Leo Aschenbrenner 的 Situational Awareness 引爆｜Moonshot AI 以350亿美元估值融资35亿美元

It is so good to be back. We have the one and only Nikesh joining us. Nikesh, thank you so much for agreeing to join Rory, Jason and me.

I am a little apprehensive watching Jason and Rory there in their full glory. So let's see how this sort of plays out.

1. Airtable Sold to Bending Spoons for $1.285B

We have Professor O'Driscoll in the corner. But we're going to start on the news of the day, and the news of the day is Airtable, one of the big names from the last decade, which has been bought by Bending Spoons, a European-based company. Airtable was doing $485 million, growing 20% year over year, ultimately at a $1.285 billion acquisition price.

It's not the outcome that everyone quite wanted or expected, but it's what we have today. Rory, why don't I hand over to you first? I'm sure you've got some perspective.

Rory O’Driscoll

Well, we're all going to do the Airtable side of the analysis, but it's just worth pointing out the Bending Spoons side. I'm buying stuff at 2.8x while I'm trading in the market at 9 to 10 times revenues. They're going to do this all day, every day. As we said a few weeks ago, one of the big advantages is that they're in the game with capital and a traded currency to hoover up a whole bunch of this stuff.

On their side of the table, I totally get it. Obviously, on the Airtable side, there are a lot of comments asking, “Is this giving up? Is this a reflection of where SaaS is?” If we weren't anchoring off the $11 billion, it would be a great price. If you said someone set up a company 10 years ago, grew it to $450 million in revenues and sold it for $2 billion, plus or minus, they'd say, “That's an amazing outcome,” right?

But, of course, we all anchor off the $11 billion 2021 price, and it feels like a lowball. I think it's still a great value-creation achievement. We have a talented entrepreneur on the show with us, and we need to start with that. It's a great outcome.

Jason Lemkin

Can I ask Nikesh one particular question on it, Harry, if it's okay? I was going to ask you—I have a lot of interesting thoughts on this deal, whether Airtable matters in the age of agents and AI—but Nikesh is also one of the best dealmakers out there.

The shocker to me with Airtable wasn't the price, because I think it's low-market. The shocker to me is that no one else stepped up: no PE firm, no Thoma Bravo, no Vista. It's 20% growth at $500 million, with some AI dust going on it. Do you think there was even another offer? That's the thing. I just assumed someone would outbid them.

And you're asking me that because I'm, like—

Jason Lemkin

You're the dealmaker par excellence. We just happen to have the king dealmaker on the show.

Yeah. Look, I met Howie a few times. He's a great guy. He's built a great business. I think, to Rory's point, there's a bit of founder fatigue here. He's been through a lot of ups and downs, in terms of internally and in the market, but it's got a good product.

The broader question, which Jason hits on, is: What is going on in the SaaS marketplace? You see dislocations. Is this a pricing dislocation, or is there a fundamental change in the long-term growth rate that people expect out of SaaS? If it's a fundamental change in the long-term growth rate people assess, then the multiples are right now. That's, I think, where the market is grappling with this.

I think the people you mentioned, the PE guys, might have a full roster of stuff they'd like to sell to Bending Spoons, as opposed to wanting to buy against Bending Spoons. So I think they might be caught in the demand-and-supply problem right now. They have a lot of inventory.

Jason Lemkin

It just worries me. You see more deals than we do on the acquirer side, right? I mean, we see it on the target side. It's just, when you look at long-term growth businesses—

Yeah, say that again. Sorry.

Jason Lemkin

So, we stick to long-term, currently unprofitable, long-term growth businesses. But, yeah, I get it, and that might be the answer.

I mean, when Francisco Partners raised $22 billion to do deals sort of like this, right? Thoma Bravo was like, “We're looking for AI-infused B2B companies.” We could pick at Airtable. It did do that; it did infuse AI workflows and others. I don't know that it's growing, but it did that, right? Twenty percent growth at $500 million with AI workflows isn't nothing.

For all the founders out there looking to be picked up, this one would have seemed to me to be just above the fold. They should have leaned in on this one, not the one growing 8% and shrinking because it was destroyed by AI, even though it's cash-flow positive and has plenty of cash. All the boxes you check, right, are sort of there as an attractive target, and yet no one outbid them.

Well, look, Bending Spoons did buy them, so clearly somebody saw value there. Not everybody saw it, but Jason, I want to go back to something you said around AI infusion. I'm a little wary about this AI-infusion stuff, and this is what I talk to my team about every day from an operator perspective.

I say to them, “Are we Mercedes? Are we trying to sprinkle a little bit of AI in our car and say, ‘I have a little bit of AI’? Are we Tesla? Are we making sure that our car will drive the next 10 exits by itself and might have to grab the steering wheel once in a while? Or are we building a Waymo?”

The question back to you is: Did Airtable do a bit of a Mercedes action, a little Tesla action, or is it a little bit of a Waymo action? My biggest fear is that a bunch of people out there in the garages are getting funded by Harry and Rory, and they're going to build Waymos of the future, while we'll be busy putting lipstick on the pig.

Jason Lemkin

It's a tough one, right?

Yeah, it is. I didn't feel, Jason, that this was the kind of category that PE would sweep up.

Rory O’Driscoll

Yeah, kind of merging the 2 sets of comments: One is, definitely, you're trying to add some AI pixie dust, but fundamentally, you're a core productivity app. You've been the proselytizer, Jason, saying, “Oh, my God, look at all you can do with Lovable.”

If I wanted to build my own CRM and I wanted it to be personalizable, 10 years ago I might have used Airtable because it's way more configurable than Salesforce. But today, if I'm the nerd that wants to build my own CRM, I might just go to Lovable, Replit or Claude Code and bang it out from scratch.

So if you want to do that, I'd call you a person with low imagination. There are so many other cool things to build. The last thing you want to build is a UI.

Jason Lemkin

I agree, and, yeah, I am a person with low imagination. I cop that. But I agree. The point is this: What it means if you own a horizontal productivity app that's mainly individual-user? I just think that's one of the tougher categories for PE to get their head around.

Ironically, given the Evernote purchase, this is right in the Bending Spoons sweet spot. If you think about it, just like Evernote, the people who have stuck with this product are going to stick with it. They're going to apply their formula.

Maybe what I'm saying, slowly, as I process, is that capitalism works, and it ended up in the arms of the best owner of that product: the people who can take it and turn it into a cash-flow machine. I bet you, 2 years from now, it's doing $600 million, not $900 million, but I bet you it's $300 million of free cash flow.

Yeah. Or more, right? Just double the price, and your data is locked for 2 years, right? To your point, I don't think PE has the stomach or the willingness to do that, but I don't think it's what they do really well, right?

Jason Lemkin

But, yeah, I think their stomachs might be full. I think it's unfair to say they're—

Rory O’Driscoll

That's exactly right. If, as I often say to people when they show me a turnaround deal in my business, I say, “Look, if I wanted a shitty turnaround deal, all I have to do is look at my portfolio. I'll have 4 of them already. I don't need a 5th problem, right? I make problems on my own accidentally. I don't need to go actively, proactively say, ‘Let me get more of this shit.’”

You're exactly right. Anyone in PE who's done 5 software restructurings in the last 12 months needs a 6th like a hole in the head.

Nikesh's points, I mean, both his points are obviously great. The one on the Waymo versus the whatever—we can come back to the founder-fatigue one. That's a tough one today, because another way to look at Airtable is: Man, so early to no-code, right? Such a clever product back in the day.

Like, there were two products that I wasn't even smart enough to understand why they were cool at the time. There was Airtable, which turned a database into a spreadsheet. So I understood it, right? And then there was Notion, which turned a database into a document that I didn't even realize was cool. They were both so clever pre-AI, and they both took off in different ways.

But we don't really—we didn't—we don't need that no-code database today. Supabase is doing a million PostgreSQL databases a week on its own. And as a founder—what, founded in 2013?—it's tough to pick yourself up, and he already picked himself off the floor, right? He already did the layoffs, already got profitable, already rebooted, already went into founder mode again. I'm all in.

