Tomas, welcome to the show.
Pleasure to be here, Turner. Thanks for having me on.
I know it's kind of funny. We had never met before last week and then we were at the Samil thing and then we're recording this podcast and then next week we're going to be at the Beyond Summit. So it's like three times right in a row.
Three-peat. Let's go. Let's do it.
And so, really quick for people who don't know, what is Theory Ventures?
We are an early-stage AI-focused venture firm. We invest anywhere from 1 to 45 million, typically in B2B software and infrastructure companies.
How would you kind of, I guess, summarize sort of the state of AI today? A little bit of an open-ended question, but how do you kind of think about everything that's going on?
All-out sprint. That's the way it feels. I mean, okay, so why do I say that? The first is there aren't enough GPUs for anybody. So people are sprinting to buy GPUs or rent them. I think the second thing is model improvements. A model only remains state-of-the-art for about 41 days, even though it's several hundred million or a billion to train, maybe less.
And then there's also an all-out sprint for customer acquisition. Buyers are the most open they've ever been to trying new things. And so if you can capture many of them, you'll have a big business. And then the businesses themselves are growing at unprecedented rates. So I think everybody is sprinting.
Yeah. I guess maybe the first one you mentioned, the GPU prices. What does that even mean? Like for somebody who is not super familiar with, I guess, any of this, like how would you just explain that to a smart person who is hearing this for the first time?
Yeah. So to run a machine learning or an AI model, you need a GPU, which is a particular kind of chip that does lots of calculations in parallel at the same time. Matrix math, it's called. And if you have a MacBook, you have one. In fact, you have an excellent one. Maybe one of the best that you can buy for your computer. And you can run small models, which are effective there.
But if you're a company that offers an AI product, you can't just buy a whole bunch of MacBooks. You need to buy servers that have lots of these GPUs in them. And those GPUs are hundreds of thousands of dollars most of the time. And the prices increase every week because there are not enough of them. And so you can either buy them and run them in your own data center or you can rent them from other people. And there aren't enough because one, the models that we're building are much bigger than we thought. They're now trillions of parameters. The demand for those models is much bigger than we thought, primarily because of things like OpenClaw and agentic tool calling and coding. And there's not enough memory. There aren't enough CPUs, which is really important, and that's mainly because most of these chips are produced by a company in Taiwan called the Taiwan Semiconductor Manufacturing Company, or TSMC.
Yeah. But you know, it's coming back, like, you know, like Nabisco. Do you know that Nabisco is an acronym?
National Biscuit Company.
Yeah. And there was this wave of like, you know, American Motors, right? I remember I met this gaming company, it's a total tangent, but they were called the Brooklyn Packet Company. It's like, that is an awesome name. And so anyway, we're starting to see some of these like locality-based names come back again.
Well, it's either that or you pick a, you like make up a word like we were at a point where people, their name would be like computify.io, like they just like, because you had to like make something up to come up with the name.
Yeah. To buy the domain name at some reasonable price.
Yeah. And now the same thing's happening with GPUs. The prices are going up.
Yeah. That's right. And then the other dynamic is just power and land. You know, can you find a place to build a data center? And it takes three to five years, maybe seven years to build a new power plant or to buy a jet turbine to power your data center. And then you need to build the data center itself, which takes 18 to 24 months. And so there's all these lead times, right? Atoms finally are starting to be really important in the world of software.
This episode is brought to you by Numeral. Numeral is the fastest, easiest way to stay compliant with US sales tax and global VAT. It's easy to set up and they automatically handle all registrations, ongoing filings, and their API provides sales tax rates wherever you need them with all the integrations you need. Numeral supports over 2,000 customers in both the US and globally, and they pride themselves on white-glove, high-touch customer service. Plus, they guarantee their work, and they'll cover the difference if they mess anything up. They're fresh off a fundraise, raising a $35 million Series B from Mayfield, which they're going to reinvest into building an even better product. If you want to put your sales tax on autopilot, check out Numeral at their new domain, numeral.com. That's numeral.com for the end-to-end platform for sales tax and VAT compliance.
This episode is brought to you by Flex. It's the AI-native private bank for business owners. I use Flex personally and I love it because they use AI to underwrite the cash flow of your business, giving you a real credit line. The best part is 60 days afloat, double the industry standard. Flex has all the features you'd expect from a modern financial platform like unlimited cards, expense management, bill pay that syncs with your credit line, and their new consumer card, Flex Elite. Flex Elite is a brand new Ramp-like experience for your personal life. A credit card with points, premium perks, concierge services, personal banking, cards and expense management for your family, network tracking across public and private assets, and a whole lot more. Fully integrated with your business spend, one card for your businesses, one card for your personal life, one card for everything. To skip the waitlist, head to flex.com and use my code Turner to get an additional 100,000 points worth $1,000 after spending your first $10,000 with Flex Elite. That's flex.com and code Turner for $1,000 on your first $10,000 of spend. Thank you, Flex. And now, let's jump in.
