Skip to content

By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Sourcery with Molly O'SheaSourcery with Molly O'Shea

Anthropic Hits $2 Trillion in Historic IPO | Altimeter's Four Seasons of AI

Anthropic just made history, briefly surpassing a $2 trillion valuation in its IPO.

Apoorv AgrawalguestMolly O'Sheahost
Nov 20, 202651mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

See how

Transcript

  1. 0:00 – 0:42

    Apoorv Agrawal, Partner at Altimeter Capital

    1. AA

      You know, there's four seasons of the year. I call them Anthropic, OpenAI, SpaceX and Google, and honestly, the Anthropic S-1 has been a total treat. As a growth investor, the biggest things I care about is, is this a great business and is it available at a great price? So far, it has been a game of emperors, and we've seen the biggest players with the biggest balance sheet win. Fortune 500 customers that spend more than $100,000 on Anthropic say that it is less than 0.1% of their total expenses. We're in inning one of this whole thing. Adoption is early. Chips are gonna win. Data centers are gonna win. Clouds are gonna win. Models are gonna win.

    2. MO

      [laughs] Everyone's gonna win. [upbeat music]

  2. 0:42 – 3:40

    How big is the AI boom, really?

    1. MO

      Apoorv, welcome back to Sourcery.

    2. AA

      Thanks for having me. This is becoming an annual tradition.

    3. MO

      I know. Hopefully, we can make it, like, quarterly at some rate. I'm just gonna need to get my editors [laughs] to be on board with the increase in volume. [laughs]

    4. AA

      Let's do it. Let's do it.

    5. MO

      So we talked over a year ago. The markets were completely different then. They're changing every week, and they're gonna change again every week for the foreseeable future. I really enjoy talking to you and the team at Altimeter. We recently talked to Brad at the All-In Summit, but mostly around, you know, like Invest America and-

    6. AA

      Yeah, that was great

    7. MO

      ... the CAC scans-

    8. AA

      Yeah

    9. MO

      ... which was awesome. Um, but I love talking to you guys because you have a deep sense of the markets, seeing it both from the public side and the private side, the really high-growth analysis and then, you know, you also see the emerging players. So for our conversation today, I really wanna focus on, how do we make sense of this market? How do you eval- evaluate it as a private, close-to-public markets investor?

    10. AA

      Happy to do it. Uh, well, thanks for having me. It's a treat to do this again and, and chop it up. You know, there's a lot to talk about. Um, I'm sure we'll get into all of it. Um, but I thought we'd start today just, you know, 20,000 feet view of, like, what's, what's, what's going on in the state of AI, um, and maybe to frame it, um, for people who are not living and breathing AI every day, uh, we start with, how big is this AI thing as it compares to everything else we've done in the past, the technology, um, uh, all, all the technologies that we've built in the past? Second, um, where's all the money going? We hear big numbers on screens, big CapEx numbers, big data center numbers. Uh, where's all that going? Uh, and then third, what is it being used for? Is it being used for good, good ideas? Is, is there positive ROI to that? Uh, and frame sort of the, the, the, the AI landscape. Um, so I'll start with that. You know, the first thing I'll say is that the big tech CapEx on AI, it's a big number. Um, it's not only a big number, it's also growing fast. It's about $860 billion this year, um, over tripled in the last two years. Um, and to put that in perspective, it's, um, now rivaling the US defense spend, um, which is a big number. And, you know, to put that in context across the last 100 years of human history, um, if you look at every prior boom, uh, railroads and oil and gas and telco and, and shale, we're already way past what we spent in those, um, booms as a percent of GDP. Now, if we can share these slides on, on, on the pod, uh, readers might ask, "Hey, are we at the, um, top end of, of, of that railroad spend or are you- are we still gonna go keep growing?" And I think that's the right question to ask.

    11. MO

      Mm-hmm.

  3. 3:40 – 5:46

    Apps and models grew 100x

    1. AA

      So that takes us to the second question, which is, hey, let's look at what's happening with these dollars. This-- I mean, this is a phenomenal chart. This is, um... Dollar for dollar, you can see that there's a trade happening between the big tech CapEx and large semiconductors. And, you know, to an- analyze whether this AI build-out looks similar to some of the other technologies we've built, you know, internet, mobile, cloud, you know, we studied all of them. And, and the good news is that the start of most technology super cycles looks like it. Um, so here's a chart that talks about what is the revenue that's happening at the semis layer, the infra layer and the application layer. And, you know, we've got the cloud economy on the left side. This is a software economy that we know and love, you know, Salesforce and Adobe and ServiceNow together make up about, um, uh, $600 billion of revenue across, and that is supported by an infra spend of about $300 billion. These are the clouds, um, and a semi spend of about $80 billion a year. The AI economy on the right side looks completely different, right? And, um, on the top, you've got models and apps that, you know, make up about 140, 150 billion of revenue, uh, but a much, much bigger semis number at, at, at, at the bottom that supports them. So, so you might ask the question like, "Hey, is this-- how does this compare to prior, prior super cycles?" Um, and so, you know, the good news is that, that, you know, this is an example of what the mobile era monetization looked like, and you see the first inning, a lot of value accrued into semiconductors. Uh, the second inning, a lot of that in infrastructure and, and enabling devices. Uh, and then finally, inning three and onwards into the software and apps layer. And so this is something similar to what, what's, what's playing out in the AI economy, which is you're seeing, you know, while the economy has grown 6X in the last two, three years, um, the apps and model side of it has grown about 100X in the same time. So the economy's grown 6X. The apps and models has grown over, o- o- about 100X. Um, let me pause here. Does that all make sense? Um, any follow-ups to that? If not, we'll talk about where is all this money going? What are people using it for?

  4. 5:46 – 8:47

    From chatbots to agents

    1. MO

      I mean, I think it's kind of wild to- Put the numbers out there. People have been trying to measure, and we'll see these measurements over and over again in different ways-

    2. AA

      Yeah

    3. MO

      ... of how AI is modeled out and what this economy means, and where are we in this cycle. I remember, like, Brad and the team at Altimeter back at Coatue's East Meets West, like maybe two years ago or something, just talking about the early innings. And it-- this was like ChatGPT's inflection point.

