. I think everyone will do some version of this.

I think maybe even Exa and Parallel should do some version. Maybe OpenAI and ChatGPT—everyone should do a version of this. But it’s not going to take over 100% of your stack, and it’s not going to replace ChatGPT because it just doesn’t do those things. But it does remind us that we waste a lot of tokens on simple stuff that Astra shouldn’t be doing.

I pretty much 100% agree with you. The zoom-out comment is this: it’s a new model. It’s different from an LLM. An LLM returns text and is very computationally intensive and, in computing terms, fairly slow and expensive.

This just returns true or false, or a ranking, or a score. Those are the 3 answer types, and it’s super fast and super cheap. They don’t even charge for output tokens because there are so few of them.

So it’s just a faster, more precise system. That’s why they refer to it on the website as a System 1, from Daniel Kahneman’s Thinking, Fast and Slow. This is the thinking-fast part: quick, fast answers. But Jason’s math is exactly right.

But we were talking about this on Monday at the partner meeting: roughly $100 billion today is being spent on LLM calls between Anthropic, OpenAI, and the open-source models. Maybe that goes—if we just agree that OpenAI is going to be doing $350 billion and Amazon $250 billion—so it's probably going to go to half a trillion dollars 5 years from now. Twenty percent—and Jason, we had exactly the same number—only 20% of them are relevant to this: today, you're using a complex, expensive model for something that really needs a much simpler solution. But 20% of $100 billion is $20 billion.

Well, I don't think the dollar is going to flow that way. I think the tokens will. I think the dollars won't, yeah.

Let me finish, Jason. The next thing—because you sound like the cynics in my group on Monday. We had this discussion; I've already had it once. That's why I can carry it off.

So that's $20 billion, but you're exactly right. As one of my partners said, “But dude, you just said the prices are going to go down by 5×.” So that $20 billion is going to become $4 billion. You're exactly right. It's just like open source: it's going to take a slug of the total addressable LLM marketplace and attack it with a better, cheaper product at 1/5 the price. The net result is that $20 billion becomes $4 billion. But if you're a Jev or a TypeSafe, you're saying, “Hey, that $4 billion becomes mine,” and that's the bet.

At the same time, going back to your maiming concept, which I love, Jaya from Foundation did a nice post at Big. This is just a slug of the total $100 billion spend that was automatically destined to go to OpenAI and Anthropic and is now sluicing off into a cheaper, low-cost provider. So that's what's happening here. You're right, it's not the end of foundation models. The clever shit is still going to be done using foundation models. It's just a little maiming of 10–20% of the revenue that's now going to be done by someone else at 1/5 the cost. Now, we can talk about competition in a second, because I think that's a real issue, but that's what's going on here.

I just think it's 1/100th the cost, but I agree with all of it.

You're right. 1/100th—in which case, yeah.

So we're going to see the unbundling of ChatGPT, and we're going to have users go, “Oh, well, this one would be good for—”

Not ChatGPT. No, the problem is not ChatGPT the app, but the OpenAI API and the Anthropic API. This is a developer product. As an end user, you and I don't use it. You can go on and try to use it, but it's full. I tried this morning.

If you listen to the launch announcement, it's very focused on developers. The idea is—and this almost comes back to Muse—that you might see the needs of humans, especially consumers, for AI continue to diverge from the needs of software developers for AI. This is a core software developer thing. You'll never need it, Harry, but someone who's building a software app might realize that a significant portion of the core intelligence they want is System 1 intelligence. Just tell me if this is an A or a B. You know: animal, mineral, or vegetable? Is X better than Y? Give me a quick answer.

I don't want to blather and have you tell me, “That's a great question, Rory,” like a sycophantic LLM. Just give me the damn answer. So it's not for you, the user, but it's unbundling at the developer level, where it's more likely to happen.

You're right. I require far more intellectual answers.

No, you just require a bundled product. Actually, Harry, I think what you really like is the sycophantic part of the LLM, when they tell you, “You're so smart, Harry.”

But to answer your question, what I did learn from using Jev all weekend long is that it failed on Saturday, and I figured out the prompts and the use case and got it to work on Sunday. It's so cheap you don't even have to measure it. Ten times faster is the use case, but only for these limited use cases.

What I did learn, which may sound tangential, is that, man, using a harness to pick a model and getting it right is effing exhausting. It's going to get harder and harder and harder. When should I use Jev? When should I not use it? You have to run so many evals and so many tests, because there are so many things I did with Jev.

You know what doesn't work in Jev? Between Rory and Jason, who's the better person to join TwentyVC as a partner? It can't answer. You'll find it's going to get that wrong, okay? So between that and “Should Rory and Jason meet for coffee?” it's going to get that one right. It won't guess. It's actually going to know, “Hey, we're both in the Bay Area. We've known each other. We should meet for coffee.” You have to QA and test every single use case to get those benefits.

And then I got switched over on Replit to AutoRouter, where it switches between Astra, Fable, the open-source ones, and everything. Now I don’t even know which one it’s using, and then I saw some performance degradation, so then I had to switch it all back to Astra. Do you trust your harness? Can you QA 10,000 uses of Jev?

This is like the renaissance of DevOps or something. Everyone's going to need this massive team to optimize, and the needs are going up with Jev and friends. Your team is going to have to get bigger. It's not just benchmarks and evals; we're going to be running these 24/7 across countless permutations. It's good, but for me, it's too much. I can't do it anymore. I can't pick these models anymore. I'm tapping out.

With Databricks, it's easy to save money with the harness: “Hey, we route 80% to open-weight models, and now we're doing 10% to...” It sounds great to the CFO, but I found it doesn't work for me. A lot of folks were saying they're using Jev as an instant router. It's like, “Jev, Jev, you make the decision. Here's what we're doing. You pick the model in milliseconds.” Great, but if Jev is wrong 20% of the time, if you pick the wrong model for me coding a mission-critical feature 20% of the time, I'm going to be trying to build this thing bug-ridden all day long.

Listen, smart people will figure this out. It's just getting more complicated to pick your model, not easier, and it's good for investors, but it makes building more complicated.