Yeah, I know like a lot of people are making these projections like we need 10x more next year and then the following year we need whatever the number is, 10x more from that. Like, can we even keep up? Like, what do you think is going to happen?
No, we won't keep up and we will overbuild. But okay, so capex spending, so those are dollars that are spent to build out data centers, will, I mean, we'll be at like 1.2 to 1.4 trillion this year and out of US GDP that's about low to mid twos as a percentage basis. So it'll be one of the... right now it's there's World War I, World War II just in terms of like large infrastructure projects. The railroads peaked at 7.7 and then there was the Eisenhower development of the national highway system and I think this year we will exceed that. So the question is, will we beat 7.7% and get to around 2.1 to 2.4 trillion in a year or two? I mean, yeah, it's definitely very possible.
It's definitely very possible. And you can see Google is outspending Microsoft in GPUs even though GCP is significantly smaller than Microsoft. And that tells me that there's some very sophisticated math to justify those buildouts. And as long as that math works, we will continue to build. And then the electric grid will have to be reimagined because it was never conceived to handle these kinds of volumes.
Do you ever reach a point though where this kind of plateaus, like with the railroads and with the national highway? Like it's just kind of, we built the highway. It's like, it's there.
Yeah. But I mean, the highways continue to grow, right? I mean, I don't know how, you know, 101 in San Francisco, they keep adding lanes and LA I-5.
But yes, I mean, we will keep going, keep going, keep going and very likely overbuild because it's impossible to determine when the economics either change or the demand changes, right? You could... why would the demand change? Well, we have the... every year I think Google for the last two years has said they generate 80% more tokens per GPU hour than they did the year before, which is doubling productivity. So that's a really big deal. Then you have segmentation. So you can say like, I don't really need a super fancy model to update my CRM. I can use a model that's running on my computer. So you could actually have a lot of the workloads going on your MacBook and that will happen.
But for now, I mean, I don't know what AI penetration is as a percentage of ultimate penetration, but I have to admit it's less than like 2%. You know, I have to estimate it's probably something like that, which means we can grow 50 to 100x from here.
So, do you think what's kind of going to happen then, it sounds like, is we're going to sort of hit a wall in how much we can expand the infrastructure capacity? So, we'll have to just get more efficient essentially with what we have.
I think that will happen. I mean, you have this token maxing era. Token maxing is putting a leaderboard in your company and seeing how many tokens you can use. And that's a lot of fun. I hit 250 million one day. It literally did everything through an AI. And you know, after burning a couple thousand dollars, you're like, 'Okay, that was fun.' Once.
What did you do? Like what you were telling me about that before, the token max, like what did you make it do? And then what did you accomplish from spending the $1,000?
You can't do it just by querying ChatGPT. There's no way. You might get to a million tokens that way. Parallelization is absolutely essential. So you have to create a plan for what you want an AI to do that particular day. Coding is huge because it's reading large existing code bases. And then anytime, I think the best technique is anytime you thought to do something with a computer, do not start with the browser. Do not start with your email client. Go to the AI and try to figure out how to do it with the AI. And then you can ramp. The challenge now is many of the clouds will stop you from doing this because they're like, 'Hey, you're doing it.' They'll hit you with a rate limiting error of 429 or 529 and they'll say, 'Too much, too much for you. Too much.'
So this could be like, you could say like, go and read the entirety of Twitter and just give me a summary of the best tweets or like all of Reddit or something, like just give it an insane task.
Read all my emails, listen to these 50 podcasts, transcribe all of them, download these 10 GitHub repositories, install them and see if they work, benchmark these four local models, which one's faster, download a bunch of startup presentations and analyze them, extract, you know, like all kinds of stuff. Find for me the 10 most important academic papers in the last week. And you're just... anything that comes to mind, like for the last two weeks, download all of the earnings transcripts of every public technology company, draft 10 blog posts and pick a best one and then critique it like an editor. And you know, you just... the hard... actually the harder part is not getting the AI to do it, the harder part is creating a workflow with that where you anticipate multiple steps and forks. And if you can do that effectively, you can go through a tremendous amount. And for those who are, you know, coders, there's this beta feature within OpenAI's Codex called SLGO where you just tell it, 'This is what I want you to achieve,' and then it will just continue going. And I was chatting with a friend who said he had his going for 18 hours and it worked.
