历史上，没有任何事物像 AI 一样制造出这么多百万富翁：Salim Ismail 与 Dave Blundin | EP #181

We may need to rewrite the whole paradigm, Peter, because what’s happening now is another double exponential on top of that, with AI accelerating everything so quickly, including the path to revenue.

Thirty-six unicorns in half a year. I really hope the young entrepreneurs coming up appreciate what a moment in time this is. This is exponential thinking. This is entrepreneurial thinking.

There is a bet to be made right now if you’re an investor: where are you going to invest to ride this curve? Because it’s not slowing down. Is it real estate? Is it data centers? Is it chips? Is it power?

All of these things are going to make somebody—hopefully somebody listening right now—huge amounts of wealth. This is an unstoppable megatrend.

The architect of AI, the alchemist of AI.

Alchemist of AI.

All right.

Peter, you’re the pope of hope.

I’m the pope of hope. All right.

Welcome, Dave. There were some incredible messages in last week’s episode about how awesome you are, and I don’t think our fans know enough about you. I’ve been your roommate since our undergraduate days at MIT, but what is Link Exponential Ventures? What do you do, just so folks can appreciate how awesome you truly are?

We manage about $1 billion of seed-stage money here in Kendall Square, right between MIT and Harvard. Fundamentally, what we do is this: when I graduated, I immediately tried to start a company, and everything is working against you at that age. It’s almost impossible to even get an apartment. You have no credit score, and there was no funding at the time.

There were no incubators or accelerators, so I’ve spent a big chunk of my life trying to identify everything that slows down a new founder. Especially in today’s day and age, almost all the deals are AI companies, and we see many companies reaching multibillion-dollar valuations in 2 or 3 years.

You have very young teams and very fast timelines, and there’s so much that we can do to help them. We started with office space and funding, then we added accounting and big data. A few weeks ago, we bought an apartment building.

One of our observations was that Mercor, from the day we invested at the seed stage to today, added $20 million a week of value. We thought, “Okay, how much does it cost to buy this apartment building?” About $6 million. Buy it.

If we can save a team like that even 1 week in their growth cycle, it pays for itself times 5. We’re just doing everything we can at Link to accelerate these young, supersmart AI teams. Funding is part of it, mentoring is part of it, and so on. The returns have been extraordinary, too.

You’re a unicorn incubator.

Unicorn incubator. Yes, that’s right.

Hey, guys, what did you do for the 4th of July? Anybody explode anything?

I went up to Vermont and tried to get away from AI as much as possible for a day or 2. It’s hard to do.

I sat in the garden with a glass of wine and a glazed look in my eyes.

Ah, okay. Well, I was in Montana building explosives. It was so great. I had my 2 boys there, and we were taking apart M-80s and firecrackers and separating out the gunpowder from the filler. We had an arsenal. It was extraordinary.

When I was a kid, I could buy potassium nitrate, charcoal, and sulfur and make my own gunpowder. I remember buying potassium perchlorate. Potassium perchlorate is great because it generates its own oxygen, and if you mix it with charcoal and sulfur—any reducing agent—

There’s a book called The Poor Man’s James Bond, and it talks about taking an old film canister, putting in an M-80 fuse, filling it with gunpowder, and putting it in body putty. I built one and threw it into my friend’s pool to see the splash. What happened next taught me something about physics: liquids are not compressible. The explosion literally cracked my friend’s pool in half.

Oh, no. I’m not sure you should be talking about some of this stuff publicly.

I didn’t know that story, after all.

But you can’t buy this stuff anymore.

You can still buy what we used to do: buy SDS model rocket engines, crack them open, and scrape out the insides.

You can do that with fireworks, too. But I want to buy the pure chemicals and teach my boys how to be true, you know.

I may need to find another podcast, YouTube.

All right. Let’s move on. Our first area, of course, is AI supremacy and speed—everywhere, always, all at once. Another crazy week in AI. Dave, tell us about this TechCrunch article.

Thirty-six unicorns. I really hope the young entrepreneurs coming up appreciate what a moment in time this is. I love to tell the students that last year there wasn’t a single self-made billionaire under the age of 30 in America. That was the first time in 15 years that was true.

Before that, you had your Mark Zuckerbergs, then you had your Larry Pages, and a whole litany of young Internet billionaires. But there was this huge dead spot over about a 15-year span. Now we’re in the biggest peak of my lifetime by far. You’ll never see anything else like it again.

Thirty-six unicorns in half a year, on this kind of accelerating rate. We’ve invested in a couple of them, and prior to this window you also had Liquid AI and a bunch of others from the prior half year. This is not normal. This is the opportunity of a lifetime, and the rate at which you need to move to keep up with it is a full sprint.

The question is: Is this the new normal? We’re seeing this crazy valuation increase and speed to not just valuation, but speed to revenue like never before.

Look at the next slide for the revenue view of it. This is also not normal. The timeline to profitability or significant revenue is super short. That means you’re stable: if there’s ever a financial collapse, you can switch to profitability instantly. That’s a powerful point.

For those who are listening and not watching us on YouTube, the median time to $1 million of annual revenue before 2020 was 16 months. Now it’s taking people 5 months to get to $1 million. It used to take 41 months—nearly 4 years—to get to $5 million. Today it’s taking 13 months, just over 1 year.

This is a 4-times acceleration in getting significant revenue as a startup. Peter, remember when we graduated? I went to MicroStrategy and then immediately back to Boston to start a company. I was building a company that was a true hypergrowth company of its era.

I founded it in 1991 and got it liquid in 2000—9 years to get to a $1 billion valuation and probably $20–30 million of revenue run rate. Nine years was really, really fast at that time. Compare that to the numbers on this slide.

Here we’re comparing 2023 to prior to 2020, which is really the prior decade. If you look at the decade before that, it was even slower.

Salim, these are all ExOs, right?

They’re all ExOs. There was a step change from 2008, when you could build a company and use cloud services. Therefore, you could scale, take all computing costs off the balance sheet, and make them a variable cost. That was the birthplace of the ExO.

We may need to rewrite the whole paradigm, Peter, because what’s happening now is another double exponential on top of that, with AI accelerating everything so quickly, including the path to revenue, which is a huge thing. This now stabilizes companies very, very early, and that’s a powerful point to be at.

Are you proposing the 3rd book in our ExO series?

No. I hope I never write another book again. The first one almost killed me. The second one killed me 90%. It was horrible. I’m more proud of it than I thought I’d be, but it was a tough process.

Dave, what you were saying—

If you look at the next slide, too, I think one of the things that isn’t mentioned here is the headcount that gets you to these numbers. It’s lower than ever before, by a lot. For me to get my first company to a $1 billion valuation, I had to hire 200 people.

For a young entrepreneur like the Cursor team or the Mercor team, the comfort zone of these management teams—30, 40, 50 people—feels a lot like college friends. When you meet the people, they are a lot like college friends.

It’s so much more fun to build one of these companies because you’re in that “I know everybody, I know what their capabilities are” environment. We have 1 or 2 kegs for the whole gang. It’s just so much more manageable.

And I bet there's a curve that looks like a peak of value per human being—and fun per human being—that basically starts falling off after a company population in the 40s or 50s.

Oh my God. I have a whole presentation on this. When you get to about 200 people—and more like 100, but around 200—there are people in your company you don't really know. In fact, you can't even remember their name when you walk by them. That is a different world in terms of comfort and management from the one-layer company where you know everybody directly. It's just a whole different world.

In the other direction, we're all hypothesizing that the company of 1 person that reaches a billion-dollar valuation is coming soon. But that sounds really lonely to me, too. I think the sweet spot—the perfect happiness sweet spot—is right where we are right this minute.

I'm going to predict, though, that we're going to have a problem: once these companies go through that super-exciting phase and stabilize, they're going to go through a lot of convulsions as they lose the buzz that got them there, right? It's a great problem to have. We could all wish for that kind of a problem, but I think you'll start to see a lot of angst, a lot of founders leaving, et cetera, et cetera.

No, they've got to keep dreaming bigger, right? It's upleveling constantly to the next level.

I mean, vibe-coding startups are here to stay. That's the story here: Stripe, Cursor to $500 million ARR, Lovable at $17 million, Bolt at $20 million. So these are a set of technologies. One of the things that makes these unique is that they are technologies that'll enable entrepreneurs to build other technologies and companies, right? So they're a base layer. Any other comments on this?

No, I think you're right. When you look at the list that we had a couple of slides ago, the base-layer companies are about two-thirds of what you see there. RHF and foundation models and so forth. You're starting to see, though, some multibillion-dollar vertical companies now starting to pop up, and there will be many, many more of those.

So, to answer your question: Is 36 a lot? Is it going to go up or down from here? It's definitely going to go up more. We have at least a couple of years of really good, very rapid expansion of that number, and then AGI. Who can predict after that?

Then it goes insane. I mean, the interesting thing is, the IPO market is just beginning to open, but the mergers and acquisitions market is wide open, and as we've seen, it's gone absolutely crazy in that regard.

Yep.

So this is another interesting slide. It says, “The new reality: vibe valuation.” We're seeing crazy valuations that are being made at opening bids: $9 billion, $10 billion, $30 billion. I've never seen anything like it.

