Why the most expensive Seed deals are the cheapest | E2299

23:50Tomasz Tunguz:I think it's a different manifestation, which is the board and the voting construct, where you see founders having tremendous voting control over a business rather than shrinking the size of a board. You still need an audit committee chair and a nomination governance and then a compensation committee chair. So there's just a certain number of people on a board, but you definitely see founders with tremendous control over the board. And that, I think, is going back to a point that we were talking about before, a sign that echoes, say, 21 of how much control they have over a business. Alex, you bring up a really good point, though, which is kind of interesting.

24:25Obviously, we've seen over the last five to seven years the entry of new capital sources. So the crossover funds, some strategic funds, even large sovereigns coming in and participating in the later stage rounds of these companies. And so, you know, to Tomas's point earlier, it's possible that the role of the VCs and or the composition of the later stage VCs, you know, is kind of a moment in time when founders have the options to load up and, you know, go to the next level of capital because there's more of those sources, be it sovereigns or crossover funds or whatever it may be. but it may just kind of change the choreography of how they scale these companies.

25:09And there's some pretty specific examples, Cloudflare being one of them, where it's a disproportionate amount of the capital that's come in is not from venture. Yeah, we'll get into why that's the case in a second. But you're telling me that essentially Tiger is not dead and that the crossover story is not over? Because I feel like for 1824, 36 months there, the idea of seeing all this quote, quote, tourist money coming into tech was written off again. Well, there's the crossover guys. There's your T. Rowe Price Fidelities, you know, others that you mentioned, there was BlackRock, Blackstone, and then there's the Mubadala's G42s and, you know, the milieu of kind of sovereign funds and sovereign spinouts that are getting wiser, smarter, and more aggressive about getting involved earlier.

25:53So I think that changes things a bit. I'm glad you said Mubadala, because if you can't say Mubadala or Tomasek, you pass the shibboleth test and therefore you can't come on the podcast. You have to be able to pronounce them correctly. And I learned that in a Tomasic conference room once when I butchered it and it was corrected by every single person in the room. So now you know if you're listening to this show. All right, here's the thing. If people are raising money to go out and amp their token budgets, right, to cover their token spend, why do we need VCs at all? Why shouldn't Denspropic and OpenAI just meter out tokens in exchange for equity?

26:24Cut out the middleman. Paige? I thought that one of them was doing a program quite similar to that. I mean, like we have partnerships. Yeah, yeah, I think that's very interesting. I think we're starting to see more experimentation around spend because there's like a more of a clear line of return, I would say. um so yeah i think it would be interesting to see open ai move more deeply into that i mean they also do have quite good partnerships programs like we have partnerships with open ai and anthropic and that allows our portfolio companies to access certain amount of tokens and i would say they've been pretty aggressive about that for a good reason as those companies grow larger and spend more on tokens later on it's a great acquisition for them the thing that I missed that everyone just reminded me of is the open AI pitch to Y Combinator companies offering $2 million worth of tokens in exchange for equity.

27:21Essentially, it's a SAFT, a simple agreement for future tokens. There's this notion of tokens, AI tokens for equity, the financialization of tokens, but then there's also the financialization of pure compute that's happening. And I don't know if you guys have seen this, but there's been a couple of funds who announced that basically, you know, they'll say, oh, I'm investing$20 million in a company. But half of that 10 million is actually in the form of compute. And so in a world where - Like GPU hours, Michael? Like exactly, GPU hours for equity. And so, you know, this is really interesting because when you're kind of raw, you know, materials for what you need to actually build your product or deliver your product becomes the currency, the financialization of compute and or tokens, you know that could create a very interesting environment and bring in some different participants for sure so then the closer to the metal you are the better of a vc you can be because if nilabs are going to dole out compute gpu hours for equity and open a is offering tokens i guess beneath that's offering electricity access for equity at some point in time like how far can we go down this rabbit hole yeah eventually if power is the bottleneck you know potentially This is what I love about the current moment in technology time.

28:37Everything seems completely unsettled and shifting. And if you go back to the SaaS era, it felt like it is entirely solved. Like you wanted to triple, triple, double, double, double. You want to have rule of 40, blah, blah, blah, blah, blah. And now everything feels upset. And so I guess, Tomas, is that why we've seen founders recently airing a bit of their venture capital dirty laundry in a way that in all of my years of hanging around this world, I hadn't seen. It feels like founders are almost less afraid than they used to be. And I wonder if it's these dynamics that are leading them to be a little bit more fearless when it comes to sharing spicy anecdotes about your team.