And I did this on Wednesday. I was frustrated with a dictation app and so I just told Codex, 'Just replicate this app so it works.' And then, you know, whatever, 45 minutes later it said, 'Here you go.'
Yeah. And so that cost $15, but I'll only pay that $15 once.
And now you have this dictation app that you can use whenever.
Yeah. Dictation is another way of driving a lot of tokens.
Interesting. And yeah, because I do feel like we've all seen those headlines at this point where like, you know, Meta actually had to get rid of the leaderboard, I think was one of the most recent things because people were probably doing something sort of similar where you're like gaming the leaderboard.
Yeah. And I mean, you know, you can use jet fuel in your car and it'll go faster, but like, you know...
Will it actually go faster?
Yeah. Yeah. I mean, you know, racing car fuel octane is the... so this is fun. So hydrocarbons form all of propellants, right? And there's hexane, which is, I think it's C6H6, and then there's heptane, and then there's octane, which when you get, you know, is 87, 89, 91, and that's the amount of octane that exists within gas. And then if you get racing fuel, the octanes are much higher, they're 91, 95, 105. And so the amount of energy per unit volume is significantly higher and your car will produce a lot more horsepower if you put racing car fuel in it because there's just... the explosion is stronger. So you actually will go faster.
Yeah. I mean, at the top end, sure. I mean, you know, provided that everything in your engine holds together. You know, this is not an endorsement of like putting nitrous into your Prius and seeing if you can break 200 miles an hour.
Yeah. But I guess it's kind of the same thing with AI. It's like you may have a super powered model or you're like running all these different parallel tasks, but are you even doing them properly? And like, have you set things up right to make that even worth it?
Yeah, exactly. So let's ground this in some numbers, right? So a very small model might be a few hundred million to two, three, four billion parameter models. Those are great for dictation. They're great for grammar cleanup, transcription, those kinds of things. Then you have the next range of models, which I would put at like the 25 to 35 billion parameter models. And there, almost anything that you can do with a computer aside from coding, you can now achieve with one of those models. And they will be faster on your laptop than they will be talking to cloud or OpenAI, just they're just faster at it. And then there's another class of models that's like 120 to 150 billion. Those are not that often used. And then you have the state-of-the-art models which are trillions of parameters and they can do architecture and implementation of very sophisticated code or novel math discoveries.
And then those are the ones. So it's basically, I think I've seen when you follow like what sort of some of Dario and Anthropic's kind of positioning, it's like it costs a ton to train these models and as the revenue starts ramping, you know, you start getting profitable on these models, the older models, but then they're training the new ones which are even more expensive to make, which makes it so it looks like they're losing money, but the revenue gets even bigger and it's kind of like these like stacking super expensive to train workloads that eventually we get to a point where they just start making a ton of money. Is that kind of how this is going to go or...
It's, I think it resembles pharmaceuticals more than it resembles software, where you might spend three years researching a drug and then you... I think in pharma, I'm not deep in pharma, but I think you have 17 or 20 years with an exclusive patent on whatever the next statin to reduce cholesterol.
But with AI models, like we said, you have 41 days to be state-of-the-art. And you can see there's a company called OpenRouter, which is an open-source router of model calls. You can see the share shifting pretty significantly. So in November of last year, Grok, which is the xAI model, had pretty significant share above 15 percentage points. Today it has a lot less. And then you can see the share shift as a result of the subsidies from OpenAI or Anthropic on their different products or new models, right? Whatever, GPT-5.6. And so you don't have 17 years to recoup your investment cost. You have to keep running faster. The other dynamic that's really important is as these models and the training data becomes larger and larger, there's a great paper that talks about how the model performance will ultimately converge. And we're seeing this at the beginning. You could see like GPT-4 was significantly better at agentic tool calling. And then I don't remember exactly what the Claude model was that kind of caught up. And then Gemini was really strong in this particular domain, maybe it was math. And another one was great at Humanity's Last Exam, which is a knowledge retrieval benchmark. And now they've all... they're adding more and more benchmarks and they're all more complete.
And the differences between them are increasingly subtle.
And so ultimately then the advantage is just, do you have people using it? Like it doesn't even matter how good the model is sort of because they're all the same and it's more so, is it like the behavior or like, have you captured the workflow in some capacity or...
