20VC：AI泡沫将破裂：一半Neocloud将消亡｜中国：是否应禁止芯片出口、并对中国开源保持警惕｜Mag7：谁会消亡、谁能胜出：为何Meta乏善可陈而Microsoft强势

If there is a dislocation, no one is better prepared to survive it than hyperscalers. Let's take neoclouds. Neoclouds right now—there's a whole bunch of them. I think at least half of them go away within 36 months. Fireworks is making a lot more money than Baseten. The more you customize the model, the more the token changes its value.

Do you know what? I think I genuinely have the best job in the world because I get to sit down with people like you. I'm dumb as rocks, but I get to ask questions that normally I wouldn't be able to ask, and I get to learn from the greatest minds. So thank you so much for joining me for a second time, Jerry.

I'm happy to be here.

I want to touch first on something that you said to me before. You said if the Iran war continues to fester, then you expect to see a correction, and depending on the depth of the correction, the AI bubble will burst between October 2026 and March 2027. Can you help me understand your thinking here?

1. Credit Markets Threaten The AI Boom

Sure. If you look at what happened in 2001, at the end of the dot-com bubble, innovation stopped for a while, and then new innovations came in—the LAMP stack, which led to the build-out of websites. Google started taking off. In 2008, cloud computing was very slow to take off. This is because these financial disruptions slow things down. The stream of commerce gets disrupted.

Right now, with AI, debt is a huge part of this. It's so unique compared to previous cycles: There is so much debt. All the hyperscalers have taken on much more debt than they ever have before, and the challenge becomes whether these guys will get disrupted if the credit markets have a disruption—which would certainly happen if we had a problem in the overall capital markets.

The concern for you here is that we'll have a credit-market disruption caused by the global conflict, which will then impact the ability of these hyperscalers to borrow cheaply?

Well, that's one potential disruption. I see several that could occur, and I think if it's going to happen, it's going to happen if this Iran war continues. We can't have it just continue. The reason I say this right now is because of complacency. We've got tremendous red lights that have been going on for a year or more in the credit markets, and there's just complacency. It's like, "Ah, we're fine." I don't think people are accounting for risk.

If you think about it, those guys who are in the private-debt market have spreads that are too narrow between real risk and not so much risk, and they're not really accounting for that. So I'm concerned about that as one sector, but there are multiple concerns because this AI revolution is incredibly complex, with massive amounts of dollars being spent on it globally.

What are the signs to you that we're seeing a cracking in the credit markets?

Complacency is the first thing you look at. When it happened in 2008–2010, there was a handful of people—there have always been documentaries about these guys who made money betting on shorting the housing market. But everybody else in the world had no idea what was really happening. It was complete complacency.

What you had was a really ugly situation where the people who were supposed to be keeping an eye on things—the credit-rating agencies and the credit-risk departments of the big banks—were asleep at the wheel, and this terrible thing happened. There was a malaise around the risk that people didn't recognize, because historically there had never been a huge default problem with mortgages. That was the kind of thinking that was there, and they didn't realize the underlying problems associated with credit.

I see the same thing today in the credit markets: There are many opportunities for there to be problems. In the late '90s, you had long-term credit blow up. We just had Leopold [?] blow up because he wasn't accounting for the leverage that he put on his fund. I still see the little warning signs, but complacency is the biggest issue.

In Japan, there's another problem, right? This is the second time the U.S. has bailed out the yen. Why are they bailing out Japan? That's because Japan holds $1 trillion in Treasuries. If they have to unwind that—if they sell $100 billion worth of Treasuries—the market could absorb it. But if they sold $300 billion worth of Treasuries, a third, in order to be able to buy dollars to support the yen, we would have a real problem on our hands immediately. An immediate global problem.

I do a lot of backcountry skiing, and you recognize the avalanche conditions when they're worse or better. You recognize that things can be easily tipped over. The war in Iran, if it gets really ugly—which it hasn't yet—if things collapse and inflation comes back, these are disruptors. There are multiple different opportunities for disruption, but all based on the war.

When we look at prior credit-market cycles, like you mentioned there, was the challenge not that the underlying assets were of poor quality? When we look at the hyperscalers today, yes, Meta is having a bond issuance that's priced higher than expected, but Meta's core business is throwing off hundreds of billions of cash flow. They don't really need to borrow; they could do it off balance sheet. It's an optimization game. The assets are good.

I think you just saw that the free cash flow is the lowest it's ever been in the history of the company, number one. Number two, back in the dot-com bubble, there was a lot of fiber that got laid in the ground, and that fiber was always valuable, but the companies that laid the fiber and stuff all went bankrupt.

When you're really heavily dependent on debt and there's a dislocation, the underlying value of the asset declines. It may not decline forever, but it declines pretty sharply in a very short period of time, and that's when you have margin calls. That's the way it goes.

What should they do from here? They should not take out such levels of debt. How do you expect this to play out?

2. Hyperscalers Outlast Neoclouds

Well, look, people are making decisions based on the risk they see to their business. If there is a dislocation, no one is better prepared to survive it than hyperscalers. All the hyperscalers have enough ongoing business, and they've been very consistent. That's why they're worth what they're worth.

The Magnificent Seven is there because they've been doing this for a long time, and so they know that they could absorb this. If it happens, it'll be good for them because everybody else gets wiped out, and then assets become cheaper for them to acquire, and they're still in good shape. The demand for AI compute is not going to change. That's not going to go away. The issue is the ability to fund it in the short term.

Let's take neoclouds. There are a whole bunch of them right now. I think at least half of them go away within 36 months, and if there's an economic disruption, a lot of them will go away right away.

Can you help me understand that? I can't parse that. What will separate the neoclouds that go away and become valueless from those that retain value and become even more valuable?