    4. AA

      Yeah.

    5. MO

      Very early on.

    6. AA

      Yeah.

    7. MO

      Even last year, like I had Elad Gil on the episode, uh, in like October or so, and this was before Claude really took off, before these agents took off.

    8. AA

      Mm-hmm.

    9. MO

      And so we're not, we were not even in the, like, inference-

    10. AA

      Mm-hmm

    11. MO

      ... zone yet.

    12. AA

      Yeah.

    13. MO

      And so things have been evolving faster and iterating faster, and we're, like, at this precipice of, like, ascendance-

    14. AA

      Yeah

    15. MO

      ... that has really-- like, it's interesting hearing the railroad-

    16. AA

      Mm-hmm

    17. MO

      ... analogy. It's like, how do you layer those time horizons against each other?

    18. AA

      Yeah.

    19. MO

      And, like, where they, they layer off in, in terms of the build.

    20. AA

      Yeah. And you know, that's, uh, that transition is what happened, I would say sometime, um, last year. So you're right. You know, obviously the modern AI era kickstarted with the launch of ChatGPT in November 2022. All our listeners, you, me, use this product. It's a phenomenal product. Um, and you know, the number one use cases, as in the slide for this, call it chatbot era of AI, is find the answers, knowledge retrieval, plan your trips, et cetera. That's about 80% of the use case. But sometime end of last year or the start of this year, we entered, um, what I call the agentic AI era. You know, these are long-running agents, um, swarms of agents. Um, and the primary use case for those is, is very different. It's coding, agentic automation. And by the way, this is not just for technical users, you know. This is for sales, marketing, support teams, investment teams. You know, Altimeter investment team uses, uh, swarms of agents. Um, and it's, it's a dramatic shift, and now this makes up about 80% of the use, u-uses for, for agentic AI. And so I think the phase shift, um, is happening. Um, you know, we have-- there's some can-- I, I, I don't think this is the last phase shift we'll see. I'm sure there'll be a phase three of, of, of AI that people use. You're starting to see some of that. We have some i- candidates of what those modalities might be. You're seeing proactive agents with, with news and instinct and, and dots. Um, uh, you're seeing really long-running agents and, and swarms of agents. You're seeing, um, agents and particularly larger models be very good at performing surgeries on smaller, uh, models.

    21. MO

      Hmm.

    22. AA

      Um, so lots of, I would say, ideas for what this phase three might look like. We'll, we'll see where it settles, but, um, I think there's definitely a phase shift happening. Um, and then to, to your, to your earlier question, Molly, which is, um, you know, where are

  5. 8:47 – 11:33

    Top AI users grew 17x

    1. AA

      we on that, uh, call it inflection point as it relates to are you at the peak of this or are you, are you at the, uh, start of this? We think we are at the start of this. Um, I framed it in the slide. I call it the Tale of Two Cities, um, which is this is, uh, data from OpenAI that they've publicly released. Um, over the last year, year and a half, top 10% of their customers have grown their usage 17x, while their median customer has grown their usage 2x. And so the gap between a frontier firm and a median firm using AI is widening. And so what that implies is for, for, for investors like us and, you know, those in the audience, are your companies at the frontier? Are you, are you seeing acceleration into the trend? Or, um, are you, are you lagging behind? And you know, one of the questions that we debate a lot on the team is, are we accelerating into this, uh, AI adoption, or are we plateauing? It's kind of like if, if, if, if you ask me, "Hey, are we accelerating or, or plateauing on internet adoption?" I'd say we're, you know, it's near fully penetrated.

    2. MO

      Mm.

    3. AA

      Right. And we think on AI, we're, we're, we're very early. We're still accelerating as, as, as, as this chart shows. Um, and then we double-click into it. Um, and you know, we ask that question at the individual level. Like, okay, the enterprises are showing that, but the same is true in the individual level. Um, this is data from, from, uh, Coursera that they've publicly released. Um, top 10% of users make up two-thirds of all tokens. And the usage gap is wide. There's a power law u- in usage, but that is what pr- creates the opportunity for a catch-up for the rest of the population, um, as AI adoption grows. And I would say that's the biggest gap in AI right now, which is not everybody is using it like the power users are. Um, so I'd say that's, that's, that's how we think about the opportunity ahead. That's how we th-think the-- that's why, that's why we're, we have the confidence that we're early in the super cycle, uh, with long, long ways to go. One of the other proofs that we're very early in the adoption is Fortune 500 customers that spend more than $100,000 on Anthropic say that it is less than 0.1% of their total expenses. Adoption is early.

    4. MO

      Mm-hmm.

    5. AA

      We're in inning one of this whole thing. You know, it's gonna go from 0.1% of their expenses to 1%, and you know, that's how you get the 10x gains. Anthropic's been growing 10x every year for the last three years, and this is, this is, this is how it's done. They've got 1,900 customers that spend more than a million dollars, so it's not that there's one or two or three very concentrated customers. It's a very diversified customer base, and even the largest customers, for them, it's a small expense item.

    6. MO

      That was one of the craziest things I saw from the S1.

    7. AA

      Yeah.

    8. MO

      Hopefully, this goes out

  6. 11:33 – 13:31

    The four seasons of AI

    1. MO

      before Anthropic IPOs or who knows what happens now with the market, so [chuckles] .

    2. AA

      Yeah.

    3. MO

      But, um, to get into that further, break down the S1. Like, what were the biggest takeaways-

    4. AA

      Yeah

    5. MO

      ... for you?

    6. AA

      Yeah. As, as you and I have joked, Molly, you know, there's four seasons of the year.

    7. MO

      Mm.

    8. AA

      Uh, I call them Anthropic, OpenAI, SpaceX, and Google. And honestly- The Anthropic S1 has been a total treat. So as a growth investor, the biggest things I care about is, is this a great business and is it available at a great price? We can get into the price in a second, but this is a phenomenal business. This is a phenomenal business. Why do I say that? It's growing at rapid pace with long runway to grow. It's profitable and hypercap- capital efficient. This mix is very, very defined and we can- I'm sure we can get into each one of these in depth with numbers. Uh, but that was my biggest takeaway. This is a, this is a phenomenal business, um, available at a price that is, you know, 11 to 14 times forward, which is, which is a, a great price. And so I think it'll be a fen- a very successful IPO.

