Theory Ventures' Tomasz Tunguz on AI value accrual, blockchain investing, and the moat question in a world of fast-moving foundation models

Talk soon. Thank you. Bye. Yeah, it's talking to these uh space folks and I mean everyone we talked to is amazing, but it it's it's awesome. I was looking up at the moon with my son, the four-year-old, recently, and he was like, "We we got to go there. " And I was like, "Yeah, I want to go there.

I I think we'll be able to do it one day. " He's like, "We'll need a rocket. " I was like, "Well, I know some people that are working on that. " And it's it's amazing being able to be like, "Yeah, like actually like it's happening and I know some of the people that are working on it.

It's going to be a lot of work, but we're going to get there. " Anyway, uh let's tell you about numeral sales tax on autopilot. Spend less than five minutes per month on sales tax compliance. Go to numeralhq. com. Sign up today. Uh sign up today. Also, uh Ryan Log has a extremely viral post.

He says, "Beer is narcan for when you overdose on Microsoft Teams. " And I like the comment. Liked, followed every bro. You tell me your high like ratio. Insanely high. 2% riding it all the way to a million. % at at close to a million is impressive is impressive stuff. Banger. It's a banger.

Uh I don't know how much time we have. We have four minutes until Well, we have time to tell you about Finn, the number one AI agent for customer service. They're number one in performance benchmarks, number one in competitive bakeoffs, and they have the number one ranking on G2.

The Finn Milliondoll Guarantee the best performing AI agent consistently beating competitors in head-to-head evaluations and they offer a million-doll guarantee. If you sign up for Finn and you're not 100% satisfied in the first 90 days, we will give you up to 1 million of your money back, no questions asked.

The Finn guarantee. Check it out. Well, we have just a couple more minutes until our next guest. Uh, so let's talk about a new unicorn alert from Arurer Rock. Etched the world's first transformer. ASIC. I've been pushing for this for a long time. I know the Etch guys. I did a long interview with them. Uh, random.

Surely you invested. Surely you invested, John. Mog. Surely. Surely you just threw in a Yeah, a check. I should have a small check. you know when you met them seed round unannounced round 85 million at 1. 5 billion following two stealth rounds at 500 million and 700 and they're now raising at at 2.

5 they just closed another round 2.

5 we got to get the founders on one of the most interesting kind of AI like future thought I don't know like thought leaders or something just like an like can the the the founders can just noodle on how everything in artificial intelligence plays out they've done a bunch interesting demos running uh Minecraft fully generatively.

So there's no underlying game engine. It's just taking the inputs and generating the visual outputs. Uh absolutely insane team uh teal fellows um and they've been on a tear and there's a bunch of insane stories where the founders at some dinner in Taiwan chatting up like a TSMC uh employee to get more line time.

like the hacks that this company has gone through to make this a reality has been fantastic. Uh, and he's the one that told me uh this idea of like the like LLMs being human simulators and so the expectation should be you get like the median human and maybe not someone that wants to kill everyone.

Um, and so he has a very low P doom but he he but he uh walks through that very very uh very in a very interesting way. Uh, anyway, how did you sleep last night? Are you are you a big Eight Sleep fan? After our photo shoot and video shoot yesterday, we uh we we're filming advertisements now.

Not we're not just telling you about Eight Sleep. We're putting it on the line shooting video ads for I cannot wait to air this. I got a 76, which is basically the new Oh, I got a 79. I had a rough night, but I still beat you. Play that sound effect, Jordy. Thank you. John's on off.

Basically, my new range is I just am squarely sitting in the 70s. It's brutal. It's brutal. It's brutal. Anyway, uh you got to get on eight. Go to eightleep. com/tbpn. They got a 5-year warranty, 30- night risk-free trial, free returns, free shipping.

Uh and in other sponsorship news, Shane Copelan uh announces that Poly Market has made Times 100 most influential companies. Turns out that people want the truth. Basically, fantastic. They built effectively a crystal ball. He did. Shane built a crystal ball. is our crystal ball.

