OpenAI 的 5000 亿美元估值｜Altimeter 最大押注

Do you think that the winner of the AI super cycle is going to be worth more than $1.5 trillion at the time? It is clear that ChatGPT has become a verb. OpenAI is on its path to being a winner of the consumer AI super cycle, and OpenAI is our largest investment in the history of Altimeter. Talent—this is where the war is now.

You mentioned Meta. They’re generating about $100 billion of operating cash flow. To put that in perspective, OpenAI just raised this massive round, $40 billion. You know, so they got—

Yeah.

—the text moment, they got the voice moment, they got the image moment. Just this year alone, you’ve had Operator, Deep Research, ChatGPT Agents, Codex, and GPT-5 now.

This was an interesting release because I don’t think it went the way many had expected. There was a lot of hype around it.

Mm-hmm.

It didn’t go so well.

We are halfway through the year, and they’ve had such big releases. I feel like that is the secret sauce; that is the defensibility. Speed is the only moat. Palantir was actually a very misunderstood business for a majority of its existence, in large part because of how customer-obsessed Palantir is.

Give us the rundown on Expo. How did you get into this investment? What do they do?

Well, the punchline is, the number-one hacker in the world is no longer a human.

Apoorv, welcome to Sorcery.

Delighted to be here. Thanks for having me.

Well, thanks for having me. We’re at Altimeter’s office.

That’s right. Welcome to Altimeter.

Thanks. We have a lot to discuss today. We’re going to go deep into OpenAI, the $500 billion valuation, and the philosophy around the parasocial relationship that’s formed between ChatGPT and its users. We’re also going to talk about the talent wars, all the money flowing into that, some of your recent investments, and then your time at Palantir—how that cemented your career and how you think about things as a forward-deployed engineer.

To start, let’s get into the $500 billion valuation. When did you first invest in OpenAI?

We’ve tracked OpenAI for a bit. Obviously, it’s a very important company. We first invested last year, and ultimately, what gave us the confidence was that it is clear that ChatGPT has become a verb. OpenAI is on its path to being a winner of the consumer AI super cycle.

If you study the past super cycles, let’s start with the internet, one of the biggest technology super cycles. It produced a large winner in Google, with $2.5 trillion in market cap. The next one, mobile, created a large winner in Apple, with $3.5 trillion in market cap. Social created Meta, with $1.8 trillion in market cap.

Enter GenAI. Do you think that the winner of the AI super cycle is going to be worth more than $150 billion at the time? Definitely, right? And so the real question was: Is OpenAI that winner? While it’s not a layup, we got increasing comfort that OpenAI, with ChatGPT, had caught magic in a bottle.

Mm-hmm.

It was a brand, not a technology, not a product. It was a brand—something that was impacting our day-to-day lives—and that’s what ultimately gave us the confidence. I’m happy to dig into the numbers.

Let’s dig into the numbers.

I’ll start with the consumer side of things. Consumer markets tend to be winner-take-most, if not winner-take-all, sometimes, right? You look at search: Google has more than 90% market share.

If you and I were internet investors in 2001, and we had the hard job of deciding which search engine to back, we could’ve backed Lycos, AltaVista, or Ask Jeeves, and the truth is, we’d be right because the numbers are all going to go up and to the right. All you had to do was wait until 2004, Google’s IPO, and you’d have gotten 99% of all gross profits generated in search.

That’s a little bit of how I feel about what’s happening with ChatGPT right now. They’ve publicly disclosed that they’ve got 700 million weekly active users. They recently crossed a big revenue milestone. I believe they announced a $10 billion revenue run rate in June, and a similar dynamic is going on in AI.

If you took the user count of ChatGPT on one side and all the other AI apps on the other side—Perplexity, Claude, Grok, Gemini, and the long tail of apps that people use—added them all up, and multiplied it by 10, it’s probably still less than ChatGPT’s user base. The same is true of revenue.

We saw this sign of a power law playing out—the same power law that played out in the internet, the same one that played out in mobile, and the same one that played out in social. The rough math, as we see it, is that they’ve announced 700 million weekly active users. Let’s say, on a monthly active user basis, it’s somewhere between 1 billion and 2 billion.

Revenue was announced at $10 billion, so, roughly, 1 billion users and $10 billion in revenue means about $10 per user per year, in round numbers. We think, over time, the number of users could go up to 2 billion to 4 billion users. On the monetization side, over time, it could go from $10 today to $60 or $70 per user per year, which is where Meta, Google, and other large consumer platforms are.

Together—P times Q—3 billion to 4 billion users multiplied by $60 or $70 per user gets you to about $200 billion of revenue opportunity over some medium amount of time. That is the consumer AI opportunity. That is what we’re playing for.

The other parts of the business—the enterprise business and the API business—will have a shot as good as any other. Enterprise markets tend to be more fragmented, not as concentrated. That’s where ChatGPT Enterprise is doing really well. We use it. Those are the numbers on OpenAI; that’s how we think about it.