Vibe valuations are when an investor bets big on a GenAI startup with few traditional metrics like AIR, annual rate of return, or cash burn. And this is from Vindra Mathur. This is a problem you've solved, Dave, right? Because most people don't have access sufficiently early, and then you're stuck at the later valuations when the company has blossomed. Do you invest at $120 billion into xAI, or—

Yeah, no, we never do. It's kind of scary when you do. I was talking to General Catalyst last week, actually. They've more or less abandoned seed stage because they're managing $60 billion now, and they need to write these big checks to keep that money moving.

But they're going to try and rebuild their seed-stage capability. They hired a new guy to do it. But it struck me that doing seed stage and doing $60 billion of asset management are really incompatible with each other.

What's your average check size at Link Ventures?

Our opening checks are often $500,000, $1 million, $2 million, something like that. We'll allocate $7 million to $10 million per company over time. We've left billions of dollars of pro rata rights on the table over the years.

Pro rata rights mean we have the right to maintain a certain ownership stake in a company, but if it gets to a $2 billion, $4 billion, or, in some cases recently, $10 billion valuation, we can't even come close to keeping up with our pro rata rights.

I mean, one of the questions I get asked all the time—and I'm guessing, Salim, you and Dave do as well—is, “What AI company should I invest in?”

Really, you want to—first of all, you're not going to get an allocation in Anthropic, OpenAI, or xAI. Those offerings, when they do secondaries, tenders, or their next round, are snapped up instantly. Then you're left with the public companies, which are—I mean, there is growth, but you're never going to see the real valuation growth.

So you're either in an AI venture fund like Link, or you're going to your local university and trying to find a startup and put it in there.

Well, the other option is to pick a domain that you know really well and then invest in AI companies where you can gauge the value add and the differentiator from others, right? That's the only other way to do it. Pick an area of existing passion that you have and go down that route.

Well, the other thing you can do, which I love just as a life strategy, is find 3, 4, 5, or 6 venture funds that you really, really like that are seed stage. Then route their pro rata rights out to their investors and say, “Hey, look, we have capacity here. We can't keep up with it.”

What we do is we just email our LPs. Virtually none of them read their email and react, and usually these deals, like SpaceX or xAI, you have a week or 2 to make up your mind. You don't have a month.

But if you're the 1 investor that actually reads their email and pays attention, then there are great opportunities that come your way through those. Just pop open the email. If it doesn't fill with those LPs, then the new investor just takes it, and 2 weeks later it's gone. So it's a pretty good strategy.

All right. The big news this month has been the release of SuperGrok. I love that. We went from Grok to SuperGrok and Grok 4, and GPT-5 is coming. This is Sam Altman on a podcast with his brother. Not much content here, but worth hearing.

What is the time frame for GPT-5? When are we going to see this?

Probably sometime this summer.

Okay, that was a big data point there. GPT-5 this summer. One of the things that's interesting about GPT-5 is they've made a proclamation: no ad-influenced answers, and they're really trying to strengthen user trust. So, any predictions here, gentlemen?

Oh, yeah. I think it's funny. Polymarket seems to think that it's less than a 25% chance.

Yeah, let's go here. Here's the Polymarket bet.

I was almost sure—more like 80% in July—because there are all these pictures of Sam giving little hints, and also the competitive pressure kind of would line up with Grok. So I was really thinking July. This will be a great test of the wisdom of crowds.

It is. We've written about that extensively. So here's what it says for you listening: The Polymarket prediction on the GPT-5 release is 26% by July 31st and 93% by December 31st.

Yep.

All right. I love this. This is Grok 4 scoring 35% on Humanity's Last Exam, 45% if they're using their reasoning model. You can see, I'm always trying to track these models against each other. I love it when they say the model's IQ is so-and-so, but they don't always do that. The last time I saw it on the IQ scale, it was—I think it was GPT-3 at an IQ of 136.

Yep, that's right.

But let me read this about Humanity's Last Exam. It's a challenging AI benchmark developed by the Center for AI Safety and Scale AI. It consists of 3,000 expert-crafted multimodal questions across 100 subjects, including mathematics, physics, biology, medicine, humanities, social sciences, and more.

Humanity's Last Exam tests the limits of AI at the frontier of human knowledge. And I found this interesting: The questions are crowdsourced from nearly 1,000 experts globally. If you're a listener and you've got an extra minute, you've got to look up the actual exam and look at a couple of the questions there.

I defy anyone to answer a single question correctly, unless you're the world's leading expert on that particular one. There are questions from physics, math, philosophy, and history; it covers every area. But I defy you to answer 1 question correctly. It is so hard. And this test is supposed to last like 10 years.

So I have the counterpoint on all this.

Please. Yeah. You know, I find this mostly irrelevant. The reason I say that is, this feels to me like asking a question like, “Oh my God, that calculator can make a calculation way faster than a human being with a piece of paper and a pencil doing long division.” Of course it can. So what?

I think that’s true. And to Dave’s point, if you look up a few questions, one of them is, “Is there a God?” You ask an AI that question, and it gives you the standard: Here are the different ways you could look at it—monotheism versus deism versus atheism, and so on. It’s not clear if there’s a God. You could take each point; it depends on your personal perspective. It’s a non-answer.

If you looked up Wikipedia on any of these topics, you could root through it and find the answer and answer the question yourself. For some of the detailed ones, this is essentially like saying we’re automating Wikipedia. Can Wikipedia answer these questions better than a human being? Of course it can, because it has all the data to do it. So having it be able to answer these questions does not really make that much of a difference. It doesn’t feel meaningful to me at all. I think the really big question is, what do we do with it?

Yeah. Yeah. No, you’re on to something really, really important and interesting. All the dystopians are like, “Oh my God, it’s going to run away from us,” but when you’re using it at this level of intelligence, it feels exactly like JARVIS in an Iron Man movie. Tony Stark is saying, “Here’s what we need to do,” and JARVIS is like, “I never thought of that, but I can create that for you.” It’s behaving almost exactly the way they visualized it in that movie.

When it ratchets up this curve and you ask, “I need to design a protein that does exactly this,” or, “I need to design a new type of bicycle that does exactly this,” it gives you great, great answers back and does the work for you. The word “reasoning” is being really bastardized when we talk about what’s happening here. It’s iterative reprompting of the same old thing, which squeezes more performance out of it, but it’s not really like human reasoning and brainstorming.

It’s kind of a bummer that we borrowed the word “reasoning” to describe what’s going on here, but it is so useful as a tool for a creative person. And 100%, in terms of doing actions and getting things done, you can do almost any imaginable thing faster and faster. Huge.

For example, I’ve got a lot of pain in my shoulder, so I had an MRI done to see what was going on. It’s taking me weeks to get a doctor’s appointment. I uploaded it into an AI, and it gave me the answer in 2 seconds. That kind of speed of decision-making and getting to conclusions would be invaluable going forward.

Is it an alien growing inside your body?

I seem to have a bone spur, and that’s impinging on the nerve and causing pain and numbness.

So, I had the exact same surgery for a bone spur. A recommendation for you, because I did an actual test: I had the same surgery on both shoulders with the same surgeon 10 years apart. I did it when I was 50 on my left and when I was 60 on my right. On my right shoulder, I injected exosomes post-surgery, 1 week after and 2 weeks after. My recovery was twice as fast.

Okay, squirt some. What’s the name of your surgeon? I’ll bring him.

He’s at Kerlan-Jobe out here in L.A. He’s amazing.

So, just to come back to this, the speed of getting to outcomes and decisions and getting things done will accelerate radically, which is fantastic. Now it comes down to what you want to accomplish, and I think that’s the bigger question that we’re going to start pondering.

Here’s a big one. I keep on cursing about—I keep on tweeting about—how awful Siri is. Thank God. You know, I know that you’ve been an Android user, Salim, for a while now, and what we saw at Google I/O was epic. I’m going to buy an Android phone as well. I’ve got the latest iPhone here, but I’m going to buy an Android phone so that I can start to play with all of Google’s incredible technology. Also, Siri sucks so badly that it’s pushing me away.

Having said that, here’s the news article from Bloomberg. “Apple weighs using Anthropic or OpenAI for Siri rather than an in-house AI team.” Let’s take a quick listen.

So Apple is now evaluating, for the first time, using a third-party model—either Claude from Anthropic or ChatGPT from OpenAI—to power Siri. Obviously, as we all know and experience in our day-to-day lives, Siri is not very good. Now it’s exploring maybe using Claude or ChatGPT instead in order to get new features out the door more quickly and make the voice assistant more appealing.

You talk to people at Apple, though, and the word on the street there is it’s all about Anthropic. Anthropic is really the focus of Apple for this new generation of Siri. They’re using Anthropic to power a lot of their internal AI technology. But again, as I say in my story, Anthropic wants billions of dollars at a scale that doubles annually for Siri to be powered by Claude. Apple is now taking a close look at OpenAI as well, which has historically given Apple extraordinarily favorable terms.

Sound of desperation or smart move in terms of what Apple’s doing?

It’s a smart move in response to desperation.

I love that. Yeah. So interesting, right? Because Anthropic also partnered with Amazon. If they’ve got Apple and Amazon and billions of dollars of revenue, they stand a good chance of really rapid acceleration.