That's the same question: which hedge funds are going to go away and which ones aren't. If you looked at Leopold's returns, you'd think he's never going to go away, and he's probably going to survive this because he still has a good return for the year. But people are going to be a little wary about his risk-taking capabilities.

Ultimately, it's the people running the company. What's going to separate one neocloud from another is who is running it and how they're organizing it. You and I and everybody else can't see under the covers how that company is being run.

I can tell you, if you look at inference providers, I think Fireworks was making a lot more money than Baseten. You look at the efficiency there and think, “Oh, well, Baseten's raising money at the same valuation.” Well, it's not the same business. I'd bet on Fireworks over Baseten. It's a 10-times-better business, in my opinion, because they're more capital efficient.

That's purely based on capital efficiency?

Capital efficiency and their willingness to make profits on the business. I think Cursor was pretty smart in that the Baseten contracts from last year with Cursor—I don't think there was much profit in them for Baseten. They just got revenue and scale from them, but they didn't get a lot of earnings. If you're not making a lot of money and you're putting up a lot of money, you're at risk. You're absolutely at risk.

You mentioned Fireworks there. We had Lynn on the show, on an amazing panel, where she said that specialized intelligence would actually be the future, and that the majority of companies would have their own models trained on their own data, and that would be very important.

Do you think we have a world of millions of specialized models in this way and a couple of frontier providers? How do you see that?

3. Specialized Models Expand AI Demand

There are 2 things. In the big overall view, if we say that models are there to provide intelligence, and we look at the world and see that we've got 7 billion people, how many intelligent people do we have in the world? In some ways, you're going to see that models are going to replicate humans in the sense of being specialized and being able to do a specific task in a specific way.

Someone who's cutting a gem has a certain intelligence about how to do that work that's pretty specialized. I think you're going to see that, in the early days of these models, it's all going to be about specialization and the ability to customize.

What's happening is that you can't customize Anthropic models or OpenAI models right now—not the big frontier models. You're not allowed to do that. So that's just giving an opening, I think, for open-source models to be tuned.

As I mentioned on our last call, I thought that open-source models and ASIC chips were going to be part of a tsunami of their own. If you're looking at a frontier model with a double-digit cost per token, and you're looking at an open-source model that's 10 or 11 cents per token, while all tokens aren't created equal, it's still enough of a difference that there's going to be massive adoption.

It's not like AI is only demanded by enterprise customers or a few consumers. It's a global demand by every business in the world today, even though people haven't quite acted on that demand. It's like websites at first, right? In the 1990s, only a certain number of companies had websites, but the building out of websites has not slowed down. It's massive. The desire for that and the need for website building continues to this day. It's endless, and I think it's the same thing when we look at intelligence.

The demand for it is going to be endless, by endless numbers of people, and this is helping create the opportunity for open-source models and ASIC chips.

I totally get you on the cost efficiency of open source compared to frontier models. But what everyone says is that you're seeing the token traffic go toward open models, and you're seeing the dollar traffic—the revenue—go toward frontier models. Is that how you expect it to continue? Will frontier just be paid a lot more for harder problems, and open source take the majority of the easy ones?

That's a great question. I'm going to give an answer, but I want to caveat it this way: there's an opportunity in the answer for short-term disruptions that last from 3 months to a year, where economics appear to have leveled out.

As long as the frontier model companies—and I'm convinced they have goals for continuous learning and, ultimately, lifelong learning in these models—can execute against those goals over the next decade, the demand for those models will never cease, and so they'll just continue to grow.

But because the global demand is so massive, I'd assume that we're probably at single-digit demand fulfillment today—low single digits—and I suspect that open source has a long way to go to fill in the need. Of course it's going to be low cost, and of course most of the dollars are going to go to the people who can afford to pay for them.

Teslas were really expensive at the beginning, and only wealthy people could afford a Tesla at the beginning. Now that's changed. I think that's the way it's going to work: only the wealthiest companies and people can afford these models in the early days.

Open source is going to be playing a big catch-up game in terms of revenue, but they're going to get a lot of money and a lot of things very, very soon. It's coming.

“A token is a token” is actually what Gavin Baker said the other day, and Jensen doesn't give a shit whether you put it on a frontier or an open-source model. He wins at the end of the day.

Right.

Do you agree with that perspective, and how did you analyze his open-source evangelism with his letter?

I disagree with “a token is a token.” That may be true at the moment with pretty much all frontier models, but I disagree with it because companies like Fireworks and others are helping companies customize. The more you customize the model, the more the token changes its value, because the more you customize what's being produced.

Some models talk a lot more than other models, and so they produce a hell of a lot more tokens, just like your guests, I assume.

Essentially, you're saying that the efficiency gains that can come when you work with a provider like Fireworks mean that one token goes a lot further than another token in a lot of cases.

There are 2 things. The more each model is customized, the more it's going to have a particular, I'll say, style to it. Whether it's a verbose style or a style of brevity, that's going to matter a lot in the overall cost.

I think that what we're going to see, as enterprises and people with money have more time to understand these models, is that more and more customization is going to occur to do specific things. When you talk about agentic systems, it's natural that you think, “Hey, you're going to use an open-source model for coding, because they've got that figured out.”

But for customer service and onboarding, you can see open-source models being really ideal because it's a highly specialized task. I think that highly specialized tasks are going to be something that pays off a lot quicker and a lot cheaper.

If you only have $1 million to spend, you could spend that $1 million on customization and getting specific tasks done a lot more efficiently than you can on a frontier model.

How does this not cannibalize the frontier models' business? I'm not one who wants to see OpenAI and Anthropic challenged. Their thriving is good for all of us. But I don't understand how that doesn't cannibalize their business, making their TAM smaller.

Don't forget, we're talking about intelligence. You want more of it all the time, and you want different flavors of it, right? Humans have emotional intelligence. They have all different styles of intelligence. We know that, learning-wise, some people have more visual intelligence, if you will. I think we're going to see the same thing with models.