    9. MO

      So why do you think it's a great business?

    10. AA

      Well, as I said, you know, it's-- they're growing at incredible pace. They've grown 10X every year for the last three years. Q3, they did about $17 billion of revenue. It's the fastest scale-up in history. It's-- And it's not even close. Um, they've got about 12,000 customers that pay over $100,000 a year, and they've got about 1,900 customers that pay over a million dollars. And even despite all of that, they're early, right? Customers with a net revenue retention is 900%. Um, customers who pay more than 100K, as we discussed, and, and Fortune 500 say it's less than 0.1% of expense. So, you know, this is a very rare sauce, like I've never seen this before, a business that's growing this fast with no customer concentration and a long runway to grow. Um, now one of the biggest fears that everybody had was, "Hey, is this-- how

  7. 13:31 – 16:10

    The most capital-efficient scale-up ever

    1. AA

      much are they burning?"

    2. MO

      Mm.

    3. AA

      I would say that's my favorite thing from the S1, is this is a deeply profitable and capital-efficient business. Um, take a guess on how much they burned getting to, you know, whatever, 70 or so billion of revenue run rate.

    4. MO

      16.

    5. AA

      Less than 10 billion of re- uh, cash.

    6. MO

      Damn it. I thought it was 16 billion. [laughs]

    7. AA

      Almost. I mean, even 16 is like-

    8. MO

      Yeah

    9. AA

      ... wild.

    10. MO

      Mm-hmm.

    11. AA

      Right? Like they, they burned less than $10 billion to get to, to get to that number, and I mean, that has got to be the-- like this will go down in the muse- like history, like the books of history. Like I, I, I think it'll be very hard to beat this level of capital efficiency, um, while being deeply profitable. I mean, serving a dollar of revenue has gone down from 88 cents of compute to 15 cents. That's why they got, you know, 85% margins. Um, and they're monetizing their compute capacity, which is a scarce resource that's gone from like, you know, 16 a watt to about 60 a watt.

    12. MO

      Mm.

    13. AA

      Um, you know, a portion of their compute is used for inference and that, that, that portion of the compute that's used for inference is incredibly efficient. So, you know, they talk a lot about that in their S1. They talk about how they're orchestrating heterogeneous compute from TPUs to, to the Trainiums to the NVIDIA, uh, GPUs. And I think there's a lot of, uh, hard work that has gone into, uh, eking out maximum compute efficiency, um, which I think, uh, gives them a lot of, um, moats in the business.

    14. MO

      With comparisons to GDP and expenditure to defense budget, um, how-

    15. AA

      Yeah

    16. MO

      ... I guess the question would be like, where is all the money going?

    17. AA

      Mm-hmm.

    18. MO

      Like what-- how are you seeing this play out? W- how do you get to the second level?

    19. AA

      Yeah. I would say the, um, you know, the, the, the chart that we discussed, um, which is the number one destination of this capital is the build-out, is the data center build-out. This is, um, a, a good chunk of that, more than half of that goes into the, you know, IT equipment. This is your chips, uh, memory, et cetera. Um, and then the remaining into actually getting the, uh, uh, data center, uh, off the ground. Um, and so I think Chase from Crusoe had a great breakdown, uh, in, in, in my class about the exact split of the dollars. Maybe we can, uh, flash that up for-

    20. MO

      Yeah

    21. AA

      ... for, for those who are curious, and he did an incredible job of mapping it down to, to every single dollar. Uh, but that's a rough split. Um...

    22. MO

      Mm. For those that don't know, Apoorv taught-- And do you still teach

  8. 16:10 – 20:00

    Inside Apoorv's viral Stanford class

    1. MO

      at Stanford?

    2. AA

      Yeah. It's an, it's an annual thing.

    3. MO

      Stanford Business School? That's pretty cool.

    4. AA

      Engineering School.

    5. MO

      So, okay, explain some of this. Who are some of your guests that you've had on? I don't-- I-- We're gonna like take a little tangent, but like-

    6. AA

      Sure

    7. MO

      ... explain this.

    8. AA

      Sure.

    9. MO

      'Cause these have gone viral plenty of times, and I, I remember reaching out to you and being like, "Can we do it at Stanford?" Like [chuckles] your, your talks-

    10. AA

      Yeah

    11. MO

      ... keep on going viral on many-

    12. AA

      Yeah

    13. MO

      ... different like topics that you've been doing.

    14. AA

      Well, you know, that's, that's a labor of love. I, um, wanted to, um, recreate what I, what I wish I had when I was an undergrad. Um, you know, so what we do is, um, it's a, it's an intro to AI course that we go through the entire stack of AI, as Jensen calls the five-layer cake. Uh, we start at the chips layer. Uh, we then talk about data centers, infrastructure, models, applications, and agents, and we spend about two weeks on each part of the stack, uh, at the, at the class. And for each part of the class, we have the good fortune of being joined by, uh, leaders in that space. Um, so we had, uh, Chase from Crusoe talking about the data center build-out. We had, uh, Sunny from Groq talking about the, the chips and near chips. Uh, we had, uh, Ali Yazdani from Databricks talking about the infrastructure layer, um, folks from Anthropic and Chai on the models and, you know, so, so we, uh, do it from Baseten. Um, and so it, it, it's a, it's an incredible lineup and the goal is to, you know, for, for the 150 or so students that, that come to the class, which is, "Hey, you're about to dedicate your career to this. You're about to either go build a company or, or, or join a company or, or invest in a company, um, in a totally new economy."

    15. MO

      Mm.