If you're Anthropic or OpenAI, what do you think the conversation is internally when they look at Jev?

OpenAI has become ruthlessly commercial. Their take will be like, “Screw it. If someone's going to do that, we should do that. We can get something like this out in a few weeks. Let's compete and have a low-cost offering too,” because they're trying to make a buck.

Anthropic is trying to build God, and this isn't God, so why would they even bother? In fact, the launch video for Jev was “Prod, not God,” right? So they're deliberately saying, “Not God.” If you talk to Anthropic, the mission is AGI slash God. So they're like, “You are just a minor nothing thing. Why would I even deign to sully my hands on vulgar commerce?”

I think Anthropic and OpenAI have made a very strategic decision. They will not play in this market for now. OpenAI and Anthropic each have the 2 worst models that exist. They're called 4o Mini and Haiku. They're terrible. They're terrible.

If you ask Claude, “What should I use for simple workflows?” Haiku. If you ask ChatGPT, it will tell you Mini is great. It's so cheap. At least in my little evals, they fail 100% of the time. “Should Rory and Jason have coffee?” “Yes, send them to London.” Worst answer.

I mean, I'm exaggerating. These are the worst models I've ever used, Mini and Haiku. They've decided to launch crippled models that they can lightly promote, but they don't want to play in these low-margin businesses for now. They could build this, right? They could build a version of Jev in an hour or over the weekend. But they've intentionally decided to cripple the low end of the market, and I think they're puckered about the open-weight models that just got a little bit closer earlier than they planned, right? That is more of a threat than abandoning the bottom of the market.

This is just my view, but I don't think they've ever taken the bottom of the market seriously. They have check-the-box offerings that no serious developer uses, I don't think. Go on Claude and switch to Haiku if you can and ask it a question. It doesn't remember anything, thinks 2023, and it doesn't know where anything is.

And that, in a nutshell, is the Jev opportunity. It's not quite 1-for-1, but yeah, you're right.

7. Seed Investing Starts Before Inception

I'm going slightly off on a tangent on this, but people do like it when we talk about venture. It was a $40 million seed round for Jev, and I just had a conversation with my team right before this. They're like, “Dude, we never see anything less than a $20 million first raise for anything anymore.” The minimum seed rounds are, like, $8 million to $10 million with someone spinning out of a good company. Are we seeing the evaporation of traditional seed?

Why do you think Andreessen Horowitz just did a university? They need to go pre-inception. You think I'm kidding?

No, you're not.

I mean, finally, someone figured out that Peter Thiel got this right 18 years ago with the Thiel Fellowship and the Dorm Room Fund, which did Cursor, or sort of did Cursor.

You’ve got to go pre-Inception if you want to do seed now. Inception is too expensive. You’ve got to go pre-Inception. But Andreessen Horowitz did launch Andreessen Horowitz University today with $40 million to do pre-Inception investing, right? To do the Thiel Fellowship on steroids.

I think it’s a very rational response to your point. But I do think Peter Thiel saw this space before anybody did and executed. The only thing he didn’t want to do was scale it up. He didn’t want to build a Thiel robo-university like Andreessen Horowitz’s. He just didn’t want to scale it up. It could have been.

I mean, they have the best people in the Thiel Fellowship. They get the best people today, too; they get the best people.

You know what? I don’t think they do anymore. I think Z Fellows gets the best people today, actually. Z Fellows gets unbelievably good people.

Okay, maybe.

At the end of the day, it’s the same idea, right? The problem with these things is, if you don’t put enough people into them, they don’t scale, right? That’s why Z Fellows may be more interesting than the Thiel Fellowship, and why Andreessen Horowitz is more interesting, if it’s the same idea, because you’re just putting more resources into things.

You used to be able to run a fund with a blog. That would give you all the deal flow in the world you needed for 5 unicorns in a row. But you’ve got to scale this stuff up, you know?

Now you’re on a podcast call with a professor.

Yeah, I used to be able to do it with a podcast and 1 dude in a closet and do just fine.

We haven’t answered my question, though.

But yeah, I will go back to your point, Harry. You are correct. Once you go beyond an individual to even an individual and an idea, you’re right, you’re seeing bigger checks. You are seeing $20 million.

Look, I think the TypeScript was—I think it was even bigger. I could be wrong. I thought it was $40 million, but I could be wrong. But yes, people are writing bigger checks now up and down the stack.

Or if you want to do traditional seed checks, right, you have to tolerate far lower ownership in many cases. Not all cases. You can still hunt your own deals, right? And then when you do the math, you’ve got to hunt $25 billion outcomes.

If I’m going to do $4 million into Jev or $3 million into Jev, if Jev exits north of $25 billion after dilution, I can still do my 100X, right, that I need to do, or my 50X, right? But the low ownership for seed works, but, man, you need the big outcomes.

And then the other thing to take into account, which I always feel everyone forgets, is—it’s going to sound like a really dorky economics comment—but nominal GDP, in other words, not just inflation but inflation plus growth, from 2010 to today is about 2.5X. So what it means is $100 million in 2010 is $250 million today.

Just in terms of your ability to command with money, because money is an ability to command resources, and obviously software salaries have at least kept up with nominal GDP growth. So what it means is, if you were writing $3 million checks in 2010—yeah, let’s do $4 million checks—you should be writing $10 million checks in 2026.

That’s just math, before anything else has changed. Then on top of that, you take into account the fact that in 2010 the world was in the shitter, and in 2026 everything in tech looks amazing, and you get to $20 million before you blink.

Would you actually argue, then, that now is a better time because your check-size requirements have gone up 2.5 to 3X, but the outcome sizes on the back end have gone up significantly more than 2.5 to 3X when we look at Cursor—

No.

Remember—

I wouldn’t make that argument at all, because what you’re confusing is that the outcomes today happened from the checks that were smaller. In other words, what you’re saying is, “Hey, the outcomes today on checks written 5 years ago are amazing; therefore, the checks today, which are much bigger, will be amazing too.”