Yeah, I think we're going to get to a place where you reach a minimum viable intelligence where if you work at any company with a computer, there's what, I don't know what it will be, but some minimum, let's say it's a 30 billion parameter model in late 2026. And if you have a computer that you can run it, that's good enough. It's just like we're not giving everybody inside of a large company a state-of-the-art laptop. Like you have an i-whatever. You might have an IBM PC that's pretty good. You might have a MacBook Air that's pretty good, but you're not like, you don't have an M3 Ultra with 512 gigs of RAM to every... there's like a minimum level of performance that's good enough. And then you kind of upgrade every two years or three years depending on your company's policy. You can imagine we get to a very similar place with models where you say like, okay, like I have the current Gemma model from Google, it's 31 billion parameters and I can do most of my things on my laptop and that's fine for me. And then the frontier models, then they push into the domains of like high-performance computing, math research, materials research, chemistry, and really pushing PhD-level analysis further and further.
And you have some companies, Dow, Corning, pharmaceutical companies who are willing to pay a huge premium for that, but everybody else will use the sort of a mid-range model.
When you can build anything, Amplitude lets you know how to build the right thing. Use human language to get complex answers about your products. No more manually selecting events or building charts or dashboards. Just ask. Use agents to sense changes in customer behavior. Decide what's causing them and ask you if it's okay to fix it continuously in the background while you work. Get the answers you need while building directly in the tools you are already in like Claude, Cursor, Lovable, and more. And for the first time, understand if your agents actually work. Measure quality, debug failures, experiment, and measure their ROI with agent analytics. Amplitude. With AI analytics, all you have to do is ask.
You had a post recently, 'AI at Discount' is the name of the post, and I guess essentially the premise is Anthropic, if you actually look at it, looks like it's actually trading, depending on how fast, like if you take the rate of growth, it's actually trading at quite a bit of a discount considering how fast it's growing. I think on the other side though, everyone's like, 'Oh, these AI companies are so overvalued, blah, blah, blah.' So what is actually going on with how these companies are being valued by the markets?
Yes. Okay. So when a company is growing really fast in the public markets, many, many people, not everyone, but many people value it on what's called a forward revenue multiple basis, which is fancy words to say estimate the revenue in the next 12 months and then you take the market cap and divide it by that estimate of the revenue growth. It's called a forward revenue, EV to forward revenue multiple. Anyway, most software companies, and you can benchmark the fastest growing software company today at scale aside from a pure model company is Palantir, and they're growing at like 60, the projected growth rate is 68%.
Which is impressive for their size.
Yeah. That's like a midsize venture scale business five years ago, but this is a publicly traded, you know, billion dollar plus revenue business growing at 68%. And they trade at a very elevated multiple.
And then if you look at Anthropic, okay, so Anthropic year over year grew 30x and now it's closer to 43x. They went from a billion in run rate to 43 billion in a year, right? It's just like absurd. I mean, it just breaks all laws of business ever. Like it's just like impossible for that to happen. You'd think they were like committing fraud or like scam or just like it's fake.
Yeah. They added in the month of April, they added all of Snowflake's revenue plus all of Palantir's revenue in a month.
Yeah. I mean, just monstrous. So anyway, so let's say they, you know, they grew 43x. Okay. What do you think they'll do next? It's almost absurd. What do you think they'll do next year? But even if they're at 43 and then they get to 100 and they're valued at 900, well, they're kind of valued around like single-digit forward revenue multiples. And then you look at Palantir and it's valued at like, I mean, 30x, 35x. And so Anthropic's actually trading at a discount, which is kind of wild because the growth rate is 80% versus 4,300%.
Yeah. And is that like, does the market not expect Anthropic to continue to grow that fast? Like, is it just people saying, 'Okay, this is like not sustainable. It's still growing really fast, whatever.' But like, we're going to assume that this slows down? Or is it like a private market thing? Like, it's just because it's harder to get access to it and technically Anthropic can just price it whatever they want really. Like, is it just not a fair price that's like just out of whack or...
One, it's very difficult to project forward revenue. The revenue is non-recurring. I mean, some of it is contracted, but it's unclear. The third part of it is at some point the revenue growth will be limited by just total amount of GPUs. So Anthropic and SpaceX AI signed an agreement so that 25% of the Colossus data center, which is focused on training, will now be allocated to Anthropic. But, you know, at some point there just aren't enough GPUs. And so what happens to a business that's growing 43x in a year that starts to grow at, say, 30%? Which is still, I mean, a $43 billion revenue base. You're talking about adding 12 to 15 billion of revenue a year. I mean, it's like they just added that in a month and now they're going to add it in a year. Like, it's almost like unrealistic to think it's going to slow down that much.
Right? But you don't. So what are you underwriting? Like, what do you think it'll be? I don't know. And so then there's also the capital intensity. You need to build out these data centers. Do they need to raise debt? What does that look like? How much dilution are you taking as an investor? So there's a lot of unknowns. And then there's this trope, trees only grow so big. Have you ever heard that?
No. But it makes me think of like the law of large numbers or just like this company could never get that big.