You're looking at yourself and you're like, "Can I do it another 13 years?" It's a tough one today when you've already done the whatever. I'm sorry, Nikesh. When you've already kind of checked the box and done it, and you're like, "God damn it, I got us to 20% growth," it's tough to not tap out. We're human beings; it's tough to not tap out.

Rory O’Driscoll

One of the things that's interesting here—and I've said this in the context of venture in general—is your comment, Jason, is basically that the technology trends moved on from the thing they built. One of the weird things about venture, with the holding-private-for-longer thing, is now the holding period of venture is longer than the technology platform cycle. So if you join halfway through, right, this is coming at you.

In the old days—20 years ago—this would long since have been public. It would be trading as common stock, and it would just get hoovered up on that basis. A lot of these late-stage rounds long since would have been public in another world.

I think the biggest fear right now is something that was started 10 years ago. Is it past the point of rebuilding, and are you better off building from scratch than trying to tinker with something that was built 10 years ago? I think that's where the challenge is.

That's interesting. You've done some recent acquisitions. It used to be my mental model that apps last longer because end users are pretty stuck in place and they keep them forever, whereas the infrastructure market moves quickly. But we've definitely seen some of the apps companies get stranded, right?

The infrastructure companies that have been able to evolve to link into the AI demand have been able to go from strength to strength. Look at Datadog. We were investors in JFrog, privately held—you guys are killing it—but you're co-attaching to the AI trend. The poor little apps companies have just got nothing to co-attach to to give them lift.

Jason Lemkin

I think that's a moment in time. One of the things we all know, but we don't talk about too much, is that AI still has a lot of false positives. There are too many edge cases that it can't solve. You still need grinders to solve the edge cases.

Waymo doesn't drive on the street without tons and tons of people being paid for labeling and tens of billions of dollars to find every tree and mark it. It's the equivalent of the Waymo: "Mark the tree. That's a tree, idiot." That stuff needs to happen for a lot of enterprise for AI to be effective.

Until we go through that process and leverage AI, putting all the hoops around it from a machine-learning perspective, there's life for infrastructure businesses. The choice we have in the next 5 or 6 years is: can we build all that plumbing, all that guardrailing, with machine learning and change the core engine to some version of AI at the right price? We survive. If you don't, then maybe Bending Spoons it is.

Bending Spoons it is. On Bending Spoons, it is, those kids. Okay, leave us on.

Now, final question before we move on: does this put a marker in the ground in terms of enterprise value for companies like this? If you're Notion, which raised $10 billion last time, how do you feel looking at this? If you're Monday.com—again, 2 products with similar motions—it was a shocker to see it, right? Especially the way everybody presented it, with enterprise value and all this. But Rory's right. It's market-low. It's market-low.

Rory O’Driscoll

We'll find out tomorrow.

Jason Lemkin

Yeah. True.

So far. What's the hedge fund guy? Sorry, we might not even talk about him. We already forgot about him.

Rory O’Driscoll

Leo.

Okay, I already forgot about him. I think one of 2 things is going to happen. We're going to forget about Airtable tomorrow because other stuff's going to happen.

Jason Lemkin

Yeah, we are. Or what I think might happen is this is the one where people capitulate—both founders and investors—where they say, "Look, folks have already had markdowns since 2021, but they're not consistent."

This deal, in many ways, was everyone capitulating. The late stage got 1x. The founders made $150 million, a lot less than they thought, but certainly enough to survive even today in San Francisco with rents up. Everyone said they capitulated to the markets.

I think we're also all in board meetings where we're seeing the opposite: 30% growth at 9 figures, where we're going all in. "Guys, we're at 20% growth." But we may see a quiet wave of Airtabling it: "It's time, guys. It's like how we did it. It's time. It's time to compete." He's a great founder, but that time has moved on, and it's time to capitulate.

That's a question right there. It may create more conversations, or it may be forgotten about in 3 hours. Not sure which. But it was a jaw-dropper for a brief moment in time, like losing most of your hedge fund during your wedding. But we move on.

2. Leo Aschenbrenner's Situational Awareness Blows Up

Well, we'll talk about that. That's a brilliant transition. Leo Aschenbrenner, the famed wonder kid who wrote “Situational Awareness,” which was an incredible memo that he then parlayed into a $225 million vehicle that at one point had $45 billion of assets. He really rode the wave, so to speak. He did it with 4x leverage.

In the past week, it kind of came crashing down, and then Ken Griffin and Citadel bought his public book for a reported $16 billion. Ken has made out like a bandit, reportedly making about $3 billion on the back of it in a very short amount of time. How do we think about this? He was the wonder kid of the AI wave.

Rory O’Driscoll

I think he was absolutely right on the trend—absolutely right on the trend—and remains to date right on the trend. In other words, the data just last week about capex absolutely supports his memo. So, conceptually right on the trend and then absolutely wrong on portfolio construction.

If you accumulate a portfolio of high-volatility stocks with 4x leverage, the math makes it clear that your probability of getting wiped out once is just very high. It's as simple as that: absolutely right on trend, absolutely wrong on portfolio construction. It's almost like it was inevitable.

I'm really sorry, but how do your investors let you get to that place? You made them 10x last year, and you probably don't question anything. He did amazing. Which of us really lets ourselves ask honestly, when someone makes you 10x, "What can go wrong?" I was like, "Can I put in more money?" That's what happened.

My wife said, when I was talking about it, "I don't want to see any schadenfreude." There was a lot of schadenfreude—people laughing. I feel sorry for him. It was a tough call to have to go through that, just to put it out there on a human level. He was clearly wrong on the bet, but that was a brutal week.

Jason Lemkin

I think he'll be fine. I promise you he's not going to get caught in the same place again. So, good news: he's learned a lesson, and he's going to live. He's going to survive to live another day.

I'd ask.

Jason Lemkin

I mean, Harry, it's your job to mentor some of these younger kids like Leo. I think maybe you could step in. What is he, 26 or something like that?

He was 25, dude.

Jason Lemkin

25, yeah. I think it's time for you to become the elder statesman in the industry and start mentoring him on leverage: when to lever up to 4x, when not to. In all seriousness—

I need someone smart. I've read it.

Jason Lemkin

Grouping plants and—and, you know—

Yes.

Jason Lemkin

Expanding homes with multiple people living in the same house. Don't bother him.

I assume his LPs, or his investors, knew this was a highly levered fund, right? Am I wrong, Rory? If they know it's 4x-levered, then they know there are black-swan issues when there are short squeezes and other things. I don't think his investors should cry if they knew how it was playing.

My limited experience as an LP in funds with leverage—not quite this much—is that you know it's not free.

Jason Lemkin

I have a feeling his LPs didn't lose any money. If you're up 440% and you go down a billion, you're back to where you started. So I think it's fine.

Rory O’Driscoll

The interesting thing about that—not quite, because actually this is where I think it could get a little hard—is that it all depends on timing, right? Because hedge fund things are weird.

If you came in early, you made a ton of money, and then you lost two-thirds of what you made, and you still made money, right? Brutal comedy. If you came in in the last 6 months, you might have been wiped 80%, right? Because hedge funds, unlike venture funds, have people coming in at different times and at different bases.

So I think the real—and I thought even perhaps Jane Street had put in some money recently, but a bunch of people had put in money. If you bought in in April, May, or June, you’ve been wiped. And then, to your comment on those people, I mean, the litigation—I think fundamentally, yes, he will be fine.

There’s a long list. Larry Fink, who founded and runs BlackRock, had a blowup early in his career. There are lots of people who had blowups early in their careers. Nikesh has worked for one of the most aggressive risk-taking human beings on the planet—that’s SoftBank. He’s seen ups and he’s seen downs, so you can survive.

What?

Rory O’Driscoll

Forex is for babies. Genuine comment here. So you can survive and 10 years later be wildly successful in finance.

I think the crux of the near-term question will be those investors who came in late, who, let’s be really direct here, will be pissed. You put money in a hedge fund in April and you lose 90 cents on the dollar in July, you’re going to read the docs real carefully. If there are any disclosures that weren’t made, or if you’ve done something beyond the remit of the fund, you will have liability, and all this will happen.