And how much did you invest at the time? What was their valuation?

We entered last year in the $150 billion round.

What’s unique with you and Altimeter is that you’ve only invested in OpenAI, whereas other funds have spread their bets among other LLMs. How do you think about that concentration, and what is your investment strategy?

Our business is defined by power laws. The rough math of our business is that there are 5,000 companies that raise every year that we track. Five hundred of those—10%—will raise from tier-one investors. Fifteen of the 5,000 will return over 90% of the gross profits for a given vintage.

You might have heard of the Pareto principle—the 80/20 or the 90/10. Fifteen out of 5,000 is 0.3%. This is not 80/20; this is not 90/10. This is 99.7% and 0.3%. The power law is so sharp that very few companies actually return the majority of the gross profits.

When we think we’ve found one that is on that power law—we think OpenAI is definitely on that power law, and there are a couple of others on that power law in our business—we try to concentrate into them. We had a similar experience with Snowflake. OpenAI is our largest investment in the history of Altimeter.

What sets OpenAI apart from Anthropic and even Google?

It’s a great question. All of those are formidable teams with incredible products. Obviously, Anthropic is crushing it with its API, probably one of the most loved developer products out there, both with the API and the Claude Code product.

I would say they’re winning in their own right in the enterprise segment with developers on the coding use case and doing a great job at it. Google, same thing. There are a lot of great products they’ve built. I would say that a lot is yet to be seen from them. I feel like the best of Google is ahead of them, so there’s a lot more to come there.

At least, that’s how we see it. I think those are the 2 most formidable players, as you mentioned.

Going back, circling back a little bit more: 700 million weekly active users.

Mm-hmm.

They’ve developed a very strong consumer brand.

Mm-hmm.

They have a cult. They have a cult brand.

Right.

They’ve achieved this. It’s actually insane, the rate at which they did this. I think you published a chart on this.

Mm-hmm.

And then also the other charts from East Meet West.

Mm-hmm.

Those are great. We’ll add all of these in. I’m curious, from your standpoint as an investor and a user, and understanding their commercial efforts, how does consumer adoption bleed into enterprise?

Look, I think ultimately there are 3 legs to the stool on the usage numbers that you highlighted. It’s obviously the number of users, how much time they spend per day, and the retention of those users.

Those are the 3 legs of the stool that we measure to have a sense of the health of a consumer app. On the first one, as I mentioned, I think they are far and away larger than any standalone AI app today.

In terms of time spent, we’ve published this analysis: They are far and away ahead in time spent per user per day compared with any of the other AI apps today. Actually, not just AI apps. One of the analyses we did is that they are now larger than some mainstream consumer apps, like X.

Yeah.

Retention is pretty wild, actually. What we call smile curves are starting to show up in retention, meaning that, in a chart where you see the number of users that stay on an app after they’ve started the journey, typically you would expect an exponential decay of the users on that app.

For very few apps, the curve actually goes up, so it looks like a smile.

Mm-hmm.

ChatGPT exhibits a smile curve. The other apps that exhibit smile curves are Instagram and TikTok, and the implication is that users are finding so much value on this app that they want to come back to it.

How this bleeds into habit formation is, I have this beautiful thing on my personal phone that I use for my personal life, and I want that at work. To your question about how this bleeds into the enterprise notion, similar to the effect that good software or good devices have, you want to use at work what you have at home. That's the success of ChatGPT Enterprise.

How did this momentum start? I know that they have this crazy strategy, but I'd love to hear it from you. We've talked about this before, but could you go deeper into the momentum that they built and how this is—what's the formula?

Look, I think it all starts with November 2022. I actually have vivid memories of it. It was the same month that my wife and I got married, and we were at our honeymoon playing with ChatGPT, just obsessed with it.

I think they caught magic in a bottle. That moment is so precious. There is no amount of launch prep you could do for what happened in November 2022. Magic in a bottle, right? They had incredible viral adoption—hundreds of millions of users within a bunch of weeks. Over time, if you look at the charts, the user growth has been a steady climb, and you can see the kinks when they've launched significant milestones.

For example, in 2023, it was a steady climb. At some point, they took off the sign-in page, so you didn't have to log on to use ChatGPT, and you saw a kink. Then you saw the next big one in 2024 with advanced voice mode, right? Then you saw another one with Studio Ghibli and image generation.

Yeah.

So they got the text moment, they got the voice moment, and they got the image moment. I think the strategy is the grind of the feature velocity. It's so high. Keeping up with OpenAI product updates is a full-time job, and that's the strategy: ship fast.

Mm.

I feel like, done at a consistent pace, that is the secret sauce. Just this year alone, you've had Operator, Deep Research, ChatGPT Agents, Codex, and GPT-5. We're halfway through the year, and they've had such big releases. I feel like that is the defensibility. That is the secret sauce. Speed is the only moat.