We had a crystal ball in We had an actual crystal ball. Shane built a crystal ball. I say it's tomorrow. We got the crystal ball here. Tomorrow's headlines today. Do not drop that. Ben Jensen dropped uh No, Masa dropped it and it was very ominous, but we have the crystal ball here.

We highly recommend every technologist get a crystal ball. If you don't have a crystal ball, you can go to polymarket. com. Crystal balls underrated as a as a desk object. You got to have a crystal ball. You got to have at least at least one. Put that right there so you can see it. Yeah, there you go.

Uh but congratulations to Shane and the whole Poly Market team. They are on an absolute tear. They are on a tear as is our next guest. Welcome to the stream. Tamas, how are you doing? Pleasure to be here. Thanks for having me on. Thanks for hopping on. Boom. Welcome.

Uh would you mind kicking us off with a little introduction on yourself and your firm Theory? Absolutely. Yeah. My name is Tamsh Duz. I'm founder and general partner of a firm called Theory. We build really concentrated portfolios in early stage AI and blockchain companies.

We're about 10 people strong and uh we've worked with nine unicorns in the past and uh just we're about two and a half years old so far. So you were at Redpoint for 14 years. Talk to me about the decision to launch your own fund. What strategy you wanted how did fundraising go?

I just want to hear the whole story of Yeah, absolutely. So we raised fund one in late 22 which is when the Fed had raised rates 500 bips must have been not not a great time to raise but we were able to raise fund one. We raised about 235. We were over subscribed. Yeah.

And um and then we raised fund to about 400 a little more than 450 million two years later. And congratulations. Oh, nice. Thanks. Uh and yeah, so we're off to the races. We have a head of AI who was head of AI at three unicorns and our CEO built the healthcare practice at Palunteer.

And as I said before, we invest in a small number of companies and work really hard to help them grow by building out their sales teams. You're the head of AI. The way Yeah. the way you're talking about the firm, it feels like it's like you're building it like a a company. What talk about that? Yeah.

Well, I think there's a pretty big change in I mean, if you look at like hedge funds and private equity, a lot of them have restructured into basically operating companies. And I think the same thing will happen with venture capital. We think about the products that we sell as financial products, right?

The when I started in the business in 2008, series A was a financial product where we invested four for 25% of a company. And that product doesn't get any buyers anymore. Uh because uh you know the US venture capital asset class has grown from about 8 to 250 billion during that time and so there's a lot more capital.

Uh and so we have to differentiate our capital somehow and the way that we differentiate is by depth. So we operate every Monday we get together we talk about different market maps and what what are research projects and I'll give you an example. We had this amazing intern last summer.

He walked in one day and he said,"I think the future of AI is not with GPUs, it's with AS6 or specialized circuits that are focused on AI. And if you believe that, then here are all the investable opportunities. " And so we debated that. So we go through different market maps every Monday.

And then we have this team that has we have different departments, right? So there's a team that builds internal software team that builds the the sales teams for our portfolio companies. There are people who manage a buyer network 200 people strong.

And the idea is, well, if we can also have an investing team and we're looking at elephants in the dark and each one of us has a flashlight, then maybe with that composite view, we'll have a more refined perspective on an early stage software company and ultimately make better investment decisions.

How often do you see someone else's market map and kind of laugh at it? No, we do our own work, but uh but we use I mean, it's always great to see other market maps out there and and have different perspectives, right? Because we're all trying to figure it out at the same time. Yeah.

Uh on the AI ASIC question, how do you think about that as a disruptive innovation versus a sustaining innovation? Uh NVIDIA is obviously super set up to work with TSMC at scale. At the same time, I've talked to some of the founders we just mentioned, etched, uh building transformer as incredible teams.

I wouldn't bet against them and they've been on an absolute run. And so uh I'm in this weird scenario of like kind of being bullish on both, but you know, there has to be some way this plays out. Have you dug into it at all in that? We've dug into it some.

I mean, you've seen I mean, Google's been investing in TPUs for a really long time. You have the Amazon in differentiation, Microsoft. Yeah. Have their own, I think. And then Nvidia announced this week or at least there was a Wall Street Journal article talking about how Nvidia is pushing into managed cloud.