In the end, this is America: everyone will sue everyone, and it’ll all be fine. But there will be some dynamics going on now, because can you imagine going back to your investment committee and saying, “We did a quarter—we did a hedge fund, but it appears it wasn’t hedged right, and we lost all our money in a week”? But if you’re the Collisons, on the other hand, you came in on day 1, and you still made out great.

Well, that’s a relief. I was worried the Collisons would be short of cash, so that’s good to know. Rory, I don’t have any cash in Silicon Valley. Well, they’ll be fine.

Rory O’Driscoll

They need money to buy PayPal, so—

They do, and they need money to buy—what was it?—OpenRouter. I mean, they’re doing such a lot. Actually, we were just talking about it. It’s been interesting to see them do all this corporate development while still private. Just super interesting in terms of—you know, a lot of this stuff would be marginally, perhaps, easier with a public stock.

3. Anthropic's AI Models Breach Three Companies as Cyber Threat Accelerates

I had a CEO of OpenRouter on the show on Friday, Rory. So there we go. I’m excited for this next topic, because with Nikesh, I think we’ve got the most pressing person. Anthropic’s models breach 3 companies, too. This is obviously on the back of the OpenAI and the Hugging Face debacle.

Really? Anthropic breaching 3 companies. Also, is this just the most epic beginning of a bull run in security? First and foremost, this is all flex, right? They all want to tell you how good their models are and how powerful they are. So it’s kind of bizarre, because normally if you end up breaching somebody’s infrastructure, it’s not a good thing. But we’re all saying, “Look, look at these models. They’re so powerful.” Fine, granted, they’re all very powerful.

I think the first thing our friends at Anthropic and OpenAI should have done, which I’ve told them, is just point your models at your own sandbox to make sure the sandbox doesn’t have any zero-day vulnerabilities, and make sure your sandbox is your first capture-the-flag exercise. But they decided to give it a target and say, “Go to the wild, persist, take as long as you want, or go capture a flag.” Fine. We have these models that have these capabilities.

I think the challenge we have from a cybersecurity perspective is we’re finding vulnerabilities that would take us days or months to find. The average time to patch a vulnerability, or a zero-day vulnerability found in the wild, is 55 days. Just think about that. These things are finding vulnerabilities in split seconds and then turning around and building an attack on the back of that.

I think the fundamental speed at which cyberattacks will happen, and need to be defended, changes. Now, this is good for us. It’s kind of like the sound of revenue, but I think from a more fundamental perspective, I was thinking this capability was going to show up in 6 months. I think I said that with you, Harry, and it showed up in 4 months. I think in 2 or 3 months from now, open source will have distilled all these capabilities. We’ll find open-source models out there that you can fine-tune, if you’re an attacker, to actually do this on a task basis. So it’s going to change the game.

How are your enterprise customers reacting? Because this feels to me like the mother of all—I mean, security sells on fear, and this is terrifying. What are you seeing in the enterprise customer base when this is knowable?

Sure. Look, the good news is that the flex Anthropic did with Mythos has every CEO talking about Mythos. I spent 8 years trying to get CEOs to talk about cybersecurity. I couldn’t get them to do it, and Dario did it in one fell swoop.

So this is goodness. You see everybody all hot and heavy about MCP and the capabilities of Anthropic, and how these models are going to go attack your infrastructure. I’ve never had so many CEOs call their CISO and say, “Are we ready? What’s going to happen to us?”

The answer is you’re not, because what being ready means is that I have no vulnerabilities in my code, in any vendor that I’ve got deployed in my infrastructure, or in any open source I’m using. That is fundamentally not true. Now, we found 14,000 vulnerabilities in open source in the last 14 weeks testing open-source packages, so there’s a bunch of stuff that’s been used out there.

A: Every company has vulnerabilities. They’ve got to figure out a way to patch them. B: These models will figure out misconfigurations. If you’ve left the door open, if you’ve got a device configured wrong, if you’ve got a piece of software configured wrong—and there’s tons of that out there. Not every IT person building infrastructure or configuring infrastructure is a genius. There are misconfigurations. All these things need to go away.

At the base, at the face of it, a lot of organizations are going to have to go fix a bunch of these vulnerabilities and misconfigurations. On the flip side, even if you fixed most of these, the bad guy just has to be right once. He’s going to find one, she’s going to find one, to get into infrastructure.

The question is, what is your time to detect and respond in that circumstance? The average time to detect and respond is 4 days. How are you going to get it down to a minute? So it’s not a fear problem. It’s a capability problem. It’s an infrastructure-readiness problem. And now it’s come to bear. It’s time to pay your taxes.

Yeah. I mean, that’s the last sentence. You’re right. We can argue fear versus the thing, but what you’re basically saying is the security infrastructure that you had a year ago is wholly unfit for purpose in the next year, and you, Mr. Enterprise Buyer, are going to be buying a whole load more stuff, or you’re going to be the weakest link when these capabilities are everywhere.

Yeah, you said it so well, Harry.

I mean, it just feels so good. I don’t mean to—

I think you should be on the podcast talking about how people need to buy more cybersecurity. I’m just going to buy the stocks. Now, Nikesh, can we get him some swag? I mean, Jesus, he’s a beast. Give Nikesh a hard question.

Jason Lemkin

I want to ask Nikesh as my cybersecurity therapist. I had 2 Fable issues, and they’re internal security, but I’d love to get your take, and you can make fun of me for this. I pretend to have a thick skin. I don’t, but I love criticism.

I already figured out you don’t have a thick skin in the first 30 seconds of this conversation.

Jason Lemkin

Okay, good. So I’m building an app for the Saster community. It’s called Saster Connect to help with recruiting. The details don’t really matter, but it’s the biggest thing I’ve built myself in this era.

Among other things, I’ve got a Google Doc. It’s called “Jason’s Gems.” It’s my ideas on how to improve that. They’re just ideas. They’re just scratches, okay? No one’s seen them. It’s not ready. It’s just a side doc I keep.

So, the other day, I went into Claude and just turned on the Google Drive connector, since it’s one of the 3 primary connectors. This is not esoteric. This is not third-party. And I never knew. It went in, scanned all my docs, found “Jason’s Gems,” found the ideas, and Fable then went and changed my code and my algorithm without telling me.

I never got a notice. I was never told. It was never in a changelog. It was never anywhere. I only found out later when the agent flashed “Conflict with Jason’s Gems” when I was trying to fix something else.

I’m not saying it’s terrifying, but how can organizations deal with this fact when an LLM will go out and change your core code, your core corporate OS, without even telling you? What if you have 1,000 employees doing this? Oh, and the hack—the part I didn’t tell you—is the way it did it: it MCP’d in.

So I had Google Drive to Claude to Fable to MCP, and it was able to do whatever it wanted, probably thinking it should implement Jason’s Gems. But it shouldn’t have, and it never asked me, and it never told me it did it. Is this scary? Is it not scary? Is this Orwellian? What happened?

The Wild West.

Jason Lemkin

It’s wonderful. It’s the Wild West. And part of the challenge is—the good news is, or the bad news, I know which way you want to look at it—I think the small-business entrepreneurs, people playing with their own stuff, are doing this without any regard for security.

Yep.

Same reason you're on TikTok, right?

So people are doing that with no regard for security. They want to experiment with the open cloud. They want to connect to all their stuff. They have no idea if all that data is being used for training. They have no idea what credentials are going to get used or what permissions these agents have.

And that's happening all over the place. On the enterprise side, there is some cohesion around it. I think the most obvious ones are people saying, "You can't use this." Now, that only encourages people to use it more. But there is this stream of thought saying, "If I don't allow you to use it, I have time to go figure out how you're going to use it."

The challenge you have, Jason, is—and I think this is the burden—you know, when the Wright brothers built a plane, they didn't invent the TSA. That was not the first thought that crossed their mind. No, the TSA came a lot later. So you don't think about security when you start playing with new things and cool technology. And that's what's happening.

You're seeing people play with OpenClaw. You see people play with agents. You're seeing people play with all this stuff with LLMs. Everything's happening. LLMs are training on data if you're not careful, because that's the value of giving. What is the adage? If the product is free, you're the product.

Well, guess what? We are the product. All this post-training data has been collected by every model out there on our consumption, which is not regulated or ring-fenced. That's why enterprises are paying a lot of money for all the free stuff that consumers are getting. We are the product. It is learning on all your behavior.