We're absolutely going to see this thing that prohibits the cannibalization of frontier models, at least in the short term. What I've seen with every new wave of technology in my career is that the market initially expands. Then the contraction comes when there's a contraction in global markets, and you see the fallout. Then you start again with more innovation.

And it's this continuous cycle of Cambrian explosions of innovation, followed by these sorts of glacial periods when ecosystems collapse, and then they grow back again. I think as long as frontier models can continue to innovate, because they have the money and the ability to do more innovation, there's never been a time yet when the open-source model has trumped a frontier model on innovation. They might be better at specialization, but they certainly don't compete yet in being able to do complex tasks.

When you think about that challenge of frontier versus open, Alex Karp has said the biggest enterprise customers in the world don't want to work with frontier model providers. They're scared that they're going to eat their lunch—moving to that business. Is that true, or is that Alex rather self-servingly saying, “And Palantir will help implement a full infrastructure that's not that”?

4. Enterprise AI Needs Sandboxes

Well, look, Alex has a bunch of customers that, of course, are very worried about that. The CIA and all the intelligence agencies, of course, are paranoid about that. So, one, he's got a huge customer base that thinks exactly that way.

But, 2, look, there are 2 issues, right? One is the data, and yes, Anthropic has been getting all this data, and OpenAI already has, for years. So, in some ways, the cat's out of the bag, right? I mean, if you worry about it from a security perspective, Apple could already rob my bank tomorrow. They have all my passwords. Apple has everything on me if they wanted to.

I think Amazon has a lot of data, too. Maybe not the prized data, but they have a lot of data. Microsoft—Satya Nadella himself came out and said, “Look, we've got a lot of the data around the communication, how communication works inside an organization.” So, enterprises have already given up a lot of their secret sauce over the years, and they should be cautious.

I agree with Alex that enterprises need to be smart about just continuously shoving their data up into Anthropic and OpenAI. They need to practice discernment, and they need to think about what they want to keep behind the firewall. They need to continuously have an iterative conversation about that and be careful.

I think, in part, this is what's going to drive the open-source opportunity. Yeah, you know what? We don't want that stuff uploaded into the cloud. We want it behind the firewall. So, look, 2 things: 1, recognize that enterprises have already given up a lot to third-party companies, and, 2, yeah, they need to be careful going forward.

They need to be careful going forward, and security is front and center more than ever before. We're seeing hacks like we've never seen before. We're seeing OpenAI and Hugging Face. Anthropic came out saying, “Hey, mea culpa, our models actually hacked 3 companies.” And it's almost like a brag now to have—

Yeah.

—like a... You know what I mean?

Yeah.

“Our model hacked companies. Thanks.” How do you think about the golden age of cyber that's to come?

Well, there's no question that there's a heck of a lot of complacency around security overall. Even people who think they're getting the job done are complacent. They haven't really thought it through. I don't know how many developers are running their models in YOLO mode, but probably a lot. They're thinking, “Oh, I'll put the model in a container. I'll put the tools in a container. I'm okay.”

Well, containers aren't safe. You need sandboxes. This is why the big container company, Docker, itself said, “Hey, containers aren't safe. You'd better put it in a sandbox.” That's why they've had a huge success with Docker Sandboxes. That's why E2B is successful with cloud sandboxes. People don't realize how important it is to really step up the game on security. Everybody, in my opinion, is underestimating it.

When you think about investing today personally, do you want to do a lot more in security? What can I take from that? I'm a venture investor. You know this. Dude, you know I'm here to ruthlessly make money, Jerry. You know me. What should I take from that?

Well, look, if you look at Fireworks as a team, I'm lucky I invested in Fireworks versus Baseten or CoreWeave or any of the other inference providers. Why? Because that team were the geniuses who understood PyTorch in particular. So, you've got this team of people who have a different take and a different set of aspirations for what they're doing.

They're going to move up the stack. They're going to move up and do a lot more fine-tuning, refinement, and customization for people, and they're going to be the best ones at it. That's why they're going to succeed.

I think you're going to see the same thing in security, and the most important thing that's underestimated is the need for sandboxes. The truth is, there are going to be thousands of different forms of sandboxes, and you're going to need a company that understands how models look at tools and what that behavior is, and is able to take that behavior and optimize for it.

Because these agents, we forget sometimes, are probabilistic. It's not like a developer says, “Okay, I'm going to go build this little app over here, and I'm going to pick 2 different libraries. I'm going to see which one does best, and I'm going to generate my app.” No. An agent could say, “I'm going to open up 100 different sandboxes with 100 different libraries, then determine which is the best app, and use all these different tools.”

That knowledge is in a handful of companies today, and if you look at E2B and you look at Docker, they're probably the 2 best at understanding all that stuff. So, you start there, because if you don't get the sandbox right, forget everything else.

I have to ask you: You've mentioned Fireworks multiple times. Fireworks has margins in the 35% range, Lin Qiao said publicly on the show. Many companies within the AI application layer in particular have very depressed margins, lower than that—20%. Should we all just get used to a lower-margin generation of companies, and that is AI, sadly? Or should we think differently about margin in this generation?

5. AI Infrastructure Needs Real Margins

Early in a cycle with new technology, it's always a real estate game, right? When they wanted to populate Oklahoma with settlers, they just had this giant day, had all these people out there, pulled down the flag, and everyone ran and just put their stake in the ground and said, “This is mine.” It didn't matter if it belonged to Native Americans or not. They just did it.

I think, in the same way, you've got people doing the exact same thing. They're taking low margins or zero margins in some cases just to get the customers, to get the relationship, to get the real estate. It's a strategy. If you've got capital, you're going to own the real estate, and then you're going to go back later and get the margins built into the business.