Like, you know, if you look at textbooks, they'd say like there's no, like the law of physics says you cannot go from 1 to 43 billion in a year. Like it's just impossible, right? And so, you know, I remember when we had the first, I was growing up and anyway, I remember when we had the first trillion market cap company and that seemed staggering. And now we have four companies that are around the 3, 4 trillion, maybe 5. And I mean, you know, one interesting question to ask, like break the ice at a dinner party of people who really care about finance is when do we have the first $10 trillion company? It's, I mean, I don't know, it's definitely within our lifetimes. Is it 2030? Is it 2035? You have the devaluation of the dollar and then clearly these companies are growing really fast. But it's a, I mean, my perspective, it's inevitable. And so like, will Anthropic be the first $10 trillion company? But it's kind of hard to imagine, right? Like who's going to take the other side of that bet? I don't know.
Yeah. Well, especially when you consider two years ago they arguably had like no business. Like there like wasn't the thing that exists right now is just not there, right? And now it's suddenly the fastest growing of all time. So, but I think the interesting thing, I think you also wrote about this publicly recently, the strategy that they're taking is sort of similar to what Google did where you're kind of commoditizing the complements, I think is how you describe it. How do you think then about just the strategy that Anthropic's taken then with all the products?
Yeah. So Jason, I think it was Jason from SmartBear wrote this blog post in the early 2000s called 'Commoditizing the Complements.' And the idea is if you have a really good business, what you want to do is look at all the people who have businesses around you and make all of those products free so that more people end up using your products. That's called commoditizing the complement. You commoditize everything that's complementary to you. Okay? So let's make this concrete. So if you're Google and you make money when people click on search ads, you want to make it so that people click on as many ads as possible.
And what's in the... So, and I was at Google from '05 to '08, so I saw a little bit of this from the outside. But okay, so what did they make free? Well, it used to be paid for email. Okay, email was free. And then it used to be that you paid for video hosting because video hosting was really expensive, but then they bought YouTube and made that free. And it used to be that you would pay a license to have an operating system on a mobile phone. Then they bought Android and then they made that free. And then it used to be that you would buy a dedicated GPS device for you to navigate your car from one place to another. And then they end up buying Keyhole and making Google Maps and Google Earth free. And then they bought all these books and chopped them up and scanned all of them and put them in the index. So it was just driving more and more searches. Google Docs, same thing. So you're just using the internet more and by virtue of the fact that you're using the internet more and it was free, so there was less friction, you would go to Google more and then you would get more ads.
So if you're Anthropic, you can run a very similar strategy. You're Anthropic. You are selling inference. You are selling a prediction of an AI system. And then what you want to do is, okay, well, there was all this like workflow software the previous decade. Maybe it's legal software or finance software or accounting. I'm just picking categories at random here. But you don't really want to charge per seat anymore. That's silly because the amount of money people will pay per seat, maybe it's $500 a seat per month, compared to the amount of inference they'll buy, $2,000 a month, just give away the $500 seat and have them buy more inference. You'll make a whole lot more money and then you have less competition. And so, I don't know, you know, I don't have any... I'm just observing from the outside, but that's a very game-theoretically optimal way of maximizing when you have a really phenomenal business. You just want to make sure everything else is free. So there are as many queries as possible.
Yeah. And then why does that become so important then, paying for the inference that you mentioned? Like what does that even mean for somebody who doesn't know what inference is?
Inference is like when you ask AI a question or the AI does something for you.
It's like the process of like them doing the retrieval and doing whatever they do with the GPUs that they then give to you essentially.
Yeah, that's right. So, you know, all these systems are basically, they're word prediction machines. And so when you ask like, what is the capital of Italy, it's then creating a sentence where it's predicting. And Anthropic and the other companies charge by the word. It's called a token, but it's really effectively by the word. And so the more words that you... the longer the answer or the greater the amount of information you give the model, the more expensive the query.
So if you have a really large codebase and you have lots of... or if you have a really large legal case or if you have lots of PDFs and you want the AI system to analyze it, that's a very expensive query because it turns out that the input tokens or the data that you give the model is somewhere around 90% of the overall cost of asking that question most of the time or more, 90 to 95%.
And so anyway, inference is what the model is predicting to answer your question. And so yeah, I mean, you know, if yes, anyway, I'll stop there, but that's the idea. If I can just get the system to ask more questions. And it's not like, I don't want to give people the impression, it's not like somebody's sitting there and typing and asking about a particular case. It's let me create a workflow. So to analyze a startup, let's say it's like, okay, find the backgrounds of the founders, create a bottoms-up sizing of the market, map, help me understand the backgrounds of the team, compare this to other companies. And then all of a sudden the tokens that you use, the amount of information you're feeding to the model, number of words that you're analyzing and predicting explode. So really Anthropic's business model and like their strategy is just get people to do as much as possible in Anthropic products. Just like use it for things.