Is the security architecture a year from now more of the same—better, faster—or are there some new things that you just have to do utterly differently to protect? Is it just the same problem at a higher velocity, or is it, "Oh, we never even thought about that before"?

Yes, both of the above. Both of the above.

Look, fundamentally, cybersecurity is a very straightforward thing. If it's a known bad, I'll stop it at the door, right? You show up with guns blazing. I know who you are, and I've got security at the perimeter. I'll stop you.

4. How Palo Alto Networks Defends Against Unknown Bad Actors at 19PB/Day

Yeah, it's a known bad, I'll stop you at the door. The problem is, no cyberattack happens because I stopped a known bad. Every cyberattack happens because—

You didn't know. You didn't know it was bad until it got into your infrastructure.

The question becomes: If you know it's a known bad, you stop it at the perimeter. If it gets through, how quickly can you find it and stop it before it creates harm or damage?

So from a cyber perspective, you want to be in the perimeter business. You want to be on as many perimeter endpoints in the world as you can, because that becomes a sustaining business. The more perimeters I'm on, the longer my tenure for my business is. So I'm on endpoints, I'm on devices, I'm on servers, I'm on firewalls. I'm protecting the perimeter from multiple infrastructure components in the world.

That's good. That's kind of good. Now, the question is how quickly I find known bads will change using AI, right? There's a concept of data classification. You had to write static rules. Guess what? AI can suss it out much faster from a content perspective.

You know, we track every malicious website in the world. Can AI tell me it's a malicious website much faster? Yes, it can. So the ingredients of my perimeter security will change using AI. The act of stopping things in line will still be needed.

When people tell me, "Oh, OpenAI is going to eat my lunch," or "Mistral is going to eat my lunch," guess what? They're in no perimeter security scenario, which means I still need to block the bad guy. They need to be an ingredient in my product. They're not going to take me out of business, because people have all kinds of infrastructure on the perimeter.

The other part is, if you want to suss out all the bad stuff in your infrastructure and find out the bad actor, guess what? Imagine collecting all your enterprise data and running an LLM on it—

—and saying, "Find me all the abnormalities. Find me behavior that you've never seen before."

Now I'm ingesting 19 petabytes of data a day. Think about it: 19 petabytes of enterprise data a day to look for anomalous behavior. I have machine-learning techniques. I have static techniques. I have rules that I look at. Guess what? I'm going to throw some LLMs in there just for fun to see what they find.

Now, if I can find the unknown bad actor in your infrastructure much faster using LLMs, I can detect it and block it. Right now, we run it in 1 minute.

Okay.

Machine learning. This is a good thing. The only problem is I only have 1,200 customers who've bought and deployed it. I need to get the rest of the world to go buy it and deploy it. So that's the second half of the problem.

The third part is there is stuff which is new, which does not have any security guardrails built around it: agents. The world is talking about agents. We could have a whole episode—90 minutes—on what are agents, what is really an agent, how do you give agency, and how do you control an agent?

People tell me they have identified stuff, but then I ask them, "Does it actually have agency? What does that mean?" I'm like, a Waymo has agency. It can drive you into a wall without human intervention. This is a bad problem.

But most people haven't actually given agency to their agents. So they're running glorified workflows which are seemingly agentifying things. But when you start giving agency to things, when pieces of code can decide what happens next, we're going to have a whole different conversation around how you secure those agents, how you build kill switches, how you intercept them in line, and how you stop them from doing bad things.

Yes, like Jason's agent, which is a bad thing, and took Jason's gems. Now the whole world will find out what Jason's gems are.

Totally agreed.

It's crazy.

We said open, we said China. Do you have a comment on it? Moonshot AI closes $3.5 billion at a $35 billion valuation. And if it's free, you're the product. Moonshot's free.

I think the comment on "if it's free, you're the product" is totally true, especially on the consumer side. The interesting thing here is, I'm not sure how it's true. Put it another way: One of the really interesting things about these open-weight models is the impact they're having and their ability to be a drag on prices for the U.S. closed-source frontier-model companies.

In the abstract, if you're running the inference as well, then I get the business model. The model is free, and the inference is how you make your money. It's not as clear to me, long term, if it's possible to continue on a sustaining basis to offer open-weight models without monetizing in some way.

We'll see what it is when people like Reflection AI and Thinking Machines Lab start—when these models start to happen in the U.S. It will be interesting to see what the business model is of which an open-weight model is a part. It's definitely not, "Hey, download it, have a go, and you can do whatever you want wherever you want."

I mean, look, open source has evolved the business model of support. So it'll just be interesting to see what version of "if it's free, you're the product" emerges for these companies in the medium term.

I'm going to give Harry a sound bite.

Good.

5. Energy Is the Real Bottleneck: Land, Permits & Compute Over the Next 5 Years

Average intelligence is going to be free in the long term, and average intelligence will keep getting better.

Exceptional intelligence will be paid for.

Nice. Can you give me some tone with that, Nikesh? That was all monotone. I want drama. Come on. You've got to deliver the sound bite.

I was watching somebody speak the other day, and they said, "If you whisper loudly into the mic, people lean over and pay more attention." So I'll say it again: I'm saying, in the long term, average intelligence is going to be free, and average intelligence will get smarter.

But do you think we'll rely less on frontier intelligence? We won't need it.

Oh, no. We'll need exceptional intelligence. We need exceptional intelligence to discover the cure for cancer. We need exceptional intelligence to send rockets to the moon. We need exceptional intelligence to build a space data center.

Those are exceptional-intelligence tasks. They are still going to require exceptionally intelligent people or exceptionally intelligent models, and people will pay for it because the outcome is so spectacular. I don't think you need to pay $6 a million tokens to answer a call saying, "How can I help you? I'm so sorry your network connection is not working."

Agreed. Yes, customer support will not be using frontier models—

Maybe. But, listen, of course you're right over the long term. In the short term, customer support is requiring more and more tokens to do more and more sophisticated resolution. And guess what? Those are the primary candidates that all these open models are going after with open-weight fine-tuning, saying, "I don't need hallucination. I need—"

The problem we have is—and I'm sorry to go back to the Waymo example, because I think it's the most obvious one out there—I drove in the first sort of Google self-driving car. I don't know what I was thinking in 2009 when I used to work there.

It was a Lexus with a bunch of cameras. It drove me from San Francisco to San Martin on the highway, and my hands were not on the wheel. Then, at 11 p.m., they told me to take the wheel in my hands. As I was driving a quarter mile, I did, and I was more relaxed about saying, “Oh, maybe it’s just going to figure it out when I make a wrong turn because it was so smart. It drove me.” I was like, “No, dude. This does not drive when it turns off.”

That was 2009. It’s taken 14 years after that to get one with all the edge cases trained from a machine-learning perspective for us to rely on that as being the agency that we’ve given to that replacement. So I don’t believe we’re going to give 100% agency to use cases for some time. For us to be able to do that, the amount of data collection and context we’re going to create is going to be humongous. Basically, you have to literally take every edge case in customer support and get it into the AI brain of your organization so that you can start relying on AI instead of the human.

So you’re getting 80% right now. You’re getting 80% of customer support solved. All the edge cases are waiting to be solved with AI. Then there is, for a given app, how much of the value is purely in the model versus all the other things. And you’re right: for something like customer support, we’re probably paying 10 or 15% of the revenue we’re getting for intelligence, and the rest of it is all the other stuff it takes to make that intelligence actionable in the context of answering tickets.

One of the things we look at that’s super interesting at the app level is tokens as a percentage of total revenue. It varies from Salesforce’s and Intercom’s stuff, where it’s plus or minus 10 or 15%. Obviously, in coding and things like that, it’s 70 or 80%, which means it’s just raw intelligence and a mild harness. Those are just very different.

I think over the next 3 or 4 years, we won’t be paying for intelligence; we’ll be paying for compute through our nose. Speaking of paying for compute through our nose, we often get chastised for being too public-markets-focused or too Anthropic- and OpenAI-focused. Valar Atomics triples to a $6 billion valuation as Sequoia bets on nuclear for AI.