So, you don't worry about that. You're like, “It's a land grab. It's more important to put your flag in the ground, and we can expand margin later.”

Well, it's just a strategy. I would not invest in people who build a culture around that idea of low margin. That's why I probably lost out on investing in Amazon, because I just didn't buy into Bezos's idea that “Your margin is my opportunity.”

In groceries and books and things, yeah, he's right, and in compute, he was absolutely right, and he scaled it in a different way. But most people aren't thinking like Jeff Bezos in that regard. I believe that you need to build a culture that works, and if you don't want to give all your company away to venture capitalists, you might think about monetizing and generating effective margin.

You can be clever about it. If I talk about the sandbox example, the newest thing is, most sandbox people make money on compute. Frankly, that's a dumb idea, in my opinion. You want to have a model that says, “Bring your own compute, and we'll make money because we understand how to run sandboxes better. We know how to network them better. We know how to provide traces better. We know how to do all these things that are going to give you visibility into what you're doing.”

Again, innovation is what should be at the front of your mind, and that innovation better drive margin. So, even if you don't have it right this minute, if you're not thinking about it, I won't invest in you.

You said, “We need to rebuild the entire stack.” When you think about that rebuilding of the entire stack, if you start with chips, we see more and more people coming in at the chip layer, whether it's Etched or Fractile in the UK. How do you see that?

I said it last February. Look, ASIC chips are really ideal if you're thinking about model customization. If you're saying, “Look, we're at a new phase in this AI build-out, but what we really want to do is a lot of model specialization,” you don't need a GPU for that. It's too expensive. You can absolutely use an ASIC chip.

I think the number of people designing ASIC chips is growing because they recognize that trend and want to take advantage of it.

Do you need to own the chip layer as a model company today, do you think? When you see DeepSeek building its own chips, you see Anthropic now building its own, and Jalapeño from OpenAI—

I'm not sure Jalapeño was the best name. It bothers me. I love Mexican food, I love spicy food, but I'm just not sure about that one.

Well, the next iteration is called Padrón pepper.

Look, actually owning the chip is going the wrong way long term. Short term, it makes sense for larger companies because they want to optimize chipsets for models, and that's what they're thinking about.

But when I was at the Santa Fe Institute and on the board there, we had a meeting with a bunch of chief scientists from all the major AI companies. What we came back with from that meeting were 2 points. One, we don’t know how to measure AGI even if it shows up.

The second point that’s useful to this conversation is that we should focus on the complexity between the model and the agent and, therefore, the human being, to the degree that they’re in the loop. That’s where the opportunity is. Sure enough, that’s where the most compelling place for me to focus on in investing is—that level, from the loops to customization, multiple things, security. All of that stuff from the model out is where I think it’s far more interesting. Going back to the chips, in the short term, I can see why people do it. Long term, I think it’s unnecessary.

6. Venture Needs Discernment

I’m an investor in Legora, and they obviously fight intensely with Harvey. When you look at the 2 of them, you think, “God, what a competitive landscape.” What have been your lessons over the last few years, decade, 2 decades, when you have 2 very well-funded competitors like this?

Fund the guy who is the small startup right now and watch them battle it out. Particularly in the legal market, when you’re in that competitive situation, you’re going to take risks, and you might regret it. The first one of those guys to have a security leak or security problem—and it will happen—is going to wreck their market opportunity.

You think, “Oh, well, the other one’s going to win.” Well, probably not, because they’ll probably both be vulnerable. They’re looking at each other, and they’re watching what each other’s doing. So for me, I look at that situation like, “I want to go for the next innovative young company that’s maybe not trying to do all things for all lawyers, but is more highly specialized, like GetDynasty is in the trust world.” Do something very specific.

By the way, this has been Peter Thiel’s advice: Start with a niche, dominate the niche, and then grow it out. When you’re trying to take on a whole ocean like the legal system, I just wonder if it’s contrary to Peter Thiel’s advice.

Do you worry that we just throw price out of the window? It seems like we’ve never been less price-sensitive. This is crazier than 2021, Jerry. I’m investing every single day on the ground, and I consistently have founders say, “Oh, you know, we’re raising $100 million.”

I’m like, “Oh, wow, that’s—how much are you raising?” And they’re like, “We’re raising $100 million.” I’m like, “That’s the friends-and-family round?” When I said to a founder the other day, “We write $25 million checks,” they were like, “Okay, good, so you’re small and collaborative.” Have we just lost price sensitivity, and is that okay, given all outcomes can be trillion-dollar companies?

It’s evidence that we’re still in the hype cycle, right? We’re in the hype cycle because expectations are beyond everyone’s imagination. If you’re saying, “I’m going to be a trillion-dollar company. My opening round’s $100 million or something, or even $1 billion, whatever the valuation is,” you have to look at the people who are going to take that money and recognize: Has that number been well thought out, or are they just doing it because the market’s doing it?

I would argue that it looks like Anthropic and OpenAI, those crazy mega-rounds at $100 billion and $150 billion, might actually have been cheap. I think Gavin Baker probably believes that, and others believe that. So for a few companies, yeah. But how many trillion-dollar companies are we going to have?

Discernment, again, is necessary here to decide what really can have the sort of hyperscale-type growth associated with it and which ones are going to be also-rans or a slower-growth opportunity. We need to sort those out a little better.

Is slower-growth venture a thing anymore, Jerry? If we look at Fireworks, it’s 3½ years to $1 billion. It’ll be 4 years to $2 billion if they hit end-of-year targets this year. Four years to $2 billion. Jerry, do you remember when it was Slack—18 months to $10 million—and we were like, “Wow! Wow!” Game changer.

I think in the case of Fireworks, they’re benefiting because of OpenAI and Anthropic. They are the next level. Now, with open source taking off, they’re benefiting from that. They ride on the shoulders of these model builders, and they’re the next layer that needs to get developed. They can scale right behind that.

