20VC：OpenAI估值下调至1.25万亿美元｜公开市场崩跌：GOOG 12小时市值蒸发5000亿美元｜DoorDash以170亿美元收购Instacart｜Kalshi融资10亿美元，Cognition瞄准800亿美元估值

Agents are an existential risk to DoorDash because their number-three competitor is nobody. All these app platforms have a risk of slight disruption from agents.

Calci is a money deal. If you want revenue velocity right now, nothing gives you revenue velocity like inference. Physical Intelligence, a company with relatively little revenue, is worth $20 billion. Dropbox, a company with billions in revenue, ain’t worth a lot there. It just shows revenue multiples don’t get you anywhere.

This is 20VC with me, Harry Stebbings. It's the best show of the week. It's Rory O'Driscoll, it's Jason Lemkin coming together to discuss the biggest news in tech. Today, we have OpenAI's big fundraise, $1.25 trillion price down from 1.4. Apparently only $11 billion committed. Sad face. Um, then we have the AI trade hitting a wall. Alphabet, whoa, $500 billion off the market cap overnight. Intel falling 29%. Then we have DoorDash buying Instacart for 17 billion. Base10's new round. We have Physical Intelligence's new round. We have Serval's new round. God, the show just keeps on coming in terms of these venture financings. This and so much more in the show today.

Boys, we are back. Luckily, there has been a huge amount happen in the last 7 days. Let’s start with OpenAI. Its valuation got cut from $1.4 trillion to $1.25 trillion, and the round was reduced from $30 billion to $20 billion, with only about $11 billion committed. As a result, everyone is going, “Oh, no, this is further evidence of the AI wave cracking.” How do we think about that?

I would dispense with the whole AI wave cracking narrative. I think people are trying to extrapolate their desires, wishes, whatever prejudices and biases, onto the story. It’s as simple as this: the other company is bigger, growing faster, profitable, and about to go public, so you can buy it liquid at $1.5 to $2 trillion. What would you pay for the smaller company, growing less quickly, still losing money, that’s illiquid? The answer: less than $1.5 trillion.

How much less? That’s what they’re figuring out, so it’s utterly not surprising. It is as simple as that. No projection about AI, positively or negatively—just 2 securities that are going to be sold in the next 2 months, and which one do you want to buy?

But Jason told me last week that it was the most underappreciated.

I think it is; it may need to drop a little more. Don’t forget, there are going to be bumps, right? Broadly, what we call AI stocks fell $5 trillion last week. Maybe this AI round was a week or 2 before that. There’s obviously a temporal gap, but when we’re seeing this level of volatility, there should be some nervousness.

I’m surprised people didn’t talk more about it. Anthropic had a down quarter in net new revenue. It was down in this last quarter in terms of net new bookings over Q2. Maybe it would fill up at a trillion. Maybe I’m glad we didn’t put in our bid until the end, Harry; the second tranche will be lower than the first. It makes sense that there should be a little bit of wobbling at the moment.

The stock market’s down. We could argue—maybe Rory will challenge me—that Anthropic’s last quarter wasn’t quite as crazy as some folks might have thought. There should be wobbliness, and maybe it doesn’t matter. It matters where it’ll be in 12 to 18 months, but it makes sense there’s wobbliness. There’s wobbliness in the debt markets. CoreWeave debt is paying 15%. That’s wobbliness.

Broadcom having to guarantee Anthropic’s debt is a little wobbly. I know Rory took me to task on this last week, but it’s just wobbly. IPOs are always thin and wobbly. No matter how many times we say they get oversubscribed 20x, so many of them are just a little wobbly going into it.

Yeah, again, I didn’t take you to task. I think all those things are true. The way I think of it, there’s a lot of pushing and shoving, a lot of price discovery and risk discovery going on on the supply side of the AI trade. What price do you need to be paid to lend money to CoreWeave? What price do you need to pay to buy OpenAI in the private markets when you get in the public markets? I agree, all that stuff is going on.

When you go from assumed euphoria regardless to tempered euphoria, there’s this price-discovery process on the way. I think that’s fair, Jason. Those things are all true. My point is merely that if the other side of the equation—the demand side—holds up, all these things will be forgotten in 6 months. If there’s a problem on the demand side, on the growth side, then all these things will be straws in the wind of a very significant change. That’s my point.

When you zoom out to what matters, if the demand for AI is anything like some of the bulls would say, then let’s talk about the CoreWeave bonds, because everyone’s like, “Oh my God, the bonds are gonna bust.” If the demand for AI is as strong as people think, then everyone’s gonna need every damn computer, GPU, and data center they want, and those bonds will be money good because someone will buy the damn company. There’s just no risk of that.

If the demand is not there, then the first person who’s gonna suffer is not gonna be Google or Amazon, who have a big-ass balance sheet. It’s gonna be guys like CoreWeave, and that’s where the problem will come. That’s why there’s some risk in it. So that’s all that’s going on.

You mentioned price discovery. It was a tough week for Alphabet or Google, falling about 12%, its most significant drop in a decade. Google Cloud doubled to an insane number. It doubled to $130 billion, growing 103%, but search did slow to 12%. Free cash flow was negative, unusual when looking at Google. Is this just another element of price discovery?

Again, on the search side, it’s been true for 2 or 3 years that there’s probably going to be some element of cannibalization of search. It still grew pretty nicely. I think it grew 12% or so on search.

I think it was a combination of search angst, general angst, and CapEx angst. Nothing more than that. It’s been so funny to watch the last 2 or 3 years as we’ve gone from “Google is dumb as rocks” to “Satya is a genius.”

Then you fast-forward: “Oh my God, Google is the most amazing company ever. Poor dumb Satya, he missed the boat.” And now it’s switched back to, “Oh, maybe the things we’re worried about with Google are still true,” which is search cannibalization, exactly. But to be fair, to your point, their growth is there. These guys are saying they can sell compute as fast as they can make it, and that’s their number one thing. That’s where the growth is coming from.

I just think we’re in a nervous period. We’re understandably in a nervous period. I think it’s at a $120 billion run rate, growing at 100%, right? That’s about as good as it gets for the folks in the cloud. The cloud group should feel pretty proud of themselves.

I just think when any metric comes up short—Shopify just fell 15% as we’re discussing this because they came up a smidge short, even though they raised guidance for the next quarter—everyone’s just nervous as hell. I think it’s appropriate. I actually think the Anthropic IPO is going to suffer for it. There’s evidence of substantial pricing pressure. And as dominant as they are, there’s evidence that they might not be quite as dominant as they appear. I just think it’s okay. There are going to be these squiggles as most things still hopefully go up and to the right, until we can distribute out to our LPs.

What do you think Anthropic’s going to suffer for? You mean the consumer excitement won’t be there? You think they won’t have the pop they could have? What does that mean?

I think folks are just going to be nervous about a deceleration in quarter-over-quarter growth, and they’re going to be nervous about pricing pressure. It’s hard to process everything out of the S-1/A, the quarterly amendment, but it sure appears that OpenAI is basically running around selling somewhat equivalent products for 30% off. We can discuss how much CISOs want China-based open-weight models there, but they’re putting pricing pressure on. Jeff is putting pricing pressure on, and the latest models are effectively cheaper.

You have to have so much momentum to overcome a certain amount of pricing pressure and deflation. I just think the numbers suggest some nervousness, right? Quarter-over-quarter growth is slowing, and what does that mean? How much nervousness does it create?