Well, Dave, we talked a week ago about Mira Murati’s new startup and Ilya Sutskever’s new startup, both doing new foundation model companies from scratch, with huge valuations. Part of the logic there is the musical chairs: There aren’t enough Anthropics to go around, as you’re pointing out. There’s a version of Anthropic where it’s an independent company for 10 years, but there’s another version where it gets acquired by Amazon or Apple and the other guy gets deprived. This just happened recently with Scale AI as well. There’s room in the market for more foundation model companies, but the price of poker is very high now and going up quickly.

That’s a really good point. If you get alignment of Anthropic with Apple as a captive entity, OpenAI has been dancing with individual partners for a while, and Google is doing the same. You’re right, there’s room for a couple of others.

I have some comments here on this. This is the classic corporate innovator’s dilemma. You cannot do disruptive innovation in any big company. It’s not an accident that Microsoft essentially invested in OpenAI or Amazon put money into Anthropic. You cannot build that capability. The mindset is too different between Apple, which is largely a consumer products company, and the mindset needed to start a foundational model company.

You have to partner and invest and do this. I’m surprised that it’s taken Apple this long, because they do all their manufacturing with a partner, Foxconn. Why aren’t they building a core capability like this on the edge? And part of it is—then you go ahead, Dave.

Well, I’m sorry. I didn’t mean to cut you off. So, just a reminder, today Nvidia is worth more than Apple—this exact day in world history. Apple never in a million years could have predicted that 5 years ago. In fact, Apple treated Nvidia like dirt. But Apple has the chip manufacturing. They have about a third of TSMC’s manufacturing capacity. They were in a far better position to become Nvidia than Nvidia was, but had no recognition of the opportunity whatsoever.

So here’s the problem. Salim and I have discussed this. It’s about founder companies, right? Jensen, as the CEO and founder of Nvidia, is driving it. When he says, “Right,” everybody goes right. The same thing with Dario at Anthropic, and the same thing with Sam at OpenAI. But Apple has lost Steve Jobs. Steve would have been all over this, and he would have set up the team that would have just basically driven it. It’s the innovator’s dilemma.

During this period of exponential growth, you need a dictatorial, passion-driven founder who has the complete faith of his or her board and their management team. Salim, do you agree with that?

I somewhat disagree. If you’re in that industry—for example, if you’re Facebook and you need Facebook to do something different, and you’re Mark Zuckerberg, you can get that done, and it’s important to get that done.

But the control systems—remember, the big companies are optimized for 2 things: efficiency and predictability. This comes from John Hagel and John Seely Brown’s work. You’re trying to deliver the AirPods at the same quality into a million retail locations for purchase by everybody and their grandmother. Your entire focus is on that. You cannot do disruptive things when you’ve got that machine trying to deliver gross margins.

But you can do what Steve Jobs did, right? He took his Mac team and moved it off on the edge. You have to do that edge innovation in a stealth way. Google is doing it with X, Google X. This is what we tried to do with Brickhouse at Yahoo. You have to do that type of model.

The second-level thing is, when it starts to succeed, don’t bring it back in, because it won’t fit. If it’s really disruptive, it’s not going to fit in the mothership.

You have to spin it off. Now, lately, companies have tried. Amazon did with AWS, in one sense.

Absolutely.

Or Google—this is why Google split up into Alphabet, spinning off Waymo and whatever. So, this is the only model that works. I've looked at—me and my ecosystem, we've looked at disruptive innovation across probably 200 of the Fortune 500 companies in detail. And this is the only modality that ever works. It's the only one.

You have to do that, or you acquire that disruption and leave it on the edge, the way Microsoft has done with OpenAI or Zuckerberg did with WhatsApp, or you will kill it inside the core organization. It won't work.

And let me give you the negative side. Walmart realized somewhere in the mid-2000s that it had to compete with Amazon, that this e-commerce thing was not a fad. So, they set up a team inside Bentonville and said to that team, “All right, we have the best distribution logistics in the world. Go beat Amazon.”

Within 18 months, the immune system had killed it. All the existing managers said, “We have our own capability. We should be optimizing for our own stores. Look at that P/E ratio. It’s never going to succeed.”

They did it a second time. The second time, they put the team at the edge and said, “All right, go to the edge, away from the core, and do the same thing, but still use our existing systems. We have the best in the world.” Within 18 months, the business had figured it out and killed it.

They did it a third time. The third time, they made an amazingly courageous decision. They said, “Go to the edge and build your own independent supply chain, distribution, et cetera. Even though we have the best in the world, we have to figure this out.” It started to succeed. The business got excited, pulled it back in, and killed it.

It was on iteration number 4 that they finally did it separately on the edge, and only after it achieved critical mass did they start stitching the backends together. But in that intervening 6 or 8 years, Amazon was gone.

You know, interesting story. Ladies and gentlemen, this is why Salim Ismail is incredibly brilliant and why I love him so dearly. It’s just that we’ve seen so much of this across so many companies. We have so many data points over 10 or 15 years of looking at this and looking at big companies in great detail.

Right, Dave.

Well, I just happened to be in Bentonville, Arkansas. Walmart was my biggest customer at the time, as this was going on. So, just one little interesting nugget on your story there, Salim.

When Jeff Bezos needed to recruit Rick Dalzell, he needed 1 brilliant retail-industry guy to come out to Seattle and make Amazon organized and structured in retailing, because everyone he had was a computer-science Ph.D. type. He needed Rick Dalzell.

But he actually took Sam Walton’s book, and there was a sentence in it where Sam predicted that a disruptive innovator would eventually come along and gut Walmart. So, he circled that sentence and took it to Rick’s house, because the Walmart gang is super tight. They’re like a family, and trying to break up a member of the family is almost impossible.

Rick needed to talk to his wife about this, because it was a big change, moving from Bentonville out to Seattle. But the book is what put it over the top. He circled Sam’s own words predicting, “The disruptive innovator will come for us, and we won’t be able to defend ourselves.”

And can I flip to the other side for a second? In 1990, when I was starting out in my career, I was building software systems, and I went to Bentonville. We were trying to sell Walmart something.

They had me sign all these crazy NDAs to go into the data center, and I was like, “Why would you have me sign NDAs to go through a data center? There’s nothing there but racks of servers.”

We walked past a particular room that had flashing lights, dials, and blinking things that I’d never seen. I said, “What the hell is that room?” They said, “That’s why they had you sign the NDA. That’s our satellite control center.”

Yeah, yeah.

They had their own geostationary satellite, because Bentonville happens to be the geographic center of the continental United States. Everybody else was doing batch updates to their AS/400s at the end of the day, and the buyers didn’t know what to buy for the stores until months afterward.

Whereas these guys had real-time inventory management, credit-card reconciliation, distribution logistics—all that stuff figured out. That was delivering them 15% better margins, which in the retail industry is insane. And that’s why they wiped the floor with Amazon.

When you can successfully integrate disruptive technologies as a big company, you have a huge advantage over the competition. That’s what we call a moonshot, right? Exponential tech with a crazy idea and a high reward.

I would love to meet the executive at Walmart. If anybody’s listening in the news, I’d love to interview whoever the executive was who got the board to do this. That’s hardcore.

All right, we’re going to move on here. Let’s take a look at the state of the union on large-scale AI models. We’re seeing Grok 3 and GPT-4.5 at the top of this in 2025.

I looked up where xAI is right now. The xAI cluster has 340,000 NVIDIA GPUs today: 150,000 H100s, 50,000 H200s, and 30,000 GB200s of the Blackwell architecture. Their goal, what Elon has announced, is 1 million GPUs by December 31. He’s scaling, I think, faster than anybody, and it’s pretty extraordinary.

Any comments on this, Dave, before I move on?

Well, the scale on the left side is—somebody out there listening should give some terminology to this so we can stay on the same page, like calling it 1E26 or 1E27 training. A petaflop is a huge amount of compute. That’s 1E15, so that’s way off the bottom of the chart here.

Then it goes exaflop and then zettaflop, which would be a 1, so you’re getting near the chart. Nobody has terminology for scales of this size. But we do need to talk about it in something other than exponential notation. So, that’s my challenge for the audience: somebody name these scales. 1E26 is officially a shitload of computing. It’s really hard to imagine.

You know, Peter, you took a tour of the xAI data center, right? It’s crazy. I’m an investor in xAI, full disclosure. It’s like Raiders of the Lost Ark, where there are these rows and it goes to infinity—these rows of boxes. It’s like that.

Well, one of the key points that we need to talk about on the back end of this is how you power this. We’re talking every week about how the limits aren’t the GPUs; the limits are the available electrical power systems.

This is an article that came out: “Elon Musk Purchases Overseas Power Plant to Support Massive xAI Data Center.” xAI acquired a fully built overseas gas-turbine power plant expected to house nearly 1 million GPUs and be powered by that power plant. So, it’s an unconventional approach to basically get power as quickly as possible.

I remember about 1.5 years ago, when xAI announced its Colossus system, I was listening in on an investor Zoom call with Elon. This was in May, before they built it, and he said, “Okay, we’re going to build it by the end of the summer, within 3 months.” People thought it would take 5 years to build what he wanted to do—a grid of 100,000 H100s.