29:13Tomasz Tunguz:Yeah, I think it comes and goes, right? I mean, we had Valleywag and then there was, gosh, when I started, there was a website where you could anonymously rate venture capitalists. That was the funded, yeah. That was Adeo. Adeo created that site in 2006. Very spicy takes on bad behavior and meetings and so on. I think you would get sued out of existence today if you did that, like instantly. Yeah. I'm pretty sure he did get sued. Did he? Yeah. I think it comes and goes. You know, it's kind of cyclical. And I think there's a cathartic, there's a catharsis that happens. There's kind of a big release of emotions every once in a while.

29:50Tomasz Tunguz:That's healthy for the ecosystem. So this is just a dam break. We'll rebuild the dam and everyone will kind of go back to normal. If you're not using AI tools to make your team more effective and increase your productivity, trust me, you are being left behind. Sitting on the sidelines is no longer an option. No more waiting. With NetSuite by Oracle, you can start putting AI to work for your company today. NetSuite is the number one AI cloud ERP, and it's trusted by over 43 ,000 businesses. It's not just an AI add-on or a chatbot that sits in your browser. No, this is a unified source of truth that brings together all the data you need to run your business.

30:31From software to IT services to healthcare, equipment manufacturing, financial services, and many other great American industries, NetSuite delivers a customized solution for your business. If your revenues are at least in the seven figures, get your free business guide, Demystifying AI at netsuite.com slash twist. The guide is free to you at netsuite.com slash twist. Okay, because some of the stories surprised me. And I have friends who are VCs. I've been to LP meetings. I've gotten to be in a lot of rooms. And I've never seen people falling asleep mid-pitch. Like, that blew my mind. And the fact that it was like 10 different stories of napping VCs.

31:11Like, someone asked, is there like a plague of narcolepsy going through the valley? and I don't know Paige it doesn't feel like that would fly with kind of the modern founder who's in a hurry and has their eye on a pretty big prize and so I'm curious how do you manage to keep your eyes open when you're listening to founders pitch you and your fund I know I'm usually pretty excited to be on the call so I don't think I've ever fallen asleep in a pitch meeting before yeah okay well I think you know in fairness when the markets are high and and cranking as they are now, you tend to hear these stories.

31:47And when they're really low and the stories are slightly different, like, oh, they asked for a 4X liquidation preference and they brought in participating preferred. So you get the unhappy peanut gallery when the markets are terrible. And then you get the, I'm feeling my oats and I'm going to talk about bad behavior when the markets are high. It seems like the kind of natural flow of things. So venture transparency is the, it follows the NASDAQ pretty closely then. NASDAQ's high, everyone's doing well, valuations are up, VCs are on notice, NASDAQ is down, everyone's poor, founders are quiet. Okay, so essentially this is peak.

32:24Share your story mode. Okay, I want to click on one story and no one here works at Sequoia, so we're totally safe. Brendan from Merkur, one of those AI data companies that's grown to$80 trillion in revenue. Had him on the show a while back. lovely guy episode 21 59 if you want to watch that uh he said the quote sequoia scam is worse than a single horror story in the last six months i've seen a half dozen rounds where sequoia invests in two tranches everyone pretends they only did the higher valuation founders must represent this to their employees and then shop it to angels and he calls sequoias quote blended price as blatantly deceptive so i'm curious one not to pick on sequoia in particular but is this something that's happening broadly or is this a handful of examples being um aggregated into what appears to be a trend, but actually isn't.

33:07And Tomas, you're the perfect person to answer this.

33:08Tomasz Tunguz:It's starting to happen more and more. I think it recalls like 2021, where you would have three rounds of financing happen within a single year. We're starting, we see that. I mean, often you can look at it, the infrastructure as a service vendors are starting to see multiple GPU, I mean, fireworks and base 10. There are many other companies because they're growing so fast. And now because the anticipation is there, if you're an aggressive mid to late stage firm and you want to get in, well, you can structure it in a way where you can blend this valuation. So I don't think it's, I mean, I think it's like, I would guess like still 5 % of rounds, but it's not a red herring anymore.

33:47All right. Page, have you seen, oh, Michael, please. I would just say it's, you know, this particular concept is not really new. I mean, if you had been an entrepreneur starting companies, even all the way back to the late nineties and early 2000s, It was not uncommon that a lead investor would come and say, great, I want to lead the round. Here you are. We noticed that your last round, you had authorized a certain amount to raise, but you didn't raise at all. So do you mind if we take that last couple of million in the last round and then come into this round? And their blended cost ends up being lower.

