埃隆加入芯片竞赛、标普500重估，以及人类驾驶将成为非法行为｜EP #242

Without question, for me, the number one story this week was Elon’s announcement of the Terafab.

This is the most important endeavor in human history by far. In order to understand the universe, you must explore the universe.

He’s basically building a galactic factory. On the left is 20 gigawatts, which is the current global output. And just the audacity of Elon’s vision—it has tremendous geopolitical implications, as we discussed on the last pod. This could either accelerate or, more hopefully, mitigate World War III and a Chinese invasion of Taiwan. We’re going to need all the compute we can create. In fact, I’m actually kind of worried that a self-driving car uses up basically a full GPU.

When is it going to become illegal for humans to drive? I think the thing that would make it later is purely the shortage of chips. The technology will be there and the demand will be there long before the chips are there.

Figure out how to do more compute with less silicon for this exact use case, and you’ll be an instant billionaire.

I mean, we’re heading toward a $100 trillion company. Maybe the largest, most important company on and off the planet.

Can we get there? Now, that’s a moon shot.

Now, that's a moon shot, ladies and gentlemen. Everybody, welcome to Moonshots, another episode of WTF. Here with my incredible moonshot mates. DB2 in Boston. Uh AWG, looks like you're on your home base as well. I am, but without my saucer separated Enterprise 1701D behind me. Oh, yes. I've got I I contracted one of my boys to create finally finally LEGO's come out with a Star Trek uh you know, LEGO set. I'm tired of all the you know, Star Wars LEGO sets. So, yes, 1701D. Uh and it does do a saucer separation, but I'm not going to try it right now because a probably disaster may follow. And of course, we have Dr. EXO Saleem at his normal location at JFK. Saleem, how you doing, pal? [laughter] Salim is gone. Oh, okay. Well, so much for that. Yeah, it's as you know, listen, we try to be mobile. We're all dedicated to this podcast. But uh let's continue on because the singularity is not going to wait. So today, we’re working to get you future-ready. A little bit of a format change: we’re going to be having some deeper conversations about a more limited number of subjects, still covering the news that’s breaking right now—and there is a lot.

Our mission is to get you excited about the abundance that’s coming, show you the opportunities that are coming to you whether you’re an entrepreneur, an investor, a student, or a parent, and really, you know, this is the time to be paying attention to the supersonic tsunami—the most important technology in the world. Hopefully, this is your number-one podcast on AI and exponential tech as well.

Hopefully, there’s more in here than ever before. It’s bundled into themes that we can discuss, but if you just look at the raw news-story count, it’s, as you would expect, exponentially exploding. Our goal, all of us, is to make sure that as we’re talking about this on Moonshots, it’s meaningful to the listeners, gets you excited, gives you context, and helps you think about this in a different way.

Let’s jump in. Without question, for me, the number one story this week was Elon’s announcement of the Terafab. He’s basically building a galactic factory. Think of this as putting all the parts of his LEGO puzzle together in an extraordinary fashion that is going to create massive capabilities.

Let me hit these quick points, and then we’ll jump in and discuss it. The Terafab is an objective across Tesla, xAI, and SpaceX to build 1 terawatt of AI compute per year. To put this in context, the global output today is 20 gigawatts of AI compute. Again, we’re measuring AI computation in terms of power, not just chips anymore.

Elon wants to build 50 times the current production rate of the planet. He’s building two kinds of chips: an edge-inference chip for robots and cars, but also a high-power, rad-hard chip for his space Dyson sphere that is coming online.

The fab is in Austin, and it looks eventually like 100 million square feet of capacity. One terawatt in the near term—single-digit years. Long term, a petawatt gets you only there from lunar mass drivers.

Salim Ismail has joined the story. Hey, Salim, good to see you, pal.

Hey, folks. Sorry, I’m bouncing around a bit, but I’m here.

All right. In which terminal at JFK are you today? Where are you going?

I’m flying to Brazil for 40 hours.

Of course you are. Of course you are. You’re a probability function on planet Earth.

So, just to put this in context, check out this chart. On the left is 20 gigawatts, the current global output, and just the audacity of Elon’s vision: 1,000 gigawatts, or 1 terawatt, is his objective.

One thing I heard him say is, “Listen, I’ve been going to all the chip manufacturers out there and saying, ‘I will pay you for as much production rate as you will give me. I don’t want to compete with you, but give me more and more and more.’” Of course, none of them are moving at Elon speed. So he said, “Screw it. I’m going to go and build my own production facility.”