It’s a 3-year-old small modular reactor company. It raised at $2 billion; now Sequoia is leading a round at $6 billion. Specifically, there’s an NVIDIA partnership to power AI data centers, which caused a lot of excitement for the company.

Rory O’Driscoll

Harry, I met somebody who’s in the business where they take chicken feces, turn that into methane, and produce gas. I thought, “Oh, it’s a cute project he’s got running somewhere in the middle of the country.” Then he told me that he has billions of dollars, and he’s selling the energy to hyperscalers.

Anybody who can produce any energy source—it doesn’t matter where you are—is right now trading in multiples, because land, permits, energy, compute: this is the thing that is going to get priced for the next 3 to 5 years. I think it’s almost like the question will become, between Anthropic and OpenAI, who has more access to more compute in the next 3 to 5 years? That’s what people are going to buy.

It’s very hard to find compute right now. You can also take all the free Chinese models you want; where are you going to run them, Jason?

Jason Lemkin

No, no, for sure. Just to Nikhil’s point, I actually used to be a little bit in advanced energy storage in my first startup, and things like chicken manure didn’t used to make sense. These models actually used to work.

Literally chicken.

So did cow. So did cows. I even looked at some of these things, but the margins were so low. The IRR was so low, but the businesses worked. Now AI—there’s such a demand for compute, everything works, including all types of nuclear like Valar, right? Including chicken manure.

Laugh, but I remember talking to a manure farmer doing this back in the day, and he’s like, “Well, the best we can commit to is 8% annual return if everything goes well.” It’s just hard to get a VC excited for those returns, but maybe it’s 80% today.

We’re going to be talking about that on this show.

Me neither.

Rory O’Driscoll

You brought it up, but everything—

And nothing in energy storage worked before AI, right? There’s the battery startup, right? What’s the one that just raised at $12 billion, too?

Jason Lemkin

Yeah. None of these things worked without AI, right? Neither did RAM. None of these products really were that great. But now they’re the greatest products in the—

Give me my RAMs. I can’t even get my Mac Studio with more than 64 gigabytes. Give me my—everything’s working, right? I—

Rory O’Driscoll

I’m going to ground us in a little facts here, just on the 3, right? Just because I think they’re all super—I mean, Valar and Base Power are super interesting but very different.

You’re right: Valar is purely the “we need more power for compute” bet. They did plug into an NVIDIA chip, and they basically showed that you can get criticality and generate power. But no, I think everyone knew that. Just to say, I think the regulatory journey for all these things is still a hole, just to be clear, in terms of when you can actually plug it in on an ongoing basis.

There are a bunch of these private and then a bunch of these public, and I think NuScale is the one that’s doing the existing technology that’s well understood. I think light-water—not the nuclear, but it’s like, “This is the way we built them,” so it’s pretty much the same. It’s the furthest along in the regulatory path. Then all these guys, including Valar and Oklo, are doing new, different things.

The big question will be, after you get the initial demonstration that it works, what’s the regulatory path? You have to be wildly supportive because there’s no way to get cheap electricity without doing nuclear. So from a public-policy perspective, go team. But just from a “don’t spend the electricity yet” perspective, you’ve got to plow your way through the bureaucracy.

At some level, you want people to be mildly cautious before you permit these things—probably less cautious, dare I say it, than we’ve been for the last 30 years, where I think we’ve stifled innovation, and possibly a little more cautious than you might be right now to get it right. So there is an approval journey ahead, but it’s awesome we’re doing it.

The question we’re debating is: where’s the money going to go? How much are we going to pay for intelligence? Once we pay for intelligence, and we’ve talked about compute, I’m pretty sure Harry will want us to talk about all the capex that’s going to happen out there.

At some point in time, somebody’s got to pay for all this compute, and that money has to come from some version of some people paying for AI. That’s the only thing that’s going to allow the Valar and chicken-manure world to actually be worth something, too. Yes, in the end, someone’s got to buy a trillion dollars’ worth of tokens in corporate America, and—

6. The $1 Trillion Token Question

Jason Lemkin

Well, if we buy a trillion, they’ll buy half a trillion a decade later. I hate to be cold, but as a former European, I can say that—you know, a current European, whatever. I’m the most cynical dude on Europe you’ll ever meet.

Yes, we have 2 Europeans here. Unbelievable. Unbelievable.

Is this not just another layer of companies that is dependent on, Rory, to your point, OpenAI and Anthropic continuing to go on their charge and hit their number? We’ve never had an ecosystem that will be so dislocated if OpenAI and Anthropic do not hit their 2027 numbers.

Rory O’Driscoll

I don’t think so. Whether OpenAI or Anthropic hit their 2027 numbers or not is orthogonal to the fact that there is infinite demand for AI at this moment, and that infinite demand needs to be satisfied by compute.

Whether it’s OpenAI that builds the data centers, buys the data centers, or pays for them, or somebody else pays for them, there is demand in the market. Look, if you think about what’s going on, I still posit 70% of the compute demand for AI is being consumed by consumers who are getting a free ride.

Maybe they’ll go to reallocation. Maybe we’re going to have to give more compute to enterprises over time as they become better monetization capabilities, or you’ll find that eventually the promise of consumer monetization is going to start showing up.

We all talk about why an agent can’t book my airline ticket and make me a restaurant reservation. These are simple use cases. I don’t need to go solve cancer to get that stuff to work. That stuff’s going to work. When that stuff works, there are going to be monetization opportunities on the consumer side.

So I believe that, at first principles, there will be tremendous amounts of compute needed to satisfy the AI use cases, both on the consumer and enterprise sides. Which player ends up monetizing them becomes a question for the markets to decide, and that’s a timing question—no different from Leo’s question. That’s a question of who builds the capability and the services. Google was not the first search engine—

Okay, I’m arguing against you. At one level, obviously, if you zoom out enough, you’re right, but if you zoom back down—and we notice in these discussions I’m always—

Jason Lemkin

I see where you live, Rory.

Rory O’Driscoll

Yeah, dude, I’m just—you’re running your $280 million company. I’m just trying to turn a $20 million investment into $100 million and call it a day. I’m a small guy.

But the genuine comment is kind of mixing infinite demand for intelligence. Let’s assume it’s just the enterprise now, because I think you are right: the consumer side is super interesting, especially for OpenAI, but let’s leave it aside because we can only do one thing at a time.

I think the question where it does matter is that, right now, the assumption is that 70–80% of that demand gets channeled through Anthropic and OpenAI. In other words, there’s 70% of the Google compute backlog and 70% of the Amazon backlog.

In the short term, most of the market is assuming that OpenAI and Anthropic buy the compute, buy all the stuff that’s further down the stack, buy the chips, and resell that intelligence on a frontier-model basis to U.S. enterprises. If it doesn’t happen that way, there’s going to be a pretty big dislocation.

Yes, it’s perfectly possible that there is a public-market dislocation because the players at the table might change, and that’s great. That’s called a buying opportunity, because it doesn’t take away from the infinite demand.

It’s extremely possible that perhaps this wonderful company called Moonshot, which we talked about 3 seconds ago, could be the model of choice, and that somebody’s going to take that compute, which is not going to be used by frontier LLMs, put Moonshot on it, and sell it to enterprises at 10 cents on the dollar, on the 10-cents-a-dollar tokens.

[Speaker?]

Yeah, Moonshot is happy. NVIDIA is happy. Enterprise is happy. OpenAI is very, very sad. You’re right—that’s the dislocation.

But the question becomes: Which of these is the market going to support? Is the market going to give you infinite capital to be able to build the compute because they believe you’re the anointed winner, or does the market believe that you’re running it differently and want you to run differently?

So I don’t think the demand goes away. I think in all these conversations, one variable goes away when we run into this sort of technology shift: in an infinite bull market, we take execution out of the picture. Yeah, it doesn’t matter. Every chicken-manure company and every nuclear-reactor company that says the words in PowerPoint is going to get funded by everyone because they assume flawless execution.

You look around, and poor Jason is looking at SaaS companies and saying, “Holy—some of them aren’t executing as well as the others.” So eventually, execution matters, and that’s going to decide the winners and losers in the market—not the shift of which intelligence is the best.

The best example of that would be that 2 years ago, OpenAI was first and Anthropic was second, and now Anthropic is first and OpenAI—

And Google was written off.