34:21So, I mean, I've seen a fair amount of that over the years. So it's not. Is that generous, Michael, or is that predatory? I can't quite decide. Or is it both? I think it depends on the situation. If you're super excited to have that lead come in and lead that round and it helps consolidate and bring together other investors, then you may let them put in a small amount in that last round that didn't fully cap out. All right, Paige, how often do you see this kind of activity? And do you agree with Tomas that it's no longer a red herring, even if it's not kind of the standard route? Once in a blue moon.

34:55And I think the situations are usually one of two things happening. One, it's an incredibly exciting company. And there's a lot of pricing power that the founder has and interest. And they're interested in they negotiate with the lead on a certain valuation. And then they have other folks who they want to bring in the round but don't want to take that dilution. And then in other situations, it might be less advantageous to the founder and more pricing pressure from the lead saying, hey, we want a discount on this round. We'll do some of it at a lower valuation. But rare. So founders probably won't run into this.

35:35But you, as an investor, if you're being offered a round at a certain price, how much transparency do you expect the founder to tell you if they have one of these blended leads, let's say, in the same round? I think it's a very nuanced question because usually in your docs, like you'll have information access and information rights and not every investor gets those information rights. So I think it is like a nuanced question. Like obviously I would like to know and I would like ask about how the round is structured. At certain points, we'll invest before a lead's involved and so then we'll be pricing our own.

36:12But in any situation, it's like, I'm taking a look at, is this a founder we want to work with for the next 10, 15 years? And then also, is the valuation at a point that it makes sense based on what we think the potential outcome could be? This week, we saw finally Anthropic drop a version of its much vaunted Mythos model. It's called Fable 5. It's very expensive. Costs literally twice what Opus 4.8 does. First of all, who here has not played with Fable? I presume we all have, but has anyone here not touched it? I have not. Michael. Okay, so Michael's the lead on today's show. That's fine. Paige and Tamaz, first impressions of Fable.

36:52I used it, but I didn't really put it through its full paces. So I'm curious if you think it's the step function that some people claim.

36:57Tomasz Tunguz:I think it's really impressive. I mean, okay, so just to kind of set the context, you have new model releases approximately every 41 days. And most of those model releases on key benchmarks have one or two percentage points of improvement. They're a 10 percentage point improvement. So pretty fundamental. I ran it through its bases. I had it analyzed three code bases last night. Okay. And then it was using it to optimize performance and it did phenomenally well. Absolutely. There's the thanks for the model card. And so you have really some, I mean, the agentic coding going from, you know, 13.4 to 29.3 is just an enormous, enormous leap.

37:33Tomasz Tunguz:So it is fundamentally much better. It's a bit slower. You can watch it. Like it will, I think about it as a central coordinator where you give it a task and it will federate work to different agents and orchestrate them over long periods of time, manage its memory. It's incredibly effective. But also, like you said, extremely expensive. It's not the most expensive model, though. If you look at OpenAI's pro models on a per token basis, those are three to four times more expensive than Fable. But OpenAI argues that GPT 5.5 Pro is token efficient to models. Are you taking that into account? I'm not taking that into account.

38:06Tomasz Tunguz:I'm just looking at the input and output tokens. I thought GPT 5.5 was 5 per million in and 30 per million out versus 10 and 50. No, that's right. But my understanding is that there are certain pro models that are reserved for math and science and that are significantly more expensive. And there, there's a tremendous amount of thinking tokens that need to be taken into account. But yes, for general purpose models, Mythos or Fable is the most expensive. Staying with you, Tomas, you know, we've talked a lot about how in the last couple of years, AI has gotten better. Startups can do more with it, either to improve their internal operations, to make better products, better services.

38:42We see a step function here. I'm curious if you think this is going to change the quality of what startups can bring to market and therefore possibly increase their growth rates and find even better product market fit faster.

38:53Tomasz Tunguz:I do. There's no doubt. And you look at what you can build in a day and have the models operate overnight and self-improve, it's extraordinary. So yeah, I think the pace of innovation, maybe put it the other way, the expectation of the software buyer will be that the software is secure, you're selling a suite, not a point solution, and the software is improving every two or three days. This is one of the benefits of SaaS compared to package software was you were paying for an ever-evolving subscription, you might see a release a month. Now, I think the expectation is, oh, there's a bug next tomorrow morning, I think it'll be fixed.

39:30I mean, now when I see people talking about Notion, Michael, I mean, they literally like ping the founders and they're like, can you please fix this? And they're like, aye, aye, Captain, we'll get on it. I presume that's something that's now mostly possible via Agenda coding. But, you know, Notion has been, I would say, one of the leader in AI. We use Notion here at launch every day. What do you see from model improvements moving forward? Do you think they're actually going to help companies like Notion continue to improve at the current clip? Or is this more of a, it'll look the same, but just be slightly more intelligent when I prompt it?