But if you’re somebody else in the app layer, I’m just not buying that. I’m not buying it for legal, and I’m not buying it for the app layers yet. But for infrastructure, absolutely.

But I get killed for this. I say it publicly: triple, triple, double, double. You remember this. Am I glib and myopic, or am I a product of a cycle?

You are glib sometimes. Not all the time, but right now you are. I would argue that general statements don’t apply here. You have to be highly specific.

If you look at Anthropic and OpenAI, what they’ve done has never been done in the history of the world. It shows the importance of the time. If we look through the history of venture, we are definitely in a completely different era, and the frontier model companies have done something extraordinary in the history of the world.

This is definitely on the level of inventing fire, inventing electricity, whatever you want to call it. It’s truly extraordinary what the frontier model companies have done. They’ve lit the match to AI.

The companies that can follow right on top of them and not get killed by them, but can grow, solve more infrastructure problems, and help create an ecosystem around the model companies, deserve those economics. Other categories, no way. For example, neoclouds—no way. I don’t buy it.

I think 1 or 2 of those neoclouds are going to end up dominating, and at least half of them are going to go away. They’re going to go away with massive amounts of money being burned as part of it.

Does the model-routing layer carry enough value to you to be independent?

My opinion is OpenRouter has massive amounts of transactions because people are basically lazy, right? It was easy: “Okay, I need to connect to this model. I’m just going to use OpenRouter.” OpenRouter charges 5% on top of that, which is a crazy amount of money. That’s not going to last.

You’re going to see exchanges. There’s a blockchain company called Akinaki that has just launched DODEx on their mainnet, and this thing is an exchange to go out and buy inference. As part of that, all the model routing is done for you.

I think you’re going to see multiple alternatives to an OpenRouter-type product, where people hosting the models themselves will provide, through an exchange, an easy way to acquire the inference. The need for an OpenRouter-type product, particularly paying the 5% markup for the inference, won’t matter, right? That’s what those things do. They’re not necessary long term.

So if you’re on the board of OpenRouter and the $10 billion acquisition comes through, what do you say?

You say, “Fuck yeah. This is great.” A lot of people take the money when they can. Credit to OpenRouter. They were there early. Developers didn’t see another alternative. They could just go there, go to the API, and they were willing to pay a 5% markup to get their inference.

And guess what? Shame on the enterprises for letting them burn all that money. That’s a huge amount of money. I think you’re going to see a big disruption in that model in the next 3, 4, 5 months.

Actually, it’s not just Akinaki. Venice.io is doing that, and there are 2 or 3 other companies now in the process of building exchanges where you can directly get the inference you need without paying the 5% markup.

I think you very accurately said where we are today: the potential dislocation of excitement, dislocation due to external affairs, and then the reblossoming of an ecosystem, so to speak. If you think about that, and you advise me as a venture investor deploying, say, your money today, what would you say to me?

Play the game on the field, Bill Gurley-style? Be mindful; don’t spunk cash into Neo Labs at a $1 billion pre for 1 person out of OpenAI? What would you say to me?

Bill’s on the board of the Santa Fe Institute with me. His guidance is pretty smart. He’s pretty much on point with a lot of things in venture capital.

What I would suggest is that you look for impact. Look for people who are just way different from anyone else. You look at them and realize it’s not that they want to build this business; it’s that they have to build this business.

If you find that in a person and recognize that they have the commitment to it, because the commitment is all in—there’s no other option—look for that. Look for the fact that what they’re going to do has impact if they do it.

When you review the founders you’ve worked with, where was that most obviously striking?

It’s rare, right? You could say all the founder-driven Magnificent 7 companies would qualify, right? Elon, Jensen, and Zuckerberg all qualify for that definition.

But you look at these other companies and say, “Wow, Fireworks AI looks to be like that.” AtoB is definitely like that. The founder, Vignan Velivela, is absolutely gonna do it. There’s no question in my mind.

Oven [?]—this guy came out of Meta as well. His name’s Sadie Khan. But Ankush said this is one of the best CEOs he’s ever seen. Vinod Khosla was one of the best CEOs building Sun Microsystems. So when someone says that, you take it seriously.

Do you think we see a compression in liquidity timelines? We have Cursor scaling to a $60 billion sale in 4 years. Do we see venture cycles get shorter in this environment, given companies grow faster?

I think what Cursor did—the team is really smart—is that they pivoted out of the IDE space. In that pivot, they convinced Elon that they could build models. They hadn’t proved it yet, but they convinced him that they knew enough to do it, and Elon was pretty desperate to solve his problem with xAI.

It was a great fit, and they got the $60 billion, so you hit the bid. If OpenRouter gets a $10 billion bid from Stripe, you take it. I think those are not the norm. Those are events that are happening because the board and the management realize, “Hey, maybe what we’ve built isn’t a decade company. Maybe this is something that we need to move out of, and we take the win for what we had.”

I had a few companies back in the day that I wish had done that. Flipboard was one of them, and I wish Flipboard had taken the $1 billion exit, but they didn’t.

What happened there? They had a $1 billion exit on the table.

They had an opportunity, yeah. They had 2 bidders going for them at the time, and they were very, very interested in them at around, I’ll say, within 20% of that number. One of them was Twitter, and the other one was TikTok, through the founder of ByteDance.

He got advice from someone called The Coach, who was pretty famous at the time, that said, “Hey, don’t sell your company.” The Coach was, unfortunately, passing away, and the board bought into The Coach’s advice, and they stayed with it.

Now Flipboard—you don’t care about it, right? They missed the opportunity. I think those opportunities happen with a lot of companies. I have a lot of arrows in my back from this situation. So you just have to know when it’s a time to go and when it’s not a time to go.