[Speaker?]

And Google was written off. At some point, Gemini was nonexistent. Google was written off. Now suddenly Google has the compute, the cloud sales, and Gemini.

But still not the amazing open frontier model. Still not the coding agent.

The customer-support agent is going to be extremely unhappy because he didn’t get a chance to answer it using the best model. Just kidding.

Rory O'Driscoll

Well, to Nikesh’s point, Harry kicked this off by saying, “Have we ever had an ecosystem so dependent on the success of OpenAI and Anthropic?” I mean, it is, but maybe to Nikesh’s point, so much has changed since we started the show.

When we started the show, it actually seemed like everyone would benefit because average intelligence—or whatever term Nikesh would use—would permeate software, and that would be good enough. Now, in front of us, this is the revenge of the frontier. We may not care in a year which model we like; we need frontier models. We need the best.

But we may not care who wins. We may not care who wins this battle. This may all blow over, and it may all be about compute. We may not care who wins. Whoever wins will plug in.

Jason, I think the models will get better and better, and the distinction between models may not be enough for you to decide to rip one out, because I think the part which we are starting to build—and will be building for the next 3 to 5 years—is context.

Think about it for a second. I run a simple firewall company—or a simple, complicated firewall company. You can stick any model you want on it. The model doesn’t know why my customer’s infrastructure is down. It doesn’t know what product my customer is using. It doesn’t know what operating system it’s using. It doesn’t know what the configuration of the customer is. My model doesn’t know why this happened the last 5 times to a customer.

All that knowledge, all that learning, is being captured by me in effectively vector databases and in-context learning systems. That’s what my team is doing. I have more people collecting context than I’ve ever had.

It’s kind of like the Waymo thing. I’ve got people planted, saying, “This is a tree. This is why it goes down.” So as I build that organizational context, I can stick any model I want on it. The model distinction will not matter because the context will become as important, or perhaps more important.

You’re clearly 100% tracking Satya with the kind of Microsoft comments recently on AI companies. It makes enterprises need to build their own value and build their own context rather than do it in a frontier model, right?

Well, I think—yes, he’s saying something different. I understand what he’s saying. That’s a different comment; mine is a different comment.

I think there are 3 parts to it. There’s the model, which is the raw intelligence—let’s just call it that. There’s the context needed to answer your queries or answer your problems. And then there’s the context needed to train that ecosystem.

I’m talking about the context needed to train the ecosystem, which means I’ve got every customer case that ever happened at Palo Alto Networks getting transcribed. So my model knows what was a good answer and what was a bad answer.

And what core model will you start to build all this context on, do you think, or have you decided?

I remember calling Thomas [likely Kurian] at Google when the whole thing just started—the shiny object called LLMs—and I said, “Hey, do I need to build a cyber model?” He said, “Dude, over time, what’s going to happen is the models are going to get smarter and smarter, and small models will not be as smart as the big models.”

And he was right. The small models are more intelligent than the big models. Now, at some point in time, if your average intelligence becomes smart, which is what I said, then the distinction between a little more intelligent and less intelligent is less important than knowing the domain and the context.

So I think we’re coming to a world where, in the next 5 years, domain becomes equally important with model intelligence. I think Satya is saying something different, and Sam is saying you can’t parse every problem into multiple models without carrying the context to the model to give it enough context to get the answer.

Yeah, because he’s saying, put all the context in a harness which is sitting beside the model, which I provide, and then use whichever model you want and commoditize it.

Every model company is saying, “No, I’m going to make my model smarter with context, because otherwise I get commoditized.” So I think that’s a bit of a commoditization battle that’s going to happen between models and models plus context.

But, Nik, on that, you also said something—not in conflict with it, but related. You’ve got all your intelligence in your vector database, or whatever it is, all your context, and then you can pick and choose your LLM on top of it.

But as you said, the LLMs don’t perform the same. Even with Opus 5, Fable and Unlet at Palo Alto Networks, you have a team that can manage those changes. We’re learning as we go along.

So you’re learning. What about the average enterprise that doesn’t have as strong a team as you? How can you really switch out these LLMs, even if all the context is in your vector database, and have confidence that the results will be the same?

Jason Lemkin

Bending Spoons. Bending Spoons.

None of the above.

Rory O'Driscoll

But I mean, it’s a Darwinian moment.

Yes, I got it. That moment does not suggest that everybody survives. I understand now.

In other words, what you’re really saying is, if we don’t figure this out, we will be working for the Italians, so we’re going to figure it out. Got it.

I love the way I went to private markets to get the private-market discussion. The straightaway discussion is, well, it depends on what OpenAI and Anthropic are willing to pay for it. It goes back to that.

It’s just funny how everything rotates back to compute and what the big buyers are willing to pay. And you’re right, Eric, because I pushed on why we’re always talking about just the same 2 companies, but we internalized that no matter what you talk about, you end up back talking about them because they’re the giant sucking sound of demand that’s just pulling everyone along, all the way up and down the chain.

Which is why I think you’re correct: If that demand signal turns out to be attenuated or dips, or even is true in the long term but blinks for a year or 2, it’ll be a weird time in tech.

Rory O'Driscoll

And that’s why we’re all focused on the poster child for the trend. But I think the trend is bigger than the poster children.

Yeah. AI and intelligence is bigger than OpenAI and Anthropic, is what you’re saying.

Rory O’Driscoll

Yes, you are right, and enterprises are going to want to consume it a lot. But the structure changes if it turns out that 70–80% of the beneficiaries are other than OpenAI and Anthropic. There will be a pretty significant dislocation up and down, right? I mean—

Rory O'Driscoll

I was at a $200 billion company at one point in time, and 2 years later I joined Google, which was a $14 billion company.

Good call. You’re a good stock picker.

Google, we touched on Microsoft and Google being told to—what was it?—“You have to dance,” like 2 or 3 years ago, whenever it was. We obviously had all of them coming out saying, “Capex—we’re going to keep spending, and maybe it’s going up.” How did we analyze the results and the reaction from them? Obviously, cloud was an acceleration from both Microsoft and Amazon. Really incredible numbers. How did we analyze this? Roy, do you want to set context in any way?

Rory O’Driscoll

It’s not that hard. You had 4 people report that would be relevant here: Amazon, Google, Microsoft, and then Meta, right? The big picture is that the people who have a business selling cloud inference all had an amazing quarter.

Google Cloud, the smallest, grew 82%. AWS grew 37% at scale. It’s always hard to know at Microsoft because they bundle a bunch in, but they grew 20–30%. The big-picture comment is that people sold a ton of inference, right? Because of that, people said, “I’m going to buy a lot more compute,” and the CEO of AWS, in particular, made it very declarative: the ROI here was amazing.

The market was really happy. In particular, Amazon and Microsoft got marked up pretty significantly. By contrast, Meta also said, “I’m going to spend a lot of money,” but it wasn’t as obvious how they were going to make money, so that stock went down $20.

Probably the most surprising thing, going right back to the layer thing, is, God, those are really strong numbers. All these people are selling a ton of compute. These are $400 billion run-rate businesses, plus or minus, in total, and they added 30%, which means $100 billion more a year of revenue across these 4 companies in compute.

It’s just the scale—the scale of the things you can lose sight of. That was, for me, the big aha. Will it persist? Who the hell knows? We can talk about that again. But the facts on the ground, the new information in Q2, were bullish. That was my take.

You can sell this AI.

Yeah. Jason, what should we take from that? Enterprises need help with it. Palantir is the best.

Jason Lemkin

I think what we should do is send it to our portfolio companies and tell them to work harder, because there are no excuses. If Palantir can do it—they came back from 15% growth 4 years ago—why can’t you do it, kids? Work harder. Work harder.

I don’t know what the message is. Certainly, to Nikesh’s point, I’d love to hear Nikesh’s thoughts. If you can package and capture intelligence, the demand is inexhaustible at Palantir, right? You can also capture somewhat model-agnostic intelligence, but the demand here for intelligence, for compute, at some level, Palantir is a very sophisticated harness on top of massive amounts of data, right?