It’s worth pointing out the numbers in very precise terms, because the funny thing is, even after we point them out, people can still disagree about what they mean. But at least let’s get the rough-and-tough numbers. In Q1, and I’m using GAAP numbers here, not run rate, because GAAP numbers, as I say, are audited, and the CFO will go to prison if they’re wrong, so I always assume they’re correct. Roughly $5 billion in Q1, roughly $11 billion in Q2, roughly $17 billion in Q3. Those are the numbers, objective fact.

What it means from Q1 to Q2—you can look at it as an absolute number. They went up $6 billion, and they went up probably a little under that, about $5 billion, from Q2 to Q3. On an absolute-number basis, that’s one perspective. You can also look at it as a percentage. They were at $5 or $6 billion in Q1, then they went to $11 billion. They doubled Q1, so their quarter-over-quarter growth rate was 100%, which is an astronomical year-on-year growth rate.

Then they added exactly the same number in Q3, but by definition, because they’re starting at $11 billion, they’re adding the same absolute number of ARR on double the revenue size. The quarter-over-quarter growth rate has, quote-unquote, “plummeted” to roughly 50%, because they started at $11 billion and added roughly $6 billion. I’m just using rough numbers. They’re growing roughly 50% quarter over quarter, which is still 5X year-on-year if you compound it.

If you buy into percentage-based growth, you could say, “Hey, they grew 100% quarter over quarter in Q1. They grew 50% from Q2 to Q3.” If they stick at 50%, that’s still a 5X year-on-year growth rate. That’s the positive spin. The negative spin is, “Oh, they were growing at 100%, and only a quarter later they’re growing at 50%.”

Those are equally true statements of the same facts. As Jason said, another way to think of it is that, at the margin, they added a little high fives, early sixes—I can’t remember which—from Q1 to Q2, and the absolute number they added from Q2 to Q3 was slightly less than that on a GAAP basis. The most negative interpretation, which I’m not saying I agree with, is that you’re adding roughly $20 billion a quarter of revenue, in which case you’re slowing out pretty quickly.

You can throw out those numbers, and I’m sorry that was a wall of numbers. You can squint one way and say, “They’re still growing at 50% quarter over quarter, which is 5X year on year. That’s freaking amazing.” You can squint another way and say, “Oh my God, it’s decelerating from 100% quarter over quarter to 50%. Does it go to 25% next quarter?” All the money is in that bet. That’s the only discussion in the bank.

The way—listen, Anthropic obviously isn’t a startup at this point, right? It’s not even a scale-up. But if I, as a startup, saw someone that did $6 million last quarter and only did $5.4 million this quarter, I would not invest.

Hang on, let’s be clear. I understand. I’m just going to be very clear: it’s not that they did $6 million and less. They added new ARR.

But they, their bookings, I mean, bookings decelerated.

Yes, the rate of bookings growth deaccelerated. The absolute number of bookings went up. I mean, they sold more shit in Q3 than Q2 by about $5 billion.

What’s actually interesting, though, is that given the bookings decelerated, one question is whether they’re reaching TAM levels in enterprise. They’re at 0.1% of Fortune 500 spend, so there’s still so much more budget for them to consume, and the bookings still decelerated.

Yeah, that’s why I think the only—listen, there’s not enough data in here, or maybe I’m not smart enough to parse it. To me, the answer is simply pricing pressure. The demand is there. If you look at OpenRouter and others, the demand does not appear to have declined. I don’t see any metrics that say penetration of the Fortune 500 has reached 100%, but you’ve got a lot of pricing pressure everywhere.

Whether it’s a combination of folks buying from other vendors because it’s cheaper or implicit discounts they’ve added in Haiku 5.5 and others, maybe we’ve reached the point where Anthropic’s ability to command a price premium has hit a local maximum. That’s one potential thesis. I’m sure if we saw OpenAI’s numbers, they’d look much worse if they’re literally offering a 30% discount on Astra, with what appears to be a higher cost basis. They’re offering a 30% discount, so lucky they’re not going public second, right? But it appears to be working.

Thirty percent off is a big deal. It approaches the 40% that we target for open-weight models. This is my thesis: pricing pressure hit them in Q3. That’s the only thing I can point to, because everything else is going pretty darn up and to the right.

Yeah, probably true. You’re right, Jason. It’s worth stepping back and reminding ourselves that, on a GAAP basis, they had $4.5 billion in GAAP revenue last year, full year, exiting at a run rate of $9 billion, and they’re probably going to do $56 billion in GAAP revenue this year. Year on year, their growth rate is 12X. That’s just an objective fact at this point.

But the interesting thing, to your point, is that you look at the same kind of growth rates and come up with roughly $120 billion of GAAP revenue next year. They’re probably going to go from a 12X growth rate to a 2X growth rate in a year. To your point, it’s a pretty sustained slowdown to still magnificent numbers. It really is glass half full, glass half empty.

You squint one way and say, “Oh my God, this is a company doing $56 billion in revenue that can still roughly double next year. That’s freaking amazing.” You squint the other way and go, “Oh my God, it went from 12X growth to 2X growth. Is it going to go to 50% growth or 5%, 10% growth the following year?”

Do you remember when we had Stewart Butterfield in Slack and zero to 10 in 18 months was the gold standard? The gold standard.

You know what’s a little bit interesting? I know it’s just an anecdote, and people will make fun of me, but I’m trying to launch this app, Saster Connect. Hopefully it goes out in the next couple of weeks, but I’ve been working on it for a while. It’s an AI-matching recruiting app. When I switched it myself to Haiku 5.5 and Jev, the total cost went down about 70%, 72%.

There are some costs that are going to keep going up. We’re all going to run five agents a day, 10 agents a day. We’re going to run them on multiple platforms. I just upgraded to Muse Paid yesterday, right? There are trends going the other way, as we see in the Anthropic quarter. But if a lot of mature applications can see that kind of deflation from improved quality of tokens and pricing pressure, that’s a lot, even though everyone’s going to be coding 72 hours a day with 20 agents.

The decrease in effective price is material, and it’s faster. It’s faster, 72% cheaper, and better. That hadn’t happened yet, at least in my personal agentic experience since the beginning of 2015. On an as-used basis, nothing got cheaper. No matter what anybody said about evals and token prices, nothing got cheaper until Haiku 5.5 plus Jev [?]. Then I saw a 72% decline for a B2B workflow running constantly.

For the same task.

Yeah, for the same task, and higher quality. People can make fun of me, but it is interesting. For me, it is interesting for that to actually happen in the real world for the first time.

Agreed.

It just makes this hard to predict for me. It makes it hard to predict how it’s all going to play out. Twelve months ago, none of these models barely worked. Today, they’re epic. In Black October they’ve become weapons to destroy Binance and everybody else. It’s hard to know how these sub-metrics are going to play out in 12 months. I think it’s almost unpredictable.

But I think that exact statement, Jason, is why we have unpredictability, or the binary nature, in venture markets. It’s why we have volatility in public markets, which I think is just a lack of certainty around what the 5-year outlook looks like. I think the whole world does not know what 5 years out looks like in a way that we generally have for the last 30 or 40 years.

I’ll tell you, not to share too many other personal stories, but comparing it to the show, this morning I finally churned off Canva after 12 years. I don’t need it. At the start of this year, Canva was still accelerating, you know? There were some rough reports this week about its current growth. They kind of varied. I don’t know what’s true. One said growth was zero; others said they were preparing layoffs. Who knows what it was? It at least seems to have decelerated 20%. It’s that predictability, right? You just don’t need it. The agents don’t need it, so it’s hard to predict one year out, let alone 5.

Jason, what happens to Canva? Does it go out?

I don’t want to be a Debbie Downer, right? The Debbie Downer case is obvious: they missed their window.

But then what? They miss their window and they’re just a private company that grows slower? Is that it?

I don’t know. It seems to be working for Automattic, outside of a few CEO turnovers and temper tantrums.