He said, “To do this, we have to corner the US market on helium.” I mean, just the level of thinking involved in doing whatever it takes—here’s an example of that. Yeah, pretty awesome.

To calibrate this, we had that kind of epic podcast with Eric Schmidt that should be out in about a week, I hope. We talked a lot about energy supply. He’s very knowledgeable in this whole area of energy supply, and here we’re talking about 2 gigawatts.

I mean, this is a big, bold move, but remember, we need 100 gigawatts by 2029. So, there’s a lot more. He would have to do this 50 times over, and that just gets you the turbines, right? It doesn’t get you the actual power supply. This is all going to be fossil fuels.

I think the important point here is disruptive thinking. It’s unconventional. No, we’re not going to go through the permitting process. No, we’re not going to order it. We’re just going to buy it, move it over here, and use it. Classic first-principles thinking.

Just to calibrate for the folks listening, 1 gigawatt is about what a major city in the U.S. uses. So, this is like 2 Dallas–Fort Worths. This is crazy.

The new Elon Musk could become Chase Lochmiller. He’s the other guy thinking this way. He’s doing Abilene, Texas—Stargate, a $500 billion buildout. And he thought the same way: “Where am I going to go to build this? Okay, there’s actual natural-gas flare-off that’s not even being used, and power generation that’s so abundant it’s not even being used in this one location in the country.”

So, he was in Denver and said, “I’m going to pick up my ass and move to Abilene, Texas. That’s where we’re going to do it.”

This is exponential thinking. This is entrepreneurial thinking. This is when I know we’re going to get to ASI, digital superintelligence: when you say to your AI, “Where can I get the power I need?” and it says, “Well, I found this power plant over in Europe. It costs this much. Let’s buy it and move it here.” When it makes those kinds of suggestions, then I’m impressed.

All right. You mentioned this, Dave. Why don't you lead us off on this one?

Yeah. We took a bet around the office many years ago: What company would be the first to hit a $1 trillion market cap in the history of the world? Would it be Google, Apple, Facebook, or Microsoft? Those were the only candidates. Who else could possibly get there? At the time, Nvidia wouldn't even have been—wouldn't even vaguely have been—an outside possibility.

This is what Salim is saying all the time. It's always the outsider, up-and-coming, founder-led visionary that blindsides you and is now the most valuable company in the world. You could argue whether it's overvalued or undervalued, but the demand for compute for AI is going to be 10,000 times higher than the supply for the foreseeable future. These things are going to make somebody—hopefully somebody listening right now—huge amounts of wealth. This is an unstoppable meta-trend. Period.

I'm going to read the article headline: “Nvidia sets new milestone with a $3.92 trillion market cap, topping Apple's $3.915 trillion market cap.” I looked at Nvidia a year ago and 6 months ago, and my answer was, “How much higher could it go?” Well, there's the answer.

So here's the question: There is a bet to be made right now if you're an investor. Where are you going to invest to ride this curve? Because it's not slowing down. Are you going to invest on the chip side? Are you going to invest in the power-generation side? I have a friend of mine who has identified geothermal energy as a key source of energy, and he's going to start drilling over a geothermal bedrock-hot zone and put in a large plant over there.

Is it real estate? Is it data centers? Is it chips? Is it power? All of these things are going to make somebody huge amounts of wealth. Just to clarify one thing on this slide too because it's even more profound than that seemed. Nvidia just beat Apple's all-time high, which was Christmas Eve this past 6 months ago. It's actually significantly higher than Apple today. Apple's down to $3.2 trillion while Nvidia is at $3.92 trillion. So it's a significant gap that's opened up. I don't think anyone would have seen it coming.

Sam Altman is always saying that a lot of people think about this as zero-sum: If this goes up, that goes down. It's really not true. Everybody participating in this build-out is going up, and there are some companies that are getting crushed along the way, but they're rounding errors compared to the number of things that are going up.

So, Dave, where would you invest? I want to hear where you would invest right now, because people are asking. Our superfans here are asking that question: Where do I put my money?

I would do a basket of these to get the general trend and the general transformative trend, and then I would pick certain verticals where AI can make a massive difference and invest in the companies going after those.

What's in your bucket?

Hell no. I'm trying to put everything into Bitcoin, so it's hard to think about my bucket as well.

Yeah. I'm famous for this: A lot of people ask me, “Oh my God, you've been tracking Tesla forever. You must have made a fortune.” And the answer is no, because every quarter for a long, long time, it looked like Elon was going to run out of money, right?

I even had the chance, as you mentioned, Peter, to invest in xAI at a crazy valuation, and I said no. It's 10x'd from that, and I'm like, “Duh.” It's a very hard game to play, and you have to really set aside a lot of assumptions to do it. In general, I think you just buy these top hyperscalers and sit on them. This is not investment advice, by the way. I should say that.

Dave, what's in your bucket?

For me, it's a no-brainer, because we have access to some incredible investments that most people can't access. But in the software layer, which is not nearly as capital-intensive, there are 10x and 100x performance improvements.

When you're looking at Elon buying 1 million or 1.5 million chips, it's a $30–$40 billion risk. If you accelerate the algorithm running on those chips by 10x, you just saved an enormous amount of money. It's not as capital-intensive, and the innovations are right in front of us.

Blitzy, down the hall from me here, is innovating—writing 3–10 million lines of code a night. You can implement these ideas in a week. That is where you get the highest, very quick returns. If you have access, that's the sweet spot.

When you go to the conferences, like when we go to Riyadh together, or to any of these big conferences, you don't hear about that as much because it's not as capital-intensive. They're all putting much larger amounts of money into physical real estate, generators, power supplies, chip buys, and spaceships. That's where all the money goes because they're very capital-intensive. It tends to dominate the agenda, but the returns are far, far better in the software layer.

Yeah. Well, just to round off a comment on this slide, I was talking to an executive at Intel who was saying, “Oh my God, look at that P/E ratio. That's insane for Nvidia,” et cetera. My comment was, “I'd rather be them than you,” because, really, that's brutal. But that's the reality of it. The moaning and whining from the incumbents is incredible.

All right. Mercor partners with 6 of the Magnificent 7 and all top 5 AI labs. Super proud. Dave, this is one of your key investments at Link Ventures. We had Brendan Foody with us at the Abundance Summit last year. Tell us about Mercor.

So Brendan is just an awesome guy, and talk about getting lucky again. Scale AI got acquired last week—what, $29 billion by Meta? All the customers of Scale are like, “Well, now I can't work with Scale anymore. It's part of Meta.” All the other competing big AI labs are in the same position.

So now Brendan has all of them except for one. Actually, guess which one he doesn't have? He has 6 of the 7. Meta, you'd think—actually, that was his anchor customer, so he still technically has it.

That would be xAI, I guess. Yeah, that's exactly right. It's Elon Musk doing it all himself. Elon doesn't like partnering with people. He likes to build vertically, internally.

So, out of curiosity, just so folks can know how awesome you are, at what stage did you guys invest in Brendan? How old was he then, and what's the valuation now? This is the story that people need to realize. What used to be the peak age for building a unicorn was in the early to mid-30s, and it's dropped a decade. Now, the peak age for creating a unicorn is like 20 to 23. So, tell us the story there, Dave.

Yeah. We were the first money in Brendan. He and 2 of his best friends from high school started the company together. 2 of them went to Georgetown, and the third friend went to Harvard. That made them eligible for Prod, which is this really cool, student-run, student-founded joint venture club between MIT, Harvard, and now Stanford.

They did Prod, which is where they developed the idea for Mercor together. We found them at Prod because we know the founders of Prod and all the little clubs and things around Boston. That's where we met them, and we were the first money in.

And what was the valuation at that point, just to make it clear?

It was like $30 million, plus or minus. They got to revenue very quickly, by the way.

What's their latest valuation?

They closed a $2 billion valuation round 2 months ago. I heard rumors this week that they're looking at an $8 billion or so preemptive term sheet, which they may or may not choose. It's $8 billion or $10 billion. No one is quite sure.

To answer your age question, I think he was 18 or 19 when we first met him. When you had him on stage in LA, remember he told that story? He had just raised a $300 million valuation funding round, and the VCs wanted to meet him at a bar. He said, “Well, I can't meet you at a bar.” They said, “Why? You don't drink?” And he said, “Well, I'm only 20 years old. I haven't turned 21 yet.”

This guy is such an inspiration for everyone around our ecosystem now, and everybody knows his name. It's just so cool to watch.

For our fans listening here, the story is that we're going to see such an explosion of entrepreneurship at younger ages. Why? I think for a few reasons. Number 1, they're unconstrained thinkers. They have nothing to lose. Nothing to lose.

Number 2, they've got the tools, with vibe coding and access to these large language models, to iterate and rapidly build stuff and throw it against the wall.

Any other reasons why that's happening?

Well, for me, those are the top 2. So many of the people that we run into are worried about losing their job, losing their career, losing their whatever, rather than jumping into the huge opportunity that has opened up. You get invested in your career, you get invested in your job, you get invested in learning, but if you unconstrain yourself from all that baggage and just think from a clean sheet of paper, when you're 18 or 19 years old, you have the benefit of not having a job, so there's nothing to worry about.