That’s not exactly what he did in the launch industry, right? He just lapped the entire existing launch industry and the autonomous-car and electric-car industries. He’s playing his playbook over and over again.

Before I get into this data and stats, comments, Dave?

Well, this is the most important endeavor in human history by far because it unlocks everything else. No great surprise that he’s announced it at the scale that humanity needs it.

But the specifics on how you’re actually going to physically do this are unknown, because there are fundamental constraints on the number of ASML machines, EUV machines, and many, many other constraints.

This is the most complicated product ever made by humanity, and the supply chain makes cars look like child’s play. So he announced the mission. It’s the right mission. The scale is crazy. We estimated this on the last podcast at 50 times all current production of chips, and I guess our estimate was dead-on. So we got that part right.

He did allude to it last summer when we were meeting with him. What I was most curious about is how he was going to announce this and attract all the talent that he needs without irritating Samsung. He signed a $16 billion deal for production with Samsung, which is more like $45 billion if it’s going according to plan.

I guess one of the cover stories here is, “Well, these chips are for cars, and they’re also for space. They’re hardened for space, so they’re not like the other chips.”

He said, “I will buy everything Samsung can offer me, but you’re not offering me enough. So I will still build all these chips, and I will still buy everything you want to give me.”

One thing I love—and he pointed this out in his Austin fab—is that it’s full vertical integration under one roof, so he can run rapid iterations on chip design. That’s impressive.

In our Austin podcast, when we were talking to Elon, I asked him point-blank, “TSMC is being way too conservative in terms of their production of chips. They should be 10 times their fab-manufacturing capacity.” He said, “Well, you know, the industry is cyclic. Maybe they’re being conservative intelligently.”

Which is hilarious in hindsight if you go back and listen to that audio, because in the back of his mind, he’s thinking, “Well, I’m going to build something 50 times bigger anyway.”

It is crazy that Samsung, Intel, and TSMC are not racing to build 10 or 20 times more production. So Elon, of course, is going to do it instead.

Okay, can I show you guys some calculations that I found extraordinary here? Listening to the presentation he gave 48 hours ago, his target is 1 terawatt of compute per year in orbit. He said, “Mass to orbit: 10 million tons per year.”

We’re talking about an average satellite—his next-generation Starlink—at 1 ton. Long story short, in order for him to launch that much capacity, it’s 274 launches per day on Starship. That’s a launch every 5.3 minutes, which, of course, he says, “Listen, in the airline business, that’s normal.”

But just the audacity and the level of thinking that Elon takes on is amazing. Alex, you want to jump in?

So many thoughts on this. Well, first of all, I think the elephant in the room is whether Elon can indeed ramp up capacity for the Terafab in, call it, the next 5 years, which is the timescale being tossed around.

If you look at the lunar aspirations—not a terawatt, but a petawatt from the moon—if you do the back-of-the-envelope arithmetic for what a petawatt of GPU compute that comes from lunar mining would take, you run the arithmetic and it comes out to be approximately 3/100,000 of the lunar mass.

So a petawatt coming from lunar mining, with electromagnetic launches from the moon, is starting to have a material impact on the mass of the moon.

One big crater dug out of the moon.

So that’s a petawatt. As we scale, of course, to an exawatt of compute—and because why not—then, at that point, we’re talking about something like 3% of the moon’s mass.

When people think I’m joking when I talk about disassembling the moon, or that the moon had it coming, this certainly paints a portrait.

The moon did indeed have it coming. The moon is slated for disassembly to build the Dyson swarm. This is what it looks like.

I think, more broadly, there are other interesting secondary implications. This is a joint Tesla, xAI, and SpaceX maneuver. Many folks have speculated over the years, “Wouldn’t it be wonderful if all of Elon’s industrial ecosystem came together into one singleton?” This is sufficiently cruxy, with 20% of its production slated for Tesla and 80% slated for SpaceX, that this starts to look a little bit like maybe a cornerstone for some grand unification of all of Elon’s projects.

Alex, we talked in a previous podcast about the idea that Elon said this: We’ll see the first $100 trillion company. When we look at the numbers here, I want to show another set of calculations I did on what the Terafab might be worth in the ecosystem.

We’re heading toward a $100 trillion company. Can we get there? At the end of the day, I don’t know how you guys feel, but the Musk-world ecosystem here looks like it will lap, by 1 or 2 orders of magnitude, what NVIDIA has done. It may be the largest, most important company on and off the planet.