I want to be clear: was that humor?

I don’t think it’s humor. What I mean is, I think Automattic will stay private forever. I don’t think it will... It has raised its billion, right, as 40% of the web. I don’t think Automattic will become an agentic company, and I think Matt is fine with it. I think it will stay private. I genuinely think it’ll stay private forever. I don’t know what the VCs do with it, right? But I don’t think it’ll ever go public, and I don’t think it’ll ever get acquired.

The only comment I’ll make is that software companies in particular aren’t moving forward. It’s like the shark: you’re moving, or you’re going down, right? It might take an extended period of time, but you’re going down. I remember there used to be mainframe companies in the ’90s and early 2000s that were still living off their 20-year-old contracts, but the asset just becomes very terminal-valued.

So yes, you can stay private and continually cut costs and live off your install base, or do your own little Bending Spoons on your own, but it’s a tough place to be, and it’s hard to get value from it as an investor. It would be interesting to see what many of these companies decide to do. Can you clamber aboard the AI train, even if it’s a little late, and get enough traction to maybe not be the winner, but grow 5%, 10%, 20%? I don’t know. But if you don’t do that, if you decide not to play that game, I think over the medium term you’re just heading to quiet obsolescence.

I’m really picking a fight here, because Cliff always messages me. He loves the show and always messages me, like, “You guys are harsh on me.” And I’m like, “You know, Rory and Jason—I never say anything. I’m just—”

I didn’t say anything harsh. The surprising thing from all this, guys, is that obviously, if you’re doing what we’re doing, you have 21 agents running agentically, and they’re using Higgsfield and Opus and the other things, it makes sense you’d churn off Canva, right? Your agents don’t log into it, and it’s not really useful for them.

The surprising thing is that the real world isn’t buying much of this stuff anymore, right? If I’d thought Canva, Squarespace—which just announced 20% layoffs—and Wix, which is going to be bailed out by Base44, and others, and Shopify, would be doing fine because most of the world isn’t running 21 agents and Higgsfield and Musefine and everything, we’re not all tech nerds. But it isn’t really playing out that way. Obviously, some of this is because ChatGPT is so good for images. There are things that are so good. Groq is great for images, but I’m surprised that a lot of these core businesses are being so severely impacted. Maybe it’s a limit of my knowledge, and it worries me.

Instinct’s amazing for images.

Instinct’s great, but in your news, Elon just had to cut back GroqBot use because it’s too expensive. I think I underestimated what the cost of Instinct was because I didn’t understand how persistent the VM was going to be. I think Instinct could cost as much as $20 a month per user to service Instinct the way it’s supposed to work today.

Way more.

More than $20?

Yeah.

Then it’s a moronic business model at $40 or $50 per user per month. I got it wrong because the Lovable-Replit thing, which is a highly tuned, on-off VM, is about $4 to $5 a month, okay? They have large teams tuning that so you can have a little bit of AI. The VMs spool up and down very carefully. They’ve gotten it down to $4 to $5 a month. So I assumed Muse and friends could be cheaper. I assumed Muse could even be cheaper, right? And Manus might be twice that because they don’t have the resources. But it’s much higher to have these persistent agents running. It’s much higher. Do you think it’s more than $20 a month? I know you wanted to do it at $30, Harry, but how does this math pencil out at $30 or $40 a month per user? All these tokens we’re going to buy?

I admit that it seems relatively paradoxical to believe it’s more than $20 a month and to want to do it. That’s a very good argument from you. But for power users, it’s very significant, and I think for people who use it once or twice a day with limited automations, it’s very, very cheap. So I think your blended cost is probably much cheaper. But for me, I’ll be a terrible economic profile for them. The worst.

Look, I think the real truth is that, like every digital product, it always takes off at a point in time when the economics don’t quite work. I remember when Facebook was ramping up early on and people were like, “Oh my God, the server bill.” It was using scads of servers, and it was maybe going to have CapEx of $100 million. People were like, “That’s the end of the world.” The bet you take over and over again is that digital goods get reduced in price 50% or 60% year on year. If people like the shit and they get hooked on it, over 2 or 3 years the math gets better, right? That’s really what’s going on here.

YouTube is one of the best businesses in the world.

Or maybe we’re going to see these waves where, for all but a handful of folks, the subsidies just can’t last. We’ve also seen this in AI. The subsidies just can’t last. It’s too expensive, right? I don’t know whether Instinct is on that side or not. Cursor barely pulled it out. Cursor V2 would’ve died reselling Anthropic tokens at a loss, right?

If you’re going to regard Cursor as a failure, then yeah, you can win the argument. I’d take the $60 billion and call it a win myself.

I said the V2 version of Cursor would’ve died just selling Anthropic tokens at a loss perpetually. If that’s all they were able to do, they would’ve died. Great teams figure it out. I get the venture thing. I wouldn’t mind being Saragoh and investing in Manus at 50 or 100. I wouldn’t be too bent out of shape about the business model if I invested at $50 post. Don’t get me wrong. It wouldn’t bother me a lick. But I just don’t know which side. Amjad from Replit said yesterday that the era of these free subsidies for the agents is coming to an end. If Elon can afford it, Muse can afford it for now, but what’s the scale of the Muse subsidy? This is the second-fastest-growing product they’ve ever had. If it’s $10 or $15 a month for them, who knows? Maybe Amjad’s right. Maybe it’ll happen before they can raise that $6 billion at 30.

Just for context here, if it’s $10 a month per user and they’re subsidizing, that’s, call it $100 a year, right? For every million users, that’s $100 million.

This is a guy who’s spending $70 billion on CapEx. He ain’t gonna blink because the compute bill goes up by a couple of billion dollars, right? And I do have—

Yes, Meta can afford it. Meta can afford it, right?

No, I want to go back to your comment, because it’s interesting. I tend to be the boring guy who says the numbers have to work in the end: gross margin, Cursor, blah, blah, blah. But when things are moving so quickly, you can’t wait until the economics work perfectly to launch a product, because someone else armed with venture dollars will launch that product 12 months earlier, when the economics don’t quite work, grab all the mindshare, and then lean into the cost curve and hope the improving economics bail them out. That’s the bet you’re taking.

To take the Cursor example, they’re doing some version of, “Hey, look, if you don’t get a big acquisition for $60 billion, well, if what you said earlier is right, and I can get my own model and my costs come down, I make out like a bandit. If I’m still stuck on Anthropic in 2027, I’ll probably get fucked over.” But it’s still a better bet than waiting until 2027 and saying, “Oh, now that there are lots of open-source alternatives for state-of-the-art coding models, I should launch a coding company.” It doesn’t matter; it’s done, right?

The truth is, the capital markets and the venture markets are just so aggressive that waiting until everything makes sense is not an option. So, by the reciprocal property, almost everything that’s new and digital will look crazy for a while in terms of economics. The only question is, can the economics get better before the money runs out?

Look, no question, and we can move on, but it’s interesting that Elon had to cap it.

It is.

They have resources. They had a pretty good IPO. He has resources. He’s going long. He has bet his company on AI and subsidizing this thing. He subsidizes everything round trip, right? Even with his personal fortune, he subsidizes everything. This is a man who wouldn’t want to cap Grok bot if he didn’t have to.

He’s also the man who’s getting so many dollars per unit of compute from Anthropic that he’s like, “Yeah, I could do some free Grok bots, and these people could make cute cartoons and send pictures of sexy anime girls to each other. On the other hand, I could get, like, $50 an hour from Anthropic. Screw it, I’ll do that.”

Well, they’re not doing that. They’re building applications; they’re building code. Grok bot is fairly technical, unlike the instinctive Muse, right?