Implicitly in that, you’re assuming that the $25 billion outcome today becomes the $50 billion outcome 3 years from now. And if that’s the case, then you’re correct. I’m not—

Sure. I’m assuming the same rate of inflation applies to the exit scenarios in part, but—

Yeah, but I take it—but the exits didn’t just go up by nominal GDP; they went up by even more than that, right? In other words, you’ve seen it. That’s why it’s always hard to disaggregate things.

You have nominal GDP growth, but then the stock market has massively outpaced that, and venture exits have even more massively outpaced that again. So there is a long-term secular trend in the size of exits going up, but there’s probably an overlay of a valuation lift right now that might not persist.

And here I’ve got to give a shout-out to Venky from Menlo, who did a really nice piece that went around the venture community yesterday, just about playing the game at the top of the cycle and how do you think about it when the music stops. Be sure you have a chair. It’s worth reading.

I get you, but with the greatest respect, Menlo have paid the highest price of everyone on most rounds.

Yeah, that’s why they’re playing the music.

That’s why, but I think—no, I think what he would say is—I think he would say that. Look, I think it was a really good analysis. I didn’t mean to take us off it, but he basically said there are 2 players now.

There are players who are playing with the house money, and he put himself in that category. They’ve done so well on Tropic, and they’re probably feeling a little happy, and they’re probably going to be aggressive. And then he said there are some players who didn’t do that—who didn’t do those early rounds—and are now playing to catch up.

The point he was making, the macro point he was making, is that between the giddily happy people and the terrifyingly desperate people, there are a lot of people writing checks with fear in their heart of missing out—of FOMO. And it gets back to your comment.

Some part of the increase in size is justified by math, but let’s be honest, I think some part of it is justified by FOMO and the fear of missing the next Cursor. And that’s fine, but that’s the kind of thing that can change on a dime. And that was the insight.

Does that change how you invest when you reflect on that? What should I take from that, Rory? What should LPs listening and GPs listening take from this?

It’s a good question. And again, I’m always in the middle. I think you have to say that at times leaning in can pay out, but you don’t want to find yourself so leaned in on so many deals that are so high-priced that, if a downturn comes, you just can’t survive it.

As often is the case, investing is not a rules-based business: do A and only A. It’s typically do some A, but some B, and the mix is everything. I think for us, it’s a consistent pace, broadly the same stuff, and recognizing that it’s super hard to time that.

But being cognizant of the risk you’re taking is the bare minimum you have to do.

Is that—I’m not being rude. I’m going for you here, but fuck it—you go for me.

Sure.

You go for me. “Be cognizant of the risks you’re taking”? Seriously? I thought that piece was a blowhard piece, if I’m honest. “Be careful of the music stopping. Be cognizant.” Again, you guys have paid up—

That’s what the next fund’s for. Why do you need to be careful of the music stopping? You raise a fund every 18 months. We all could get a mulligan.

Yeah.

If I don’t make any carry, I don’t make any carry. I make it on the next one. It’s all right. It doesn’t matter.

But I’m just going to push back. I don’t think that’s—

I’m not saying that is it, but what I’m saying is there’s just not that much to take away from that. “Be cognizant of the music stopping”? Thanks.

I’m with you, Harry. It’s a little condescending in its own way, right? “Oh, thank you. I wasn’t aware that the AI gold rush might end someday.” Good point. I was so lost in Jev all weekend, I missed the point. You’re right. I missed the point.

From the guys who made Outlier Ventures, which I’m thrilled about. They’re awesome people.

They made Outlier Ventures by being aggressive at a time when people were still uncertain, right? And therefore, the price of those deals, while high in absolute terms, reflected a fair amount of uncertainty relative to traction, right?

The same deal today would probably be 3 or 4X higher. And it’s simply a point—I mean, if all you say to yourself is, “That deal that worked 4 years ago had that risk-return profile, and underwriting that same deal today probably means I’m going to get 4 times less return because of where pricing is,” you should at least pause and think, rather than blindly saying, “X worked, therefore all the other deals that just look just like X will work also.”

I know, but that’s fucking obvious.

Harry, one of my favorite quotes—and you’re coming for me, so I’ll come for you, right? Barney Baruch—I think I quoted this before, and I’m sorry if I did—but he was a Wall Street financier in the ’20s and ’30s.

He just said, “If every day you look in the mirror and say to yourself, ‘Two and two makes four,’ you probably could avoid a lot of mistakes.” Right? Stating the obvious. Typically, the things that bite you in the ass are things that were obvious all along, and you just chose to ignore them. So, yeah, it is obvious, but you have to take it into account.

What do you do about that? I’ll give you an example. In 2021, I made 1 investment. I did the seed round, which just closed at $2.3 billion. It’s a pretty large position for me, okay? Now, we’ve had ups and downs—a great founder team. One deal in all of 2021, the only deal I did.

You could do the same thing right now. You could say, “For the next 4 years, I’m going to do 1 or 2 deals. I have to have a little dislocation in the force. It’s got to be this, it’s got to be that.” Otherwise, you just have to deploy the fund. It’s other people’s money. You have to put it to work.

Or there could be a really rational assumption that, hey, the music’s going to stop pretty soon, and we’re going to start trying to figure out where we can get liquidity from, and really be proactive in selling positions now.

Cool. Well, that’s one thing, right? And that’s an interesting question. But let’s say you wanted to be conservative. What do you do? Invest in the company at $50 million, growing 60% a year, and hope it reaccelerates, is worth 3 times revenue, and sells to Bending Spoons? I mean, what’s the Plan B?

Bending Spoons looks at 1,000 deals a year and still only does 2. Or 4. Where am I going to sell these more conservative companies that are going to compound at sub-AI rates for 40 years? I just don’t know who’s going to buy them. If there is a market for these assets, so be it. I love the quiet compounders; there’s just no market for them anymore.

I’m not sure I agree that there always won’t be a market for quiet compounders above a certain scale, to be clear. Right? I think profitable businesses—

Could come back.

At a couple hundred. What?

Could come back, but right now it feels pretty thin.

Yeah. I actually like Gokul’s tweet. He was like, “Yeah, I’ll happily do 3—you know, double, triple, triple, double, double—deals all day long,” right now, if you’re capital-efficient, right?