Yes. And this is why OpenClaw is so strategic because what do I do? So OpenClaw is a little assistant that lives on your computer and you can create a task list for an AI and you can say, find for me the best place to visit in Italy, go and, you know, schedule this with this person.
It's kind of all the things you described earlier that you can do with Claude.
Yeah. And so, but instead of doing them synchronously back and forth, you can create a huge long list and then those tasks can take 30 seconds or they can take three hours. And that's how you token max when we were talking about, that's how you jump from a million tokens a day to 100 million or 500 million tokens per day.
And Anthropic and OpenAI, it sounds like, want people to token max technically, like the highest margin version of token maxing, which is probably like a B2B workflow in some capacity.
Exactly. So, and you want people thinking that they no longer want to interact with a computer without AI, which I think many of, you know, many people in the Valley are already there. And because you can just do so much more, because I can just enumerate this list of tasks and then Claude or somebody else, some other model will just burn through that backlog.
Is there anything that you're not using AI for right now? Like on a computer?
There's some tasks. I mean, I'm on an Android and so it won't answer SMS messages because that pipeline's broken. But no, you really want to stay. I mean, there's this great book called Flow, right, which talked about how do you get into a place where when you're working, you're just directly connected. And it's kind of tied, there was a philosopher named Heidegger who talked about the design of tools. And if you think about using a fork, once you learn how to use a fork, the fork becomes an extension of your hand and you don't feel a difference. And I think working with an AI is like that in the sense of I can just tell it, right? I can use the most native... I don't have to learn to type, which is probably a dying skill.
Yeah, you can literally voice talk to it.
Yeah, I can just dictate what I want it to do. And then if it has enough information about the way that I work and it has access to my systems and I've helped it produce its own, make its own fashion, its own tools, then it can work as if it were me. And why would I? I mean, I'll give you an example. I was on a plane going to Atlanta and they told us in the waiting area there's no Wi-Fi, right? And so half of the people are relieved because they can watch a movie guilt-free. And the other half, the workaholics, like, 'Oh my gosh, what am I going to do for three and a half hours?' And so I sat there and I tried to find a really fast internet connection so I could download a local AI model because now I look at the laptop and I'm like, 'What? You're so dumb,' right? Or it's the same feeling when you get into a self-driving car and you start operating it and then you get into a regular car because you're someplace. It's like, 'Why won't you drive yourself?'
Yeah. Right. Yeah. So I have one more question on this sort of kind of like inference topic. So I actually don't know on like a technical level how this works. So because Anthropic's business, you could say they basically like sit on top of a cloud provider and they're basically like this layer on the cloud provider. How does that actually play out in just in the sense of like how that sort of business model works? Because like, do they need to build their own cloud provider eventually because they're just kind of like a GCP wrapper or like a AWS wrapper really at the end of the day?
They can decide, right? So you can own the buildings and the chips inside what they're called, data centers, which Google does, right? So let's think about this three-layer cake. There's the data center, and then there is the chip inside the data center, the GPU, the chip that's analyzing, and then there's the model. So let's look at those three layers. Google has all three. Google manages its own data centers. Google manufactures and designs its own chips called TPUs, tensor processing units. And then Google makes its own model called Gemini and Gemma.
And that is a great business. And then you can say, okay, Anthropic does not own the data center. It does not design its own chips. It just makes a really great model. And that looks a lot like Netflix.
So Netflix competes with Amazon. Amazon has Prime Video, but Netflix runs a lot of their infrastructure on AWS. Both businesses can succeed. There are pros and cons to each. And so let's look, you know, a great sort of segue is, okay, let's look at SpaceX AI. SpaceX AI has a data center. They don't have chips, so they're missing that middle layer. And they have a model. So there it's an Oreo, right? Where they're kind of, well, an Oreo with nothing in the middle.
Oreo with a vacuum.
It's like the... It's Oreo when you take it apart and lick the icing and then you stick it back together.
There you go.
That's right. Or you put somebody else's icing in it.
Yeah. So there'll be different strategies and you need different amounts of capital in order to do that and you'll have very different margin structures. If you can vertically integrate, which means own each layer, I think you will ultimately be significantly better off because you can design the chips and the data centers for your algorithms. Whereas if you're an algorithmic or a model company, you will definitely have a say in how those chips and those systems are designed, but you are not the only customer.
Fair. And you probably need to have enough scale to justify the investment into all your own stuff because it's not easy and it's not fast and not cheap.