Okay, boys, we had another pretty big transaction that went down which, in a normal world, would’ve dominated the news. In today’s news, it kind of lasts through a couple of hours, but DoorDash buys Instacart for $17 billion, under half its 2021 private price. D1 still owns 9.7%; Sequoia owns 8.1%. Before we talk about agents, which was a disclosure, the thing of note here is an inevitable ending for a clear second-place provider to DoorDash.

I mean, “inevitable” is a bit strong, but yes, they clearly tried to do it before. It clearly makes sense. There’s also probably an element of, “Do it now while you can,” if you’re DoorDash. It all makes sense.

Listen, some of this is not my strength, but the fact that the Postmates acquisition worked was a surprise to me. I assume when Uber Eats bought Postmates, you get a little bit of a brand and a list, and you basically shut the thing down. I didn’t really believe these deals would be accretive, but they appear to be.

If Instacart has 20% of the market and DoorDash has 60% of the delivery market, and you can get through antitrust—which you probably shouldn’t, if you think about the numbers I just said—and you can get to 80% market share, this is how you dominate networks. You dominate every part of it by having 80% market share. I’m surprised you can combine these networks and they don’t end up with one plus one equaling 1.1 a couple of years later. That’s my experience, but it doesn’t appear to be true.

These deals, if well architected, appear to work, so it should be accretive. It should make sense. And I think you noted it’s half its ’21 private price or whatever, right? You may ask what the similarity is with the Braze deal at Salesforce. A lot of times, people have been circling each other for years. Tony said they’ve been talking for years, right? There are deals that don’t make sense even if they’re cheaper, and then there are deals where, like, Salesforce bought Braze. Salesforce was losing to Braze in marketing for years, and when it got super cheap—when it got to 3x ARR—they had to pay a big premium. They just buy the assets they want when they become cheap, and it may be that simple. If they’re accretive, buy them. This may just be simple things that corporate development is supposed to do.

One thing I found fascinating was related to a conversation we’ve had before: agentic usage of these platforms. DoorDash said AI agents have sent only several hundred orders against 3 million deliveries a day. Were you shocked by how surprisingly low that was? Or do you think it’s symbolic of the fact that people will always want to choose food and have that discovery process?

I think in a sense it was a dumb comment, for what it’s worth. It reminds me a little of when Brian Chesky, a year and a half ago, said there’s no need for AI for rentals and housing, or when Parker Conrad, another generational CEO, said there’s no need for AI in B2B 18 months ago. Sure, today there are only several hundred orders, but these platforms, like all app platforms, have a risk of slight disruption from agents. They all have a risk of slight disruption.

Which is why I think it’s so self-serving. I would be pushing this whether you’re Uber or not. When I had at Mac, the president of Uber, on, he was very clearly most worried about agentic usage of these platforms because it would disintermediate Uber. You could get FREE NOW, you could get Lyft, you could get price-optimized, time-optimized, and who gives a shit about the platform provider? You have it for Vinted, you have it for Uber, you have it for DoorDash. I think it’s very specifically well done because it made headlines and news in a way that he actually wants. No, humans are here, and we will continue to be here.

Maybe the better way to say it is, let’s assume they’re telling the truth. That’s the objective fact about what’s going on now, and the million-dollar question is, do you see that behavior happen more and more? And then how does DoorDash respond to it? They’re all doing versions of the same thing. I could totally see that for some of your standard food orders, it’s just not that much easier, for the record. Telling your agent to order your dinner is just as easy as clicking on the DoorDash app, but—

Well, DoorDash has two big competitors. It has Uber Eats, right? And it has the restaurant. The restaurants don’t want to use DoorDash or Uber. Most restaurants have self-pickup, at least to-go pickup, where they can avoid most of the fees, or self-delivery. An agent can figure that out for you. “Hey, Harry, did you know that Lemkin’s Curry down the street will actually deliver for $2, and you don’t need to go through DoorDash or whatever they have in London?” I did not know. Thank you, Muse. Thank you, Instinct. I’ll take it. That’ll save me $14 on my order? What’s the pub right behind your office that I really like, Harry?

The Grazing Goat.

Love that place. Five out of five, okay? Maybe they could just drop it off at the next podcast for free. I don’t have to pay 20%, plus tip, plus this, plus that. You think I’m kidding. I think agents are an existential risk to DoorDash because their number three competitor is nobody.

Can I point out a cultural difference between us? You guys pay 15 bucks for a burger, and it gets delivered at 50. There’s a federal charge, state charge, delivery charge, service charge. It’s ridiculous. In London, it’s kind of like 15 quid is 15 quid. It’s fascinating to see the difference in what you pay and what you get.

Yes, and that’s why these businesses are wildly profitable. They figured out exactly how to extract every dollar of rent from the person who wants to sit on their couch, press a button, and get their meal an hour later. It works.

It’s been a brutal business since the old Grubhub days. It’s just evolved with the internet, and agents will evolve it again. It’s been brutal. It’s been brutal. The idea is to take as much margin out of the customers or the restaurants as humanly possible.

I’m with Bill Ackman. I think Uber’s a banger of a buy right now, though. I mean, it’s so cheap.

I broadly agree, and I haven’t spent a lot of time on this, but the two worries you have are way bigger disruptions than agentic disruption. And what you’re implicitly saying is, it’s borderline fine.

Yeah. Agentic disruption is bizarre in taxi hailing. You’re going to go to your WhatsApp and say, “I would like to get an—

I agree. I don’t think that’s—

It’s a worse user experience. “I would like to get an Uber. I am at this position on the map.”

Yep. Agreed.

And then the other issue is people love food.

They love looking at it. They love pictures, scrolling, and discovery. The only thing it makes sense for is automation and scheduling, like, “Every day, drop it off at Scale’s office in San Francisco.” That’s it.

Okay, boys, we have 2 big rounds in Kalshi and Cohere. Kalshi: $1 billion at $40 billion, one of Jason’s favorites, a low-dilution round there. With 2.5% dilution, but still raising $1 billion at a $40 billion price a day after Utah charges the founders of Kalshi. And then we have Cohere. I think it’s the largest private raise for a Canadian company at this point, with the Canadian government, I believe, or Canada itself, investing up to $1.14 billion. Which one’s more interesting?

Kalshi is a money deal. Let’s be really clear here. The thing is working. It’s growing like a weed. Deep in their heart, everyone knows that we started off thinking we were doing prediction markets, which had some social utility, and we’ve long since converted into sports gambling markets to feed people’s addiction to gambling, right? So we’re all over the doing good. We’re all over the higher purpose. We’re pushing through a whole bunch of local regulation, which is why Utah is suing them, because sports betting is regulated on a state-by-state basis, and these guys are plowing through on a national basis. So there’s nothing to see in this other than VCs will put a lot of money into shit that’s working. That’s the takeaway, right?

If it’s at $40 billion, it’s had momentum. It raised at $22 billion like 3 or 4 months ago. So this is one of those classic momentum plays, right? Contrast that with Cohere. Cohere’s very different. It’s almost not a momentum play. It’s a, “We need our sovereign AI play in Canada because you are nasty people, and you’ve threatened to invade us multiple times, and we’ve got this product.” It’s one of the original 4 or 5 LLMs, to be fair. Back in 2023, you had Cohere, obviously OpenAI, Anthropic, and then all the others, like the one that Microsoft bought, Adept, that Amazon bought. All the others folded.