39:59Well, I think there seems to be, you know, breakout successes with companies like Notion, where they've been able to plug in so seamlessly to Claude and kind of orchestrate and do things in really unique and helpful ways, which sounds like it's how you're using Notion plugged into Claude and other kinds of tools. And to me, it's like a separation, you know, kind of a tale of two cities. There's the applications who are figuring out how to perfectly blend in, you know, with the LLMs and kind of core models and make their product that much more valuable. And then those who are struggling to figure out how they coexist and work with the larger models.

40:39And it's interesting to watch for sure. Yeah. I'm curious to see what people will build. I've seen the usual slew of demos like, oh, he built a horror first person video game in one shot. Oh, oh, look, it did my laundry for me. Oh, it took my boyfriend out to dinner for me. People are very impressed, but I'm always kind of curious, like what's the second week of this, coming back to Tomas's wait for the IPO for SpaceX to see? Everything always looks really impressive day one. So Paige, do you think that your portcos that use AI, which I presume most of them, are going to be trying out Fable in a production setting?

41:13Or is this more of a, dear Lord, we can't afford that. That would tank our margins and turn us into a shop selling dollars for 50 cents? I have to ask them. I am curious about this question. I think what we've seen in most of our companies is that there's like a hybrid approach where they would use a lower cost model for something that is like more repetitive. And then for higher level reasoning or orchestration, they'll use a more expensive model. So does that work as well as people say does? Because Tomas said orchestration. Tomas is nodding. He loves to nod while on mute. I don't want to interrupt, but yes, it works exceptionally well.

41:55Tomasz Tunguz:It works unbelievably well. Tell me how it works. Yeah, yeah. So I'll give you an example. So let's say you have a repetitive process for updating your CRM or answering a particular email. What you can do is you can have a state-of-the-art model create what's called a skill, which is in Markdown, which is a text file. This is how you do this. And you can pass that from, and I've done this, and I've done this where I can get 90 % of the things I do with AI on my laptop to run on a local model on my Mac. And that has meaningfully reduced my overall token spend. And as I add skills, I've gone from 65 % to say 91 % as of yesterday in terms of local model inference.

42:36Tomasz Tunguz:And then Stanford released a study yesterday, the day before showing across a broad distribution of different skills. This is very true. So I'm a huge believer in this, whether it's like model distillation or skill distillation, this will be the architecture for most applications. So we're going back on prem. We're going back. I mean, maybe because I think. No, it's definitely happening because I think also like if you look at the cybersecurity concerns of running some of that information in the cloud all the time, it does make a lot of sense, especially we're seeing that in manufacturing use cases because they're one of the biggest targets for cyber threats.

43:15So explain that to me in more practical terms. Are we talking about small language models running on air-gapped hardware or is this more just like we have a Dell computer and there we can slap Quinn 3.7 on there and just have a good time? Um, well, so for example, one of our portfolio companies is a company called Maneva and they build applied AI for the factory floor. And the founder was previously at DeepMind studying embodied AI. And so they use video to robotic action models, which I'm, I'm really interested in the continuing application of like multimodal AI. So it goes beyond just text to text input and output.

43:51You use about five, uh, terms. Five big words. Yeah. I'll use some smaller words. No, no, no, you're fine. I just want you to explain them for everyone listening who is too lazy to Google them as we talk along. So break that down into little person words, please. Okay, sure. So a factory operator on the factory floor, I'll give you an example from one of their early customers at a candy factory. So originally there was someone who had to individually check every single candy bar for defects as they went down the line and then press a button if there was a defect. And so what Maneva does is they have agents that do one specific task really well on the edge, so using off-the-shelf hardware.

44:29And they can use Maneva software to basically do that task now. And what's really interesting is the folks on the factory floor are like, that's great. Like, that's not the task I wanted to do. I wanted to help the factory run more efficiently and do higher-level work. and so they're actually coming up with new ideas of where to use Maniva on the factory floor. So we don't need to use Fable 5 to see if the Hershey's with Almond Bar is a rectangle or a circle is your point. Yeah well what's interesting is like when you deliver a model it's not going to be fully trained because you need that like actual in in real life experience to fully train a model to be great at something and this I have this like thesis around hyper specialized AI where these models are great at general intelligence, but to get them really, really good at a specific application task, they need a lot of data that is stored somewhere in a company or on a factory floor in real life.