    16. AA

      So at least you have a mental model to understand that part of the stack. Uh, each part of the stack is, is, is big and growing, and I'm a maximalist in that I think that they'll all be very good parts of the stack, and I think the chips will make money, and I think data centers will make money, and the infra will make money, and the models will make money. But just so you know the key questions on the minds of the, the, the builders of the frontier. Um, and so it's been an incredible experience. Um, I think we're gonna run it again next year and, and, you know, hopefully it's, it's, it's, um, uh, it's helping the students.

    17. MO

      This episode is brought to you by Brex, my favorite. You become what you spend on, and I refuse to spend my time on work that shouldn't exist. Expense reports, receipt chasing, and manual closes. The companies building what's next, from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call. They all run on Brex. Brex is the intelligent finance platform that combines cards, expenses, and banking into a single stack with agentic finance built in. AI agents that handle expenses automatically, enforce policy before spend happens, and close your books in minutes. That's why Sourcery runs on Brex, so I can spend time on building and not busy work. It's time to get Brex AF. Learn more at brex.com/sourcery. That's B-R-E-X.com/S-O-U-R-C-E-R-Y. Bye. From handshake to signature faster, Harvey, the AI platform built for legal work is trusted by 80% of the Am Law 100 and over 20% of the Fortune 500. Learn more today at harvey.ai. That's harvey.ai. AI needs more than chips. It needs power, land, and infrastructure. Zone develops next-generation data center campuses, partnering with AI companies, site developers, and technology leaders to bring compute online faster and at scale. Zone is building the foundation of the AI frontier. Visit zonefrontier.com to learn more. That's zonefrontier.com to learn more. With the recent US results that came out on data centers, how

  9. 20:00 – 22:36

    Who gets hurt by data center limits

    1. MO

      does this affect the whole stack? Where does this hit and hurt most?

    2. AA

      The, well, the biggest one is, uh, compute capacity that's gonna come online, you know? Um, rough- in like rough numbers, we think we'll end this year at about 35 to 40 gigawatts of compute capacity online in the United States. Um, there's been lots of estimates about, you know, adding another, call it 40 to 50 gigawatts of capacity next year. I think that that's probably the place where there's the biggest haircut. Uh, now exactly what the haircut is, exactly what the permitting and regulation will slow down is hard to say, but I'd estimate that like, you know, I don't know, maybe there's a shortfall of 10, 15 gigawatts to what we had forecasted for next year. Uh, I think that's a primary impact. Um, there's all sorts of other secondary impacts, uh, for, for hyperscalers, for the labs, uh, for, for NeoClouds, um, startups, uh, that we can get into, but that's, that's roughly how I see it.

    3. MO

      And then let's get into that. So explain how that'll affect the players on the board.

    4. AA

      Sure. I think, um, look, if you are, um, holding on to the largest pools of capacity, um, you're in a great position, right? And so those are hyperscalers today. Um, one of, one of my takeaways from the Anthropic S-1 was what a great business Amazon is, right? What a great business the other hyperscalers, Microsoft and, and, and, and, and, and Google and, um, SpaceX have been in supporting the, um, uh, AI economy. Um, so I think it's net beneficial for them because they're in a position of being the bottleneck where, where, where compute, um, uh, great, uh, uh, is bought and sold. Uh, but also, uh, as the big- as, as people with the biggest ambitions in, in, in building the largest data centers, um, their, their ambitions on adding capacity is, is at risk. Um, now you can take the other side of it, um, uh, which is, hey, what happens to, to the NeoClouds? What happens to new startups that were, let's say, not the top five in the line, but let's say 17th in the line.

    5. MO

      Mm-hmm.

    6. AA

      You know, I think for them this is existential, right? For them, this is, uh, a much bigger question than it is for the, for the hyperscalers, and so, you know, um, I think that's a, yeah, probably a much bigger impact, um, if you're at one of those businesses.

    7. MO

      So zooming back out to the Four Seasons of AI, whether it's Anthropic, SpaceX, OpenAI,

  10. 22:36 – 25:27

    Anthropic vs. OpenAI at the frontier

    1. MO

      and Google, how do you make up when it's their season? Like-

    2. AA

      Yeah

    3. MO

      ... we've, we've gone through this, and I would kind of argue, like, where's Meta in the picture? Like Muse really has propelled them back into the picture, whereas some of them might have fallen out.

    4. AA

      Yeah. Look, I think, um, it's a game of emperors, and it's clear that Anthropic is far out in the lead. OpenAI-- There's two frontier labs, Anthropic and OpenAI, and I would say that both of them have gone sort of neck to neck. Um, you know, maybe to give you a snapshot of, um, what's happening at OpenAI since we last chatted, I think we spoke about it last summer.

    5. MO

      Yeah.

    6. AA

      Um, I mean, OpenAI has shown incredible resilience and antifragility since, since we chatted, so ChatGPT continues to grow and is now over 1.2 billion users. Um, but more notably, you know, they went into code red at the end of last year. It started at, uh, at the end of-- at the start of this year. Um, and I think about two jobs to be done, building and selling. On the building side, they've, they've had an incredible set of models that have come out with Astra, uh, that have defined the frontier. Um, and in Q3, you know, I would say both the products, so the Codex product, the ChatGPT work product, um, the models, um, have shown incredible, uh, uh, performance. On the selling side, you know, um, they've added a couple of, uh, great leaders to the, to the roster. Dolly Rojek, um, former, um, CRO at Wiz, uh, one of the best, if not the best sales leaders on the planet, uh, Brian McCarthy, uh, former CRO of Courser, uh, from SpaceX, uh, and my friend Chad Beets. Uh, together, I think this is the best team you could ask for. Uh, and so all that, all those upgrades have led to an incredible, um, Q3 for OpenAI as they, as, as Sarah's publicly shared. You know, they ended September at about $50 billion of revenue run rate, more than what they'd guided for by the end of the year. Um, and so I would say those are the two labs that are on the frontier, as to your question. Uh, and then, you know, we've got SpaceX and Google that are showing, showing incredible signs of, um, infrastructure prowess. They've monetized infrastructure in- incredibly well. SpaceX, I'm sure we'll get into, and Google we'll get into. And then I won't count out open source, which is kind of the o- evergreen season that is, you know, um, you know, the rough mental model we have for open source is that about 20% to 30% of all AI spend will be on open source. You know, the majority will be on closed source, but there'll be a, a, a smaller part that'll be on open source, and that's also gonna monetize. And the place in the open source stack where the cash register rings is the inference layer. You know, this is Baseten and Fireworks and Modal and others, and we think those are gonna be large businesses as well.