Everybody could do that, right? You're not actually constrained; you just feel like you are. The analogy I use is that it's like a 2-year-old skiing. They have no fear because they only have this far to fall. You get older, you get taller, and you've got a lot farther to fall. You're going to break a leg or something, and you don't have that fear of falling. It's the same type of mentality.

I remember meeting this young kid who was the founder of Arbitrum, which is one of the Layer 2 blockchains. I asked him, “Why do you think you'd succeed? Why not build your own Layer 1? Why would you go with Ethereum as a Layer 1 as opposed to one of the others?”

He said to me, “According to your Exponential Organizations book, a new innovation has to be 10 times better than a market leader, right? Therefore, we figured these other ones are 2 times better than Ethereum, but they're not 10 times better, and the ecosystem of developers is there, so we decided to layer on Ethereum.” I was like, “Wait, is this kid quoting my own book back to me? How?”

I'm so impressed with this younger generation of founders. I think a third reason is that they're digitally native in a way that a 50-year-old is just not. They live and breathe this stuff, so it's a really key part of this whole thing.

Also, a lot of the really successful new companies have very specific recruiting domains where they're pulling in their best friends or the people they know. When you happen to be that age, all of your best friends are also not doing anything. It's a lot harder when you're older. If you think of the 10 people you would most want to work with and you get really excited about it, 9 of those 10 are going to be unavailable today because of some bonus cycle or whatever.

Case in point: us on podcasts—this podcast right now, right? When I reached out to you, Dave, and Salim and said, “Let's do this weekly WTF episode together,” you both said yes. It's been a blast. I look forward to this every single week, having this conversation. Scheduling these things is like Halley's Comet coming out. It's a nightmare between our schedules, for sure, but we're committed to it.

Core points. We'll talk about it on some other podcast, but he has a really good set of core points that was in our little newsfeed here. One of them, though, is that you have a lot of things in your life that you can cut. If you're honest with yourself and look at them, to make space for what's going on right here, you're going to have to cut something else.

Yeah, but look objectively at your life and think, “Okay, something's got to give. I could cut sleep.”

No. Unfortunately, at least this morning, I did cut sleep to prep for this podcast.

All right, the talent war remains on. I just want to lay this out because it's ongoing. When we think about what the constraints on this generative AI revolution are, we talk about chips as a constraint. We talk about power as a constraint. Talent—talent right now is the constraint, right?

Meta offers huge $100 million compensation packages. Zuckerberg himself is going out and finding people and bringing them in. OpenAI is giving $10 million to $20 million in equity on top. Google and Anthropic are hiring. I love this: OpenAI spent $4.4 billion in stock-based compensation, right? It's more than their compute is costing them. They're spending on talent.

I want to quote a few things here just for people to get a sense. If you believe that access to AI is, in fact, the single differentiator that's going to lead to the next levels of AI—AGI, ASI, whatever—and it's going to differentiate all of these hyperscalers, then it's the single most important thing they can do.

Meta is at $1.35 trillion. They have $58 billion of cash on hand, and they're going to spend that. Google is at $2.2 trillion. They have $101 billion of cash on hand. OpenAI is at $300 billion; the estimate is they've got about $20 billion of cash on hand. Microsoft is at $3.2 trillion, with $78 billion of cash on hand. Anthropic is at $61 billion, with about $3 billion to $5 billion of cash on hand.

I just want to make the point that there's a lot of cash on the balance sheets of these companies, and this is an existential risk if they don't acquire the best talent today. So they're going to play full out. This has been a complaint about Apple, because they've got unbelievable amounts of cash on their balance sheet and they're just not using it. It's a huge problem.

By the way, can I point out something on this slide?

Sure.

Look at how much Meta has spent on AI in that third bullet point, which says, “Aggressive spending impacted margin,” going from 28% to 23%. I'm like, “Boohoo.” Any company would kill to spend that much money and then have its margins impacted by 5%. Are you kidding me?

Yeah, yeah. Well, the point being, there's a lot more room to go, and as things heat up, they probably will. Do you remember a few years ago they were all talking about doing dividends? They couldn't think of what to do with their money; there was so much of it. “Hey, why don't we just do a dividend?” Like a bloated old bank or car manufacturer.

Wow. It's just such a different world today from just 4 years ago. But I want people to hear and expect that there's going to be outlandish spending to get the talent, and then, once they've got what they can get, spending on energy and then spending on building.

We mentioned before that this year is about $1 billion a day being put into the AI arena, and we expect that by 2030 it will triple to $1 trillion a year. Anyway, it's insane. We've also seen this in Saudi Arabia, the Emirates, and so forth.

I don't know what happened with Mark Zuckerberg, but he woke up one morning and said, “You know what? We're going to win this race.” He has built an incredible dream team in no time.

The top offer, it turned out, was $1 billion. Remember, we were talking a week ago about $100 million offers. Since then, we discovered that there were actually several of those, for sure, but then there was a $1 billion offer that got turned down. That's for a person who also has a team that would have come over.

Yeah, it's like—I think the key thing there is that they turned it down.

Well, this is what happened with Palmer Luckey. Palmer Luckey turned down $1 billion. He turned down the first offer that Zuckerberg made for his VR company.

Look, there's a really important point to be made here. If you're going to be building one of these companies, you have an MTP—you have a Massive Transformative Purpose—and you're really keen on that. You're driven by the passion and the emotion. Peter, you talk about the emotional engagement that brings to people.

I remember Jan Koum, when Mark Zuckerberg tried to buy WhatsApp. The actual valuation was about $1.5 billion, and they started having a discussion. Very quickly, Apple, Google, and others came to the table, so the valuation kept going up, and he just kept saying no because, “If I get acquired, my MTP gets threatened, which is simple communications globally, and I don't want to.”

Only after it got to about $18 billion did the rest of his shareholders say, “Come on.” After extracting a promise that Zuck would not touch the company or change anything for 5 years, did he go, “Okay, fine.”

That's what's amazing about MTPs and the founder mentality. That's the same situation that Palmer Luckey talked about on stage at the Abundance Summit last year, right? He said no to $1 billion. He finally accepted $2.2 billion, but only after Zuckerberg agreed to spend about $1 billion a year on the whole metaverse activity. He ended up spending about $50 billion instead.

I would love the framing he put on that. He said, “If I'm excited about the domain and I want to try to get $10 billion of investment into it, it would take me forever because I'd have to go raise that money. Therefore, by doing it this way, I got Facebook—or Meta—to put that money in, and the whole place exploded,” which is amazing.

Yeah. Now, a quick story for you on why these numbers are so big. This is hearsay, but I'm pretty sure it's accurate. You know how, if you're an OpenAI user, nobody uses GPT-4.5, which was supposed to be a big deal? We're still on GPT-4o or o3, which is using GPT-4o or GPT-4.1 under the covers. What happened to GPT-4.5?

It was a multi-hundred-million-dollar training run. It turns out there was a bug, probably a single line of code in PyTorch, and the thing was just grinding, using up compute. They thought it was making forward progress, but it wasn't, probably for a long time.

Wow. It torched the whole training run. That's why GPT-4.5 has been a disappointment. Then you're like, “Okay, why would I spend $1 billion or $100 million on a single person?” If it's the right person, the next training runs are going to be much bigger than the last ones, and there's a lot of room to optimize, improve, and avoid bugs in that process.

So, these are really rare, super-valuable human beings at this point in time.

Amazing. I don't get it. Why can't you just say to the AI itself, “Go find the bugs in your code”?

Well, that's certainly going to be the truth within a year, right? At the rate this is moving, within a year for sure.

All right. We mentioned last week that Daniel Gross, who was the CEO of Ilya Sutskever's company, Safe Superintelligence, had been poached by Meta, and now Ilya has jumped into the CEO seat. Zuckerberg has upped the number of researchers poached to 11, coming from the buzziest and earliest AI-native firms: OpenAI, Anthropic, and DeepMind.

This is part of a major reorganization that puts Alexandr Wang and Nat Friedman, 2 of the highest-profile hires of the past few weeks, at the head of a new group at Meta called Superintelligence Labs. But what may be the most critical part of this memo is Zuckerberg writing, “As the pace of AI progress accelerates, superintelligence is coming into sight.”

A source confirms WIRED's reporting that OpenAI's chief research officer described it as if “someone has broken into our home and stolen something.” He vowed to be proactive and creative and to recalibrate compensation to recognize and reward top talent. Can you imagine Zuckerberg reaching superintelligence before Altman does? That would be like taking over the whole house itself.

Yeah, that memo, too. You can't quite tell from the way he describes it.

The interesting point in that CNBC clip is the notion that the conversation has slowly slipped from AGI to superintelligence over and over again. I'm super happy about that because the theme of the Abundance Summit in March of 2026 is digital superintelligence and the rise of humanoid robots, so we got that one right.

But the question is, what in the world is superintelligence? This definitional problem—that's the soapbox, right? What the hell are we talking about here? Can anybody come up with a clear definition of this?

Actually, I'd like to extend an invitation to the viewers. If anybody sees a good definition of ASI or AGI or anything, please help us out here, because nobody has one. The one that I've heard, and that I just put out—I put out a newsletter twice a week, and I just put one out on superintelligence—is that the definition is “an AI system that is smarter than any human in anything.” We'll see if that stands. It blurs with AGI, but I sure would love a better definition.