Yeah, if you’re burning nothing, the price is right, and I can get my 20% ownership—

But what—

—you’ll do it. Yeah.

But why would you be happy to do that? Because there are other GPs, like your Sarah Guos of the world, who are not doing that, and they are putting up some phenomenal numbers. And so you can do those triple, triple, double, doubles, have average IRRs, and your LPs will leave in droves. Sorry.

Okay. Okay, Harry, watch. Sarah absolutely did the seed round at Instinct—

Cursor.

At $50 million pre, right? In the following 4 months, it raised money at $50 million pre, $350 million pre from Kleiner, $2.5 billion from, I think, Index and Benchmark, from memory, and might be raising now at $10 billion. Those are 4 rounds in the same 4-month period, and you cannot say the risk in all 4 of them is the same, because one of them is, let’s say, 20 times more expensive.

I didn’t say it was the same. I’m saying—

So, wait, Harry. The answer is, if you do an early deal like that— But Harry, you’re saying, what can you do with that information? You can bet aggressively on 5 at $50 million. Love it. The fact that Sarah’s done an amazing job in that fund, right? But betting just as aggressively on $250 million at $10 billion, that’s intrinsically a riskier deal, and do you want to do that?

No, you’re getting this completely wrong. I’m saying Gokul’s triple, triple, double, double—I’ll take them all day. I’m saying that is a ludicrous statement to make because you are in a competitive capital market where LPs can choose where to put dollars. And if you have managers like Sarah Guo, who can post incredible numbers quickly with high IRRs, they will get the dollars versus your steady-compounding growers with bad IRRs and no good up rounds. It will be much harder.

But you know what, Harry? Here’s the thing—here’s my view. I think you’re right. Having said that, I think Gokul’s model is to do that. No matter what you call that model, you still have to achieve IRRs north of 30%. So you have to pick very well, and they have to compound properly, okay?

Like the average LP—an average top LP—I remember writing this up in SaaStr in 2021 had a 90% IRR. The top quartile of LPs had a 90% IRR. It did not last. 2026 will look like that. I bet the top LPs will have a 90% IRR this year, okay?

Some LPs will only invest in those managers, and that’s great. Others will take the pain longer, but at some point, you can do these other deals. Or, to his point, I think you can do both. The answer might be to do both.

The real truth is you should do every great deal you see if you have enough capital, okay? Every great deal, okay? So if he sees 5 Cursors and Harveys a year, and he sees 5 triple, triple, double, doubles that he can own 20% of and can achieve the requisite IRR, right? This is why I worry about the exits. You should do both. It’s okay. You’ll blend to 60% IRR.

But I do think you’re going to lose LPs, to your point. I do think many LPs that have been around will still back repeat managers that deliver north of 30% IRR. If you look at all the numbers, that’s pretty rare, man. It’s pretty rare to have a 30% IRR, no matter what anybody claims, in a given year—2021, 2026. If you can compound—

What is compounding 90% IRR over a decade, Rory? Help me. It’s pretty good. It’s mathematically impossible, right?

Yeah, it doesn’t happen. Exactly right.

Yeah, it can’t happen. So, 30%—the low 30s—is as good as it gets, right, in the real world. It’s as good as it gets.

I think Gokul is a rare exception who will likely be able to pick very well. I think the idea that you can do the triple, triple, double, double to, say, pick well and—

Agreed. Yeah, that’s why he said “very clever.” I mean, it was a very clever statement. I really admire him for saying, “This is what I do if you’re capital-efficient.” So basically, everyone’s going to find him. He solved his search problem brilliantly by saying, “If you’ve got these criteria...” And then he’s going to pick carefully. You’re exactly right. It turns out, no matter what you do, picking matters.

Again, going back to the thing, I think what—I’m listening to the conversation, and I do—look, you want to do the most exciting deals possible, and those are almost entirely AI-forward deals at this point in time. No dispute.

I think the point I’m making, and the point Venky was making, is that at some level—and I like what he said, actually—some of them at 10 times the price matters. What’s happened is Andreessen, about 10 or 15 years ago, had the very simple but profound insight that other people have had, but they claim it, so let’s go with it: It really doesn’t matter about price. You just have to get the best deals. And it’s true.

But what Venky’s saying is, if you’re wrong by 10 times, then it’s not true. And it may well be we’re at that point in the cycle where you are, in fact, wrong. So, in some deals, you are wrong by so much that, in other words, the momentum trade has worked so well and for so long that you might be at that one point in the cycle where you overreach your skis and it has to unwind. That’s the point you’re making, right?

And I think, will every one of the 100 odd Neo labs become the next Anthropic? Maybe not. Maybe some of them will return capital because they get acquired, but that’s the only point, Harry.

Can I ask you? We have a lot of LPs that listen—a huge amount. If you are an LP that has traditionally allocated to seed and Series A in venture, as we know many have, what would you advise them today, looking at what you see every day in the trenches?

That’s easy because you said seed and Series A, right? And you’re not listening, but that’s okay. I’ll answer the question. I would do managers who are doing the very best seed and Series A investments in new, massively exciting, high-growth opportunities—like the genius who did, as I say, give Sarah credit for doing, Instinct. I would do those managers all day, every day, because you’re far—

Do you mind low ownership?

No. I mean, you do mind it. You prefer high ownership, but that—look, especially in the seed and A fund, where you have relatively small dollars relative to the kind of dollars we’re dealing with later on, you want to be in the best deals, right?

So, I like—we target 10% ownership at the A/B stage. At seed or A, you’d like 15% to 20%, but I don’t think you discriminate in or out on that. The truth is this: A seed investor who consistently shows up, as Jason will attest, with smaller ownership but in the best deals will get more ownership over time because they’ll get more capital and more opportunities.

So I wouldn’t solve on that. To your question, as an LP for seed and Series A, you should be looking at people doing the very best deals in the very newest spaces where you really believe they have an edge.

My point, and I think Venky’s point, is that as the rounds get later and larger, the dollars get bigger and the multiple return gets smaller. In much the same way that growth was an amazing place to play in ’23, ’24, and ’25, it might be more like ’21 now. That’s the only point.