DJX lap, right? Yeah. Exactly. Right. And so they they need uh a longer term business model. So I I think there will be a pretty significant place for A6. Mhm. You think about the different architectures of model.

I mean AMDs typically are optimized for sort of or have been in the past for mediumsiz models where the like the H100s are much better running the 100 billion or multiundred billion parameter models because of difference in CPU architecture versus memory architecture.

So I I think there's definitely a really big place and yeah and D is an interesting spot. I mean it's the biggest company in the world 3. 6 3. 8 trillion 3. 8 trillion at all time.

They they have to we were talking about it earlier they have to do more than a 2x again to surpass the Dutch East India company if you account for uh if you account for but they could do it. Yeah. Um, so anyways, Jensen still has got some work coming for Yeah.

to not be places, but at the same time, they feel like the company that's under potentially the most attack from from directly from startups and other hyperscalers building their own chips, whereas like we're not we're not hearing about Amazon taking on the iPhone and trying to build a phone to to displace the iPhone, but we are hearing about Amazon building tranium and inferentia to take away market share from Nvidia.

So, uh, it's interesting. It's like this company's never been stronger, but also, you know, that success breeds competition. Well, no doubt about it. And I mean, if you look at we went back and looked at the Nvidia PE multiple over the history of the company, and you can see there basically three spikes.

There's the first spike which was associated with gaming, the second spike that's associated with blockchain and now the third spike that's associated with AI. And you typically have these really really significant growth from say like a 25x PE multiple up to maybe like 70 80 sometimes 100 and then it falls back down.

Yep. Right. And Bill Gurley I think you know I wasn't investing during the time of hardware uh or hard discs.

He basically said the best time to sell a hard disk company was when it was ramping up because you knew that they would build excess supply and as a result of the excess supply you'd have sort of an inevitable crash. That hasn't been the case, right?

I mean, you saw Nvidia this year trade down initially and now I think it's Yeah, it's done really really well since then, maybe since February.

And if Jensen can diversify to a couple of other business lines, particularly along the inference y there's I mean I don't know what the TAM is for inference, but it's probably measured in the billions if not you know. Yeah. Yeah. Also the international stuff he's been on.

We're going to have to rename it Jensen's paradox. Jeb Jevans is gonna get it. It really is. Yeah. Sorry, Jeb.

on on on the question of like kind of AI value acral um where you know there there's a big shift away from you know the value acrew will acrue to the application layer I'd love to hear your take on that I also heard an interesting formulation this idea that in the next decade there will be uh one consumer AI company worth over a trillion dollars and 10 B2B AI companies worth over a hundred billion dollars and so you you can think about it as like they're equal market sizes, but one is much more concentrated.

And I wanted to know your reaction to that formulation of kind of like a bet on how the future or the market plays out. Does that feel directly right to you? Do you think it'll be radically different? What's your take? No, I think that's exactly right. I there's definitely a power law that governs consumer companies.

I mean, you look at the size of Facebook and Google, both of them consumer companies. The second place in both of those markets, we don't really talk about anymore. Even Amazon also power law outcome. Exactly. Right.

And so that's typically the way that those consumer markets work where you think about the distribution of outcomes in B2B it's much more say Gaussian right there's like a a bell bell shape and so I definitely think that's the case I don't you know we were I had heard initially well so like consumer investing has been a challenge I think over the last decade aside from a handful of names so we were we've been wondering about consumer investing try to understand what is the key use case I'd heard one of the most interesting ideas was the social network for AI bots that you would watch them talk to each other.

So, the three of us would train a bot. We'd each put it into a message uh a message board and then watch them talk. Uh and I I thought, you know, could that be interesting over time? But I do think we'll see explosive growth in some of these applications. Yeah, it's odd like bots are fantastic at chess and video games.

People watch chess and they watch video games on Twitch, but people don't watch bots play Counter Strike or Call of Duty. And I wonder if that paradigm holds for some, you know, I don't know, humanist reason.

I don't even know what you would call it, but just like preference towards Do people watch Do people watch live human players play against bots? I don't uh rarely.