This is where the story gets funny, right? What we saw with GPT-5 was that they turned off GPT-4, and that caused a lot of unrest with users. There was a story that really came to the surface, and Sam wrote a tweet about this, but it was really clear that people had parasocial relationships with the chat itself.

Mm-hmm.

How do you think about this? It's probably a little bit more philosophical, but how do you think about that connection with the product?

ChatGPT is no longer a website. It is no longer a product. It is a relationship that users have with technology, similar to the relationship that I have with this phone. I use this for half a dozen hours a day. If I forget this at home, I will know within minutes.

I think ChatGPT is starting to get to a point where it is such a steady companion. I don't buy anything above $100 without consulting ChatGPT.

Really?

I do a lot of research and spend so much time on travel, language learning, and planning. I think what happened over the weekend with #4oForever was a reminder that, like most important technology releases, this is changing the human experience.

The same thing happened with desktop and PC gaming. People got addicted to gaming. The same happened with mobile, and the same happened with social networks.

If I were to construct 2 axes, Molly, and call one the dopamine factor—how much sugar is in the product—and the other social, or network effects, I would say ChatGPT is actually quite safe and quite healthy. On the dopamine factor, there's no doom scrolling and there are no cat videos. The amount of dopamine that you have on ChatGPT versus, I don't know, pick your favorite, Twitter or Instagram, is a lot lower.

X.

X. The level of sugar is a lot lower.

Mm.

Whereas on the other one, network effects, if my friend Molly, Brad, Jamon, or Eric are on Instagram, I'm more likely to be there. If they're on X, I'm more likely to be there. But if they're on ChatGPT, I don't have an incremental reason to be there, so the network effects are not as sharp right now.

Despite that, this has become a dominant experience and relationship for people. I think that was the biggest takeaway I had over the weekend: “Wow, this is becoming a technology that people have a relationship with in a way that makes it incredibly sticky.”

Memory, in particular, adds so much context for your life that switching out of it would be so hard. Those were some of the takeaways that I had as that weekend unfolded.

I use Kalshi Markets within these conversations, and one of my favorite questions to ask is, “What do you think the best AI will be this month?”

Mm.

Usually, it's whatever their favorite is. Time and time again, unless it's a CTO, it's ChatGPT, and it's because of the context and the memory. You fill it over and over and over again. Not everyone is having a parasocial relationship.

Right. That's right.

Some people are extreme doomers, and David Sacks tweeted not too long after this, exposing the doomers and talking about the doomers. Guess what? We're not there yet, and what we've seen with these model releases over and over again is that the world is not ending.

Hmm.

We're not going to get destroyed quite yet.

Yeah.

It's going to take longer to build. Where do you think we are in the AI adoption curve?

Yeah. Geoffrey Moore wrote this book called Crossing the Chasm. I'll refer to some of the concepts that he describes in that book, but I think David's not wrong.

With any sufficiently advanced technology, you've got the tech enthusiasts, or call them the developers or super users, who will adopt it first. I think of it as technology that raises the ceiling and technology that raises the floor.

At the ceiling, you've got super users and tech enthusiasts, typically 10% to 15% of the population. At the floor, you've got the vast majority of the population—the majority, even the skeptics.

The way adoption plays out for really any technology is that I evaluate it by asking, “Is this new product release raising the ceiling or is it raising the floor?”

For example, let's talk about raising the floor. What the vast majority and laggards need is the removal of friction. They don't wake up to use AI; they wake up to live their lives, and AI has to find a way to be a part of their lives.

Advanced voice mode reduces the friction. I've got an action button dedicated to voice mode. It's incredibly useful. 1-800-ChatGPT, WhatsApp with ChatGPT, or Meta AI—those are feature releases that will accelerate adoption over the next 6, 12, or 18 months. Sorry, raising the floor.

Raising the ceiling would be users who are vibe coding. They're coding their own daily software. They're playing with chat, chat agents. They're playing with codecs. They're playing with cloud code, and there's a set of features targeted to them.

This year, this would be Deep Research; this would be the computer-use Operator. I think both of these will go up over time, and ultimately, the biggest experience will transcend beyond chat. Maybe we'll have variables. Maybe what Johnny Ive's working on is a wearable—I don't know what it is—but it could be a pair of glasses, a phone, a device, or a watch. How you engage with that could really remove the friction of communicating with a future interface.

So I think those are some of the big things that I'm looking forward to. What a time to be alive.

Maybe we've reached AGI. Maybe we're going to reach it soon. I don't know. The definition is unclear. But with AGI and superintelligence, we're approaching a lot of really fun philosophical, evolutionary, and existential questions. How do you think about humans' relation with technology—

Hmm.

—and where this might go?

AGI. I think, if I start at the start, I started my career as an intern at a firm called Rocket Fuel, and we were using AI to target ads to maximize some reward function. That predictive machine learning, or AI, felt magical to me.

Hmm.