45:28Who builds those? Because on one hand, you think that the customers who have the data would want to be able to take a model and then bring their data to it. But also at the same time, SaaS companies who sit on top of so much customer data want to build the AI workflows and therefore maybe also tune the models. So Paige, where does the value accrue in that setup? I actually think that it's more new companies and startups like what we're seeing. I think Vertical AI is still very early on in the commercialization stage. Like we've been following the space since 2021, but as the models have gotten better, there's been more and more applications.

46:03So I actually think that a lot of this is accruing in startups. We've seen some larger incumbents move into this space, but ultimately it's challenging because you have to almost like retrofit that software to fit with your existing company. So I think definitely like these AI native founders are having, have a strong advantage. I want to get to Michael in a second, but I need to ask Tomas a question. Tomas, does the space that makes a skill.md file different from an agent eventually collapse to zero? I don't think so.

46:35Tomasz Tunguz:I think that's the domain of the application layer. I think if you're like a SaaS application or whatever an AI software company today will be in the business of figuring out the managing a context database so like the standard operating procedures associated with something building the skills and the instructions and then selecting the models so that you a customer can operate with state-of-the-art AI without state-of-the-art AI prices and we're going to remember that it's possible to run HubSpot entirely through Glaude Fable 5, you'll pay for it, but you don't want to pay for it. So why don't we just condense that and then have an application company bundle that intelligence into a software application, diffuse it across a whole bunch of different people.

47:18Tomasz Tunguz:I think that's the future of the application. Michael, weigh in here on where you think the value is going to accrue across the application layer, models, tuned models, and private data sets. It looks like right now there's so much assembly required to really get these verticalized solutions to work in specific scenarios. It feels like the companies, the startups that can just create the kind of simplest, easiest onboarding and packaging of the orchestration, the workflows, and package it in a way to where non-Silicon Valley people can apply it are going to be the ones who are moving fastest. And I think it speaks to, you know, why is open AI and Anthropic spending so much time and money building out these like external consulting organizations with Accenture and Blackstone and all these guys?

48:13It's because, you know, there's a lot of assembly required to get this across the entire business landscape. And so I think you can't minimize that. And Silicon Valley has been great at creating companies that just dumb down and make the experience much simpler and easier. And I think that's thematically going to be an important concept going forward. Do you think we're still going to have these private equity dash AI lab partnerships in 10 years time? Or is this simply just we're going to bridge this temporary chasm in AI deployment that is simply a artifact of a immature technology reaching the market before it's fully baked?

48:57Well, I mean, it looks like we're going to have it for some period of time. But that period of time is really can we materially impact adoption? Right. Because the amount of capital that's been raised, as we all know, the amount of capital that's about to be raised from the IPOs, you have to begin to create tangible ROI at a certain point, especially after you're public. And so at that point, there's a measuring stick. People want to see the numbers. And so I think they're on the clock to be able to prove, hey, there's tangible ROI coming from this industry and this industry and these companies.

49:30And so they're just doing everything they can to load up and increase the likelihood of that adoption and kind of successful tangible ROI being validated. All right. We're going to scoot through a couple of topics really quick before we run out of time because there's a lot more I want to get you guys on. First of all, Tomas, if you look at open routers data and you see what are the most popular models in the last week, the names are DeepSeq v4 Flash, Mimo v2.5 from Xiaomi, Hi3 Preview from Tencent, and then Minimax M3 from Minimax. I view that as startups being intelligent, going back to our model routing question and kind of choosing what's the error model to guide things.

50:08that startups, even though they're very AI intensive, might already have in place ways to offload some compute away from these kind of frontier leading models. And therefore, they're not going to get whacked by the cost concerns we've seen enterprise customers screened about for weeks now. Am I correct there? Or am I being too optimistic about where startups have been deploying their AI inference in the last six months?

50:29Tomasz Tunguz:No, I think you're exactly right. You're seeing a lot of shift to open source models. I think it's why it's critical that there's a dynamic US open source model ecosystem. Google's pushing in Nemo Tron, RC, NVIDIA. I think NVIDIA's committed like 23 billion to open model. So open models are incredibly important for the ecosystem. I think they allow application companies to compete with the labs, just like we were talking about. And then if you look, we were analyzing the data about six months ago, looking at open source adoption. The very first companies to adopt open source models were the ones with business models with small gross margins.

51:04Tomasz Tunguz:Which makes sense, right? Like if I don't have a lot of money to spend on infrastructure, I'm going to go and buy commodity A. Let's call it white label A. Wait, negative gross margins are bad? Curse you're talking about. Those are great. Yeah, it turned out pretty well. But yeah, so wherever, you know, when there's a need and the market fills, the beauty of capitalism. How about capitalism? All right. Does anyone want to weigh in on this before I take us in an entirely different direction? Just one more point there, which I think Tomas is right. There's also one other part when you talk about where the value accrues, which is we're all talking about models and which model am I going to use for this and that.