I mean, you would you would watch someone like play through Mario 64 as fast as they could and that is effectively a single player game against bots, but it would be very odd to play a game of Call of Duty against bots because you could just play ranked against other people and you know all that, but there's still a lot of bots like floating around and people cheating and stuff.

So, yeah. So, I don't I mean, even in chess, right, with Deep Blue, nobody watches those those chess algorithms play against each other. Yeah.

could just watch Stockfish on Stockfish all day long and it would be like I guess technically impressive, but it wouldn't have the it wouldn't have the the I don't know the emotion, the texture of like Magnus versus somebody else.

Speaking of competition, do you lean in and try to compete head-on with platform funds based on what what you can offer or do you try to find ways to to not have to compete? We try to find ways not to have to compete. I mean, the platform funds have they run a very different strategy, right?

They're the big chip stack at the poker table, and so we can't play a big chip stack game. We do have the fund sizes to be able to write similar size checks at the seed and the series A. And that's a deliberate strategic decision so that the companies that we back are not at a disadvantage in terms of balance sheet.

But in terms of competing head-on where they're strong, I think it'd be foolish for us to embark on that journey. Yeah. Right. So, we have to find our own way of uh convincing founders we're the right firm for them to work with.

But that still feels like you're willing to say, "Okay, we're going to be competitive at seed, but we're not going to try to," but that's like one specific area, but and not, you know, holistic multi-stage. Well, we we primarily invested the series A. Average check size in fund one was say, call it 134.

Average check size in fund two is closer to 22 to 25. Mhm. Yeah. Yeah. And so we're able to write a pretty significant lead check.

Uh and and as I said, I think capital now just because of how much of it there is, it's if you're if you're a founder and you're competing in a category and you don't have a balance sheet that's at least commensurate with your competition, you're a pretty significant disadvantage. So we do want to mitigate that.

But we can offer a different kind of experience, right? We can if we're only working with say 10 companies in fund one, we can be a whole lot more active with those portfolio companies, particularly as they go from series A to series B. for sure. Um, where are you investing in blockchain onchain broadly? Yeah.

Well, I mean, we've invested, we've had a lot of success in in L1's in the past. We we uh was the second largest investor in SUI, which is a 40 to50 billion blockchain.

uh we've invested in a company called Alium which provides data to many of the largest financial institutions uh in the US including Stripe three of the top five Bitcoin ETFs and and others. So I think the database layer is really attractive.

You look at Ethereum worth seven snowflakes Q1 of last year produced 400 million in free cash flow which on a percentage basis made it the single largest producer of cash of any publicly traded software company and then on a on an aggregate basis sixth largest producer of any publicly traded software company.

So databases can be extremely valuable and one of the key themes in the last 18 months has been stable coins 15th largest buyer of US treasuries. So it's absolutely uh an essential market for us to continue to pay attention to.

And then the other thing that we're really monitoring, we believe that in the next 10 years, every major software product will have a web 3 component. And that's because today the cost of compliance across a bunch of different geographies is really challenging. You look at Germany has very specific data locality laws.

UAE has different laws. 26 states in the US have different laws.

whole lot better for a customer to custody their own data and selectively approve access and then software vendors no longer in the business of having to comply with all of these laws because they can computationally guarantee that they adhere to the policies.

So those are some of the big ideas I got some more AI questions. Uh I want to talk about uh different modes for foundation models that that are kind of playing out right now.

uh the value of data and maybe at Google they have YouTube that feels like a permanent advantage for them in V3 and yet we're seeing other models come out that are almost as good and I don't know if they just scraped all of YouTube but they got the data from somewhere then you have other narratives around we need new ideas we need the best researchers possible we'll pay a ton of money to get the best uh researchers to write the best most elegant algorithms or the code possible.

Maybe it's all in the application layer. So I'm interested in this relationship between uh the value of data going forward. Yeah, great question.

So I think initially there's a lot of value in having access to proprietary data like YouTube or maybe like cursor has with the initial or GitHub with the access copilot with the initial access to GitHub. I think over time the more valuable and rare data asset is the feedback loop.