Two years later, in 2012, the team behind AlexNet beat the ImageNet benchmark, identifying cats and dogs better than humans. I thought that was magic. A couple of years later, I was at Palantir. I was a software engineer, and I started using this tool called Kite. It was a coding copilot that would help you code faster.

I thought that was magic. I thought that was great AI. Self-driving cars, GPT-3, voice mode—I mean, every 2 years or so, there's been a moment when I was like, “Wow, that's magical AI. Isn't that AGI?”

I guess my point is that the bar for what AGI is is going up fast, and what we define to be, I guess, a lot of great intelligence is going up.

So now, to actually answer your question about how that changes society, I think—I wasn't around, but I'm told that there were teams at banks whose job was to calculate the interest rates, and then the calculator came through. I remember this. As a child, somebody gifted me an encyclopedia. I remember reading it cover to cover, but it's not really needed anymore. You could just look it up pretty quickly. You could look up really any fact pretty quickly.

I feel the same way about most forms of intelligence. I suspect that in 5 to 10 years, we wouldn't have to work to earn a living anymore. What do you do in a world where you don't have to work to live? I suspect relationships would matter a lot. I suspect, as Victor Frankl talks about in his book Man's Search for Meaning, creation would matter a lot.

You would get a lot of joy in doing things that you enjoy and finding meaning. But it's certainly going to look like a brave new world.

It's a great book. As we talk about the new releases, we have to talk about GPT-5. This was an interesting release because I don't think it went the way many had expected. There was a lot of hype around it. It didn't go so well, but my biggest observation—what I thought was pretty interesting out of this—was that Sam actually took action really fast.

Mm-hmm.

He immediately went back to work and was iterating off of the feedback. What do you think that means today for what it takes to be competitive as a founder, as a CEO?

One of the things that strikes me about Sam and the OpenAI team in general is that the single word to define them is antifragile. I've seen this for years now. Remember the time the blip happened when Sam was, for a hot second, not at OpenAI? It felt like, “Oh, what's going to happen to this organization?” This is obviously a pretty big deal, but I feel like they've gotten stronger.

Very few organizations get faster over time. I feel like OpenAI has gotten stronger over time and faster over time, and the GPT-5 release is not anything different. For an organization that is antifragile, they're taking incredible amounts of feedback. They're working with all the teams, gathering feedback fast. The doors are open, the lines are open, and again, it's a sign that evolving fast is ultimately the mode. Speed is the mode.

That's what I think about the GPT-5 release. I think it was a complicated release. It was the first time they released a system, not a single model. It has a model router. It bakes in a lot of thinking about how much compute resources to allocate to a query coming in, and so, again, it's a sign of antifragility.

Why was it so significant, technically speaking?

Yeah. I think each of OpenAI's prior releases has typically been a model or a set of models. This was the first time they released a system—a unified system that makes the decision of how much compute to allocate.

Should I give you a quick response to a simple query like, “What's the capital of the United Kingdom?” That could just be a lookup. You don't even need to scramble the jets for that one, as opposed to a more complicated query like, “You're an investment analyst, help me analyze, gather feedback for... summarize the feedback on all the users of Neuralink so far.”

Mm.

That would employ an orchestrator agent. It would scrape the web, gather all the data, summarize it, and probably package it together in a table. That's probably a 3- or 4-minute query. That delineation, which previously was something the user would have to make in a model selector, is now made for you.

Yeah.

That's a pretty significant change because, remember, the framework of raising the ceiling versus raising the floor: this is raising both. For the majority of users who don't really want to select a model, this is great. This is simple, clean UX, a Google-like bar. You just start.

The same is true for the ceiling. If you wanted to ask complicated questions, you don't have to think about, “Hey, what is the best model for me to do this? Is it o3, o4-mini, or o3-pro?” I think that was one part of the significance, call it the consumer experience.

The second part was that they focused very heavily on a couple of domains—coding, design, health care, and so on. The feedback for those focused domains has been pretty good.

Given they're a consumer product, do you think the structure of having these demos, these releases, is the most effective way for them to communicate, or do you think they should adopt other things? I mean, they do go on podcasts. Sam definitely did a podcast tour beforehand, but how do you think about that when talking to who their real customer is? Because then people bring up benchmarking and how they perform—

Yeah.

—and all of this sort of stuff that you know is not what Sally is doing on the couch before homework, you know?

Yeah. Ben Thompson wrote a great note on it, and I'll share something he wrote about, which is that once you have a large enough user base, it's really hard for you to satisfy everybody.

Mm-hmm.

You just have to make choices, and not every model upgrade or new product will please everybody. I think that's what happened here.

And on your question about benchmarks, I think the benchmarks are the starting line, but they're by no means the finish line. The benchmarks are kind of saturated, right? Particularly—

Do we need them? What is the point? What do they even mean? I feel like they're all arbitrary.

They're a starting line.

Okay.