    7. MO

      Explain that further. Why is open source evergreen?

    8. AA

      Sure.

    9. MO

      These are all different seasons-

  11. 25:27 – 27:57

    Why open source is evergreen

    1. AA

      Yeah

    2. MO

      ... but open source will forever be evergreen.

    3. AA

      Sure. Um, look, I think there's a l- uh, large enough set of customers who care about a confluence of control, um, cost, and sovereignty, meaning that, um, let's, let's pick the example of, of Cursor or, or Instinct or Decagon or Harvey. You know, pick a AI native company that wants to leverage their learnings from their interactions with a customer to train, uh, uh, to fine-tune a smaller specific model, uh, that allows them to serve their customer better, while, by the way, having a much smaller model, a much cheaper cost of inference, um, and fully harnessing, uh, uh, the data flywheel within themselves. You own that intelligence as opposed to renting that intelligence. So I think that enough AI natives today have made that choice. Um, and, um, as, uh, uh, Jonathan Ross, the CEO of, um, uh, Groq said, you know, there's about 35 companies that run 99% of all inference. You know, the top five are the labs we spoke about, Anthropic, OpenAI, Google, SpaceX, but the next 30 are AI native startups like Cursor, like Instinct, and I think this set is ma- today is making that decision, uh, and is, is wh- it's what's driving the, uh, uh, I would say the fastest adoption of open source tokens. Um, the next set is enterprises. You know, this is enterprises that you and I know, um, the, the banks, the healthcare companies, et cetera. I would say they are, uh, following suit. We are starting to see, um, some of them, uh, um, make that preference. In fact, OpenAI made an announcement a couple of weeks ago where they have a partnership with Baseten now. And so if you were a large customer, uh, of OpenAI, let's say you had a, making it up, $100 million commit on OpenAI, and you wanted to use a portion of that to open source, you could burn down your OpenAI credits. And so I think that is a proof point to your question, Molly, which is you now have the big labs, OpenAI, making the choice that, hey, you can take your spend on my models or, or, or on OpenAI and spend it on, on open source models. And I think that's another, you know, it's the reality that is too large to not, um, acknowledge, and I think it's very smart of OpenAI to rise above the level of being a lab to becoming a platform.

    4. MO

      That's really interesting. That's a, that's a very good point. In terms of the open source

  12. 27:57 – 31:50

    Open source is six months behind

    1. MO

      companies, you mentioned Baseten.

    2. AA

      Mm-hmm.

    3. MO

      So can you describe and explain the differentiation between Baseten-

    4. AA

      Sure

    5. MO

      ... and the Fireworks and the Modals and-

    6. AA

      Sure

    7. MO

      ... that kind of environment?

    8. AA

      Sure. Um, look, I think the, um, so, so the first thing on open source is a lot of people ask us the question of like, "Hey, a lot of these-- Are, are, are the open source models good enough? Are they close-- Are they, are they at the frontier?" Um, and the answer is not at point in time, but the best open s- the frontier open source model today is where the closed frontier was six months ago. So, you know, basically if, if you are willing to buy the frontier from six months ago, open source gives you that, and it has consistently, uh, delivered that. And you know, we have, we have incredible open source models, uh, both in the US and outside the US. Uh, actually just this morning, uh, you know, congrats to the Recursive team, um, launched a great open s- American open source model. Um, and so I'd say that's, that's playing out. Now, open source is not free, as open sour- you still have to go build the, uh, cluster that'll serve your, uh, model, uh, and that's, we believe, the primary place, uh, a lot of gross margin will accrue. Uh, and so that layer is the players that you mentioned, Baseten, Fireworks, Modal, all great companies. Um, um, you know, just like the clouds were, you have AWS, GCP, Azure, all great businesses, uh, no dominant player. I suspect that a similar market shape will emerge for, um, uh, the open source call it clouds.

    9. MO

      Mm.

    10. AA

      Um-

    11. MO

      The underlying conversation between all these conversations that I've been having with different investors or companies, we recently just did an interview with Fall-

    12. AA

      Mm-hmm

    13. MO

      ... and Fall is great. They're-

    14. AA

      Yeah

    15. MO

      ... a generative media company.

    16. AA

      Mm-hmm.

    17. MO

      Their point of view is that every company now has a different storefront-

    18. AA

      Yeah

    19. MO

      ... for compute, and I think this kind of goes back to that. We're like in the land of hyperscalers-

    20. AA

      Mm

    21. MO

      ... and neo clouds. How do you make sense of that? How do you make sense of that against, you know, the four seasons of AI when all of these-

    22. AA

      Yeah

    23. MO

      ... kind of end up looking like hyperscalers at the end of the day?

    24. AA

      Yeah.

    25. MO

      We saw SpaceX go public, and now they're really-

    26. AA

      Yeah

    27. MO

      ... talking about being an AI company.

    28. AA

      Yeah.

    29. MO

      They're talking about how over 50% of their revenue is related to AI.

    30. AA

      Mm-hmm.

  13. 31:50 – 35:18

    SpaceX is now an AI business

    1. AA

      and I mean, phenomenal earnings call for, from SpaceX two weeks ago. Um, where do we even begin? I mean, we thought SpaceX was a space business, but turns out it's a SI business.

    2. MO

      Mm-hmm.

    3. AA

      Um, you know, the rough makeup of SpaceX's revenue from their last earnings release is they do about, um, $60 billion of revenue run rate, half of which, uh, comes from their AI, uh, you know, their SI unit. Um, uh, about, uh, a third of it comes from Starlink, uh, and, and the remaining from their launch business, and actually this, this SI business is growing rapidly, right? It, it, it, it's, um... I think it went from, uh-- It's actually gr- I, I had a tweet about this. It added over 20% relative market share in Q3, so it's, you know, just i- i- incredible growth, and now they're guiding to great confidence in having 100 billion of run rate by the end of this year, and half a trillion. Is that right? Or more by the end of next year.