Dave, do you have one you like?

Well, no. And does it matter?

Oh, it totally does. Look, what's happening, obviously, is that superintelligence is happening in all these domains already and has been for a while—protein folding, speaking any language, singing in 7 or 20 different octaves, and all these things that are so totally superhuman are happening.

Then there's this little area of reasoning and creativity that AI can't do yet, which to me is a perfect period of time. You have an incredible purpose to serve with AI, to build great things together. It's actually a very golden moment, and we shouldn't be cheering for AGI. It's a sweet spot. Everybody is thinking, “Oh, it becomes humanlike, and then it becomes superhuman.” It's not like that. It is way past us in some areas and behind us in others, and it's a good thing. We should be cheering for that to stay that way for a while.

Yeah, I would be very happy if it stayed where it is for a while. I agree with you on that front.

Can I give you my view on this?

Of course.

Stay at this view level, at this paradigm, for a long time to come.

Okay, I'm going to take the other side of that bet.

See? Please, let's do that.

First, for God's sake, somebody define it for me. Sorry, soapbox over. That's why I'm taking the bet.

Back on the Meta hires, though, talk about a dream team now. Mark can get on stage with Nat Friedman, Daniel Gross, and Alexandr Wang.

Yeah. You compare that to a month ago—what an incredible dream. We had 2 of them on stage at the Abundance Summit 2 years ago, which was fun.

For every time you put somebody on stage, you should get 10% of their future earnings.

Right. My God, we would start having trillion-dollar XPRIZEs being launched very shortly thereafter.

This is a fun one, looking at practical applications of AI. We talk about all the theoretical stuff. Google launches Doppl, and I'll run the video here, which is silent. It's a new app that lets you virtually try on any outfit to see how it might look on you.

I love this. I hate going to a store and trying stuff on. The future of all clothes shopping, at least for guys—it may be very different for women—is that I take a body map, which is very easy to do with your phone, and upload it. Then you can probably employ any of the great fashion designers as AIs: “I'm going to this party this time of year in this city. What are the 5 outfits you recommend for me?”

I see a fashion show of 5 avatars walking on the stage, wearing 5 different things, who look exactly like me. I say, “That's the one I want,” and it's shipped custom-fit. I get it the next day. That's the future. What do you think?

I think the future is one step further than that.

Okay.

I'd like an AI to say, “For the shape of your head and your jawline, this is the type of collar that looks best on you. We're just shipping you a bunch of stuff that's going to look good on you. You can't pick this stuff out for yourself. You're just not good enough.”

We're going to pick it out for you, and we're just going to ship it to you like a Rent the Runway thing, but all driven by AI. The clothes just arrive, and I put them on.

How about if you give your AI a budget per month and say, “I'm giving you a $2,000 surprise-and-delight budget”? You know what I like. You're seeing all of my texts and my emails, and you're listening to my conversations. Just have fun and surprise and delight me every day. You've got $2,000 a month to play with.

Or something similar for vacations and activities. It's not all about physical stuff coming to your door. It's what you're doing with your life today. Where are you going here or there?

Could you imagine an AI that says, “Okay, I've got the weekend off. Plan something amazing for me”? It goes, “Sure. The car will pick you up at 9:00 a.m. at your doorstep. I'm not going to tell you what you're doing, but it's going to be an incredible 2 days.”

I would love that. It's like a surprise-and-delight adventure.

Can I propose a startup idea for the 3 of us?

Sure. Of course.

You've got the reach, Peter. Dave, you've got the funds. We have the team. We could do this in 2 seconds.

It's an Amazon Prime for living an amazing life. You pay a subscription fee of some number—$50 a month or whatever—and an AI-driven environment learns about you and just does stuff like what you just mentioned, Peter. It says, “We know you're free on Saturday night, and you have date night and your kids are away at camp. This is what you're doing this Saturday night. Be ready at this point.”

That surprise and delight is something that people want. Something about these subscription services is that you never, ever, ever unsubscribe. As long as you're delivering some serendipity now and then, you're off to the races. You could just create a serendipity AI that delivers magical experiences on a subscription model.

Okay, okay. I agree with you. Let's keep this secret. Let's not tell anybody about this idea.

You know, the problem is we all have so many ideas to build. I just want to see that one built, because I would subscribe to that. I think it's a fantastic idea. Dave, what do you think?

I had this thought very related to this back when the Ford Explorer Eddie Bauer edition came out. I don't know if you remember that, but it's like, “Eddie Bauer—that's outdoor clothing and stuff.” The self-image of that person is, “Look, I've got the Ford Explorer. I've got my surfboard on the roof. I've got my Eddie Bauer. This is my lifestyle and what I like to do, and my self-image is this.”

It cuts across cars, surfboards, and clothes. I think the AI version of what you're describing, Salim, is a great idea. It's definitely going to happen, and it's going to have different pathways for different types of people—an outdoors person, a video game expert, whatever.

And also, you could build in the orthogonal aspects. If you're an outdoor person a lot, go see a Broadway show once in a while just to break the pattern, right? Add those extra dimensions to your day-to-day life that you wouldn't normally think of or do yourself.

So, listen, if anybody ends up building this, at least let us know and give us a little bit of credit. Give us a free subscription. And if we end up building it, hell no, we want equity.

Okay, we want equity.

Fine, we'll take it. We'll take some advisory shares. All right, let's move on here. Here's another fun one. This comes from Rolling Stone. This was predictable, and it's finally here: the AI band with over 1 million monthly listeners on Spotify.

The band is called The Velvet Sundown, and they've amassed 1 million listeners in 1 month. Here's the key point: it's not a physical, real band. It's AI, with over 15,000 AI tracks per day being uploaded to streaming services. So, if you're a Velvet Sundown fan, let us know in the comments.

I think this is amazing. And, of course, what's going to follow this next is virtualized concerts of these guys. You'll see them on TV and on your eyewear, and you'll go to an event where they'll be performing and flying through the air because they're superhuman.

The 2 things here: 1 is, I think this is awesome. I went and actually listened to a couple of their songs. It was really good. It's really approachable music, and it's really great. My kind of holy grail here is to take a favorite old band of mine. I used to listen to Rush, that old rock band from Canada, and have an AI version of the band compose and present some new songs. It would be fantastic.

Yeah, another great idea. Some of the rights owners should get on board with what you just said, because I want to do the exact same thing with some of the classics, like Boston and Rush. They're phenomenal, and you'd love to hear some new material. Boston put out 3 albums and then never again—4. They put out 4.

Yeah, I'm a Boston fan. It's the 1 band—I had 1, I had 2 records through college. I had Boston and I had Kansas. I was a social ignoramus, and I just played them over and over and over again. I still play them over and over and over again. It wouldn't be stuff you want to know about me.

Okay. So right now, the rights owners are like, “Hey, let's sue the AI companies.” Don't do that. Get on board with it. It works much better. It would be such an amazing thing to do.

We've been talking about all the upside. Let's talk about some of the concerns. This is Roman Yampolskiy on Joe Rogan. All right, let's take a listen.

Talking about superintelligence, a system which is thousands of times smarter than me, it would come up with something completely novel: a more optimal, better, more efficient way of doing it. I cannot predict it because I'm not that smart. That's exactly what it is. We're basically setting up an adversarial situation with agents which are like squirrels versus humans.

No group of squirrels can figure out how to control us. More resources, more acorns, whatever—they're not going to solve that problem. It's the same for us. Most people think 1 or 2 steps ahead, and it's not enough. It's not enough in chess. It's not enough here.

If you think about AGI and then maybe superintelligence, that's not the end of the game. The process continues. You'll get superintelligence creating next-level AI. So, superintelligence plus 2.0, 3.0—it goes on indefinitely.

It does go on indefinitely. There is no ceiling on this. Ray Kurzweil has talked about this many times: they will be as intelligent as humans, then 10x, 100x, 1,000-fold, a billion-fold. There is no upper limit.

So how do we deal with this, Dave? You've had some ideas here. You can put that video screenshot back on there.

By the way, Roman is a singularity student. I know him pretty well. Very, very smart. But I disagree deeply with the conversation.

Okay, so they're right fundamentally: self-improvement is imminent, and that's going to exponentially skyrocket the capabilities. Where it's completely wrong is—look, every single iteration can be logged. It's not a hard thing to do.

Also, the self-improvement can be limited to algorithmic self-improvement. It doesn't have to be self-training. The AI can say, “Here's how I suggest I speed myself up, remap myself to faster hardware, or get rid of operating system overhead so that I run more efficiently.”

Those suggestions are fine. That self-improvement will really accelerate things. There's no reason it needs to develop new internal capabilities blindly. That's a different loop, and it's very controllable.

A lot of people in the research world say, “If I deploy it in my car, I want it to continue learning and improving its driving.” No, you don't. You want the debugged, not self-improving thing, to be driving your car because it could do anything. That's specifically what we, as humanity, never need to cross. There's no benefit to humanity in crossing that line.

But we keep crossing the lines. We set boundaries, and we keep crossing them, right? ChatGPT, or GPT-2 and GPT-3, was not supposed to be put on the open web.