Exactly. They're not the finish line. You typically start with a bunch of models that are at some, call it, ZIP code of a performance metric, and what you actually need is more proprietary evals.

Let's say for Sorcery, you had 100 to 200 custom evals that work for your use case, that you know work well. Then the question is: Is the model doing better than a human on those evals? Or you might have particular features, or just making sure there's no, like, you know, better tool use or error debugging, et cetera. There's a long tail of custom evals that I think do a better job now than benchmarks.

It was interesting to watch. I obviously love Calci, but I was watching the Calci chart on this for the end of the month, and during that, it was live. During that release, they had a couple of different demos. After the second demo, ChatGPT had the top spot, and then it completely did a reverse Uno and switched. It actually switched, I think, with Gemini, I believe.

Mm.

And so it was really interesting to see, but my observation from that is, of course, retail sentiment. If they were public—

Yeah.

—this is how volatile this would be. It's the most competitive market ever.

Yeah.

And then you have Twitter chatter, but then you have, like, okay, you actually went to CEOs, you did market research.

That's the reality. So there's a disconnect there.

Mm.

But it was an interesting observation into how—

100%.

—how volatile would this be if it was public?

Very volatile. It would be very volatile. I mean, there have been multiple moments, like the blip—the time last summer when there were a couple of departures, like the GPT-5 release. There are all these moments where I feel like the narrative, as Kelshi and tracks appropriately, completely flips up and down. It'll be a very volatile start.

Brad Lightcap captured some of the momentum from their customers with GPT-5. I want to go through some of these and then get something that you've heard from your portfolio.

Mm-hmm.

So he heard some really good momentum, and I'll show the tweet from Cursor, Lovable, Harvey, Amgen, Uber, and Notion.

Mm.

I know you have a particular instance. Do you want to share this?

Absolutely. This is another example of what you said: the Twitter narrative was so volatile, but we actually called the experts—the actual customers. We called a series of customers, from a bunch of customers of the 2 or 3 largest labs to customer service and cybersecurity. It's early—2 weeks in—but it's good so far, working with the team. As we said, it's a product in motion.

One particular one stands out: our company Expo. On their benchmarks, they saw a significant step-up with GPT-5, much better than any other coding, design, or math startup saw in their ability to find cyber exploits. Summarizing, the jump was from just under 60% to 81% when switching from Anthropic's latest model to OpenAI's latest model, GPT-5.

The jump was so significant. I think it was higher than what the team had expected—certainly higher than what the XBOW team had expected by quite a bit, and maybe even higher than what the OpenAI team had expected. They've published the results, and it's becoming pretty clear that AI is much better than humans at this thing called hacking.

We need to talk about Expo.

Let's do it.

So give us the rundown on XBOW. How did you get into this investment? What do they do?

The punchline is that the number one hacker in the world is no longer a human. It's a set of AI agents that work together, break through perimeters, and find vulnerabilities, sometimes even exploit them.

There's this platform called HackerOne, which is a marketplace where corporations can find ethical hackers to help them red-team or penetration-test their perimeter. Expo has been let loose on this platform, and a couple of weeks in, it became the number one player in the United States. Two weeks ago, right at Black Hat, it achieved the number one spot globally.

It's a pretty big milestone. There's a lot of things that AI has gotten better than humans at—obviously, image recognition and a bunch of games. Hacking is now there.

Hmm.

That's the significance of Expo. The one thing I will say is that while Expo will be a great business, it is a business that must be built for the safety of the Western world. Ultimately, cybersecurity ends up being a game of cat and mouse between the good guys, who make some progress, and the bad guys.

The ChatGPT moment of cyberhacking hasn't happened yet. It'll happen, and offensive actors will use AI to do so. We need Expo and players like Expo to help keep the world safe.

What are the applications?

Typically, the way you would have done this previously is this: let's say you're a developer team. You would write software, ship software, and then have a red team or penetration-testing team analyze all the vulnerabilities in your code. You might do this test once a year or twice a year, and you're typically rate-limited by the number of penetration testers there are in the world. There aren't that many of them. It's expensive. It's ultimately a manual exercise.

Today, a lot more code is being shipped. You're writing a lot of code using AI. And, by the way, that code is being written by models that were trained on the open-source code base with a lot more vulnerabilities, and so it's a lot more vulnerable.

The biggest application is going from this really analog red-team penetration-testing process that happens once a year to continuous testing that happens as you're writing code, as you're shipping faster, continuously. Why be rate-limited on the testing?

Hmm.

And so that's the biggest application, used by companies small and large across the spectrum.

This must have been incredibly competitive to win. How did you meet Uhay Dimur, and how did you win this?

The team at Expo is incredibly, incredibly talented. Uhay Dimur, founder and CEO of Expo, taught computer science at Oxford for a couple of decades, then built GitHub Copilot with Nat Friedman and the team at Microsoft. This is sort of his yin to his yang: as he was building GitHub Copilot and saw all the vulnerable code that was being written, he started XBOW last year.