Give Cohere huge credit. They are one of the survivors of that first group of LLMs, before Neo Labs was even a thing. They didn’t release revenue numbers, but you can see how you can make a sovereign AI story. It’s almost the exact opposite. In the case of Kalshi, the government is saying, “I want to stop you now and put you in prison.” In the case of Cohere, the government is saying, “I’m going to make this work no matter what, even if the economics don’t work.” So they’re almost like the yin and yang of investments in terms of their relationship with the government.

Does that work? Can the government just do that?

Do what?

“Well, hey, we want to make you work, and so we’re going to inject you with $1.164 billion.”

Isn’t that what our governments do all the time? They subsidize everything.

Yeah, you can do it once, you can do it twice, but ultimately, if enterprise adoption isn’t there, it doesn’t work.

Sometimes it’s worth just being in the game. It’s like putting a few chips into Anthropic in the Series D round. Sometimes it’s worth just being in the game, and sometimes it bounces your way, and you’ve always got the liquidation preference.

Jason, the D round was $16 billion, so I’d take the D round. Yeah.

In all seriousness, I don’t know how Canadian national governments work. When you talk to folks who run large corp dev portfolios at top public companies, they really just want it not to go down in the aggregate. They love the gains, right? But as a public company, you only get so much credit for the gains from your portfolio, and the losses really suck, because they swing your EPS in really weird ways, and you have to put asterisks and daggers next to it. So as long as they don’t lose too much money on the investment, I don’t know that they care.

I think there are 2 separate things. You’re right: the motivation for anyone doing this isn’t purely financial gain if you’re the government. Agreed. But Howie asked a separate question: can it work? I think the answer is, yeah, you can keep it going. Governments can make things happen for a fair while, but what they can’t do is make the private sector use the product.

You see it with Mistral, too. To the extent that there’s a willingness to spend Canadian government tax dollars on a Canadian government AI solution—or maybe if Canada goes ahead with affiliating with the EU—these guys and Mistral may be the “you’re-not-America” LLM options, and there’ll probably be some spend in that. But do you think it’ll be a compelling business that will go public at 3X this round? That’s not at all obvious to me.

In today’s world, today’s day and age, Cohere not disclosing revenue is hardly a flex.

No.

It’s a pretty bad sign.

Dude, I love it. I love it. It’s like, “ARR grew more than 165%,” and you’re like—

Since inception?

Yeah, no.

You know, one thing I was thinking about on the Kalshi thing before we leave: Sequoia led the last round, right, at $1 billion. They put in $1 billion right after felony charges were brought against the founders. I get that it’s just a state. And then we have Binance, right? $6.6 billion taken, apparently by North Korea, in a matter of hours. I had to wonder, between Sequoia and Binance, if the ball had bounced just a little differently, would FTX still be around today, and would it still be a winner? Same VC, same risks.

Listen, you obviously can’t take money from one sister company and put it in another, but I think Binance barely survived this $6.6 billion. They don’t have a transparent balance sheet. We don’t know how much money they had. What if it was $13 billion? Where would they have repaid it? I think they were one step away from an FTX-type situation. Luckily, CZ got pardoned, right? But it made me think: since you have the same investor and the felony, we’ve all just changed the way we take risks because it’s game on, right? It made me think about a parallel universe where FTX was a shining portfolio company today.

Do you know what? The fascinating thing is FTX’s $500 million position in Anthropic, which debuted—

Well, that we know. Yeah.

—would be worth $39 billion at the $1.5 trillion valuation. So timing is important, is the takeaway, kids.

The most compelling factor about the Binance thing was that they could pony up $6 billion to replace the money that was taken. What a good business, assuming that you’re not doing an FTX and swiping it from someone else. As I understand it, and I did not spend a lot of time on this one, Binance were able to make good the crypto loss. That implicitly means Binance makes enough money to be able to make good the crypto loss, which means it’s an amazing freaking business.

I think so. I just think, and again, I’m not a crypto guy, but when CZ said they could cover it but it would be tough, that suggested to me they didn’t have $13 billion. What I really thought was ridiculous is how many folks said their balance sheet was all the crypto they hold as a trustee, as a custodian. That’s not true. That’s why Sam’s in jail and CZ is out, right?

Agreed.

I don’t think they had $13 billion. I think they got out in Black October.

Yeah, you can’t take my crypto dollars and give them to Howie, and say, “Sorry, your stuff was stolen.” Agreed. Okay.

Yeah, their balance sheet. Just think how much still happened in Black October, right? And then we have this one that I think is craziest. I don’t mean to get off track, but this Ardex thing that breached WorkOS and a bunch of others. This is just since we did the last one. This Ardex thing is insane.

I saw Michael Grinich post about this 16 hours ago, but tell everyone what this is.

Ardex is something we should be scared about. It’s initially well-regarded. It’s only a couple of months old, a Chinese open-source pen-test solution for penetration-testing applications. We all need pen testing. We always needed it before. We certainly need it in an agentic era where everyone, from Binance to the rest of the world, is being breached. We’re all being breached. All the refrigerators in Europe are going down because they’ve been breached.

Black October was much blacker than we made it look. Some really smart developers just a couple of months ago in China built an open-source tool. Great. It’s free, and you could hook it up to DeepSeek and your other LLMs, and it will just try to hack you. This is what you’re supposed to do: hack yourself. But it became very easy to turn it around and attack other folks with agents. It got flipped around, and it’s become one of the greatest threats. A lot of folks are saying, “Well, you could always do this,” but now you can do it with LLMs at scale. You can just tune it against anybody you want.

So they originally built it as a defensive tool. Yeah, if everyone said, “Oh my God, this is a pen-testing tool,” that makes total sense. It’s basically like, “Guns don’t kill people. Bad people with guns kill people.” You built a gun, and you hoped it was in good people’s hands, and now it’s in bad people’s hands, and they’re killing people.

It worked great at first, right? Until you turn the gun around.

It’s probably hard to overstate the cyber dynamics of the next 12 months. You’re right. It’s one of those stories you never cover on a daily basis, but at some point something really weird is going to happen, and you’re going to go, “Come on, guys, we should’ve seen that.” The entire US Fedwire goes down for a day, and we’re like, “We can’t move money.”

Yeah. We joked about the refrigerators in Europe, which, for people who don’t know, were hacked so someone could turn up the temperature on the refrigerators to 50 degrees. Luckily, it was rectified, and the frozen peas were saved, which is a great relief to everyone in Europe. But you could see this happening with much more mission-critical things, like traffic lights or heating facilities, or you name it. So it is quite alarming. Ding, ding, ding. Mr. Lemkin, we’ve had quite a few rounds, and it’s your turn to bring something to the IC this week.

Goodness. Can I take a break for the rest of the year? We’ve been busy. We’ve deployed so much.

We spent all the money.

We’ve gone back for three funds this year.

No, no, no, we’re Andreessen. We just raised a big-ass fund, so we’ve got plenty more for you to go through. Jason, we’re a deployment machine.

Okay, great.

Which company are you going to take to IC this week? Serval? Cognition? Baseten?

Well, listen, Harry, as you know, much as the Windsurf team lamented having to sell out to Google for just a few billion, the Cursor team’s going to lament selling to SpaceX for only $60 billion, right? Cognition at $80 billion seems market-correct. To get to a 10X with dilution, a trillion-dollar valuation, I start to get nervous with the latest rounds from OpenAI and Anthropic. I do worry that fundamentally it only owns so much of its own IP.

But I don’t know if you saw Anthropic’s last Q3. The one thing that’s working, that never stops working, is coding agents, and Cognition has a lot of the field open to it. They made a lot of their own luck. I don’t know if the original version of Devin, when I brought it to committee a year and a half ago, worked. They lucked out buying the Windsurf team. They lucked out buying IP. They lucked out with open weights.