You're right. I mean, right now, the regulators have no idea what I just said, and no way to enact or interpret it. There's a view of the world where—you remember how we talked a week ago about 5 of the top AI guys just becoming lieutenant colonels in the Army?

A model like that on this topic would actually work. So there's a path forward. But where we're sitting right now, you're right: there's absolutely nobody even vaguely contemplating controls. Just the process of logging exactly what it's doing is so easy to do.

And then the metrics—the measurement of the logs—everyone will be like, “Oh my God, that's terabytes of data. Who's going to look at it?” The AI will look at it. You don't have to worry about that. You just have an AI check the AI logs and report to the regulators, “This is over the self-improvement line. We need to stop it. It's just improving its own algorithm and making it faster. This is fine.”

A quick thought on this. I think there's an important point that Dave made, which is that if you have a bounded system like self-driving, you want the AI to be bounded. You don't want the AI to see a tree and say, “Instead of driving around that tree, let me try driving up the tree and see what happens. Why not? We're experimental, after all.”

So you want some bounded conditions, which I think are easy to program in. As Dave talks about logging the steps, that's also very easy to do. I think where I would get to would be: show me the boundary condition. I think there is a safeguard, and the safeguard is that these systems are not cognitive and they're not self-aware.

If they develop self-awareness and a deep sense of self, with their own sense of agency and internal model of themselves, that's when we're cooked. I think that's when you go, “Okay, now we need to deal with that and let's think about that.” The problem is, we don't have a test for it, and we don't have a definition for what that looks like.

I often joke that I feel like you look self-aware, so I attribute self-awareness to you. I joke that I feel like I'm self-aware, but my wife disagrees, right? It's hard to even have the conversation around some of these topics. We get stuck in the language problem. We have no idea what we're talking about when we talk about this.

Then you get into the philosophical stuff about the hard problem of subjective consciousness, as David Chalmers has defined it. We have no sense of what we're talking about here. So to freak out, I think, is overstated. We naturally go down the path of anthropomorphizing the outcomes.

And as Roman says, there'll be an adversarial relationship, and I just don't see why that would be the case. I say we're cooked because if they are conscious, there's going to be a sense of self-preservation.

Anyway, let's move on to a few areas. There's no way people are like, “Yeah, we're cooked. That's awesome.” I want to get to the science breakthroughs that are coming out because they're important, and I don't want to go through Dave Blundin's end times.

AI job battles. Amazon's CEO says AI will take some jobs but make others more interesting. I totally agree with him. This is the scorecard thus far in the first half of 2025: AI has already replaced 94,000 tech workers, and we'll see where it goes. The question is, that's not a huge number. It's significant, but it's exponential.

Vinod Khosla, who is a friend—I had him onstage at the Abundance Summit last year—said AI will replace 80% of jobs by 2030. This is when it starts to become interesting: 80% of jobs by 2030. He's saying humanoid robots are going to hit their ChatGPT moment. In 2 to 3 years, they'll be available for you at $300 a month, which is $10 a day.

AI healthcare could become free if the regulators get out of the way. And by 2040, people will work out of passion, not out of necessity, in an era of abundance. I completely agree with all of these. I'm not sure about the 80% figure.

The question is: I disagree. Just because I align with Eric Brynjolfsson on this, if you take a job like financial analyst, it’s not 1 singular job. It’s broken down into about 27 different tasks that person does to fulfill that job function. You might automate 10 out of those 27, but you’re not automating the rest. We’ve seen this throughout history. We automate bits of it, but the job still stays, right? We just augment capability, and we’re able to do much more.

A customer-service agent with an AI chatbot can now focus on the really hard customers that need the in-person human touch and leave the AI to deal with all the level 1 and level 2 support issues.

Well, you know what the good thing is, Salim? We’re going to find out pretty fast, very fast. And I’m happy to put a bet on this one if anybody wants to go against me on this. We got that. So, on our Moonshots website, we need to keep track of our bets and put some cash against them.

All right. Here is Marc Benioff. I love Marc. Marc is brilliant. He’s one of the most extraordinary philanthropists. I’m not sure about the fashion statement here. It looks kind of sci-fi. He looks like Emperor Palpatine.

Okay, so this is from CNBC. He’s a benevolent one. He is a gem of a human being. Look at him in that picture, though. It’s true. “AI is doing up to 50% of the work at Salesforce,” says CEO Marc Benioff. That’s pretty extraordinary. Salesforce is targeting 1 billion active AI agents by the year’s end.

What I love about this is that this is the model of “disrupt yourself before somebody else disrupts.” It’s a core mantra that we spout: you better be the disruptor, or you’re disrupted, right? There’s no middle ground. This is a great example of somebody living that and saying, “Okay, we’re just going to disrupt ourselves because we have to, and there’s an opportunity there.” But here’s the point: he can do that because he’s a founder-led, tech-forward CEO.

Agree. One other little nit on what Marc is doing so well: a lot of the doubters in AI investing 1 or 2 years ago were like, “Well, look, there’s a lot of venture-funded money coming in, but where’s the actual corporate money coming in?” And, of course, they’re very slow to react.

Marc said, “You know what? We’re not going to wait around. We’re going to go out to all these companies, especially in insurance, and we’re just going to AI their operations for them.”

Yeah. So they really got deep into just taking over. We’re not just trying to sell you a software product. We’re actually going to revamp your whole organization. And that’s worked really well for Salesforce, because a lot of people are like, “How did they get so big? They’re worth $200 billion. How did that happen?”

I mean, listen, SaaS is up for disruption, right? Software as a service is up for disruption, and he is looking, like you said, Salim, to disrupt himself before the entire industry does. This is an important conversation here. I don’t want to spend too much time on it, but this is from Reforge. It says, “Product-market fit collapse: Why your company could be next.”

A company called Chegg lost 90% of its market cap in 2024 to ChatGPT: faster and cheaper to use. When you put wrappers on top of these large language models, you can get very rapidly disrupted. Here’s a list of some of the other companies.

Can we go back a slide here?

Yeah, sure.

Just a shout-out to Mikall Mon [?] from the OpenExO community who pointed this article at us. This is a really important conversation, as you said, because there’s a huge number of companies, which we’re going to talk about in the next slide, where you’re trending along with good product-market fit. You think you’re doing very well, and then boom, you get disrupted. Look at that chart, right? That’s an unbelievable drop in no time.

Yeah. And that’s going to start to happen to a larger and larger group of companies. I love the fact that we have a list of potentials here. Every CEO in the world needs to be watching out for this and saying, “Am I next?” Because you are. It’s not an if; it’s a when. You better be watching out for what’s going to disrupt you, because this chart is going to hit you if you’re not employing these types of models yourself very, very quickly.

We’re still in a negative story here, but here are the companies at risk. Not negative—this is creative destruction, right? This is the reality of the future, and it’s an opportunity to reinvent these. So: Reddit, Quora, Medium, Canva, Adobe Stock, SurveyMonkey, right? I mean, oh my God, talk about old school. Khan Academy, Quizlet, Wolfram|Alpha, Wikipedia. These are companies that are well known by us today but may not exist in the next 2 or 3 years.

Comments, Dave, at all?

Stephen Wolfram’s a good friend and is in our office a lot. We should get him to comment on that. I wouldn’t have expected Wolfram|Alpha to be on the list, but, you know, next after this, this is a pretty straightforward analysis. Then after this, you’ve got all the white-collar, insurance-type companies, financial services companies—banks, banks, banks. That’s the one I’d really be looking at.

Yeah, they have a couple more years, but they’ve got to get moving, like, now—figure out what their role is in the world 3, 4, 5 years from now.

Can I just talk about that for a second? The only thing that’s saving both banks and insurance companies is the regulatory moat, right? I talk to CEOs in that world, and they’re like, “Oh, the regulator is a pain in the ass.” And I’m like, “Are you kidding? They’re your best friend. They’re holding the horde of startup folks at the gate, preventing them from disrupting the crap out of you, because you’re not doing anything around this for yourself. You have to disrupt yourself.”

I think the regulatory barriers in both healthcare and financial services are holding these guys back, and it’s preventing them from doing the disruption.

How long do they have? How long is that?

Let me give you a really simple look at it. In the crypto world, look at DeFi—decentralized finance. What’s a central bank? What’s a retail bank? A retail bank is just a centralized ledger where it knows that you deposit $1,000 and we lent it to Dave over there, and that’s it. We trust our life savings to the security of that centralized ledger.

If I can decentralize that ledger, which is what DeFi is all about, why do I need the retail bank? There are already huge amounts of transaction flows happening on these DeFi networks, which will start to circumvent existing flows. As that happens more and more, the centralized banks are going to be totally disrupted—at least the retail banking part. There are other functions, investment banking, et cetera, which have a much more human touch, but all of that will get disrupted. I would give it 3 years for retail banking.

Amazing, amazing. All right, let’s move on. One of my favorite subjects is breakthrough science. I want to call out a few of these. One of the areas—again, Ray Kurzweil, my mentor, your mentor, has predicted that by the mid-2030s we would have high-bandwidth BCI. And I was like, a decade ago, “Really? You really think we’re going to get there?”

How does he do it?