We mentioned that, obviously, Cursor is selling to xAI. It seems like IPO markets are open for the rare few—for Anthropic and OpenAI when they want to, and for SpaceX. But I’m concerned that your Airtable of the world, at $4.85 billion, couldn’t IPO. You can’t IPO with less than $1 billion in revenue today. Does that concern you?

No. I just think that we’re at a moment in time where, if you want a proper IPO, you need to be on track for that. But I would think Cursor, if they had gone out last summer, if they wanted to, they could’ve gone. I mean, they would’ve been taken public, despite the fact that I think the management team was wise to realize that they weren’t quite ready for that and didn’t do it.

But they could’ve. Absolutely, they could’ve. The numbers were crazy. There’s always a banker willing to do it. The question is which bankers, and whether it’s the right thing to do.

Do you worry that Airtable is the start of a much broader generational cohort that will be sold at a mega discount to the last round?

No. First of all, there’s not that many buyers, like Bending Spoons, right? So there’s not that many buyers there, so I don’t. I think the companies, if they’re at $400–$500 million and they have that kind of revenue, the question is, are they gonna continue having the revenue?

If they haven’t already integrated AI in a compelling way, I’m not optimistic about their future at all. If you don’t have a really thoughtful AI strategy and a thoughtful AI product, I don’t believe that you’re gonna have an opportunity to do much of anything with the company in 2 years.

Are we not seeing most of the SaaS generation put lipstick on the pig, so to speak? “Oh, fuck, let’s sprinkle some pixie dust in this.” “And, oh, now you’ve got an AI copilot, and oh, there you go.”

Well, you know what? It’s a good thing you mentioned that because we’re in a new era now. We’ve gone to what I call the co-work era, where it’s really more agentic, as I mentioned with autonomous agents. For those few companies that have deployed autonomous agents successfully, co-work is becoming the new trend.

As co-work becomes more successful, more stable, and more broadly used, I would be really concerned about SaaS companies that don’t have some kind of system of record or some kind of AI strategy in place to succeed. The co-work era begins the threat. So the threat to the SaaS world is just starting right now, but it’s still early days, so you’ve got time to pivot.

If you’re a SaaS company, you’ve got time to do AI, bolt on AI, and figure out some other direction. But if you’re not doing that now, good luck.

Good luck if you’re not doing that now. Dude, I look at private equity today, and I like the PE model, but I’m looking at the Thoma Bravos of the world, and I’m just like, “Oof. Ouch.”

I really like Orlando Bravo, and he was great on the show, and I want him to succeed. But fuck, that’s a hard job you’ve got with your Coupa and Planview platforms of the world. Do we just have a vintage which sucks, and we just get over it?

Well, you know, it’s amazing. In 2001, TPG had a terrible fund, and venture firms, like everybody, did. They survived it because they had a whole lot of telecom investments that just evaporated. Forstmann Little had a lot of telecom investments, and that led to the end of the firm. The firm ended and died. No more Forstmann Little.

So, look, I suppose these PE firms that have challenging portfolios have time to do something about it now. But when and if a financial dislocation comes, that’s the problem, because they’re all levered up. The problem with the PE business is the leverage on the businesses.

If EBITDA drops, churn increases, and if it happens rapidly through a financial dislocation, there’s a margin call effectively on the debt. Yeah, it’s gonna be tough. It’s gonna be really tough.

Dude, a lot of these assets are 4–6× levered. I mean, it’s high.

Yeah, it’s high. I mean, look at Leopold. He was only 3.5× levered, and he had to sell a lot of assets. I agree, it doesn’t look good, but look, they have enough EBITDA today, and they have PE firms that know their survival’s at stake. The PE firms have time to come up with some strategies as long as the market stays up.

This is the point I was making: The global markets are critically important to what’s gonna happen in the tech sector. Critically important. If you have a dislocation—I think it was Tom Lee that called for a 10% decline, or drawdown, in the S&P this fall—if he’s right, if it’s any worse than that, I don’t know how people handle it when assets deflate and you’re levered up. I don’t know how you handle that.

Yeah, you’ve said that the 10% drawdown, and you said earlier about FIRE and inventing FIRE with the frontier models. Sam was like, “Hey, administration, take 5% of frontier models.” Do you think the answer, when you create AI, is that you have to be owned at least partly by the administration?

Well, first of all, that’s never happened in the history of the United States until this current administration. So that’s never been necessary. I don’t see why it’s necessary now.

If you look at utility providers in the UK, you have your British Gas and your British Telecom. I know they’re not now because they were sold and privatized.

Yes.

But you had your Royal Mail. Actually, the majority of utilities were state-owned.

But look, I mean, you have a history of socialism in European countries. Post–World War II, socialism has existed, and people have always supported that. I do think there was a need for the governments to get involved because there wasn’t the capital markets available to them like there were in the United States.

Obviously, in the United States, there’s been public-private collaboration. But Sam’s already built the company to this size without needing to sell 5% to the government, so why does he need to do it now? It makes sense if, like a Manhattan-style project, we did that for AI 10 years ago. Fine, do it because it’s strategically important to the country.

But today, given their size, I think the only reason you’d do it is for political reasons.

Talking about strategically important for the country, do you think it’s right that we have export controls on chips?

I think it’s important that we think about how we’re going to deal with our technology. We need to really have a strategy. I’m not a believer in regulation for regulation’s sake. Put it into context. Give us a strategy. Let’s publicize the strategy. Let’s debate the strategy. Let’s have people responsible for it.

We don’t need just some regulator to come out and say, “Let’s just do this.”

Do you worry about the dominance of Chinese open-source models and the ability for backdoors to be introduced into their models, or do you think this is grossly overestimated?

That’s a really good question. The main thing about open-source and Chinese models today is, 1, all these models are not gonna exist in 10 years. There are gonna be completely different ones.