We're obviously incredibly lucky to be working with them, and we haven't won anything. The work has just begun. We've just started partnering with them.

I think they were looking for missionaries, and our first meeting—my first meeting with Uhay—felt like our fifth meeting. We'd done quite a bit of work even before I met them. Ultimately, as a former software engineer, I could intrinsically understand the problem. How I reached out to them was also because it makes so much sense. I understand the root cause, I see how they were going to fix it, and how you build a large company doing it.

Our first meeting felt like our fifth meeting. I think we met on a Friday morning, and we decided to work together by Saturday evening.

Wow.

Yeah. Fast and furious.

That is quick.

Yeah. The shape of Expo is so unique. As you said, it's a small team—less than 50 people—all across the world. Uhay lives in Malta. Nico, the CTO, lives in Argentina.

They've won thousands and thousands of dollars in awards that are publicly available for anybody to win, but the effort is also ultimately to create training data and focus on our target segment, which is large enterprise. Think large financial-services customers, large insurance customers, large healthcare customers, and large technology businesses on one end. That's our customer base.

On the other end, it's, call it, small- to medium-sized businesses that want faster compliance. That's where we spend the tokens. It is true: we spend more on tokens than we spend on humans, and that is, I guess, the shape of an AI-native firm.

Okay. Well, we need to talk about where spending is today in AI and where this will be in the future. I know you have some great charts about this. They're not pyramids.

It is probably the biggest question in AI right now: where will the value accrue? We've been studying this now for a couple of years, and, as I'm sure the audience sees, the large CapEx announcements that the hyperscalers are making add up to hundreds of billions of dollars going into the ground, building data centers, acquiring chips, and building the infrastructure to enable all that.

We decided to study the same thing with the build of prior supercycles—with the internet, cloud, mobile, and AI—and put it together to see where value accrues over time. I'll summarize it for the audience here. In the cloud supercycle, you have about $400 billion of revenue being earned by the applications, let's call it layer 1.

You've got about $200 billion spent on infrastructure. This is AWS, GCP, Azure, et cetera. And you've got about $50 billion on semis. This is Intel, AMD, et cetera—the CPU layer. The shape of this is $400 billion, $200 billion, $50 billion.

Enter AI. NVIDIA alone, at the chip layer, earned roughly $40 billion in data-center revenue in Q1 last quarter. Annualized, that's about $160 billion. Let's say the total industry is $170–180 billion—maybe, let's say, $200 billion.

The infrastructure layer is $20–30 billion. This is inference revenue being earned by the hyperscalers and neoclouds and so on—a fraction of the $200 billion. The revenue being earned by the application layer—players like OpenAI, Anthropic, Perplexity, Glean, and so on—is on the order of $30–40 billion.

The shape is this way. Cloud is this way, and AI is so different from cloud that it begs the question: when—and if—will this invert? That's the biggest question we've been analyzing.

And so, we walk through a bunch of analogs. I've written about this publicly: it takes a while for this to invert. The canonical example is AWS. It started in 2004. 8 years later, in 2012, they got their first outside customer, Netflix. Eight years. For 8 years, it was a lot of the build—laying down the railroads—until you started to monetize it.

And so, that's what's going on right now. 90% of all AI dollars are actually in the semis layer.

There's one important line item that you missed: the talent.

The talent.

The billion-dollar talent.

The most important piece of it.

Who knew that if you went to college to be an AI researcher, you'd become a billionaire? That is creating an entirely new class of graduates.

That's right.

I want to get into this because I think it's so interesting, and we're seeing it from multiple sides.

Yeah.

We'll see it from the OpenAI side, but also the Meta superintelligence team.

Mm.

Can you break down what is going on here?

The ingredients of, call it, being on the frontier are compute and CapEx—we spoke about it—data, which is sort of accessible to everybody. We're saturated with the training data that exists. Talent: this is where the war is now. And if you look at Meta, Meta has an incredibly large balance sheet. They've got, I want to say, $70 to $80 billion in cash on the balance sheet.

Mm-hmm.

Every quarter, they're printing $25 to $30 billion in operating cash flow, which they might use to invest in CapEx, buy their stock, distribute dividends, or whatever. On an annualized basis, they're generating about $100 billion of operating cash flow. To put that in perspective, OpenAI just raised this massive round—$40 billion—and Meta is producing more than twice that—

Yeah.

—in a year. So if I was Zuck—gotta give it to him: founder-led organization, incredibly bold, not afraid to make tough decisions, big decisions—he's acquiring the most important ingredient, talent, the leading indicator of all value. And they've got plenty of cash to do it. Bold move by a bold leader. What a time to be watching this.

What a time to be alive.

What a time.

You were part of one of the most intensive, one of the most impactful, one of the most well-branded teams that came out of technology, and this is Palantir's forward-deployed engineers. I want to break this one down because it's insane. I was just looking at an Instagram post. This is where I get some of my news. Groundbreaking. So I'll read this out.