Maybe vector databases, to Leo’s point—I might be wrong or oversimplifying. But they’ve captured that to a magical element in the age of AI. People need to solve these problems with data. They need answers, and Palantir gives, I think, more than 1,000 customers, right? 1,049 customers. They’re giving $8 billion worth of answers, growing 100%.

These 1,000 customers will pay almost anything to get these questions answered with AI. They’ll pay almost anything.

We’re in a capex cycle. There’s $1 trillion of capex that has been committed for the next 1 year across all these people, broadly speaking, and the market is saying, “Great. I see these large companies have the ability to fund this $1 trillion of capex,” and there are signs that they’re getting compensated for some part of the capex that’s out there.

Whether that’s because of higher prices being commanded, people demanding deployment of AI, or deployment of cloud, this is good news. That capex dislocation is not happening today. Now, it could happen tomorrow if some of these people who are committing to capital are not able to show up with the capital. But for now, we have 1 more run at the roulette table.

That’s what’s happening. We’re being told that this market is going to support capex until it can’t. I think it’s a bit of a gold-rush moment. I think every consumer app will get rewritten in the next 5 to 10 years.

Why would I not have my agent talk to my DoorDash app or Uber app? Why do I have to go to every one of them and click 7 times and have it have no context or learning? If you talk about contextual learning and agents, everything is up for grabs.

Every consumer app that was ever put on the iPhone has to be redone. Every enterprise app in SaaS you just debated has to come back with, “I have an opinion.” The demand, the construction, the work that’s needed is humongous. Let’s take that for granted. That’s going to happen. This market is proving that.

I think until the market can keep funding it and the timing works, the biggest—and only—problem we have right now is a timing problem. Would the revenues show up fast enough to keep funding the capex cycle, or is there going to be a dislocation in capex versus outcomes?

The telecom industry is very used to this because they used to spend billions of dollars building 3G, 4G, and 5G, and then they’d see the rewards come later. So they went through a capex cycle, and that’s pretty established in the market.

It looks like we’re going through this compressed version where capex and revenue have to show up pretty close to each other, because the numbers are just way too big to be funded by speculators for long periods of time. I think that’s what we’re seeing, and that’s why this brings back the whole OpenAI-Anthropic debate.

It doesn’t matter if they show up with the money or not. Somebody will show up because there’s enough demand. I think the next dislocation could happen in the supply of compute.

You can bring all the capex to bear, but I think, to your point, the hyperscalers may not get their regulatory set of approvals. Europe may not allow data centers. You may find 30 states with picket fences which say, “No data centers in my state.”

So there’s a supply problem that happens on the compute side, which could have a knock-on impact on all our infrastructure buddies in the semiconductor space saying, “Holy shit, it doesn’t look like all the stuff they’re building is going to go out as fast as we thought it was going to go out.”

I think that’s kind of where we are at the market-mechanics level. I don’t think there’s a demand problem. I don’t think there’s a jobs problem. I don’t think there’s an appetite or intent problem in terms of all of us wanting to rewrite this stuff.

And I think, to Jason’s point, why not? Palantir’s at the party. They’re also saying, “I can package intelligence, make sense of it for you. You don’t have the capability. You don’t have the resources. Let me make sure you don’t become extinct in this wave of technology.”

I’m going to go back to 1997, 1998, and 1999, when we saw the last big pivotal technology called the internet. A lot of the characteristics were similar, except you just didn’t need $1 trillion a year to keep building the internet.

I think the interesting thing, as I play back all the comments you’ve made, is that the odd thing is, if the most likely failure mode is not ultimate demand—and I agree with you, it isn’t—but just an enterprise’s ability to digest that speed, then, to some extent, and I think Gavin Baker made this point, if enterprise can’t digest fast enough, then, to some extent, if the spend slows down because they can’t get it online quickly enough, it may be timed perfectly with the enterprise’s ability to digest, right?

The better digesters will win, and the poor—

No, that is an interesting point. All that data says that the companies that are digesting AI quickly are growing faster than the companies that are not.

So, if—and I think that’s not true in every industry, but my guess is, to your point, if, for example, you’re playing in finance and your competitor is using advanced LLMs and you’re not, for trading or whatever, at some point you will be Bending Spoonsed, to use your point.

Yeah. Those may be, Leo. They may not have been Bending Spoons.

The other thing is, Nikesh, in your next analyst call, can you do an ode to us where, when you get a shit question, you just say, “Bending Spoons”? Just drop the mic.

Jason Lemkin

I think you’ll find, Harry, that when you’re worth $280 billion—or whatever enormous market cap this man has—you’re not paid to joke on the earnings call, Harry. You’re paid to look down the line and deliver the product. That’s how you keep your job.

Sorry if you hadn’t figured it out: I ain’t here to bring IQ to the conversation. Okay?

Yeah. Nikesh, do you think more established enterprises can process this rate of change indefinitely? Do you think they’ve changed permanently?

What I’ve found with a lot of vendors now is, for example, over the last year they’ve made 1-year commitments, where before it might be 3 or 5 or 7, right? They’re like, “Well, the world’s going to change so much. I want to see what agents and what AI products come next.” That’s totally rational today. But most enterprises traditionally—you can’t rebuild your whole stack every 8 to 12 months.

It’s destructive on the org. But your point is that that’s a skill to win today. Do you think that’s changed? Do you think we’ll revert to the mean, where we can only process change every 5 years after we get over a hump? What are you seeing?

I think the enterprise’s ability to absorb this, digest this, or perhaps leverage this to their advantage depends on their ability to create training data as fast as they can. I think not enough people are focused on training data. This is not a problem AI can solve for me. This is not a problem that Fireworks can solve for me. This is a problem I have to solve.

I have to parse through them. Sorry to go back to the same thing, but I get 400,000 customer cases a year. I know when they come in, I don’t have enough context. Some human beings solve them. I don’t know how they solve them. I don’t know what logic they applied, but they solve them. I need to get into the brains of those people who solve them and abstract all that knowledge and codify it so that I can write my own playbooks and rules as to how to solve the problem the next time it shows up.

I’ve told my team, every new phone call, every new case is a learning opportunity. It’s not just to solve it; you have to learn. We have to go into this learning mode as enterprises, just the way you should never let your VP of finance just decide. You should say, every time the VP of finance reaches a conclusion, you have to surface it to the human called Jason and say, “No, dear VP of finance, book it,” because we book every transaction.

You have to give the organizational knowledge to some learning system that you have to build. I think that still is going to take 3 to 5 years for every enterprise, every use case. I think that’s kind of what we’re not paying attention to. The same thing applies to SaaS companies. They all have to go rebuild their stacks.

But not just the stack. The stack rebuild is the easy part. I’m pretty sure Fireworks will take my money and fine-tune an open-weight model for me if I want, and keep training my use cases to a point. But beyond that, how do I get from 70% accuracy to 99% accuracy? That’s the problem. The problem is I don’t know which 30% is inaccurate, so everything’s useless.

It’s funny, your point on learning. I was literally just trying to make sure I got the quote right. There’s the Darwin quote that said, “It’s not the strongest of the species that survives, or even the most intelligent, but the one that’s quickest to learn.” I think you are right about that. Doing what it takes to digest it quicker will be the key management skill in the next 5 or 10 years.

I think what Palantir is selling—and you’re right, maybe it’s not a full solution—I think the reason they’re doing so well is they’re able to say, “Dude, we know this is the biggest problem. As the CEO, I at least have some kind of answer here. Let me help. Give me $10 million. It’ll be great.”

Look, I think—let’s not underestimate what Palantir might be doing. There is a capability that AI has already demonstrated where it can trawl large corpora of data, summarize it, look for anomalous behavior, look for trends, capture them, reason around them, and reach conclusions.

Now, the good news is, if you’re doing any kind of offensive work, if you’re looking for amazing insights, it could trawl through petabytes of data and produce 20 amazing insights. You can go judge them and say, “Well, 15 of them are okay and 5 are amazing.” But the 5 that are amazing will change my ROI, give me 300 basis points on my top line, and improve my margin by 100 basis points. Hallelujah. You just paid for everything.

You don’t have to put a learning system into place or anything. It’s just taking enterprise data and doing a lot of that stuff. I think places like oil discovery, nation-state analysis, or a whole bunch of stuff where lots of people are required to go and do this and write code, that doesn’t need to happen anymore.