    4. MO

      Well, I think Elon wants to hit a trillion, and then $100 trillion market cap, so.

    5. AA

      I'm sure. I, I, I, I, I trust Elon, uh, on those. Um, but you know, I think that's the other path, right? So, so what is the other path? The other path is the, um, a- the infrastructure path. And so there were two players, I would say, who got the compute bet early and right, uh, Sam and Elon. You know, you know the Sam side of the house. Last October, Brad and Sam had an exchange on, on, on BGD Pod where Brad asked him the question of like, "Hey, you've got a trillion dollars of compute commitments, but 13 billion of GAAP revenue to support it." Uh, turns out Sam was prescient. OpenAI was prescient in procuring compute early and, and, and building those resources. Um, and, and, you know, Elon and, and, and team were the second. You know, they, they, they got a lot of great compute set up. Now, what they've done with this compute is they've started to s- um, rent it out, and not just rent it out like a neo cloud would. You know, a typical neo cloud would-

    6. MO

      Mm-hmm

    7. AA

      ... um, let you buy a lease, a long-term lease, uh, for one to three years. Um, but SpaceX is, uh, renting this at spot prices. And so the rough mental model for the prices of compute is, you know, if you got a long-term lease, uh, from, from CoreWeave or Nebius or one of the, the neo clouds, it would be available at like 10 to $20 a watt. Uh, short-term compute from, from SpaceX, that is, is roughly double that. Um, it's probably like 25 to $50 a watt. Actually, in the earnings call, um, their CFO said that it was on the higher end of that, so $50 a watt. Now, you might ask the question how-- who can pay for that? Well, Anthropic can, um, because Anthropic is monetizing data at, you know, 50 to $100 a watt. And so if you are in the, in, in the position, uh, of having a leading model that monetizes so well, uh, and you're out of compute, it's a win-win-win. It's a win for Anthropic. You have a product that is, you know, leading and you, you, you can make gross margin, uh, serving this. Uh, it's a win for SpaceX. They get to recoup their, uh, CapEx in less than a year, uh, whi- which is, which is incredible because most neo clouds take about two to three years to recoup their, their, uh, payback on their CapEx. Um, and, uh, you know, I think SpaceX has shown incredible, uh, innovation on that, on that model.

    8. MO

      Another thing we were talking about with SpaceX was how they've built out their business-

    9. AA

      Mm-hmm

  14. 35:18 – 37:28

    Elon's mastery of atoms and bits

    1. MO

      ... in atoms and bits.

    2. AA

      Mm-hmm.

    3. MO

      And you provided-

    4. AA

      Yeah

    5. MO

      ... like really good, uh, like understanding of this and, and, uh, examples.

    6. AA

      Yeah.

    7. MO

      So can you break that down?

    8. AA

      Elon's probably the best entrepreneur we know today, um, in orchestrating, um, atoms. We've seen, we've seen that with the SpaceX business. We've seen that with Tesla. We've seen that with Neuralink and Boring Company and, and a series of his, um, businesses over the years. Um, but now we're also starting to see that with, with atoms. You know, this is the SpaceX SI business. Um, and so on the bits side of the house, they're using their prowess in atoms, you know, their ability to build large-scale infrastructure, um, not only to build that, but to get it up fast, uh, and then business model innovation on top, uh, to earn max rent. Um, now their guidance of half a trillion to a trillion of revenue next year will, you know, we'll see, we'll see how that plays out, but the rough mental math on, on how I think about that, I mean, this would make it larger than the frontier labs that we spoke about if they, if they achieve this revenue, right? Um, how is that possible? I think it comes from their prowess in being really good at orchestrating bits and really good orchestrating atoms, um, and a combination of those. Um, you know, I- We estimate they'll end this year at about 3 gigawatts of compute. Um, they're guiding to ending next year at about 13 gigawatts of compute, maybe s- more than that, maybe less than that, we'll see. Um, and at the monetization that they have, you know, top end or, or even more than $50 a watt, like that's, that's how you get to over half a, a trillion in revenue run rate. Now, there's lots of questions in, in whether that can be achieved, of whether you can continue to run short-term, uh, prices. It's like, you know, renting your home on a long yearly lease versus renting your home on Airbnb. Like, um, there's obvious questions with that, um, whether will, will Anthropic renew and so on, but, uh, I guess we'll find out.

    9. MO

      And then where does Google come into the picture?

    10. AA

      Yeah.

    11. MO

      TPUs, let's talk about that.

    12. AA

      Yeah.

  15. 37:28 – 41:15

    Google Cloud is growing 103%

    1. MO

      How are you seeing them take up their season? Which season has-

    2. AA

      Yeah

    3. MO

      ... like where- how do you define which sea- Are they months? Like what are they? [laughs]

    4. AA

      Yeah. You know, I, I-- Google is one that I-- every time I think of Google, I say, "Don't forget Google." Um, I would describe Google right now as a tale of two cities. Um, on one side they've got a search business that is, you know, slowing down and, you know, it's a, it's, it's about $250 billion of revenue run rate growing at 12%. But you've got this cloud business that is half the size of search, but growing 8X faster. So I mean, what are we even talking about here? This is a business that is 120 billion of revenue run rate growing at 103% year over year. That is insane. And so honestly, I mean... And, and you know, by the way, this cloud business is what incorporates, um, TPU, their, their, their TPU product, their cloud sales. Um, Anthropic is a customer, uh, amongst o- others. I think there were some rumors about OpenAI considering the, um, TPUs as well. Uh, Oracle, I believe, was the first outside cloud provider, uh, and then now Anthropic has this, um, commitment with Broadcom to, to lease TPUs as well. So, you know, there's a, there's a growing, um, business inside of, uh, Google that is starting to look quite a bit more lucrative, uh, certainly in terms of its growth, um, than people had thought. Um, so that's why I call it the tale of two cities inside of Google. Um, and, um, a- as, as it is for the season, you know, I don't know. There's-- it's probably a change of guard and change of seasons inside of Google.