51:38That's obviously going to be abstracted away for the vast majority of people. And you're going to show up and say, I have this job. I want to do this thing. And whoever that solution provider will be, open router, you know, might say, great, this is the lowest cost and best model for you to use for that. And by the way, here's the compute that is the most regionally best placed and available and the lowest cost for you. And so, you know, normal humans are not going to think about these things. It's like, what spark plug do I want in my car? It's like, I have no idea. Just give me a spark plug that works.

52:10Well, this is why I think that the open router value add or the moat that it has is its auto switcher that chooses models for you and different providers for you. As an open router user, I love that because it takes that off my plate. But it also means that it becomes not just my gatekeeper, but also my tour guide into the world of AI, which I think is going to be a really important door to hold on to. Unsurprisingly, they just raised, someone help me out here, 113, something like that. Yep. In the last month. I forget the exact number. All right. Turning the page. Seed prices. Now, I've been a journalist covering venture capital since I was in college, which is getting to be pretty long ago.

52:45And if there's one thing that everyone agrees on is that for my entire career, seed prices have been unsustainable, too high. They're breaking seed economics, and no one can make money anymore in seed investing. And then people still do it. So if you take a look at this chart that I now have on your screen, this is some data from our friends over at CARTA. And as you can tell, we have reached a new era of seed pricing. If you're on the audio version, imagine a chart that's kind of flat but trending up that then goes parabolic in the last couple of quarters. And what it shows is that the 95th percentile for seed rounds in the U.S.

53:15that CARTA can see now have a valuation of$174 million. 90th percentile,$94 million. And those are up from about 66 and 50 back in 2022. So is this what finally breaks the seed market? And Paige, how are you managing to find entry prices in the companies that actually are attractive enough to work for your fund economics? Great question. I mean, I think like when I think about it, I think about understanding valuations on a case by case basis. So when we think about the exit potentials of some of these businesses, there are markets where companies that may have been able to charge one price in software days because they're now doing the work can charge three to seven X.

53:59And so that means that down the road, there may be an exit outcome that's three to seven X, like what we've seen before. So I'd say we, we take like a very case by case approach to investing. I think if you're looking at companies in the same pools that everyone else is, the prices will definitely be higher and we've seen them continuously go up in, in the past. Michael, your fund backs other managers to some degree. So I'm curious, how are your horses in this race dealing with seed prices that, to me, look not just unsustainable, but just uneconomical for early stage investors? Yeah, so here's what we're seeing.

54:40At the pre-seed level, which is we're mostly in these emerging managers that are writing the very first check at day zero into these companies. From a pre-seed basis, you still see great managers getting in at low valuations. At the seed stage, I mean, my take on this is if we track the companies and we see who's doing what, is that that median valuation that you have on the chart perhaps is a little overpriced based on our seasonal place where we are in kind of history right now. But I'm going to take the slightly more provocative angle here that that top 1 % or top 5 % is likely underpriced because the scale of the opportunity and where we are at this moment in time means that these outcomes are big.

55:31We already know that these companies are scaling revenues unbelievably quickly with less resources than ever before. Well, I didn't think that's most scorching take of the show would come from the other guy in a suit. But here we are, Michael, doing us all a solid. Okay, so putting that in perspective, Paige says that we're seeing outcomes get larger. You're saying that the leading companies might be underpriced. The implication being that the exit they're heading towards is going to be truly staggering. And if I could take that one step further, that the fact that we're looking at three roughly trillion dollar plus IPOs this year, therefore won't be an anomaly.

56:06It'll actually become more the norm down the road. What we do know is that in each one of these movements, be it the late 90s, 2008 to 2015, or now where we sit at this moment, the kind of destination point in terms of valuation, the outcomes are always way larger than what we anticipated or what we saw in the last run. And obviously, we're seeing this now with, you know, a$1.7 trillion IPO that's happening in two days and, you know, a$985 billion round that Anthropic just did. And so, you know, we're seeing this in real time. So you have to think that, you know, valuations are going to level set to accommodate and or just reflect that the outcomes are bigger.

56:46I would just argue the bigger issue here may not be valuations and it may not be, oh, are we paying more for the same type of companies or the same kind of outcomes? The bigger issue is with these IPOs that are happening and all the liquidity that goes back into the market, we know that the typical kind of LP and early stage VC funds in all of our funds, family offices, high net worths, et cetera, just spent a disproportionate amount of their VC allocation in late stage secondaries over the last two to three years. They're now going to get generational returns for doing that. Are they going to reinvest in small early stage funds that go for 10 years?