Yeah, and I don’t think Elon wants to unify all of his projects just for the sake of having one unified company. I think he wants to unify the capital raising and the capital leverage with his massive, multitrillion-dollar IPO and the massive joint mission, unlocking an unprecedented amount of capital, which is what it’s going to take to do these fabs in parallel at this scale.

Because that’s the thing that’s holding back Samsung, Intel, and TSMC. They actually could do it. He needs $25 billion initially to turn on the Terafab and get it—the buildings started, so to speak.

Right. I saw that in the analysis, but $25 billion is just 1 fab. Here, we’re going to 50× the U.S.—or the world’s production. Fifty times the world’s production. So, he needs 50 of those $25 billion investments to achieve this mission.

Yeah, there’s a lot to do. I thought I had 2 or 3 thoughts. One is: talk about the patron saint of exponentials. This guy thinks at scales that very few people do, and it sounds incredible. Classically, the future of anything he’s looking at looks vertical, and the past looks flat and boring.

What I thought was great was this amazing exponential logic, because he’s going exponential at the bottlenecks. You stop competing—you’re redefining the game. You’re challenging anybody to dare to come with you. I think that’s the amazing part of this.

The launch cadence is unreal: every 5 minutes. I think that’s exactly right. It forces the operating model to completely change, and it’s forcing everybody to rethink that, including all the engineers and all the infrastructure.

If this is done in any kind of normal industry, one thing to point out is that his predictions on timing tend to be about 15% to 20% accurate. But it doesn’t matter if it takes him 3 times as long—who the hell cares? The fact that he’s thinking at that scale and he’ll get there—that’s the fact that he’s putting his plan out.

Yeah, he’s directionally correct. The idea is, you shoot for the stars, and if you get to the moon, who the hell cares? We’re landing somewhere amazing.

That’s Elon’s plan. So, here’s the next question: a rapid schedule to disassembly of the moon, I think, is what’s on the table.

Okay. Good thing I don’t have a glass of wine.

Everybody, you may not know this, but I've got an incredible research team. And every week, myself, my research team study the meta trends that are impacting the world. Topics like computation, sensors, networks, AI, robotics, 3D printing, synthetic biology. And these meta trend reports I put out once a week enable you to see the future 10 years ahead of anybody else. If you'd like to get access to the meta trends newsletter every week, go to diamandis.com/metatrends. That's diamandis.com/metatrends. So, a terawatt of compute per year, if physically achievable—is this aspirational? What time frame? He says 5 years. That’s still 50×, which is crazy, but if he’s able to achieve that, what happens to the terrestrial data centers and to the investments made in terrestrial data centers? Gentlemen, questions on that?

Every chip, every investment in power and data centers is going to pay off tremendously. They won’t cannibalize each other. We’re going to need all the compute we can create and so much more.

In fact, I’m actually kind of worried that the self-driving car is going to get cannibalized. Driving a self-driving car uses up basically a full GPU. By the end of this year, a full GPU can also do brain surgery, or it can discover new math or new physics. It’s not clear that driving somebody around is going to make the price cut as the demand for compute goes to near infinity.

So, I think the terrestrial data centers are going to be critical for national security for every country in the world, because if something goes wrong in space, you have to fall back. Your whole society will be running on these GPUs.

Yeah, you can’t have an outage. Anyone investing in this does not have to worry about one thing cannibalizing the other.

The other thing you’re going to see—I think it’s later in the deck—but all the different process nodes, all the fab process nodes, are going to get used, even the older ones. The 3-nanometer and 5-nanometer nodes are going to be running full throttle now, even if they’re not as good as the new 2- and 1.6-nanometer nodes. It doesn’t matter. We need all the compute we can crank out.

It’s going to be an all-hands-on-deck race, and Elon is just documenting the upper bound of what we can achieve as humanity.

My expectation would be that, in the process of—I think Elon likes to call it “production hell”—realizing the Terafab over the next 5 years, we’re going to discover some new, not semiconductor physics, but more materials physics and process engineering.

It seems improbable to me that Elon will just build the Terafab based on the existing stack, as, say, TSMC did, of ASML plus the existing optics plus all of the conventional semiconductor-processing techniques. If he really is looking to disrupt the space, he’s going to want much more disruptive unit economics.

Maybe some of these technologies for semiconductor production and fabrication that have been waiting in the wings for their right time in the limelight—maybe, and this is purely speculative, he’ll look, for example, at alternatives to photolithography.