No, it takes... I mean, it might take you... I don't... I mean, Google has been developing the TPU since 2012, right? Amazon has been developing their own chips called Trainium and Inferentia, I think, for the last five or six years. And it probably takes seven to ten years to get to a place where you are at state-of-the-art. You have executed enough cycles to really be there. And so at some level of scale, sure, if you're one of the five most valuable companies in the world, Apple has its own chips, all the M1 to M5 silicon that you and I run on our computers. That's proprietary and it's a big advantage.
So then the play is probably if you're Anthropic right now, like maybe at some point you need to start doing that, but it's really just get as much adoption as you can, get as much usage, get as much revenue to have cash to work with to now fund all this stuff. And to your point that you started talking about this, it's just a sprint. Go as fast as you can to get there.
Yeah. If you have a significantly better model, you will win share. And the opportunity cost is so huge and the willingness to spend is enormous because if your model is meaningfully better, you might add a hundred billion to your market cap in a quarter.
So I think it begs the question, where do you think is a good place to be investing in AI today? Like, is it over because Google is vertically integrated going to win everything, maybe Anthropic and OpenAI win on the edges? Like, is it wide open for startups? Obviously you're investing in startups, so maybe this is like a loaded question, biased question, but like, what do you think the opportunity is today investing in AI?
There are certain markets that are uninvestable because they are on the direct roadmap for the large companies that are incredibly well capitalized, right? So agentic coding, I think if you were to start an agentic coding company today, it'd be very difficult because it is probably the most important market and you have so many businesses whose roadmaps are pointed in that direction.
Is it the most important because it is so tied to that like inference thing that we talked about where there's just so much inference is flowing through that?
Yeah. Okay. So why is it... great question. Okay, so why is agentic coding such a phenomenal product-market fit with AI? The first is there is a lot of spend in software. So the market today is really big. The second reason is software engineers are largely very expensive. So there's a lot of labor spend as well. So there's technology spend and there's labor spend. Both are very large. The third is the demand for software, I would argue, is infinite. You and I, as we age, and all of us, will only use more software. We won't use less and it will become increasingly sophisticated building on the previous software. So you have labor spend, software spend, and a very fast-growing market with infinite demand. And then the last thing is it is a set of tasks that an AI can test whether or not the AI's answer is correct because it's just so objective, rule-based, and it's just like you know if you got it correct or not.
Yeah. It's like math. Either the equation resolves or it doesn't. And if an AI system can test that itself, well then sure, you can just let it spin overnight until it has satisfied all the different equations, right? Or all the different parameters that you've defined for the piece of software. So that's called a closed-loop problem. You can just have the machine spin faster and faster and faster. And so the combination of all those four, it makes it really great. Makes these systems perform exceptionally well in software.
Where is that not the case? Well, let's say, you know, we asked it to paint impressionist art. I mean, you and I can debate like, are Monet's lilies a zenith of impressionist art? You know, you can say no, Pissarro is. And so it's subjective. It's open loop. It's the blog post. How do you... when we summarize this great episode that we did together, there's no like objectively best blog post. And so that's not a closed-loop problem. And so AI has a much harder time because you can't just let it spin. You have to say, okay, this is... you have to apply judgment as a person and say that's enough.
So, and the reason that people, that all the big, the biggest AI companies are going after agentic coding is because it's ultimately like the biggest TAM and the biggest opportunity. So then you're basically, you're almost like accepting that you're maybe settling, quote unquote, for smaller, less interesting markets, but then there's an opinion to be had of like, well, these are actually still very big markets or they may be very strategic for these other reasons.
Right, yeah, it's like, you know, I mean, after like Google in 2006, would you have started a search company?
Probably not. Maybe DuckDuckGo. I don't know. I think DuckDuckGo maybe started around then and it's still alive, but yeah, like I don't know if I would have invested in it.
No, it's just really tough because you don't attack your opponent in the area they are strongest.
So, do you think that there may be some jockeying where, I don't know, a company that's not in agentic coding that we all know of and hear of every day just suddenly kind of emerges and has created a position to kind of, you know, ladder themselves in there or something?
Well, I mean, you know, Cursor, right? There are all the dynamics around Cursor and the brilliant business that they have built. So that's definitely an interesting one to watch. And you have Poolside, which is releasing a US open-source model. So now sovereign AI, AI that is limited to a particular country, has become a critical geopolitical issue. And so you have companies that are building models for India and companies that are building models for Japan and the United Arab Emirates. And so maybe there's a market segmentation. You say, I want to be the best agentic coding system for India. Well, okay, there may be a market segmentation there that makes sense. Just the way that you might have a vertical search engine to compete with Google that was focused on travel, right, for a long time, and that was a standalone vertical. So it's not to say that you can't segment and then compete within that segment. It's, I don't think you can just go and say, 'Okay, I want to win the United States agentic coding market as a model provider.' That'd be tough unless you really have a meaningful scientific advance or mathematical advance.