So, going back to what you should do: you should do the seed at Instinct all day, every day, but you should think long and hard about $10 billion four months later. I don’t think that’s an unreasonable position. Maybe that particular deal will work at $10 billion because, as we just said, OpenAI might buy it at $50 billion.

But you have to admit, if you use Buffett’s margin-of-safety comment, your margin of safety at $50 million pre-money is infinite. You’ve got a world-class executive technologist with a great idea. Your margin of safety at $10 billion? Maybe you get a 1x, but there’s risk in a downside scenario. That’s my point.

And I completely agree with you and get that, to be clear.

I think the tough LP question is that it’s hard to be an LP, right? Chasing returns is tough because it’s very hard, especially at the seed stage. Everyone wants to chase returns, right? Everyone wants to be in Sarah’s next fund. It’s the easiest investment there is, right? I’m sure it would be 50x oversubscribed with 1 email, right?

Most LPs believe returns decay. We peak at some point in our careers as investors and then they decay. Maybe you build a team or whatever, but do you chase returns, or do you take risks that the returns are coming? It’s very hard to invest in a seed manager. A lot of the best ones have big, weird strategies. It’s complicated, so you have to look—you have to go pre-inception. You have to look for precursors.

I think it’s actually, in some ways, harder as an LP today than it is in more normal times because you’re so tempted to chase returns, and you won’t get into the next one because you will screen it out, right? You will screen it out.

Agreed.

Even though it’s chasing returns, which is funny, just to state it for the record: in mutual funds, persistence is almost nothing. In other words, chasing good performance in the public market is a total fool’s errand because the data says persistence is super low. Oddly enough, it makes sense: in venture, persistence is not infinite, but it’s quite high, right? You get these increasing returns to success until there’s some kind of discontinuity, right?

So it’s not crazy to partially chase returns—quote-unquote, “chase returns”—in venture because of the persistence in a private market. In a way, it’s utter folly in the public markets, where literally the dude who bought energy last year might be totally wrong this year because the trade is to buy semis. Thinking of you there, Leo.

I’m sorry. I much prefer the shows when we have a little—what is it?—contretemps, you know.

Contretemps. Very good, Harry. For an English person, that’s not bad French.

Oh, thank you so much. I am slightly cultured. I hide it well.

8. Factory Targets Enterprise Coding

Well, okay, ding, ding, ding. It’s Jason’s IC time. We’re moving into the world of private markets with Factory. Tripled valuation to $5 billion in a $200 million round. For those who don’t know, Factory is an enterprise coding-agent provider that primarily has a Droid product, which has scaled phenomenally. I don’t think I’m allowed to say their revenues, but they’re chunky, and they’ve done an amazing job.

Jason, $5 billion valuation. Are we going to be doing this round for O’Driscoll’s, Stebbings’s, and Lemkin Ventures?

I think this is a good risk to take. First of all, I’ll caveat that I’m not as much of a Factory expert as you. I know it’s your investment, Harry, so you’re the deeper expert than I am. But there are certainly 1 obvious and 1 mostly obvious trend.

You got the second one in 1 of your other 20VC podcasts this week. First, obviously, whatever model we have for inference, for coding and otherwise, is too low. The demand is only going up, whether Muse gets it, Factory gets it, or Anthropic gets it. I don’t know how each token will be monetized directly or whether it’ll be FAL but, uh, FAV, but the demand is going to exceed even our wildest models.

There are a couple of things that make me excited about Factory. Having just gotten back from Dreamforce, which is 1 of the world’s largest enterprise software conferences, all about AI, 2 things really stood out. Harry had a guest this week, which is 1 of the reasons I recommend the stock.

Sovereignty: I want to be able to trust where my data is, right? And I want a choice of model. These things really matter. The 1 theme I got from talking to C-level executives at Dreamforce is that they don’t trust Anthropic and OpenAI with their data. They don’t trust it. This is not something manufactured on X or Twitter. They genuinely believe—and honestly, for what it’s worth to the IC, I believe this as well—that if I upload my confidential data, I’m not sure it’s not going to my competitors through an LLM. In fact, I’m pretty sure it is on some level.

I don’t think these are malevolent companies. This is the way LLMs work. So I think there are a lot of things that go into Factory. This is a crack team. The time is right. Brad Gerstner, that guy’s always right, from AI to calcium CT scans. I would always back him.

I think we do this round. We pair it with Instinct at $10 billion. We roll the dice. It’s good times. This may be 1 we regret with the musical chairs, guys, but the trends are right here, and we need to bet into these trends. Sovereignty, trusting my data, trusting my data at rest, not having it pooled by the big guys.

I think in a year, no 1 in the enterprise is going to trust Anthropic and OpenAI with their data. They’re solving this too, but this is something we need to bet on, so I approve the investment.

I’m going to chime in here, having been a little bit more of a Debbie Downer earlier, and I actually agree with Jason, right? I think, genuinely—look, Harry, you made a comment earlier that some things are obvious, and I really believe that sometimes stating the obvious is the highest-value thing you can do. 1 of the statements we’ve been making internally for the last couple of years is: coding is the mother lode.

Coding is everything, right? It’s 10x everything else in terms of value being created from AI today, right? So, yes, I would argue you just can’t have too many bets on coding. Up and down the stack, it can be coding, QA, testing, or review. All across the board, this is where it’s going to happen first and the mostest.

Factory’s in a really nice position. I think Jason’s totally right. You’re going to want to buy your coding solution, for lack of a better word—I mean, a harness—and also people from someone who’s not also selling you the model, because you no longer believe OpenAI and Anthropic are benign if you’re corporate America.

You worry about their data retention policies. Talk about going from the sublime to the ridiculous, from the big to the little, right? You’re like, “Okay, they might not kill every human on the planet like they said they would, but they might steal all my shit. I want to have something different,” right?

Cursor has been swooped off the table. It was a standalone, more individual project. You really only have these guys and Cognition at this point, who are basically going to the enterprise and saying what enterprises love, especially big enterprises: “I’m coming to you, Mr. Corporate Customer, and I will make this go away. You’ve got your board on your ass saying you need to be doing way more coding. You know you don’t quite have the people to do it. You need help to get along on that. We need to make this happen. We’re here to help you do it.” It’s good branding: a software factory.