But this is something with the kind of momentum that I think we can’t sit out. We have a little bit of money left in the fund, a couple billion, and we need to deploy the next one. I say we close out the last fund with $2 billion in Cognition. We’re already up 5X on paper. We can’t lose, so we might as well put the last $2 billion in Cognition. And if we get a little recycling before the round closes, we toss it in too.

Interesting. Just for the people listening, that was a much better question than you normally ask, Harry. Normally you say, “Here’s a deal, Jason. Tell me why you should take it to committee.” Honestly, and respectfully, I say that’s kind of stupid, because all you’re saying is, “Give me a sound bite,” right?

The reason this was a good question, and much more interesting, was because you have three deals in front of us, and it’s worth pointing out to the listeners what they were. Three really amazing companies. You were basically asking Jason to pick between Baseten at $30 billion, up 2X from June, an inference company; Cognition at $70 billion to $80 billion, up from $48 billion in September; and Serval at $3 billion, up from $1 billion in December.

Three different companies, very different businesses. One’s an inference provider, one’s a coding agent, and one’s a very hot next-generation AI app in the ITSM space, right? So three different businesses, all of which have had a round and are getting a second re-up within this 12-month period. As I say, because it forces Jason to choose, that’s a much more interesting question than just, “Tell me nice things about this company,” if you know what I mean.

So, for the editors, can we please cut that segment? That’s the first time Rory has actually congratulated me on a framework.

It’s okay, because I think it happened by accident.

I’m going to play that to myself later just to give myself more confidence.

Got it. But genuinely, it really was a super interesting question. Would you like to do an inference provider at $30 billion, a coding agent at $80 billion, or an AI app in ITSM at $3 billion? Good question.

No shit.

Which would you pick?

Serval. Are you kidding me?

Okay, so you think your upside in Serval at $3 billion is better than your upside in Cognition at $70 billion?

Oh my God, the tide is changing. ServiceNow is old guard. Bill McDermott is amazing, but he’s like Slootman. He’s a sales-led CEO. We’re in a product-led generation. Jake and Tatiana at Serval are incredible. The product’s amazing. The enterprise adoption is better than I’ve ever seen.

And the upside, just pure multiple on cash, is way better here. There’s an addendum and a star that we can go back to. The chance of getting a 10X here, I think, is very high. The chance of getting a 10X for Baseten at $30 billion—God, you’ve got to squint. The chance of getting a 10X at Cognition, to Jason’s point, gosh, we’re nearly hitting a trillion there. So, yeah, I think the multiple is clearly much more evident here.

The challenge we have with Serval that we don’t have with Cognition or Baseten is, if we’re an AUM play just putting cash out the door, they only raised $150 million. If I’m Andreessen, I want to put $500 million, a billion out the door. And so doth butter no parsnips, as we say in the UK. I need to do Cognition because I need to move money. But for the multiple, it’s the easiest decision you’ve made.

Interesting.

You know what? The question about Serval—I really don’t know it at all. I doubt the founders don’t know it. My question to the IC is: does it expand the TAM? Am I bounded, for real, not for hand-waving, by a subset of ServiceNow’s TAM today? Every founder will expand the markets and add concentric circles, but is this really expanding the TAM?

I generally don’t know, but I’m personally not excited about disrupting existing players if you don’t expand the TAM, because I think we don’t get anywhere at the end of the day.

I think the argument they’d make would be some version of: ServiceNow is just a list of things that IT has to do, and Serval is actually a product that does those things for you, so you can command more value from it. You’d validate that by looking at their ACV on a like-for-like basis. In other words, there’s a mid-market company that might spend $200,000 on ServiceNow. Would that same company be willing to spend $300,000 or $400,000 on Serval? That’s what you’d need to see to get TAM expansion.

I know one of my colleagues had talked to them in the past. This is an amazing company, an amazing team, and a great market. It’s always worth remembering that, until recently, ServiceNow was, I think, the second-largest market-cap SaaS company. You had Salesforce and then ServiceNow. ITSM is a much bigger market than you think because it kind of becomes the general corporate to-do list for service functions like legal and accounting. All the back-office folks were using ServiceNow to keep track of their tasks.

So it’s been a way bigger business than I would’ve guessed when we saw it, what was it, around 2010? A replacement for that is a big opportunity. But I hear you, Jason. Can you expand the TAM, or are you dealing with a 10-year grind-down of ServiceNow where you do it forever?

Whereas, Harry, to take the other side of the argument on Cognition, the beauty of the coding agents, especially with Cognition, is that they’re coming into large enterprises and basically saying, “We will help you effectively rewrite your entire code base and enable your team to do agentic coding,” and getting huge dollars for that. In terms of money velocity on the table right now, Cognition has the ability to command way more dollars.

I think Serval probably has the compounding effect of 10 or 15 years of grinding down ServiceNow to a bigger outcome. So I don’t know. It’s kind of velocity but capped versus great long-term megatrends, but is it going to take a while?

Look, it may well resolve tickets that ServiceNow can only track. It may be an epic play. I’m just trying to think it through. Serval aside, I think 2027 will be a year where we find out a lot of vertical AI plays are nifty little startups but didn’t, in the end, expand the TAM enough. They’re fighting for a slice of ServiceNow, a slice of Salesforce, a slice of Atlassian, but they’re not expanding the TAM. Personally, I don’t want to invest in any of them because I just view it as features in five years.

I don’t think that’s the case with Serval. I think they are expanding the TAM, and I’m sure they’d push back very vigorously on the vertical description. I think it’s very cross-functional and cross-vertical.

It’s an adjacent point I was trying to make.

Okay. Sorry.

But I get you. Yeah.

I’m not arguing.

I think in vertical, we’re going to see the biggest tears, because a lot of folks are going to make investments assuming that, especially for an SMB or an [ARVR?], you can charge three to four times the size. And we’re going to find out, you know what? You can charge 1X the size, and now there are five or eight times the competitors in the market, and the leader isn’t dead, right? They still have a pulse. They may not be growing at the rate they were before, so you’re just slicing up the same old effing TAM 11 times. I just don’t personally want to make those bets because it’s exhausting. It’s exhausting.

And then it’s been funny. As I said, that’s why it’s such an interesting, genuinely interesting question. I’m literally thinking of it as an investor here rather than just blathering on. We’ve kind of gone between the $80 billion one of Cognition and Serval, and Baseten in the middle. If you want revenue velocity right now, nothing gives you revenue velocity like inference. But from your little face, Harry, are you implying the angst there is just competition? There are five or six great companies in that space, and that’s a worry?

I would 100% invest in Baseten too, by the way.

But the whole point is there is no two. You’ve got to pick.

Well, yeah, but I would not choose it compared to Serval or Cognition. And I think Modal’s product suite is unbelievably strong. Eric is incredible for that, and Lynn and the team at Fireworks by far have the best inference product, I think. So I would choose Modal and Fireworks, and it’s super clear for me on Serval.

Okay. That was a fun question.

Rory, did I change your mind?

You made me think. No, a genuine comment. This made me think, right? These are three very different, interesting dynamics. It makes you think about what your bet is. Is it very horizontal inference, but with lots of competition and the risk of commoditization? Is it the last surviving standalone coding agent that’s cracked the enterprise, but at a very hefty price? Or is it an app that I think will be the—if there is a next-generation ServiceNow, the probability of it being Serval is probably 70% plus, maybe 80%. Those are very different bets, and I’m actually literally thinking. They’re all, for the record, pretty interesting bets.

The one nagging comment, and I don’t know if this is correct, is that the good thing about Serval is you can say, this is your point, Harry: the old-school company is worth—I can’t remember. I used to know the number—$40 or $50 billion.

It’s worth $110 billion.