    5. MO

      How good is AI at real professional work? Mercor's AI Productivity Index, or APEX, tracks it with every major frontier model release. Built with partners like Harvey and Cognition, see where the frontier actually stands at mercor.com/apex. Today's episode is sponsored by VC-X by Fundrise, the public ticker for private tech, allowing investors of all sizes to invest in venture capital. Learn more at getvcx.com. Public is a multi-asset investing platform where you can automate your investing with AI agents. Create your own custom agent with a simple prompt. Agents can monitor markets, execute trading strategies, and manage your cash across stocks, bonds, options, crypto and more. Wanna buy at market close and sell the next day's open? Just tell your agent. Automate your investing at public.com. Founders scale faster on Deel. Set up payroll for any country in minutes. Hire anyone anywhere, get visas handled fast, and get back to building. Visit deel.com/sourcery. That's D-E-E-L.com/sourcery. Is the winner in AI cloud? Is it chat? Is it agents? Is it cloud?

    6. AA

      I, I think the answer is all of them. I think the answer is chips are gonna win, data centers are gonna win, clouds are gonna win, models are gonna win.

    7. MO

      [laughs] Everyone's gonna win.

    8. AA

      I think everyone's gonna win. But, but I think, I think the, maybe the question behind your question is like the relative scale-

    9. MO

      Yeah

    10. AA

      ... of winning and, and, and, you know, I think, you know, there's, there's lots of smart people who say everybody's gonna win. I agree, uh, but I think it's a relative scale of winning, and I think, um, honestly, so far it has been a game of emperors and you've seen the biggest players with the biggest balance sheet win. But I'm, I'm very hopeful, you know, as I, as I call them the blue triangles, I'm, I'm very hopeful for the app layer, for the model layer, for, for the startup ecosystem, um, as compute efficiencies start to kick in, as models get cheaper, uh, that more and more value will accrue, uh, at the top of the stack, which is, you know, in models and applications and, and new startups.

    11. MO

      So this trillion-dollar class of company is far

  16. 41:15 – 43:10

    The new growth benchmarks: 10x, 3x, 2x

    1. MO

      different than any other-

    2. AA

      Mm-hmm

    3. MO

      ... that had ever existed, and obviously companies that came from zero to a trillion in like five or six years. But I'm curious, so with this set of trillion-dollar companies that all are kind of servicing the same markets, how do you think about market comps-

    4. AA

      Yeah

    5. MO

      ... especially when there's new entrants that are just growing really fast-

    6. AA

      Yeah

    7. MO

      ... and we don't know when they plateau, if they plateau.

    8. AA

      Yeah. Look, I think, one, what was true in software, which is that, hey, you could triple triple and double double, uh, double, is no longer best in class. The-- I, I don't know if somebody's done a better benchmarking exercise, but like roughly speaking, I think we just went through Google, like if you were doing $120 billion of revenue run rate, there's now an example or two of businesses that are growing over 100% at that scale, so that's best in class. At the double-digit billion range, you've got, you know, uh, OpenAI ended last year roughly at about 20, $21 billion of revenue run rate, w- you know, projected to end-- we estimated to end this year at 70, 80. That's more than a 3X of that at double-digit billion dollar range, so, you know, triple at that scale. Um, and at the single billion scale, you've got several, several startups. You've got, you know, Cursor, um, uh, rumored to go from 1 to 10 this year. Cog in a similar, uh, level of scaling. Uh, s- you know, Baseten has, has grown 10X, uh, for the last two years at that scale. So I feel like at the single-digit billion range, that's the benchmark, right? And so it's- 10X at single billion, 3X at double bil- double-digit billion, and, and, and maybe doubling at, at, at triple-digit billion is probably best in class. But maybe there's somebody who's done a better, better market comps exercise of, of growth at, at, at that scale.

    9. MO

      Those are the metrics, or at least the measurement that you're using when you're evaluating new companies that are coming

  17. 43:10 – 44:49

    Why founders are the leading indicator

    1. MO

      to you, because we're seeing this every week. It's either... I can't tell the difference between... 'Cause I, I'm, I don't sit on this side anymore. Like, I don't-

    2. AA

      Yeah

    3. MO

      ... do this every day, but I can tell the difference between pure momentum plays and pure, like, freaking crushing it, we don't know how they're taking up all market share-

    4. AA

      Yeah

    5. MO

      ... type of companies.

    6. AA

      Well, look, it's ulti- ultimately, I mean, m- m- maybe take the question of how we evaluate startups, I think it all starts with, uh, a great-

    7. MO

      The founder? [laughs]

    8. AA

      ... team and founders, and-

    9. MO

      It's like [laughs]

    10. AA

      ... that is the leading indicator of all success.

    11. MO

      Yeah.

    12. AA

      Um, and by the way, I would say, I would say the same for Cursor, right? Like, Michael Trul, probably one of the best founders I've, young founders I've met in recent history, and, um, not a surprise that, you know, he's built an incredible business. And, you know, there's a series of others. We just partnered with one, um, um, my f- you know, my former colleague Bob McGrew from Arda, who I think is a phenomenal founder, um, uh, who, you know, started his career at PayPal, um, was then the second engineer at Palantir, um, led engineering, co-developed the four-deposit engineering notion, led research at OpenAI, and has now started this grand ambition which, you know, the guy's got a track record of just picking the most ambitious bets. You know, I think it's, it's talent like that that is a leading indicator. Obviously, the numbers are a lagging indicator.

    13. MO

      Mm.

    14. AA

      It's hard to foresee that with the, with, with, w- you know, uh, but, but what we can, you know, uh, at least look for is, is, is just great, great founding teams.

    15. MO

      So we've covered the four seasons of AI. We've covered the state of the AI economy.

  18. 44:49 – 47:56

    Why cyber is AI's next big market

    1. AA

      Yeah.

    2. MO

      What is left there to talk about? Like, what is, what are the biggest-

    3. AA

      Yeah

    4. MO

      ... questions you have?