57:31Or are they going to say, hey, this late stage pre-IPO thing is the way to go. And I'm just going to continue to really look for those kinds of deals. That, I think, has more of an effect on the market than the fluctuating valuation, because it means the source of capital that kind of feeds that seed stage, the pre-seed stage, compositionally may not be there in the same way it was in previous years. The numbers are getting so big. I feel like the Mendoza line for technology poor just keeps going up. It's kind of staggering now what constitutes wealth. Even in my friend group, the people who worked for Anthropic for a while, they carry themselves differently in group chats, it feels like.

58:14They just have more swagger to them. But Tamaz, the idea that these highly valued seed rounds are not overpriced because of potential outcomes being so large really does put a lot of emphasis then on selection. Because if you back one of these and it's not one of those outcomes, you're going to overpay dramatically. So does this mean that we should see greater differentiation in seed stage returns based on essentially GP discernment?

58:41Tomasz Tunguz:Well, I think so. I mean, I think the venture capital market's evolving a lot like the public markets did where you have indexes and then you have, you know, concentrated funds. We're clearly in the more concentrated category. Both strategies can work very, very well. But ultimately, selection is what matters is power law underpinning all of this. And, you know, I agree with what Michael said. You look at, I think, Vencap published. I think Venkat published a study, you look at rolling five-year periods and then 75th and 90th percentile or 90th and 95th percentile exits, and you can see them going up in each year much faster than inflation or even venture inflation.

59:21Tomasz Tunguz:So I think that's definitely true that the returns are there. I think the, and what we're seeing with this, I mean, these IPOs is just the liquidity is tremendous. I do wonder what happens to the secondary markets. We've seen huge inflows into the secondary markets. Do the secondary markets actually shift to the next, say, top 20 companies? And will investors want access there? And then the other question is around M &A, right? M &A has been historically very difficult within AI. Now you have a lot of national security concerns. And so you have many, many companies with large valuations, the total number of buyers who can afford, say,$50 billion.

1:00:00Tomasz Tunguz:exit is probably fewer than 10. Right. Yeah. So what are those dynamics look like? TBD. Okay. Just for fun, because you kind of brought it up. Do you think that we're going to see any nationalization of the major American AI labs? This has been discussed by both. This is a great list. Bernie Sanders, Sam Altman, and Donald Trump. And I'm not quite sure if that's the coalition I expected to see forming, but I'm very opposed to this. I'll just be honest. But I'm curious if I should actually be afraid or not. I don't know what nationalization really means. I mean, do you notice this government took a position in Intel and that's done very well, right?

1:00:34Tomasz Tunguz:There's been some talk of a sovereign wealth fund. We will see what that will happen. It's like nationalization, the creation of government appointed monopolies, like in the case of alcohol distribution and also telephone networks. I really would not like to see that. I think there's a ton of regulatory capture that exists there. And, you know, the history of Silicon Valley is tied to the dual use technologies where there are technologies that are used both for the government and the private sector. Bell Labs notably coming out of that. So I do think it's important that these major labs do have relationships with government.

1:01:08Tomasz Tunguz:So I don't exactly know what nationalization means, but on the whole, you know, being a capitalist, I think less regulation and less government involvement in the evolution of technology is a good thing. All right. We can boil that entire answer down from Tomas to hell no. All right. Now, I have one question for each of you because I picked out my favorite of your portfolio companies and I want you to brag about them for a moment. This is your time to put the founders in the spotlight. And, oh, Michael, you're first. So there is a war going on in the Middle East and there was a helicopter that went down and it was captured.

1:01:39Sorry, the pilots were saved by a drone boat from Sironic. I believe it was a Corsair and I believe you're an investor of this company. So tell us why Sironic is the coolest thing. Yeah, absolutely. We're an investor in Sironic via one of our fund positions, which is Silent Ventures, an incredible pre-seed defense tech focused fund. Sironic is just, I mean, incredibly impressive. They've executed like nobody's business. You can see in the valuation of the company and the rounds they've done just how fast that business has moved. So, yeah, it was pretty cool watching the news last night and they talk about the Apache helicopter that was shot down and that, you know, immediately two of these autonomous boats were sent out to pick up the crew out in the Strait of Hormuz.

1:02:26I mean, it's kind of a perfect sales video for Sironic. So, yeah, we're thrilled about the company. I mean, it's a no-brainer, that kind of product. But you can start to also see, I mean, just FYI, how these defense tech-focused companies where the demand and the instant revenue for them is coming strictly from defense, how that application and how that value proposition can be applied in many types of ways. So yeah, it's a phenomenal, phenomenal company. Shout out to them. Also, Vatten Systems, Andrel makes some sea drones and Blue Water Autonomy, I think is also in the mix. So it's one of those sectors, one of those startup niches that I think is really, really awesome and more deployed in the battlefield than I thought.