So you can say there’s a thing that exists today that’s 30 times the value of this; therefore, I can have that money. Whereas for Baseten and Cognition, you have to say there’s no standalone coding agent worth $80 billion. So it’s good news that it’s a new market. The bad news is, what are you mentally taking over? You know what I mean? The same is true of inference.

No, no, the bad news is you’re fourth in the market. You’ve got Codex, Claude Code, Cursor, and then you’ve got them. Geez. Now, don’t get me wrong, I’m all for independence. I’m an investor in Factory as well, and I totally buy that, but oof, that versus Serval versus ServiceNow? That’s a bit of an easier one.

Okay.

So wild. Rory, you’ll be surprised to hear that my girlfriend is still with me. I know. My mother is shocked too. But in the legal world, whoa, mega news. One of the big law firms, one of the prestigious OG brands, Weil Gotshal, shut down very quickly and weirdly because of a partner dispute, whatever. AI law firm Crosby, which has raised $85 million from Sequoia and the like, bid to take over its assets and its brand. Anything to say here? I thought that was fascinating.

I mean, there are two separate stories. There was Weil Gotshal crashing, which is an American law story, not a tech story. You lose a couple of key partners. These firms, even though they don’t have a lot of liabilities other than leases, are just a bunch of connected rainmakers, and sometimes—

That’s a flimsy way of looking at it, no offense, Rory. A couple of partners leaving does not mean a 90-year prestigious law firm—

Yeah, but if you take enough of the book of business with you, it leaves.

It’s a small number of rainmakers. Absolutely. It’s not like a firm—

It’s worse than if you leave a VC firm. The VC firm dies, but over 20 years. If the top rainmakers leave, there’s nothing to pay out at the end of the year.

I mean, if you look through it, I’ve read all the accounts. I don’t know Weil as well. As I say, it’s four or five key lawyers, each of whom ran a key part of the business, and there seems to have been a clash involving a couple of them. I think the guy who ran all the PE practice, which is probably 20% or 30% of a company like this, left and took his book of business. It can go wrong pretty quickly. But that doesn’t matter, because that’s not a tech business. That’s just, “Wow, that’s interesting.” And I did admire Crosby for just saying, “Hey, if there’s some attention to be paid here, why don’t I put in a bid for the business?” That was a bit of a “fuck you,” but I was impressed by the chutzpah. I think that’s the relevant word. That was chutzpah.

I loved it. Ryan just said, “Yeah, we’ll take it.”

Yeah, I think that’s it: “If you don’t want this 100-year-old franchise, we’ll take it.”

I don’t mean to be too negative on the Crosby thing, but ultimately, if you’re providing slightly higher-value services, it’s going to have the same issues. People want to keep all the money; the partners, the rainmakers, want to keep all the money. Now, Crosby probably has no rainmakers, right? And it’s doing transactional stuff. But I don’t know that this idea that lawyers will give up a bunch of the profits they could make for shares in a company that hopefully will IPO is sustainable. I don’t know how sustainable the business model is.

For the record, I don’t know if I fully buy the model, but I disagree with you in this sense: I think it’s actually the same story. What Weil Gotshal says is there are some parts of law where the partner you’re dealing with as the customer is so important, because you value his or her personal advice, that when that partner leaves Weil and goes somewhere else, you take your business with them. That’s like, “I’m doing a complex PE deal. I need this guy. I’m paying him five grand an hour, some absurd sum, but he’s my guy and I trust him. I’m running KKR, and he’s the guy I want.”

What Crosby is saying is that there’s a whole bunch of other law where you really don’t care about the name behind the thing, because it’s still pretty standard and transactional. That business can be done with Crosby Law. I actually think they’re a very nice counterpoint. If I’m doing an NDA or a Series B stock purchase agreement, I don’t need the five-grand-an-hour guy. I don’t need to know the name of the guy doing it. It’s just, you know, do the usual stuff, check the usual things. So I think—

Yes, but even at the low end, incompetence hurts you. It damages you. Have you ever done a routine transaction where some idiot lawyer cost you a lot of money? I have. I’ve been damaged seven figures by incompetent junior lawyers.

But I think—sorry, I’m just going to push again because I’m grumpy. That’s exactly the Crosby argument: it’s less likely to have that kind of mistake if it’s automated, if you’ve got AI for the standard stuff.

Okay. Well, let me flip it around. Let’s imagine I go to Crosby, I make $170,000 a year, and I get 0.01% of the company for the IPO that I’m not even sure is going to happen. Then I can go to Weil and make $250,000. If I’m actually good, I can—every great partner now makes $10 million, and every good— The entry-level ticket at Weil was over $200,000 for an associate, for a soul-crushing, horrific job, mind you. I get that Crosby— I just… My only point is, and I’m not being critical; I haven’t talked to them, but I’m not convinced this is as stable as a lot of VCs might hope it would be. It’s too human-based.

Okay, I agree there are issues with the kind of human-plus-AI delivery model, and I have questions about that. But I’m just going to push back on your comment about why I’d go to Crosby. That’s easy. For every lawyer who wants to be a rainmaker, work 24/7, and make $10 million, there are lots of lawyers who are like, “I want to do 40 hours of good law. I don’t want to be a rainmaker. I don’t like that. I’m just a good lawyer, and I want to go home. Maybe I’m dealing with kids. Maybe I’m dealing with a parent. Maybe I just want to take it easy.”

There are lots of people for whom being a lawyer, earning a well-compensated, nice upper-middle-class salary, and working 40 hours a week with control of your own time for someone like Crosby is a perfectly good product. So I don’t have a problem; they’ll be able to get lawyers, right?

I just don’t know if I can see a $10 billion outcome here.

Different question. I’m sorry, I’m just keeping—yeah.

For the…

I don’t think I can take this to IC. It’s a lot that has to happen—

Agreed.

—for the corporate company, its corporate affiliate, to be worth $10 billion, because it has to suck up so many billions of dollars in profits that the lawyers could otherwise take home as partners.

I agree with that. I think that there will be a coterie of law firms five years from now focused on the kind of work Crosby does. Crosby will obviously be one of them. They’ll do really well. We like the space.

But I agree: the question is, if you look at the billings of Weil, Gotshal as an example of American law, and 70% of the dollars come from the high-rainmaker, irreplaceable-by-AI stuff, then all you’ve got left is the 30%. And then you have to halve the value of that to get the business, so you’re left with the lower-margin parts of the business.

I do agree. That’s the question: how big can it be? But can you build a business there? Absolutely, I think you can. The question about me—

Yeah, you can build a small business, for sure. It’s just that for our IC, $10 billion is the lowest possible exit where we’re allowed to bring it to IC.

Agreed. As far as I can see, $10 billion is—

It’s 2026, boys.

Yeah.

$10 billion? I thought—is $10 billion the round size?

Look, Rory and I are a little more conservative. We’re okay with a $10 billion exit as long as the entry is appropriate. We can’t make the numbers work for the whole fund, but we can make our own little slice of the fund work. We can hit our individual 6x goals with those exits.

Guys, what have I missed? Creandum, probably the most exciting fund Europe’s produced in the last three years, raises $750 million. Physical Intelligence is in talks to raise at a $20 billion valuation. They’ve also had rumored acquisition talks in the last six months. That’s with Avenir and Lux. Dropbox cuts 9% of staff, and the CTO leaves.

Unlike the last group of three, these things are literally a grab bag of shit. Physical Intelligence, a company with relatively little revenue, is worth $20 billion. Dropbox, a company with billions in revenue, isn’t worth a lot. It just shows that revenue multiples don’t get you anywhere, right? It’s all about the future. These are just non-comparable things.