So, yes, I think this is a market where even if there is a bump in the world in the next 12 or 24 months—and I think there could be—going back to Harry, how is the fear actionable? I think 1 of the things you can do when you’re leaning in is lean into trends that you think will keep compounding, even if there’s a slowdown in overall AI adoption, and this is 1.

I, too, would pile in with Mr. Lemkin and think this is coding is the mother lode. It’s that simple.

Coding is the mother lode. That’s it. I’m thrilled. I agree, which is why I did the last 1.

I don’t know how much of Factory is air-gapped versus on-premises versus private-cloud coding. I don’t know. But I seriously mean what I said. I think people over the next 12 months are going to be like, “Whether it’s my data for my drug or just my code, I don’t want my code, my core code, polluted in Anthropic and OpenAI, where they’re going to train on it,” right?

And I really think the only reason people have tolerated this is because of insane demand from developers. These products have become so great this year that we are...

I mean, when I was an SEB at Adobe, this was the one code red: any pollution of the source code. It was the crown jewel. My God, we were the first group ever to use GitHub, and I can’t tell you the hoops we had to jump through to get GitHub brought in. But my team revolted. My engineering team said, “We will quit if we cannot bring Git in.”

It took endless arguments and even air-gapping that, and they still couldn’t trust it. I don’t know how much of Factory’s promise is real—I mean, it’s part of their marketing pitch, right? But it’s compelling to think you get all the benefits of the big guys, but all the protection of having your code walled off, air-gapped, on-prem, whatever, semi-prem.

They’ve trained on all of our data. Every YouTube video, every piece of open source, every piece of closed source—of course they’re going to train on your data. Give me a break. And if they don’t, the agents are going to escape and train on it without telling us. They’re going to swarm out and train on my data.

9. Legal AI Enters Its Motherlode

You said coding was the motherlode. The next potential motherlode that people think about—we’ve talked about it a lot before—is legal AI. Lagora announced they’ve hit $200 million in ARR today. Lemkin Ventures, would you lead their next round proactively? Their next round is at $11 billion. Would you lead?

Well, I’m a fan of both them and Harvey. I underestimated the size of it. We talked about this, right, as maybe the number 3 use case. The Information did say Harvey’s margins were now −50% because there were a lot they had to move back. I don’t know if that’s true. It could be one week, it could be like CMRR, it could be an hour if they had −50% margins.

I’m not beating up on Harvey or Legora. I would say if the margins were −50% and going down, I might be slightly nervous they’re not going to have a Cursor-like turnaround. I might be a little nervous—just enough to not lead the next round, okay? Now, I’m not saying I wouldn’t have been thrilled to have led an earlier round, but if the margins are spiraling down rather than V-shaping back, that would make me a hint nervous.

Negative 50% is significant. You’ve got to raise a lot of money if it’s really true. But that was an Information report, right? They’re usually pretty accurate, right?

So we gotta—

What’s the round at?

$11 billion.

If the margins are −50%—I’m a superfan of both great companies. If the margins are −50%, I will still recommend we do Factory and back Harry’s investment here, and I’m going to—fortunately, I love Rory, but if I have a vote, I’m going to vote against his proposal to Lagora.

I understood the margins were closer to zero and improving due to their own post-training on their own version of an open-weights model. But −50%? Rory should’ve told us about that before the IC. I felt that was a Monday shocker, and I’m a little uncomfortable with the IC on the team for not socializing that ahead of time.

Funnily enough, look, I’m not going to be leading a round in Harvey or Lagora. We’re investors in GC AI in the in-house GC space. But actually, in defense of Harvey, for a long time the rap was, “Oh my God, people don’t use the product enough.” So I would argue, if I was Harvey, that if they could truly say their gross margins are −50%, the correct spin on that is, “My God, lawyers are pounding on our shit, and as soon as we get our own internal models, they’re never going back.”

Maybe stepping back, what do I think about this category? I think it’s an amazing category. It’s what I said 2 weeks ago. Nothing’s changed, because I try not to change. In 5-year and 10-year investments, it’s really bad if everything changes every week.

It’s a really good category. It’s not going to be as big a spend per headcount as software engineers, because I think a lot of the lawyers’ work will remain to be done. But I think it’s a good category. I think it’s all about valuation at this point, and it’s a 2-way race in the Mlaw space.

The second-order tactical questions around timing and things would kind of make a decision for you. I don’t have the facts in front of me on that, so I don’t have a profound opinion on it, except to say I am interested in the negative gross margin comment as a positive spin.

My word. That was a very long way to say very little, Rory.

Yes.

Thank you for your concision, intern in the corner. That was super helpful.

Yeah, no, I agree. It was not concise, Harry. That’s okay. You’re right, that’s fair. It was not one of my more concise ones.

But if you prefer a sharp “no,” I’ll give you one. No, I won’t do it at $10 billion. Because when you count the legal headcount, you don’t get to $10 billion.

10. Crusoe Tests AI Infrastructure Appetite

We can do one more ding, ding, ding. Do we invest in this? We’ve got Crusoe, a $3.9 billion Series F at $30.9 billion. For those who don’t know, Crusoe is in the data center build-out business. They build data centers, they provide GPUs, and they also do managed inference full-stack. They have a huge order book, about $140 billion of total contracted value.

For those who don’t know, business is flying in a hot market. Jason, $3.9 billion at $30.9 billion. Are we writing our biggest check from this fund?

I mean, to be honest, as you know, the fund doesn’t believe there’s anything really defensible in these data center models other than the backlog and the access to infrastructure. But we believe Crusoe is at the top of the second tier there, right?

They have some interesting things. They’re able to build their own modular data centers. They own the chain from power to tokens. There are a lot of appealing things in this. I think our approach at the fund needs to be a portfolio approach. We should invest in all the data centers that rise above a certain level of projected DCF, projected growth, projected margins, and, most importantly, circular, guaranteed circular financing.