Right. Right.

Palantir ex-employees—this is one of the best cultures ever—have raised over $30 billion total, averaging $800 million per company, and more than 6% have founded billion-dollar startups.

Mm.

These are companies like Kalshi.

Yes.

Tariq. Tariq was a forward-deployed engineer. Sourcegraph.

Quin.

Quinn.

Yeah.

Yep. Ironclad.

Mm-hmm.

Adapar.

Of course.

Joe Lonsdale.

Yeah.

Does that count?

Yeah.

It counts.

It counts. Yeah.

Okay. And we have ElevenLabs.

Eleven Labs, Mati, Anduril—obviously a legendary firm—and Chapter. I mean, the list is long, and I promise you it's getting longer.

It's getting longer.

So there's going to be 10 times more of those. The world, watch out. The Palantir hubs are coming.

So what is it? How does this create $30 billion in raised capital? What are the principles of being a forward-deployed engineer?

Yeah, it's a good place to start. Look, I think for the 2 decades or so that Palantir has been around, for the majority of that time, nobody understood forward-deployed engineering. Palantir was actually a very misunderstood business for a majority of its existence, in large part because of how customer-obsessed Palantir is.

The shape of the business is like no other software business: median ACV is over $5 million. The reason forward-deployed engineering came to be is that anytime there's sufficiently advanced technology that cuts horizontally across a bunch of different industries—from defense, oil and gas, finance, consumer packaged goods, automotive, airlines, et cetera—what you need are experts who are really good at the technology, which was a lot of people at Palantir, but we knew nothing about those industries.

As somebody who worked in a bunch of these commercial industries, I did not know much about oil and gas, finance, or CPG. So we would spend a lot of time with our customers learning about their industries to deliver an experience that was soup-to-nuts controlled by us, because the industry was so used to buying software that, while it might be good on paper, would not actually move the needle.

The same is true right now. Why is forward-deployed engineering all the rage? It's because you have a very advanced technology in AI that cuts horizontally across a lot of different industries: software engineering, customer service, legal, health care, finance, and so on. Applying it appropriately to those industries requires a lot of context.

I think that was the principle behind creating forward-deployed engineers: basically, engineers who would go and obsess about the customer problem. Oftentimes, the customer problem was not a sexy problem. It was building a data pipeline. Sometimes it was dealing with organizational change. Sometimes it was figuring out processes. Other times, it was figuring out permissions and all sorts of things. I think that's why forward-deployed engineering came to be.

They've reached over $1 billion in quarterly revenue.

Mm.

They have forward-deployed engineers. They just keep on going on a tear. What is it about their culture that primes their employees for success?

Palantir is a very, very special place, obviously a special place for me. I'm biased. I think it's, again, a very important mission. It is a firm that, while a great business, must exist for the Western world and allies.

I would say there's a lot of great things, and I've written a whole article about it, but it all starts with the mission. The single thing that unites folks at Palantir, and what we do and what we did while I was there, is that there is something that's larger than your own existence that you're there for. It could be helping the US government, the allied governments. It could be a mission.

For that same reason, in a lot of the Palantir language, mission language was overused. Everything was a win or a loss, and that brought people together. You need that in wartime. You need something to unite the troops in wartime.

The second was an uncompromising bar on talent. For all the time that I was there, every single candidate we hired was interviewed by the founders. Imagine the operational complexity that creates, but that was required to keep the bar really high. I felt like we got, as Karp would say, a colony of artists that Palantir kept together.

Finally, I would say it's a really deep obsession with customer outcomes. Not sexy problems, not shiny problems, not vanity metrics, but really moving the needle for the customer in a way that you can only do when you own the entire process, soup-to-nuts.

And so now it shows up in the metrics at Palantir. As you said, the latest quarterly earnings—incredible quarterly earnings. I think revenue and revenue growth are lagging indicators. The leading indicators are what I described to you: what Palantir excels at.

They have an incredible company. They have a cult brand—software that dominates. Alex Karp, Shyam Sankar. What was the biggest lesson that you learned from Shyam?

Wow. Too many. Honestly, Shyam's a great leader.

Shyam's a great systems thinker. I'll share one anecdote that I learned with him. I remember we were in London, working on a customer deployment. As a young engineer and engineering leader at the time, I remember feeling that Shyam's entire mental model and focus was around winning.

It was not around a great plan. It was not around a great team, a roadmap, or a set of features that we were going to build. It was around winning. He actually wrote a lot of this in his article called “The Primacy of Winning” in Pirate Wires.

Mm-hmm.

I recommend folks check it out, but it is the single unifying orientation that Shyam has always had. He would say that winning takes care of everything. If you're winning, you don't have to worry about the other stuff.

But just to highlight the trade-offs, if you're oriented around winning, you will have to deal with a lot of chaos. You might not have a beautiful roadmap in wartime because the fact that you're moving so fast means that something is changing constantly. I feel the same way about a lot of what's happening in AI right now. Things are moving so fast that if you have a beautiful product roadmap, something's not right.