Look, on the Physical Intelligence thing, I think robotics is going through an amazing explosion. We have a small investment in Generalist. I think these next-generation models are amazing when you see the demos. I’ve done a lot in robotics over 10-plus years. When you see what the new models can do, it is truly amazing. You go, “Wow, the level of dexterity—the old models just don’t have that.” I think it’s a real breakthrough moment.

Two other comments, though. I think deployment of robotics is just way longer than deployment of classic LLMs, and people forget that. Somebody’s starting to price on an hourly API basis, which makes sense. So you pay, maybe, if you’re replacing someone, three bucks an hour for access to the model.

There are only 300,000 robots in the US, so to make the math work at around $20 billion, which is where PI is, you’ve got to have a shit ton more robots. The technology is amazing. The business model has a lot of developing to do.

But I think the last sentence you said is the real one. I think all the big foundation model companies will buy one of these things, and I think the reason is—either they build one or buy one. It’s going to sound like I’ve said this before, but the really simplistic reason they do it is this: I’m writing my prospectus. There are 81 million knowledge workers in the US. We have an LLM, and we sell to them. It’s a great business.

There are 87 physical workers in America—not knowledge workers—working in blue-collar or physical-world jobs. Now we have a robotics foundation model, and we can address them. I’ve doubled my TAM. It’s a bit simplistic, but I think the lure of robotics for the foundation model companies, and the desire to say, “Oh, we do robotics, too,” means there could well be some acquisitions in the space.

It goes back to something Jason said. It’s one of those markets where, if you just try to be logical about the steps between here and deployment, you probably get caught up in the minutiae. I’m often guilty of that: it’s going to take a long time, customers are going to be slow, BMW isn’t going to roll out 100 humanoids in a week.

But if you say to yourself, “The technology capability is amazing, the teams are amazing, we love our position in Generalist, and this is going to change the world,” somehow you’ll make money out of it. That’s probably the investment thesis.

Jason, any final thoughts on this?

The only small thing—and this one is really minor, almost not worth bringing up, other than that it’s something I’m tracking—is a tiny amount of anxiety that Hugging Face has drawn some DOJ scrutiny. I don’t think there’s been a formal request for information or anything. As investors, it would be a bummer for us if M&A isn’t really back, right? As we discussed, we need $10 billion deals even to bring deals to IC. If this $12.9 billion deal for Hugging Face isn’t going to sail through antitrust, it would just be a bummer.

Politics aside, it just seemed like we were able to do normal M&A deals, with normal tax benefits and all that. If that doesn’t turn out to be the case, then we’ve got to go back to doing convoluted team-and-IP deals or slow-down deals. It’d be a bummer.

And look at the other thing: Salesforce is on an epic M&A roll. They’ve Hoovered up everybody this year, and I don’t know that they would do that if the antitrust environment—again, no politics allowed—was where it was 24 months ago. It’s good for venture that they’re just Hoovering things up. Who did they just buy this week? Cresta, right?

They just bought Cresta for $2.7 billion.

Yeah, bless Benioff and friends, right? We need these deals to happen every two weeks. For venture to be at least de-stressful, in my opinion, we need M&A to work like clockwork. Otherwise, we’re just stressed over these goddamn IPOs.

I worry a little bit about this because we want the Salesforces and the ServiceNows and Microsofts—we just want them buying companies every couple of days for a billion and up. We need it to liquidate our portfolios. If something happened here, I’d be bummed, but it’s just something to track if it gets anywhere. Salesforce says the opposite. They’re game on for M&A, man.

Dude, I agree. Bless the team there.

It’s smart, too. He’s buying his way into the future, into the present. There’s some point where it’s too much.

Is he with Cresta?

I think what’s happening, if you just look at all the deals that have been bought—I saw this back at Adobe, so I’m dating myself—but obviously there’s a little bit changing in the world with AI and B2B, right? Basically, everybody’s getting a punch card. They bought some defense startup for a couple hundred million bucks, I don’t even know. You go around the table with all your SVPs, and they come up with their best idea.

Whoever was in service or support or whatever, Cresta was their best idea. Whoever was trying to get these support and sales agents faster, Fin was their best idea. You just keep going down the best ideas, and then someone goes out and figures out if they’ll actually sell, because some folks won’t sell, and you just make the best idea happen.

People know their jobs are on the line, so none of these deals are Marc betting the future of the company. They’re SVP chips. I’ve been in the room when you went around the room: “Rory, what’s your best idea? Harry, what’s your best idea? Come back next week. I want your top three ideas.”

It costs you nothing, in a sense, because it’s assumed some will fail. Basically, they’re getting around the table and saying, “We’re going to spend this much of our market cap and cash to do these deals because we’ve got to do something.”

The interesting question is, do you do a Cresta, a more mature company that’s using AI, but whose original stack was probably pre-LLM, and that has probably transposed its stack since the LLMs came out in the last four or five years? In return, they have some scale and some enterprise deployments.

Or do you buy something new, bright, and shiny, where it’s all LLM-first, it’s two guys from YC in 2024, and they’ve got to $10 million and they’re exploding? That’s the interesting thing.

Well, the interesting thing for me is that a relatively similar price is being paid for them versus Listen, but you’re going to have a 10x increase in revenue for Cresta versus Listen. Listen was at $20 million and sold for $2 billion. Cresta is well over $200 million in ARR.

And given that Sequoia’s in both, it’s not like they missed the fact that they’re getting a different revenue multiple. So, on the other side of the table, you’re right: the seller said, “Yeah, I get why I want a high revenue multiple for Listen, because it’s got growth and it’s new.”

And I guess why I’m still getting a good, but not stellar, revenue multiple for Cresta is that it’s a mature—more mature—company. You’re right: it’s the buyer deciding which bet to take. And thank you for reminding me. Listen is a very good example of that. Do I want the ultra-modern deal? But the problem is, I’m going to have to pay a scary revenue multiple. When you’re trading at 7X, to be fair, it’s really hard to spend a lot of money at 40X. You can spend maybe a couple of billion, but it gets scary.

I think the Cresta deal is actually pretty good. You’re paying 10X.

I agree.

Great.

No, I agree. Exactly. I think the Incon deal—I think they’re smart deals for Salesforce. You get critical mass, you move the ball forward. Maybe it’s not as bright and shiny as the newest bright and shiny thing, but your Service Cloud business is your second-largest, I think largest business now, and you have to make progress on that. I agree.

What you’ll find if you talk to a lot of the big players is that if they want to do an M&A deal and put their resources behind it—their GTM machine, their platform—they’ll all tell you this: they want at least $100 million in revenue. They’ll wait, take the meetings, and get excited. What’s the one? Listen Labs, right? That’s a tech bet, and most folks will tell you those are the dumbest bets. Either buy a small team and pay less, or buy something with nine figures.

You only do a Listen Labs deal in times of stress, because they’ll all tell you that’s their lowest investment performance. That’s what it says. But it’s a reason, if you’re confused as a founder about why deals happen: a lot of them just decide that if it’s not $100 million, it doesn’t have scale. “I can’t put my guys on it. I can’t put my team on it.” I’ve heard this from a lot of teams. It just is what it is. Even with Fin, their core agentic business had to hit $100 million. The overall business was larger. It’s the same story. Then all of a sudden it’s like, “Okay, now I can see how I can make this material in my $50 billion business.”

It’s probably true. It’s a shame, but it’s probably true.

I don’t even think it’s a shame, but the Borg otherwise just can’t assimilate you. You’re too small.

Oh, yeah.

All right, boys, it’s time to go make some money. Back to our day jobs.

Well, that’s our message.

Let’s do this.

Let’s do that. Okay, Harry, thank you very much.