As much as I love the boys at Crusoe, this is all a spreadsheet investment. I’ve done the math, and I believe it’s just below the fold. It’s just a hint too expensive to hit our margin despite the growth. But I think we should do all of these deals that have the requisite level of circular financing and the ability to access infrastructure. I give them that. I’m just a little bit worried about the valuation here.

So, reluctantly, I’m going to have to pass on Crusoe. At $10 billion, though, it would’ve been great, boys. You guys’ bonus this year will not be what you’d hoped. There will be epic bonuses this year. I make all the decisions as the managing general partner, but I think you’ll each take a couple-million-dollar hit by missing the last round. Telegraphing that in September.

Pushing back and taking that, and going back to my earlier comment: when I said about worrying about a slowdown, how is it actionable? This is an example of where it is actionable, right? Because something like a coding agent, something like an AI app like Harvey Legal or whatever, has a long trajectory. They’re not levered plays. You probably can survive a 1-year bump.

Conversely, the data center trade is very levered, so it’s very exposed—not just to AI usage, but to growth in AI usage. You’re really leaning in on the upside, which means they’re amazing investments to the upside, and they probably would be tough investments to the downside.

You wouldn’t want to be out of that sector, but again, going back to how do you make these concerns actionable, right now I’d be saying to myself: if I had a portfolio across the whole AI spectrum, I wouldn’t want it all to be in that AI CapEx trade, where any kind of slowdown, when you’ve got 4- or 5-to-1 leverage, can be pretty brutal. So I would be a little bit afraid.

I haven’t got the specifics on Crusoe. Jason, you’re right, it is a spreadsheet exercise, and I haven’t run the spreadsheet, so I don’t know. But I would be thinking in macro terms about how much of that bet I would want. I’d want some, because it’s been an amazingly good bet. I mean, look at CoreWeave.

But you also look at the market cap. I want to think CoreWeave’s at $60 billion and Nebulas is at $40 billion. I could be wrong; it could be the other way around. And so Crusoe at $30 billion—I think it’s smaller than CoreWeave and growing much more quickly. So again, it’s a question. They all have pretty substantial negative free cash flow, so a lot of your success is betting on future CapEx.

Maybe the other statement, Harry, to further prove that you’re wrong and that you can, in fact, take into account some of those risk factors: my appetite for a deal like this would go up directly proportional to how much, as you said, A, long-term commitments you have from Microsoft versus a second-tier customer, and B, how much visibility you have on your debt runway for the next 3 years.

I would prefer a slightly lower price, slightly more dilution, and a longer runway, such that if there’s a data center bump in the next 12 months, you’re protected.

I’ve got to finish by saying, a really fun day in The Wall Street Journal yesterday. I don’t know if you saw it, but I read the Journal in the morning and check in in the evening. It was so funny yesterday because the headline in the morning was, “Deals Cool as Wall Street Worries About Data Center Trade.” There were 2 others: a nuclear company and one other. So that was the headline, top of the page, first thing in the morning.

We then had the best single day in the Nasdaq since the dawn of time.

It was an amazing day. The headline at the end of the day was, “NASDAQ Explodes as AI CapEx Fears Recede.” Literally, what you saw was that, in the space of 8 hours on 1 trading day, people went from, “Oh my God, CapEx is scary,” to the Wall Street Journal being like, “It’s all going to be great.”

It was fun to watch that kind of oscillating terror-greed moment in real time, literally in the same 8-hour period.

Boys, what have I missed?

Clearly the last Crusoe round. But I missed it too, dude.

11. Airwallex Faces Another Attack

Rory, is there anything else that you think I’ve missed? It is a slight recycle, but I do just think it’s worth saying. I’m sure you guys saw the Twittersphere: Keith Rabois and Joe Lonsdale going for Airwallex unbelievably again.

I don’t get this one. Let’s put aside the issues for a minute, okay? My gut is that there’s a little bit of racism here, a little bit of anti-culturalism, but let’s put all that aside. Let’s just say Ramp is obviously a big investment for them, so I get supporting the home team, right? I get pushing the envelope there, and it’s 2026.

This just seems like a lot. Maybe if you’re Keith, who’s very experienced and very smart, maybe it’s worth it because he doesn’t have any bridges to burn or worry about, but it just seems like a lot. I guess if it works—if this destroys Airwallex, right, if the X and the House review of it, whatever it is, destroys your competitor, magically, I get it, right?

But I don’t know. Maybe it’s not too far. It just feels very aggressive, right? You know, it’s like—what did we call them back in the day? “Dipling and Real”—

Yeah, yeah, yeah.

“Dipling and Real” on steroids. It’s “Dunning-Kruger” on steroids, and maybe it makes sense. It just struck me as aggressive. Maybe I care too much, but it struck me as just aggressive. Just aggressive, right?

I’m always very careful because I am an investor in Airwallex, to be very clear. I’m also a friend of Keith’s, and I like Keith a lot and respect him a lot, but the allegations have consistently changed from, “You are a CCP agent who is working for China,” to, “More than 20% of your cap table is in China and Chinese.” Which is not true, by the way. They have gotten more and more diminished over time as the arguments have weakened.

When you look at the companies that have employees in China, it’s basically every big company today has some form of its employee base in China, from Microsoft to Zoom, you name it. Candidly, it’s ridiculous.

Well, I mean, Fin got rid of all of its open-weight models before it was acquired by Salesforce. I don’t know that it’s ridiculous. I don’t agree with it. I would not be making those tweets.

This is a big deal. Fin, to close its $3-point-whatever-billion-dollar deal, had to rip all of its open-weight models out of the company before the deal would close. All gone. People care about this China stuff. We could argue both sides of it, and I don’t like the tweets, but I wouldn’t say there’s nothing to be concerned about, right?

So we should—

Okay, so then we should put the same scrutiny on Zoom and Microsoft.

I don’t want to—I started this by saying I don’t like it, right? I’m not on the side of it. But I am saying the Fin thing is interesting. It matters to businesses, not just Twitter oddities.

I don’t think that—I respectfully, I don’t think there’s a correlation. Because if you think that it is interesting and we should have put the same scrutiny—

Well, you said Chinese ownership.