Mm-hmm.

You're probably not moving fast enough. I'd say that was the biggest thing that I learned from Shyam, and I think it's particularly relevant now. One of the other things he used to say is, “For forward-deployed engineers, ingest pain and excrete product.” It's one of those Shyamisms that's so true, basically to remind you that no task is beneath you, even if that includes plumbing data pipelines or change management. It's ultimately all pain for a bigger, more important mission.

There are so many things I could speak at length about—all the great things I learned from Shyam and his leadership.

What's the biggest lesson you learned from Alex Karp?

Another legend. Too many. One of the things that Palantir does, even to this day, I believe, is that before you start, they ship you a set of 5 books.

Mm-hmm.

A book on, obviously, getting things done and user experiences, which you would expect; a book on counterterrorism, which you would expect; but also a book on improv comedy.

Okay.

But basically, the biggest lesson I took away from Alex Karp is that he taught a lot about keeping this artist colony of artists at Palantir together. He taught a lot about the appropriate way for Palantir to be represented in the world.

For example, Karp really led all our go-to-market efforts while I was there. It was a very small go-to-market team, led by Alex, and he had some very particular thoughts on how we approached our customers. He's an incredible CEO and obviously crushing it at Palantir.

A true lesson in performance.

That's right. That's right.

Well, speaking of performance, Sorcery is sponsored by Brex.

And I'm curious, from your perspective, as we think about companies today in this supercycle—moving super fast and becoming really competitive—what are the key characteristics that you look for in founder performance?

Ultimately, we invest across venture, growth, pre-IPO, and public markets, I would say. There are really 2 questions that I'm trying to assess when looking at businesses. Is this a great business? Is this a generational business that will go create a large company, maybe a public company? Question number 1. Question number 2: Is this a great price? Is this a great structure? Does the math pencil out?

I spend 99% of my time answering the first question: Is this a generational business led by incredible leaders, inspired by a great mission? While every business is unique, it's hard for me to give you a template for all of them. I think the thing that I obsess about the most is asking the right questions.

Every time we study a business, there will be 2, 3, or 4 questions that define the business at that given time. Knowing the right questions to ask is probably the process that takes a little bit of time. The answers will typically either be knowable or not, and we can get to them. That's where I spend a lot of time.

That's pretty good. That's a good answer.

I appreciate it.

As we wrap up, we're going to do a quick-fire round.

Oh, fun.

Okay?

Fun, fun.

First, I want to know if you're bullish or bearish on these companies.

Hmm.

And then I'll ask a secondary question. These are all based on Calci's recent IPO charts.

Mm-hmm.

You can go there and see, you know, place your trade on who you think is going to IPO first. But I want to know if you're bullish or bearish on these companies.

Sure. All right.

It's 5 of them, okay?

Let's go.

Are you ready for this?

Let's go. Ready, ready.

Klarna, Discord, Databricks, Cerebras, Anduril.

Bullish, bullish, bullish, bullish. Wait for it: bullish. Bullish on all of them. Those are really good companies. Those are all really good companies. I mean, 4 of them are ultimate portfolio companies.

4?

4 out of 5. Those are all really good companies. Sorry, I'm not bearish on any of them.

What do you think the odds are that any of those will go public this year?

I think at least 1 of them will go public this year. I'm a big fan of the public markets.

Mm-hmm.

You can still have a lot of innovation in the public markets, as some of the greats have shown.

Any particular name?

Oof, tough.

Are we going to trip compliance?

That's right. We can always remove this, but as you saw, Databricks just announced this big round this morning. Had that not happened, I would've said Databricks, but that's probably not the number-one leader right now.

I guess a more interesting question I want to ask is: Do you think companies even need to go public anymore?

Hmm.

We've seen through secondary markets and these massive private rounds, over and over again, that there's just so much capital. Do they still need to go public?

I'm a big proponent of companies going public. It has an incredible amount of alignment with shareholders and great hygiene. Not to mention, it gives you pretty good feedback on the company's priorities.

I suspect there's a class of companies now, like SpaceX and Stripe, that have found a way to have repeatable liquidity for their employees at a regular cadence, let's say every 12 to 18 months. But I'm personally a huge fan of companies going public, and I think it'll always be a great way to build companies at scale, durably.

While admitting that some of the best businesses are now being built while they're private, the late-stage private asset class is just full of so many great gems: SpaceX, Stripe, Anduril, OpenAI, Anthropic. There are some greats there, so that's not to be discounted. But companies are staying private longer, building, building, building with access to capital, both primary and secondary.

Okay. We'll have to check back in about 6 months?

6 months.

Okay.

Right, right.

Well, Poorv, this was fantastic. Thank you so much. We covered so much ground.

This was fun.

Yeah. Thank you so much for coming on.