1:03:11I thought Soranek was still bouncing around the harbors just showing off their cool tech. I did not realize we had enough deployed that two of them could go save some pilots. So I was very impressed by that. Shout out to them. All right, Tamaz, you're next. Open source. In the AI era, you are an investor in MotherDuck, which is the commercialized version of DuckDB. I actually got to meet them at a recent MCP event in New York. I got to talk to their head of the AI, I think. So tell me about why MotherDuck is the right choice in the AI era and why open source will not lose all of its value to vibe-coded infra from the AI.

1:03:45Tomasz Tunguz:Yeah, great question. So MotherDuck is a company that commercializes an open source technology called DuckDB. DuckDB is a very small analytics database that can scale to just as big as the very large analytics databases. But because you can have many small databases, it's perfect for agents. So you could spin up a million different agents. Each of them could have their own DuckDB or Mother Duck instance, and then it's all controlled from a central layer. Awesome. And how's the company doing? My friend Carly works there, so I've been to them. She's awesome. We just had the event at Snowflake Summit where we had dancing ducks outside the Jewish Contemporary Art Museum in San Francisco, right outside of Moscone.

1:04:26Tomasz Tunguz:And just a phenomenal setup. Oh, I know exactly where that is. You know where it is? Yeah. The funny shape building. Yeah. Anyway, so the company is doing phenomenally well. Product is expanding quite a bit. We just launched interactive charts and dashboards and have some more product announcements coming. They're all AI native. I'm disappointed. I have a mother duck swag item, which is two mechanical keys together with ducks on them. And I brought them home for my kids to play with. And I literally set it on the counter to bring out so I could show it to you. And I forgot it in the house, damn it.

1:04:58Tomasz Tunguz:That fidget toy is so fun. Dude, more startups should do that. Good marketing technique. Hand out fidgets to nerds with ADD because we will take six and we'll never let them go. They're fantastic. All right, Paige, to close us out, I want to hear about the progress of actual American re-industrialization. I know you're a backer of Knox Metals, one of my favorite startups in the entire nation. Talk to me about how this is not smoking mirrors and we're actually going to get some damn cold rolled steel back in the country. Oh yeah, we are going back to the factory floor. So Knox Metals is a next generation, next day metal servicing platform.

1:05:33I had never heard of the metal servicing industry before I talked to the founder Zane, who I met four years ago, and we reconnected when they went through IC. But basically, like, there's multiple decabillion dollar businesses in the space, both public and private. And what Knox says is they were like, okay, like, if we're building this new defense technology, space technology, we're going to need a supplier that is meeting the demands of these companies that want to move faster and innovating from a hardware perspective. um and so yeah they are they have built an incredible suite of products that have helped them push metal out the door faster so they're in detroit they just expanded their facilities there and have been cutting steel if you check out zane's twitter it's it's really cool because they're posting like videos of their band saws and the team on the factory floor and they have big announcements coming out next week.

1:06:30So stay tuned for that. Michael, thanks for coming on. I really appreciate it. Where can people find more about your firm in case they want to get in touch? Castalia.capital. Very simple. Tamaz, what is the TheoryVentures URL? TheoryVC.com. TheoryVC. Why not Theory.VC? Was it take them?

1:06:48Tomasz Tunguz:We have that one too, and Theory.Ventures, but TheoryVC.com is a.com. Man, talk about traditionalism in the venture realm. Jeez. I thought TLDs were free range now. All right, Michael, Tomas, Paige, thank you for coming on. Twist is back on Friday. My name is Alex. I'll see you then. Bye.

1:07:28Tomasz Tunguz:Already have traction? The Launch Accelerator invests$125 ,000 and connects you with 500 plus investors to help you raise your next round. Apply at launchaccelerator.co. If you're an accredited investor looking to gain access to quality deal flow, apply for Jason's Angel Syndicate at thesyndicate.com. We find two to three deals a month. And check out This Week in AI, Jason's experts-only roundtable with top AI founders and operators every week. Find it thisweekina.ai. Check out The Twist Ticker, our daily newsletter, at thisweekinstartups.com slash ticker. Thanks again to our sponsors for making today's show possible.

1:08:08Tomasz Tunguz:Follow the show on Instagram, follow the show on x.com. This Week in Startups publishes three days a week, Monday, Wednesday, and Friday at 5 p.m. Central Time. You can submit an audio or video file question by emailing it to thisweekin.com.