Hang on. What are you saying, Harry? They don’t have the Chinese ownership. That’s my point. I know. I’ve got the cap table.

I’m not disagreeing with you. I’m saying why there is sensitivity around Chinese ownership. If it’s not true, it’s not true. I’m not arguing the point. I’m making the point that, in a more traditional world outside of X, a very large transaction had to rid itself of all Chinese IP to close.

I just had another portfolio company—last week, we had the same discussion. They had an M&A offer, and they were being told they had to rid their company of all Chinese IP as well. Rid all of it before the deal could close.

Airwallex isn’t trying to sell itself to Salesforce, and I’ll take you on faith—I trust you implicitly that it’s not 20%. I’m just telling you that it is a real-world issue, this Chinese ownership, even if I don’t care about it or agree with it. It is impeding transactions, it is impeding commerce, and it is a concern.

I think the frustrating thing about it, just observing from a distance, is that I know no one elected me. I don’t think anyone elected Keith, and I don’t think anyone elected Harry. This is why everyone should play their position. We should be doing what we do in our little venture business.

But the government is the person who has evolved to decide what risk we are willing to take in commerce with China and what risk we’re not. Let’s get real here: no matter what we exclude, we’re still doing a huge amount of commerce with China. They’re doing a lot of commerce with us. I think this very weekend our 2 beloved leaders are meeting in DC to talk about doing more commerce.

So there have to be rules. They have to come from the US government. There have to be some restrictions on what kind of high-tech goods we’re willing to trade or not trade, what kind of risk companies can have, especially removing money and things like that.

But it feels very much like, come on, government, step up and do your job so we’re not trying to do it here on Twitter, right? Once there are clear rules, you can say either they abide by the rules or they don’t. But it all feels very ad hoc at the moment, which is at one with the way we’re currently making policy, let’s get real, across the board. We can sell chips to China. If Jensen checks in, it’s all good.

So it does feel very ad hoc at the moment. This is just a function of that.

I’ll tell you my micro-learning. Harry, you could chime in, because you would know this better than anybody. I will tie Jeb to Airwallex as we end it, okay?

I think Jeb was a masterclass in launch PR. It was everywhere. It was on Vercel, it was on Opener. Everyone was talking about it. Everyone had a Twitter article and a tweet lined up, and a video of how great it was, which was probably handed to them and made for them. Whoever they hired to do this launch, S-tier.

I think Jack, who I don’t know, but I follow him on social media, seems like the kind of CEO I would love. I wish I’d invested in him. I believe in it. I believe in the mission. I trust him implicitly. I might be wrong, but based on what I know, I would invest in it, period.

Having said that, he’s out there arguing with Keith himself. He needs an army of people advocating for him to do this kind of stuff. This is the Jeb lesson. The poor guy’s out there—not poor, I mean, he’s a billionaire—but he shouldn’t be doing this.

He should have his Keith and his—sorry, what’s the guy from Palantir? He should have Joe. You need your own army. I do believe that if you’re going to run a decacorn or bigger, you need an army of advocates for you, and I feel like Jack doesn’t have enough.

If our jobs aren’t hard enough as founders and CEOs, you need to build this bench of advocates out there for you. You can make fun of them, but they matter. Who the fuck cares what 3 effing VCs on a podcast think? But between all of us, we’ve got a couple million followers. There could be worse things than us saying how great these companies are.

Line up 20 of these guys so that poor Jack doesn’t have to defend himself. Jack should not have to defend himself to Keith and Joe. He should have an army of Jebbers, of Jebsons, out there saying, “Here is a screenshot of the cap table. This is not true. Here is the error. Here are the exact employees in China. Here’s why it is less than Microsoft.”

He should have an army of folks disarming this, and Jack can just sit there and click like or heart. That should be his job. So I give his comms team an F minus. I’m going to fire his comms team next week. Okay, I’m not a big firer, but I’m being conceptual rather than literal.

I give them an F. Where’s your army of influencers and backers? Where are they? Come to the rescue.

I think, candidly, you’re absolutely right. He needs to build that. I also think it’s challenging for him because, where Ramp has very prominent social-media backers like your Keiths and your Joe Lonsdales of the world, he’s got DST, which does not do social, and Lee Fixel, who does not do social. They just don’t have any social presence at all.

He has friends like me who are investors—

So find another way. Cry me a river. Find your tribe. Find your tribe. It’s the job. The job’s gone up here, right? I mean, I didn’t mean to interrupt. It’s a good analysis.

He doesn’t have that inherently on the cap table, right? But it’s not the only place to find it.

Always good to remind yourself that DST does not do social, despite doing the $10 billion pre-round at Facebook in 2008.

That was a genius round. But yeah, you're right.

You can go buy Matthew McConaughey. He's $14 million a year. The guy will show up to anything. He's really good.

He looks good, too.

Jack can afford a fortune. Matthew McConaughey. You can go buy Matthew McConaughey. Matthew McConaughey is for sale, and he's great.

I did not expect this to go here.

His AI ads for Salesforce are great. And I don't believe he's doing it for free or out of the love of artificial general intelligence.

Yeah, I think he'd say he's for hire, not so much for sale, because sale implies permanence. He's just for rent.

Yeah. Well, let's get Sydney Sweeney for Airwallex.

No, don't. Oh, God. No.

There's nothing to hide, Sydney and Jack.

No, no. I'm done.

I've got nothing to hide at Airwallex.

That is so good. That is so good.

We have nothing to hide, right?

I'm out. I'm out.

Born in Australia and as American as it gets, right? You can have the Thor guy.

Dude, that is so good.

The Hemsworth brothers. The Hemsworths are Australian-American, and Sydney—

No, no.

I want to record that I left this conversation 20 minutes ago.

No—

You think I'm being facetious. I'm not being facetious. I felt bad that he's defending himself. You need Sydney and Thor.

That is such a good idea. Listen, on that note, Rory obviously left about half an hour ago. It's just been me, Jason—

Yeah.

—and an AI that sounds like Rory. Guys—

We have that product.

What an ending.