If there are backdoors today, they better do what they're gonna do now because those models aren't gonna exist in 10 years.

What do you mean by that? Like, Kimi won't be a dominant model in 10 years?

I mean, all the open source models that we're using right now won't be used. They'll be replaced by something else. First of all, within 10 years—I believe, and I think some people are thinking 2 or 3 years—continuous learning models will come into existence. That means that every generation, every model we have today, dies and goes away.

Two things: What is a continuous learning model, and why does that mean every generation dies, for those that don't know?

7. Continuous Learning Replaces Static Models

Because today, one of the most important goals of the model builders, particularly frontier model builders, is continuous learning so that models can do more complicated tasks. In theory, just like humans, we're continuously learning. Robots—so, physical AI—will need to have some ability that maintains their memory so they can continuously do complicated tasks, learn, and deal with dynamic events that come into their environment.

These models will be fundamentally different from the models that have been trained to date. Continuous learning models will come in, and once they're deployed, there'll be a whole new breed of open source models based on this new capability of continuous learning. Then those will evolve into what's called lifelong learning, which is truly more how human intelligence works. But those models, in my humble opinion, will replace every model that exists today.

Does continuous learning and potential lifelong learning not denigrate the value of frontier models?

Well, they're gonna replace frontier models. I don't think you can bolt on continuous learning to an existing frontier model. I think they're gonna try, and the early stages will look like that, but ultimately, I think it'll call for a new form of architecture and completely new training.

When you train a model today, it's kind of static. It's trained, and then you go out there in the world. Continuous learning models, I think, will be architecturally different.

Would you have done SSI at $30 billion, Ilya's company, which is supposedly coming out with the first version of its continuous learning model at the end of August?

I don't know him, and I don't do model deals like that unless I know them or someone I trust knows them, so I can't say.

You said there about training being kind of a one-shot-and-done process. I'm an investor in Mercor. I think data itself is much harder than people give it credit for in terms of acquisition, cleaning, and deployment. How do you feel about data-providing companies as a commodity or as a valuable asset?

Well, data keeps changing. The thing about data is that it's not static. There's, of course, value to context, right? Data gives you context and memory.

I do think that if you're an enterprise business, your data and the way you use it will be different from the way someone else uses it. Take hamburger companies: Shake Shack's data is gonna be utilized differently from the way Burger King or McDonald's will use it. There are gonna be highly unique use cases for data that are really important, but you have to have a system where the data continues to evolve and change, and the underlying utilization of it can change as the data changes.

I think I buy Lynn's thesis that you'll have specialized models for companies, and part of the training for those models will require additional surplus data. Then you'll see the likes of Mercor go from purely selling to frontier models to selling to enterprises and even the mid-market, which need specialized data that they might not have. That massively opens the TAM. That's the $200 billion opportunity.

Right now, models are largely task-driven. Frontier models are doing some levels of creativity, but we really want deep creativity. We want models to go solve climate change.

You're gonna need a diversity of intelligence. You think about board levels and management teams: You need diversity of intelligence to be able to solve really difficult problems. The market is gonna change from, "Let's just solve tasks and do it a great way. Let's do minimal creativity with writing and visual arts," to, "Oh, we really need a lot of diversity of intelligence to be creative enough to solve the problems that matter and the problems that we're gonna get paid for."

I find everything that we talk about today so exciting. I just have one pullback in my mind—

Yeah.

—which is that a lot of wise people say you always overestimate what you can do in a year and underestimate what you can do in 10. Is that the case here? Am I massively getting ahead of myself when I think about a lot of what we've spoken about? Should I actually calm down, kiddo—it takes longer than you think?

My feeling is that continuous learning feels like it's 2 to 3 years away. Maybe it's 10 years away. We don't know. We've been thinking we're on the cusp of solving cancer for the past 15 years, and we got a little bump with new drugs like Keytruda to do cancer-solving work with the immune system. But we haven't solved cancer yet. We're managing it better, but we thought it'd be over by now, and it hasn't happened.

I think I apply the same thinking to continuous learning models. We are getting a little bit of success with sample-efficient models. Sample-efficient models mean that when you get just a little bit of data, a small sample, you can extrapolate enough beyond that to become useful and to learn from that small sample. Learning from small samples is starting to happen, but when those models actually become robust enough to be useful, I don't know.

My feeling is that these things take a breakthrough from where we are. Just like with cancer, we need more of a breakthrough, and the complexity that they're trying to solve is huge. I'm not gonna bet against them, but I would say what could send us into the valley of disillusionment is a combination of a global financial event and a failure for the models to continue to grow and evolve.

We're gonna get to a place where, if we don't solve sample-efficient models and we don't solve continuous learning, we're gonna feel like our inflated expectations are somehow not being met.

Jerry, I'd love to do a quick-fire with you because I could talk to you all day. Who goes out first, OpenAI or Anthropic?

It appears like Anthropic.

Over or under: NVIDIA will be a $10 trillion company in 5 years.

Over.

Why is it so mispriced right now, then? It's been flat for the last 12 months despite numbers going through the roof. I don't get it.

I think it's because the market doesn't go like a rocket ship forever. You're gonna have these plateaus, and I think there are areas where things aren't really accelerating as fast as you think they are.

We don't know, because right now there are these circular transactions that are obfuscating real growth in the market, as the hyperscalers are trying to get ahead of the game. But we don't know. Demand is gonna have a lot to do with people having money in their pocket. If you have a 10% or 15% dislocation in the markets, people are gonna feel poor, and that's gonna affect the credit markets. That's gonna affect everything. That's gonna affect demand. It always does.

I can't believe I'm about to say this to you, Jerry, but fuck it, we've known each other a while. In the UK, we have a game called Shag, Marry, Kill. I'm gonna apply it to 3 companies, and the application is: Kill is short, Shag is buy quick but you'll probably flip it, and Marry is you're in it for the long term. You've got Meta, Google, and Microsoft.