What do you actually do with that with a prompt? Like is is the video of the polar bear with the Coca-Cola actually the right one? That feedback loop is incredibly important. And the more of them that you have, the better the models that you'll produce. Um, and I think that's also particularly true distill.

I mean, with distillation, you look at what happened with Deep Seek, right? they were just pegging a bunch of different APIs and then copying copying the results and able to to copy really quickly which was why it was really surprising to me actually to see the deep research API come out of open AI this week.

I thought after the deepseek moment there would never be a deep research API aside from perplexity which had launched it basically before uh because of the risk of distillation for some pretty valuable and expensive models. It's really I think it's in those feedback loops.

That's what I've learned uh in machine learning. And so the larger the number of users you have and the greater the telemetry you have there, the better the advantage.

So yeah, my my my like read on that is that um it's amazing that V3 has this cornered resource in YouTube data at Google, but I've been really really struggling to actually use that product on a regular basis because I hit rate limits even though I'm paying $500 a month.

It says come back tomorrow and you can do another three videos and I'm not really iterating on it creatively and then we we we were testing midjourney video and it renders much faster and there's a lot more collaboration and feedback there and so I mean Google scale is like just so massive that I I don't know if you should ever bet against it but it is interesting that I wonder you know there's there's folks in AI who are like scale pill there's other folks who are maybe feedback loop pled and I'm wondering if Google's just not feedback loop pilled enough because I don't I I don't care how on fire the GPUs or the TPUs are like provision more Google is my is my take but I don't know if I'm way off there like what's what's your read on it?

No no no I think you're exactly right. I mean they launched Google launched the Gemini uh command line interface which competes with the cloud code and I was trying to use it for two days and and I could not.

So I think I think they're t I mean and in their public earnings both Microsoft and Google have both said they're limited on the on the capex side. Basically, they can't produce the data centers fast enough. That's crazy. Um, but Google, I think Google is aware of this feedback loop phenomenon, right?

You look at the way that they subsidize fine-tuning for Gemini. The way that with the Gemini command line interface, the the free tier is substantially more generous. They're trying to they're trying to acquire as much usage as possible because they they understand those dynamics, but the data centers aren't there yet.

And then actually, you know what? Did you guys follow what um it Microsoft gave up some of the core weave lease and then Google took it on for the open AI training workloads which was really interesting. Okay. Did you see that news?

I yeah I I I I I saw the I saw the Satcha clip about you know he wants to be a leaser and then I did see they they uh they gave up the core lease. I didn't know that uh that OpenAI picked it up so quickly. Yeah. Through Google. So I think Google went and Yeah.

And so it seems like there are some at least training workloads going from OpenAI to Google. At least that's what I read. Interesting. Yeah. But I mean if OpenAI is starting to use Google servers, you can imagine there'll just be continuously increased demand and load. How much Oh, go for it.

Uh uh what are you following in reinforcement learning right now? How much of the RL story should be a foundation model story versus a new startup story? I think it's a foundation model story. I mean I a reinforcement fine-tuning is probably the technology or test time compute whatever you want to call it.

It's a technology I think has a has a really interesting future and the main reason is I can dial it up and down right. So today I can kind of control the classic models without reasoning.

But in the future you could imagine a CEO might have 10 or 15 or 100 times the reasoning budget that a junior employee might and on her case basis say I really want you to think a lot about this problem maybe it's you know strategic research project whereas if I'm just looking for a summarization but I want to make sure it's correct I'll give it a very very small budget.

I don't know in the product how we expose that to users in a way where they can do it intelligently. Well, it's already happening. I mean I mean it's already happening. It's just instead of money, it's time. So if I have if I have one minute, I'll hit 40. If I have 10 minutes, I'll hit 03 Pro.

And if I have 20 minutes, I'll hit deep research. And and time is money. And so I'm I'm I'm making that economic calculation right now. All under a $200 a month umbrella. But yeah, in the future, this could clearly just be uh like you put a price on it, right?

And you could say, I want that deep research query, but I need in five minutes. Am I willing to spend $100 on that query? Yeah. And that's and that's kind of what midjourney is doing with like the different tiers don't get you necessarily more total inferences.