How do you think then about what are the opportunities that are interesting? Like, how do you figure out, is this a safe market, this other market, this non-incumbent market that they've already kind of captured? Like, how do you figure out what's worth going after?
Well, okay. So let's think about the markets where clearly they've demonstrated an interest, the incumbents, right? So agentic coding is one. The second one is health. Clearly OpenAI has a great team pushing health products, right? You have Anthropic launching a collection of skills on Monday of this week tied to finance and the automation of finance. So that'll be important. There's legal work that's associated. So the legal market is definitely in scope for them. Anything around like infrastructure and software automation is definitely core. And so those are some of the markets. I'm sure there are more. Security, clearly they will push. I don't know if the model companies... I doubt they will dominate that market in its entirety. They will be a supplier more than an individual competitor.
But there you have six markets where the direct competitive dynamics of the largest AI companies you must consider. And you can either invest and say, I'm going to... I believe a company is sufficiently far ahead that one of the incumbents must buy or partner with them, viable investment strategy. Or you can say, okay, there are ten markets they really care about and I'm not investing in any of those. I'm going to go pursue markets eleven through one hundred. And then I'm going to analyze each of those market dynamics. How many competitors are there? How many venture-backed competitors are there? How likely is it that the customer population adopts software, right? Like if you're a longshoreman, the odds you adopt AI, I think, are pretty low. But if you are, say, in the business of like back-office automation and you are like an insurance company or a third-party logistics company, pretty high. And then the question is, do the model companies care about that market or not?
So then what's your lens for thinking through kind of this whole like SaaS apocalypse? Like it's, I feel like we've gone through these waves of like people are like, oh, every software company's dead. And then I don't know if now if it's flipped or it's like they're not all dead. I'm not sure where we're at. It's hard to keep track. But in terms of that side, like if you're a mature software company, how do you think about the defensibility?
Yeah. Okay. So the public markets value growth. It remains the most important factor as an input to valuation. It's a 50 to 60% correlation.
You're saying the growth rate of a public company, 50% of its valuation is just depending on how fast it's growing?
Yeah. 50% is, yeah, correlation, it's explained by it. Yeah.
So which are the three fastest growing segments in the public markets? The first is security. The second one is data, and then the third is core systems infrastructure. All of those have tailwinds from AI.
The slowest growing ones are vertical software companies and then like productivity apps where some of them are seeing negative growth. And then I forget the third, but there really is a distribution. It's not — you can't look at it as all publicly traded software companies. There's a distribution. The faster growing ones are doing fine. And then the ones that are slowing or contracting will be punished.
One really interesting question, actually, this would be fun with you, Turner, is: okay, imagine you are in 2001 and the dot-com crash has just happened and you're looking at all the venture-backed and publicly traded software companies. They were building on-prem software. So you would have a CD and you would get a box of software at a store and then you would install it, right, as you're the head of IT for your company. And then after 2002, some number of companies moved to the cloud. Which companies were big during the box software era that transitioned to the cloud that survived, maybe even thrived?
Uh, so I was born in 1991. So I was about 10 or 11. So I'm trying to give you the perspective I would have as like a public market investor in 2001, 2002 at the time. I'm just trying to think of how I would even size it up. I guess looking back in hindsight, maybe Adobe.
But that is not really what they did in 2001, right? Like it was — they slowly transitioned to the cloud over the past 25 years, but I mean it probably didn't start in — it probably started in like 2005 or something.
Yeah. No, that's right. Okay. So Adobe is a great case.
Salesforce is post-cloud. So they launched directly on the cloud and then their banner was 'No Software,' which meant no on-prem software.
Okay. Maybe Oracle, but I don't know how fair that would be to count.
Yeah. So you're on it. So you have Adobe, clear market leader with Photoshop and InDesign and all those things. You have Intuit.
TurboTax and all that stuff. They made the transition, dominant in their category. You have SAP, right? A 50, 60-year-old software company.
That is like a common one though. The AI stuff is all going hard at SAP now. I feel like...
Yeah, that's right. Yeah.
So, can they survive again? We'll see.
Anyway, you keep going through this exercise and we were able to name about seven to eight companies that navigated that transition. Out of how many?
I have no idea. How do you know how many there were?
I mean, hundreds. I think it's order of hundreds.
So, are there characteristics of some of these? Like, is it that they had a very specific customer that they served? And were they like — did they have management teams that took the cloud seriously? Maybe like that feels like a big component of it.