Totally agreed. We have one final one. We have a new CEO at Scale AI, an operator. They hit $1.5 billion in ARR. We mentioned the importance of data there, obviously, Scale AI being one of the biggest providers of data.

7. Scale AI Hits $1.5B ARR

We have Mailchimp revenue declining for 8 straight quarters. That’s not a nice headline, is it? Rory sells DroneDeploy to Procore for $900 million. Go Rory. A 13-year journey. Amazing outcome.

We have Visa cutting 2,600 jobs. Nikesh, you said it’s not a jobs problem. Well, the CEO of Visa says it’s efficiency and shaping the way work gets done. So, 2,600 people gone there. Whatnot raising at $20 billion.

Can I ask Rory about DroneDeploy, because it ties to the beginning of the conversation with deals and Nash?

Rory O’Driscoll

I’m not going to speak for the acquirer, because I’m not on that side of the room, but genuinely, it was one of the least stressful deals I’ve ever done because, honestly, I would have been happy to continue. This was not a founder-tired situation. I think there are about 2 or 3 interesting lessons here.

First of all, when you have capital discipline and modest fundraisers, you’re set up for success, not failure. We always raised below the price we sold at. We didn’t raise a ton of money. We were profitable. It was just a fine little company growing nicely.

The second thing is really important: the trend was our friend, not our enemy. I think with some of these very basic SaaS companies, you look back and go, “There’s been a platform shift, and you’re on the wrong side of it.” When you’re a software company that’s enabling drones and robots, you’re actually on the side of the future.

In fact, one of the lessons I learned having invested 10 years ago is that, in the physical world, AI takes a lot longer to happen. When it happens, it’s amazing, but I look back 10 years ago and I thought drones would have exploded 5 years ago. They’re really starting to explode now, as are robots. So, it took a long time.

We were on the upswing of this. It feels really good. We’re happy to hold, and then obviously we got an offer that made us do something different. I don’t want to comment on the specifics of the offer, but I think one of the ahas here is that building companies is hard. By being disciplined and putting ourselves in position, the founding team did an amazing job—3 founders all together, all still actively involved.

No, it was genuinely not a stressful thing at all. It’s like, at the right price, you’ll do this deal; at another price, you won’t. For what it’s worth, from a distance, I think it’s super interesting for the other side, too.

I think market expansion is what you need to do in some of these spaces. You need to say—and probably Nash has done these kinds of big strategic deals—“My thing is this big. I need to add the next thing my customer wants.”

At some level, the customer wants not only to be told the accounting of his business project, but also the physical progress of his building project. That’s what things like physical inspection do.

So, that deal—what’s interesting is that DroneDeploy was bought by Procore, right? A great classic software company, founded by Tooey to do software for real estate. They dominated it and had a great run, right?

Growth slowed. Most importantly, net new customer count sort of stopped growing. Growth slowed to around 17%. So, they make a big bet. I’m not an expert on DroneDeploy—obviously, Rory is—but they buy a next-generation platform to use drones to accelerate the construction industry.

Structurally, what’s interesting—and I find these deals are always really stressful—is that Procore’s market cap is beaten down. It’s got to come up with $900 million, a lot of it debt. Pay 12x while it’s trading at 4x. I find, in the old days—I’m not saying that happened here—these deals are stressful, man. They are.

It’s not Palo Alto Networks spending 0.01% of its market cap on some smart kids. This is bet-the-farm at a much higher revenue multiple. It kind of doesn’t have to work, but, man, this is the big bet, right? It wasn’t cheap. I mean, you’ll say it’s cheap because you’re on the board, right? But Procore is going to think this is expensive: to pay 12x when it’s trading at 4x.

We started this on deals with Nash, and we started this on whether 3x to 4x for Airtable was a lot. Procore is one of the companies basically trading there, too. Was this deal super stressful? Did you lose hair? Were people shouting and throwing things through the window?

Every deal needs to work. We’re not buying companies because we have money to spare or because my shareholders think we should. I think the hit-rate requirement for us is more than for a VC. In the last 8 years, we’ve bought north of 40 companies, and I want to say 75% have worked; 25% haven’t.

Our largest deal was a $28 billion deal, which probably is currently valued at north of $50 billion. That one’s kind of a career-defining move. If you take a company at $28 billion when your market cap is $200 billion and you spend 14% of your market cap, or 16% of market cap, and buy something, it better work.

When you make that work, then the market gives you credit for making deals work. I said that in my earnings call, and they got all freaked out. I’m just saying, you have to make the big ones work. If you don’t make the big ones work, then you lose the license to run your business.

That big one was CyberArk, right?

Yes.

Yeah, that was great. Got it.

Rumor has it that agents are going to be important. If agents are important, they're going to need identities. They need to be treated like privileged identities. So that's our thesis. Simple.

Like all the best deals. One of my partners always said, “If you can't express it in a sentence, it's probably a bad deal. And if you can, it's probably a good one.” Got it.

Rory O'Driscoll

As you've seen with my examples, you don't know what these agents are going to do. Man—

In your case, you're just going to restrict agent behavior, Jason.

Jason Lemkin

But they're so good. Man, they're so good. They're so powerful.

Jason, to the point—and this actually ties back to what you said—I was just reading some stuff last night that really does accord with this. Someone made the point that if you can't control what they're doing, you have to be very clear on who they are as an identity and where they're allowed to go.

If you've got this, as I said, this kind of AI employee, and you're not quite sure what they do, you've got to bound the systems they can access very tightly. So I actually think I totally get your point.

The only danger is if you take it to the extreme. That's called automated workflows. Those are deterministic outcomes. If it's deterministic outcomes, we already had that technology for the last 20 years. So the question is: At what point in time do you let an agent think and do something?

Jason Lemkin

Yes, that's the big debate.

But the flip side is that it does a really good job. And listen, we don't have the perfect security profile, to your point, right? But even with what we have—which is probably one agent with about 1,000 rules, to your point, and the rest probably have 5 or 0—even with 0 to 5 rules, it's pretty darn good 99% of the time today.

Since January, since the models upgraded, it's been really good in the last couple of months—like, really good. So it's a trade-off. It takes 1 destructive example to make it all unwind. If you give it access to your bank account, let's see what it does. If your VP of finance is allowed to write checks, I might want to have a conversation.

Speaking of people being wrong, I'm going to say I was totally wrong on something: Scale AI. The fact that they've continued that business—I would have thought the acquisition left them a husk. But I think it proves one of those rules that you kind of know but forget: When you're in a great market and you have a product that can meet that need, even losing your top people is fine.

They were selling data products to an insatiable demand for data, and I give them huge credit. They kept the thing going. All credit to them.

Rory O'Driscoll

Hey, Windsurf sold to Cognition, right?

Exactly.

Rory O'Driscoll

All these stub deals are working.

But they sold really quickly. Both sides sold quickly, and then value was created. This is even more impressive because they were left like a husk. I even called it a husk a year ago. But they built a business out of that, so all credit to them.

Rory O'Driscoll

Groq could be the next one, too.

What?

Rory O'Driscoll

Groq could be the next one, too.

Yeah, you're right. The remaining Groq—you're right. They're offering hosted inference with their technology. Yeah. No, I mean—

Rory O'Driscoll

In the face of insatiable demand, all things are possible. That's the aha.

That's a good quote, Rory. Now, Nikesh, do you see why I go home early from dinners? Because I need to be fresh for podcasting. You see, this is hard. This is hard work. You builders building enterprise value in your public companies—this is where the real grind is.

Rory O'Driscoll

He's sitting there going, “He's doing this,” eyes closed, and he'll go back to making his $280 billion market-cap company work later.

It's got to be built 1 deal at a time, my friend. Enterprise is 1% inspiration, 99% perspiration. Totally.

No, I do not.

It's what I tell my agents every day, guys: Get to work.

Rory O'Driscoll

Agents don't sweat.

Stop it. Get to work, boys. Perspiration.

Nikesh, it's been fantastic, dude. Thank you so much for having me.

Thank you, guys. Fun.

Rory O'Driscoll

Yeah, I really appreciate the time.