It's more about the where you are in the queue and the speed of your response. Exactly. That's fascinating. That's a good point, Jordy. Sorry.

How do you think about the convergence of all software that the the the feeling that there's a convergence of all software companies into into just the the sort of chat interface? We've seen Wix make an an $80 million acquisition of this bootstrap company.

We saw Air Table launch a like you know text to app product which makes sense as a database company.

Uh, Figma launched their product earlier this year and they they they all make sense in different contexts, but do you think um do you think some of this is driven by uh like Wix for example just like FOMO and and disruption uh or do you just think that that is where software is going which is basically we have certain data in some places and there's platforms and everybody's just going to make instead of buying software you're just making it in different places.

Yeah, it's a great question. I mean, all of the existing companies are looking to retain their users and expand. And I think the dynamic here is like historically you've had like the last 10 years all been subscription and now you have AI usage, AI credits.

I would guess just given our usage on some of these platforms that the AI credit usage is probably two to three times the size of the subscription revenue fully deployed.

And so you need as a software company in order to hit some of these growth rates and I'm sure you guys saw like the Vzero chart, you need exposure to those kinds of use cases and those credits just to be able to to compete. And then you want to retain those users.

And you're right, like I mean, you know, there's the chat interface. There's like three or four interfaces that everybody's building. There's a chat interface, there's the app builder interface, and then there's like the workflow builder interface, the SDR equivalent. Like what how do I enrich a lead?

and you have like the project management companies building this the website internal tools they're all kind of trying to capture as much of that spend as possible I think like strategically if you're a young startup then the question is okay well how do I differentiate myself against one of these I can do it all platforms do I do what Salesforce did and take out an individual workflow and optimize end to end or do I actually need to build broad and horizontal uh and that I think and that's why a lot of these companies been pushing into verticals, right?

Whether it's like health care for uh prior off authorization or whatever it is. Yeah. And so there's a big strategic question there like of ultimately how do you compete?

The other dynamic here which I don't think has gotten enough press is you have the model layer, you have the UI layer and in between you have this sort of like context layer and the better an enterprise has its context like these are the kinds of customers that I engage with. This is the way my data is structured.

This is what happens in particular use case. the better that's structured, the better the ultimate performance of an AI product. Historically, that's been embedded in Salesforce and Workday and all those kinds of companies. The question is, does the enterprise actually want to control it because it is so valuable? Yeah.

Yeah. Do you think do you think uh memory it feels like memory is not a massive moat today? Maybe we're starting to see it with chat GBT. Do you think it can become something that's durable? How how do you think about moes broadly in the in the future?

It makes sense in the context of you know healthc care applications or or even in banking and things like that where there's like hardcore regulations and and you we can talk about network effects but but I'm I'm curious how how you see modes evolving well memory in the consumer ca in the consumer use case will be really significant right the more time I spend with chat GPT or Gemini the more likely I am to spend time so I think there's a reinforcement effect there I think within the enterprise guys, I suspect it will be the same thing, right?

Where uh especially across individual teams, maybe not across the entire enterprise, but it's pretty critical. So, it's it's it's essential.

I just can't tell whether enterprises will be okay giving that memory to foundation models and say, I'm we're going to standardize, we're Coca-Cola, we'll standardize on Chat GPT and give them the memory, or they want the ability to move across clouds and across models.

And that I think that materially changes the way that software will be built depending on who ultimately controls that memory. Yeah. It's interesting to think about the uh the the modes that exist due to the work required like the switching cost, right?

When you think about something like payroll where nobody's like, "Oh, I'm super excited. I'm we're switching payroll, right? It's like so much work. " But then you can imagine a world in the future where you have like agents that are just good at migrating payroll providers.

And if that workload becomes something that you can run in a 24-hour period and this this sort of agentic workflows that do all the heavy lifting that would have taken a team of people like thousands of emails and all these weird edge cases. It's like where where are the where are the moes? Where are the moes? Yeah.

Um I I have a followup on uh kind of startup metrics. Um we've seen a lot of charts of folks getting to 100 mil arr really really fast. Um there's been this like lurking overhang of like will there be massive churn.