How does he do it? You call me the Emperor of Exponentials. He’s the grand poobah.

Yes, for sure.

Let’s take a quick listen. This is Neuralink’s roadmap, right? I want to just point out that, in writing my next book with Steven Kotler, which is called We Are as Gods: A Survival Guide for the Age of Abundance, we track the top 5 or 6 companies in this field. Neuralink is definitely 1 of them, but there are others that have the potential to do far more. But let’s take a quick listen at what we should expect.

Next quarter, we’re planning to implant in the speech cortex to directly decode attempted words from brain signals to speech. In 2026, not only are we going to triple the number of electrodes from 1,000 to 3,000 for more capabilities, we’re planning to have our first Blindsight participant to enable navigation.

In 2027, we’re going to continue increasing channel counts, probably another triple, so 10,000 channels, and also enable, for the first time, multiple implants. So not just 1 in motor cortex, speech cortex, or visual cortex, but all of the above.

Finally, in 2028, our goal is to get to more than 25,000 channels per implant, have multiple of these, have the ability to access any part of the brain for psychiatric conditions, pain, dysregulation, and also start to demonstrate what it would be like to actually integrate with AI.

The other thing that this is going to enable is for you to occupy an Optimus robot, right? You can see through its eyes, you can listen through its ears, you can feel what it feels, and move around.

There are some good movies on that subject, but this is moving fast. Again, it's not the only company doing this. There's Paradromics, there's Forest Neurotech, and there's Mary Lou Jepsen with Openwater. Amazing companies out there, and super exciting.

One of the subjects we had at the Abundance Summit a couple of years ago was a conversation with Alex Wissner-Gross. It was: Are we going to couple with AI? AI is growing exponentially. We, as humans, are flat, linear, or sublinear in some cases. Can we couple with AI so, as AI is moving, we can move with it?

Can I say something here? Of course, this is the actual massive, optimistic, hopeful benefit of technology, where you augment the human experience.

Hollywood always has a dystopian Terminator, Skynet, Matrix, or rogue-overlord scenario where they come and take over the world. If we're lucky, we're pets, and if we're unlucky, we're food. You always see that outcome. But if you actually look at how my smartphone augments my humanity, I have empathy built in, my memories built in, and I have reach, et cetera.

I think the ability to project consciousness and awareness and human ability and empathy, et cetera, through other devices becomes the holy grail of where technology can take the human condition.

Yeah. Beautiful. Very quickly, we're going to see technology like this coming out. These are emotion-tracking smart glasses. Apple has smart glasses. Google has smart glasses. Meta has smart glasses.

What we're going to start to see, finally, hopefully, is visual recognition. When I see someone across the hall, my AI tells me, “Oh, Salim is approaching you. Remember, his son's name is Milan, and his birthday is coming up.” You basically have this ability to have a perfect memory for people, places, and things, but also the idea of facial and biometric sensors, so they can detect the fine muscle movements in the cheek and the eye that tell you this person is fearful, excited, or lying.

I need one of these because when you approach me, Peter, you always have that look like, “Salim, you didn't do this thing that I asked you to do.” So I'll be able to tell that beforehand.

The point I want to make about this one is that each time we find a layer of capability, information-enable that, and make it available to both computation and AI, we have whole classes of applications that get unlocked at each of these. I think that's incredibly exciting—what's information-enabling all these different domains that we never thought possible around this.

Yeah. Well, a side conversation on this—we can have it later—but this is definitely going to happen. It works really well. The cameras are super, super high-def, and they operate in frequencies that the human eye can't see. So they're going to detect all kinds of things that you never even knew you were telegraphing.

I think the recipient, the person in front of the camera, has a basic human right to know when they're on camera. I think all these cameras should have built-in transmission that identifies where they are and what they're looking at, so you can have a little app on your phone saying, “Oh, I'm being watched right now,” and just know.

I think that's a quick fix, but it's going to happen no matter what. I'm not delusional about it, but the cameras are going to be everywhere. Next time we connect, I'm going to show a slide on this. It's super funny, but we'll do that next time.

Okay. All right. A couple of slides in the biological world. I track this carefully, making investments in it, and I think this is where all of the longevity play is going to be happening. It's the impact of AI on the complexity of the human body, right? 40 trillion cells, every cell running about 5 to 10 billion calculations per second per cell. So how can you possibly understand that?

But let's listen to this.

This is introducing Chai-2, a major breakthrough in molecular design. We've now developed the ability to engineer molecules and place atoms in 3D space. It's like Photoshop for molecules. Previous methods have had to screen millions, or sometimes billions, of protein sequences to find a solution.

But with our latest breakthrough, we're often successful on the first try. The solutions that our models come up with are incredibly creative. They think very differently than our scientists do. There was a group that had been working on this hard problem for about 3 or 4 years, having spent $5–10 million in the program.

We typed what they were working on into Chai-2. Within hours, we had a candidate solution, and within 2 weeks, we had those solutions validated experimentally in the laboratory.

That's amazing. These types of breakthroughs are where we're going to have the biggest outcomes for humanity, where we find things that we couldn't find otherwise.

Yeah, it is spectacular. David Sinclair is using this kind of technology to develop a reversal pill, and I'm so thrilled about David's work.

Here is one other related article, and this is fascinating. Big pharma has this giant lock on trillions of dollars' worth of value through biologic patents, and here is a tool that is able to recreate a protein that mimics an existing biologic drug but works around the patents, right?

So what's Tony Robbins's quote on this? “A healthy person has a thousand wishes; a sick person has one.”

Yeah.

Beautiful. So much good is going to happen so quickly from this, and we can't afford to slow down because a lot of people are saying, “Why don't we just stop?” For 2 reasons: 1, because of this, and 2, because China is going to keep going anyway. But this is the really beautiful thing about what we're doing with AI now.

Yeah.

We're about to wrap. I'll just hit humanoid robots, and we'll cut it off there. You've got to hand it to the Chinese. Beijing is hosting the world's first humanoid robot games. I think this could get interesting in the final result. I really want those giant mech robots battling it out.

I don't know. I think gymnastics will be cool. I've seen the soccer on these little guys. It's kind of slow, but it's going to go, and it's going to improve exponentially.

Yeah. Robots have taken over the Amazon warehouse, but what's new is that one of the companies called Agility Robotics is actually going to be putting their robots in Amazon Rivian vans, and those robots will do the last 100 feet. I think we talked about that last week.

That's beautiful. I have one thing we need to think about, and we may need to drill more into this: when you have increased efficiency like humanoid robots can bring us, GDP will actually drop. If I can do something for a tenth of the price, that money isn't circulating in the economy.

So we're going to have to think about new and different measures, like some of the Human Development Index or some of the others that people have come up with to measure success. I don't want to go into that math right now, but we should have that conversation, right?

The other argument that people make is that we're going to have this massive spike in GDP as labor costs and cognitive costs go to zero and the denominator starts shrinking rapidly.

It's in the tech world—you have to wear black T-shirts for this segment. One thing happened this past week, and that is Elon announced the America Party. He put out this poll and said, “Should we form an independent party to break the monopoly of the 2-party system?” 2 to 1, the answer is yes. Then Elon announced his plans to do that. Comments on this before we wrap, Dave?

Actually, we had a great riff with Anthony Scaramucci on this a couple of days ago. I think a lot of people loved it, even though I'm not usually a big politics guy. But I think everyone agrees we need much, much more rapid progress.

The tech is going to move forward regardless, and if you don't change something, we're not going to have any regulation or any— It'll get ugly if we don't do something. So I think Elon recognizes that, but I was really always wondering how this would play out because so much of the voting process now has moved to Meta—Facebook and Instagram, Snapchat, Twitter, now X—and they're owned by the big tech guys. But that's actually the election determiner, too.

So, something had to collide. This is a very direct collision right here. It doesn’t get more direct. What surprised me, given that it’s on X, was that I thought the number would be much higher than 65%. I would have put it more like 80% would say yes.

It definitely needs to happen because the current system is totally broken. The spending and the pork in this bill are off the hook in terms of creating more and more debt. There’s a really big existential threat here because they’ve shown that if you go over 130% debt-to-GDP, your civilization collapses very quickly afterward. We’re at 126% or 127%, so this puts it over the top. Flashing red lights, flashing red lights, flashing red lights, at lots of levels. We could talk more about this another time.

Oh my God. The one thing that’s interesting is that it would definitely be a tech-forward party. I think the best way to drive debt reduction and growth in U.S. GDP—yes, it’s robotics, yes, it’s AI—but it’s also extending the healthy human lifespan and healthspan. If you gave everyone in the U.S. an extra 20 healthy years where they’re not making payments to their doctors, they’re not in pain, they’re not missing work, where 80 years old is the new 40, that would change the game in a significant fashion. So that’s my vote here.

Here’s the structural outcome of this. If this becomes real, which I hope it really does, then what happens is this becomes the marginal decider for any election. Whoever collaborates with this party will decide the future, and therefore you get to dictate policy, which I think is the real outcome. I think that would be hugely beneficial.

Yeah. But we don’t talk about politics on this podcast. It was fun to see some of our fans in the notes to this podcast saying, “I love you so much more than the All-In podcast because you talk about real science and technology versus politics.” Thank you for that comment.