Because of the scale, I'm gonna say long term for all of them.

No. Really?

Here's why: When Meta's got 2 billion users with all their things, and when Google's got 2 billion users with all their things, there's a kind of stability in that because consumers are really slow to accept new changes and things.

Because those 2 companies have such substantial consumer businesses, and Microsoft's consumer business is good too, it acts like a buffer, a stabilizer that gives them time to catch up. Let's face it, all 3 of them have failed on the coding-agent side, but I don't know if they're gonna fail forever.

You'd have to say this massive customer base gives them incredible time to catch up with problems, and that's why they're gonna be trillion-dollar companies for at least a decade. They may not be as important as they are today, but that's not the question you asked. As an economic buyer, would I hold their stock for the long term? Yeah, I would.

Even Microsoft? No model, relatively shitty AI products, and you'd still be a buyer?

Yeah, here's why: They control communication for global enterprises. Microsoft email—as dumb as it is, I mean, Exchange, whatever you wanna call it—that's not going away. That thing is a money machine that cannot change. It cannot just disappear.

By the way, you know what's really interesting? We talk about the speed of AI and stuff, but they won't be able to control agent communication. Human communication, that's not going away, and they're gonna be able to monetize that forever. You can't get rid of it. It's not like your cable system at home, where you can say, "Fine, I don't need Xfinity anymore."

Get rid of it. You're not getting rid of Microsoft anytime soon. It's going to be around. Sadly, Facebook is the same. You're not going to get rid of it anytime soon. These guys have built these kinds of businesses because of the scale that supports their underlying business. The consumer is, in my humble opinion, the thing that's keeping those companies afloat more than anything.

You've got to short one of the Mag 7. Which one would you bet against?

It would be Meta because that's the one that may become boring. It may become like the telephone company. It'll just be this malaise, like owning AT&T or something. That's what you think about it.

Is that so? I push back: It's the largest ad business in the world. It's got WhatsApp, and it's got Instagram.

Yeah. WhatsApp—they control human communication on WhatsApp in a very meaningful way. They haven't monetized it yet, but they're going to find ways to keep the users. You've got 2 billion-plus users; maybe there'll be 3 billion in 5 years. I don't know. But you've got half the world using your application. I'm sorry, that's something that's stable. That's a stable thing.

Like I say, it might be boring, and it'll just generate dividends out to you. As an economic buyer, you don't necessarily always need growth. You can take big, fat dividends and just clip the coupon. At least for the next 5 years, it's going to be considered a safe haven. The Mag 7 is the Mag 7 because people say, "Hey, I'll put my money there. I might be underwater for a year or 2, or get less of a return than I had." But long-term, those things are going to be there, and they're going to benefit with every positive cycle in the markets.

Are Apple's AI strategy mistakes unforgivable, or is it genius patience, waiting to see how a developing ecosystem plays out?

The answer to that is something that's going to be hard to know. We'd have to go sit down with Tim Cook, now that he's retired, and maybe he'd tell us. What is the culture around AI at Apple? How are they thinking about it? What are they doing in there? I know they're using Claude Code—they're a huge, massive Claude Code customer. But how are they thinking about it? Do they have anything innovative to say? Have they been intelligent, watchful observers, or are they just dumb consumers? If they're consuming, they're in big trouble. They're going to suffer. But if they're watchful observers and they've got something up their sleeve, then we could be surprised by them.

Which PE firm will navigate the next 5 years best?

Ooh, that's not fair. That's a tough question. I wouldn't want to make that bet.

Which will navigate the worst?

I don't know who's worst, but my company, Insight, has a very small PE portfolio. Very, very small. It almost doesn't even matter. So I feel the best about them long-term because they've been really intelligent about how they deploy the capital there. But the people that are all in on PE all the time—I just don't know about that. I think they're highly at risk to any kind of financial dislocation.

Which venture investor do you think has fared the best in the transition to an AI world?

Good question. Look, I think there's 5 or 6 firms that have just done phenomenal. Firms like Menlo, which invested in Anthropic early—early enough. They're going to do great. Benchmark, while they missed Anthropic and OpenAI, has done amazing with Factory and a bunch of other great ones that we've talked about. So despite the fact that they missed the big frontier models, they have a great portfolio.

I think Khosla is also phenomenal. Vinod with Avan and his companies—they did amazing. I think Khosla's in that group. They're just going to crush it. Khosla, Menlo, and Benchmark have all just done amazing.

I have one final one for you, and this is kind of the beauty of what we do, I think, which is seeing the future hopefully ahead. What seems crazy today that you think will be quite obvious in 5 years' time?

Ah, that's easy: blockchain for agent payments. Blockchain is in the trough of disillusionment right now. It's really at a bad spot. IBM's CEO came out and said Bitcoin is at risk of being hacked in 3 to 4 years. Tom Lee came out and said Bitcoin is at risk of being hacked in 2 years. Google said Bitcoin is at risk of being hacked in 2 years.

That greed element of blockchain, which is exactly what the Bitcoin thing is all about, in my humble opinion, is about greed, and it's bringing down the whole blockchain thing. While Solana and Ethereum are looking like they have long-term potential, I do think that new things like Venice or Aki-Naki, or what Robinhood did with Robinhood Chain—great, they got a billion in revenue, probably—are going to be real innovations.

Whether you're tokenizing stocks, doing payment rails, or using blockchain for AI inference, like on Aki-Naki or Gonka, those things are going to be real innovations. People are going to be blown away that blockchain has found true utility other than some supposed form of utility.

Jerry, I've absolutely loved having you on. No, seriously, I learn so much from you. I love what I do because of shows like this. So thank you so much for joining me, and you've been amazing, dude.

Cheers. Take care.