Is this all exploratory budget from Fortune 500 companies that are going to wind up just going with like the Microsoft option ultimately or something? Or someone has some mandate to test every AI tool in every corner of their company. So they spend a bunch of money and then it doesn't stick around.

Do we need a new metric for that combines kind of like short-term churn with ARR to kind of give uh to help underwrite at the series A phase? Well, we're looking for the net dollar retention equivalent, right?

So net dollar retention in the past is just value of an account over a year and top quartile would say 125% after 22. Um I think that's the longer term metric.

The hard part and the reason it's so hard in early stage investing is you use that 14 months or 24 months to look at longitudinal net dollar retention and now these companies are growing so fast. Let's say a business goes from 0 to 100 in 12 months.

You might have a quarter maybe two to look at the net dollar retention and you can't determine whether or not you can annualize it but you have to value the company. And so that's really hard right that's a new muscle for early stage software companies. Yeah.

The other dynamic is sort of like this credit dynamic where you buy access to AI credits that's older. You look at Snowflake with the remaining performance obligation. Same idea today and people are annualizing.

It's not really ARR but at scale you have a sense for the overall consumption patterns of your your customers and so you can you can annualize it. But the first one is really hard and so like rewind 10 years a top company would go 015.

So at the time when you were valuing a company at 4 or 500 million, you had 24 months of longitudinal data and you could look at those patterns. Today, if you're valuing an AI company, it might be a quarter into its history before it hits 3 4 500 million in valuation. Yeah.

So you're basically taking the same risk uh at that round as you would be at the seed because you have the only additional information you have is whether there's great customer adoption. Yeah. Right. But you don't know whether how long it will last. Can you unpack the credit dynamic a little bit more?

I remember hearing stories from the dot boom. How uh internet service providers were investing or offering loans to internet companies and and that was kind of distorting the market a little bit or was like kind of mispriced potentially.

And then we've seen hyperscalers, they make balance sheet investments, they offer credits. What is the current state of affairs in the relationship between the the big players and the startup ecosystem that may be distorting or maybe something we need to keep an eye on in terms of valuations?

And can you can you explain in like more granular terms that you've been Yeah. Yeah. No, great question. Okay. So, in the dotcom era, you had this dynamic where you had these circular ecosystems.

So one company would raise venture capital dollars and then they would pay to advertise on another company and then that company would raise venture capital dollars and pay to advertise on another company. And so all of those companies as a whole would rise and fall with that little economy. Right?

You could take a uh you can make a similar argument which is Amazon invest in an AI company and says we'll give you a billion dollars but we'll give you 600 million of that in credits. And so we knows that you're going to take 350 400 million of that and immediately turn it around spend it on AWS.

So is that really an investment? Is that a closed economy? Can you value either the startup or Amazon web services based on that revenue? I think there's a big difference here which is just like within the dot era that ecosystem was really small and truly closed, right?

It was just a bunch of companies sending eyeballs back and forth and companies candidly were valued on overall traffic. Here Morgan Stanley ran a survey the enterprises half of AI spend is from existing software spend and then half of it is brand new.

So there's a lot of new GDP coming into the ecosystem which tells you it's it's more more resilient. Makes no sense. I'm going to say something and then I want you to react to it.

So in in the dotcom era, we had a bunch of companies valued at in billions of dollars with with uh huge losses and that became a problem over time. Now we have a bunch of AI apps that have a lot of revenue in a lot of different cases.

But then we have foundation models that are valued in billions of dollars, many billions of dollars with huge, huge losses.

And I've been trying to unpack, you know, how many of these companies the capital markets can support and for how long because we have some foundation models that generate a lot of revenue and then there's some elephants that aren't right now and they all have plans or concepts of plans for how to generate a lot of revenue.

Um, but they're not doing it yet. you know, uh not not to call anyone out, but you know, between SSI who's like not focused on launching, you know, a a product a commercial products immediately to Grock, which doesn't have consumer dominance or seemingly enterprise dominance right now.

There's there's a variety of players that are going to need to put up some uh you know, tremendous growth pretty quickly. you know, they they've