    5. AA

      One of the biggest things I think about in addition to what we've spoken about, I think, um, the four seasons of AI, that's right, you know, Anthropic, OpenAI, SpaceX, Google, with an evergreen of open source. The other thing I think about is, um, w- what is the biggest industry that is consuming AI, right? So software engineering is the biggest workload, um, as we know about. Uh, financial markets actually, you know, um, the, the quant firms spend a significant amount of money on, on AI labs, um, but that's very concentrated. I would say the biggest emerging market is cyber. As you reminded me, Elon said, "Zero day, every day."

    6. MO

      Yeah. [laughs]

    7. AA

      Um, and I think a lot is happening in cyber that is worth paying attention to, um, in that cyber is a cat and mouse game, meaning that of- of- of- offense- uh, offensive actors are now, they now have this new toolkit, um, which is incredibly potent, that can go check every single door and window and make sure it's, i- if it's unlocked, find the most critical vulnerability, uh, and exploit it. Um, you know, one such vulnerability was, you know, um, uh, Uhadimoor and the Expo team found, uh, in Google Chrome. Now, Google Chrome is used by billions of users. About a billion users use it every day.

    8. MO

      I use that.

    9. AA

      I use that. Everybody uses that.

    10. MO

      What happened?

    11. AA

      And, you know, they found, um, a vulnerability that would allow an offensive actor to take control of your computer with just you clicking a link. You don't even have to have press any buttons. You don't have to provide your number or, or place a call, and that's scary. Now, this is just one event. There's a series of such events that are happening and, um, and, and I, and, you know, ba- And I think that that is... B- because the offensive activity has gone up a lot, I think defensive activity will go up quite a bit, and I think that that is, uh, an increasing area of, of, of, of use of AI tokens. You gotta match fire with fire.

    12. MO

      How come there hasn't been a breakout winner?

    13. AA

      Well, I think the edge that you get in using frontier models is so high that I think it'll be the frontier companies. I think it'll be the, you know, the AI frontier companies, um, today that's OpenAI and Anthropic, uh, that are, that are likely to serve that demand. Uh, now, there might be harnesses, uh, around them, uh, you know, um, uh, uh, like Expo and others that'll, that'll help you, uh, shape, uh, your cyber strategy and, and, you know, make sure that you're not, you're not doing, like, your once-in-a-year annual test-

    14. MO

      Yeah

    15. AA

      ... um, but do a continuous test, uh, as you, uh, update your software. Um, but I think, uh, capability matters quite a bit, and so you will want, most, most people want frontier capability.

    16. MO

      As we close out, what are you most looking forward to in the next 12 months? I always,

  19. 47:56 – 51:34

    Why the Anthropic IPO will succeed

    1. MO

      like, regret saying this question when I say it because it's gonna change in hours, like-

    2. AA

      I can't tell you about the next 12 months-

    3. MO

      [laughs]

    4. AA

      ... but I can tell you about the next 12 days.

    5. MO

      Okay.

    6. AA

      I am, I am, I am most looking forward to the Anthropic IPO-

    7. MO

      Mm-hmm

    8. AA

      ... uh, as I'm sure is, is many, many people, uh, in Silicon Valley and, and, and beyond. I think this is the best business I've had the good fortune as a, as an anthropologist, as a, as a student of the markets to study, uh, and I can't wait for, for them to go public and, and, you know, wish them all the best.

    9. MO

      Is anthropologist a pun?

    10. AA

      Good, good eye.

    11. MO

      [laughs] Well, uh, I'm curious about this before we go. Have you done scenario analysis on what happens if the IPO goes one way or another? Obviously, the market is contingent on many different macro factors, whether it's war or cyber attacks-

    12. AA

      Mm-hmm

    13. MO

      ... or credit or whatever. But- Anthropic has absolutely-- Anthropic and OpenAI, their premiums have absolutely propped up the private markets.

    14. AA

      Mm-hmm.

    15. MO

      People all think now they'll become a trillion-dollar company.

    16. AA

      Mm-hmm.

    17. MO

      I have no doubt they're a trillion-dollar company.

    18. AA

      Mm-hmm.

    19. MO

      But what happens if this hits the markets and it becomes-

    20. AA

      Sure

    21. MO

      ... volatile?

    22. AA

      Yeah. Look, I think may- maybe to answer that with a couple of numbers to frame the conversation, Molly, it's... Y- you've, you've seen their numbers that have come out. They're, they did about $17 billion of revenue in Q3. We think next year they'll do, let's say, 150, $140 million, billion of GAAP revenue. The rumored IPO price range, as you said, is about 1.5T to $2 trillion. At the low end, that's 11 times forward GAAP revenue. Uh, the high end, that's 14 times. Like, it's not a heroic multiple. It's completely reasonable based on current market comps. And so I think in the contemplated pricing range, I think it'll be a phenomenal short-term success, and certainly long-term. I think these are-- This is the private Mag 7. This is likely to be the, the leading companies, um, for the next decade and more. So I think we're looking at a very successful IPO. Um, now, in the short term, you might have choppy waters, but, uh, I'm, I'm quite excited.

    23. MO

      Apoorv, thank you so much for joining us. I know we're in this really funky studio space, so I normally will come up, and we have a studio at the New York Stock Exchange up here in SF, if you didn't know, like, New York Stock Exchange is here.

    24. AA

      Mm-hmm.

    25. MO

      But, uh, this is a lovely interior design studio called Lark by, uh, Pamela Jewell, and she helped design Revel. I don't know if you know Revel, but Revel is a software for hardware company.

    26. AA

      Yeah.

    27. MO

      They're based in LA. They're-

    28. AA

      Very fun

    29. MO

      ... opening an office up here. Um, Scott Warren's the founder, and so I was like, "I need a space." I'm like, "I know Scott's getting a new office. Where's the office? Blah, blah, blah. Does it have whatever?" And he was like, "I'm gonna connect you to Pamela." And then Pamela got us this space, and it's, like, amazing. So if anybody out there needs-

    30. AA

      Thanks, Scott

Episode duration: 51:34

In the uListen app, the transcript lights up word by word as you listen. Get uListen

Transcript of episode 0kFwIg3Q__4

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.