It’s not like, as a civilization, we don’t have lots of alternatives.

Exactly. And he’s got superintelligence to help him get there, to design new systems.

You’re going to need the humanoid robots to be building the fabs. We don’t have the workers.

Salim, I think this is a really important point being made right now, Alex. Thank you for this. You think about the secondary technologies and the benefits, like all the carbon fiber that came from the space industries and cascaded down to everyday life. The secondary inventions that will be needed here will be massively beneficial to humanity.

Well, this is one of the most exciting things in tech history—in fact, the most exciting thing in tech history—is what Elon was talking about in Austin: laying down single atoms using some kind of self-organizing process.

I feel like Alex is exactly right. Something will get discovered in the next year or 2 using current LLM AI running on GPUs that will then dictate a very different, non-lithography future. But it’ll probably be 5, 6, or 7 years before those start getting manufactured at the same scale as lithography. It’s super exciting to watch.

Dave, I asked Claude to give me an estimate based on all the data that we got from Elon during his talk on the value of future Terafabs. It makes the point here that initial capex, as you said, Alex, is $20 billion to $25 billion, but the real capex for buildout is going to be on the order of $150 billion, at a minimum. It may be half a trillion dollars.

Then the model here looks at the annual cost savings for his captive opportunities. If he’s building the chips himself and putting them in Optimus and then Cybercabs, and then there’s external revenue, and then there’s an implied enterprise value, it’s on the order of $1 trillion to multiple trillions.

TSMC is valued at $1.7 trillion, and what we talked about last week was that the Terafab is expected to produce on the order of 70% of TSMC’s output. So, we’re layering on another multitrillion-dollar opportunity.

These numbers seem low to me.

Yeah, if you’re generating 50× the total output of AI chips on the planet.

This is operating at a different scale. It’s still a car company, and it’s not.

Yeah. Well, not only that. Right out of the gate, this is not just doing what TSMC does. This is TSMC plus NVIDIA. NVIDIA is worth $4.5 trillion.

Yeah, but even that’s ridiculous as an analysis. We’re talking about 50 times the production of the world’s current compute. Out of the gate, you would take TSMC plus NVIDIA and multiply by 50 to get a starting point for an estimate. So, this is off by over an order of magnitude—well over an order.

Any of you concerned about a monopoly here?

No.

If you're following the prediction markets for the SpaceX IPO, this is already starting to get priced into the SpaceX IPO. The SpaceX IPO was originally going to be $1.5 trillion. Now prediction markets favor a $2+ trillion SpaceX IPO, pricing in the SpaceX portion of the Terafab.

In some sense, again, I'm not quite clear on what the government structure here is going to look like or how clean it's going to be. But to the extent it falls mostly in the SpaceX bucket, the SpaceX IPO—not investment advice, obviously—could end up being, as you say, SpaceX plus Starlink plus NVIDIA plus ASML plus TSMC all rolled into one.

I just, as a piece of advice for entrepreneurs out there, want you to understand the level of audacity that Elon is looking at. He's building at multiple orders of magnitude beyond anybody else. From his first-principles thinking, he's looking at where the blockages for his growth are.

We had this conversation: People are not generating enough chips. I need to build a chip fab. And then he doesn't just say, "I'm going to buy Intel or build a chip fab equivalent to what TSMC is building in Arizona." No. If I'm going to build a chip fab, I'm going to build something that's 50 times bigger than the world's supply. Amazing.

He's also thinking two moves ahead in the big chess game. Two moves used to be 20 years; now it's 6 years. He's two moves ahead of everybody else. Alex and I were talking earlier this week about fusion energy, and Elon doesn't talk much about fusion energy. Why? Because he's visualizing solar in space.

Solar on Earth is a great stepping stone to solar in space, and it requires panels, batteries, and cooling, but it doesn't require turbines and fusion reactors. So he can skip a couple of hard steps and go straight after the next move in the big chess game. It's really interesting to watch how those timelines have inverted.

The Dyson sphere is now right on our radar, and even Google is talking about it, which means everyone's talking about it. That came to the forefront really just in the last month or 2, and so the whole timeline of humanity got shifted. The Dyson sphere will come before anyone even figures out how to get licensed for a fusion reactor.

Can I be cynical just for a second?

Yes.

The plans are ridiculously grandiose, and if anyone is going to achieve them, it's Elon. But it's curious that the timing of this is just leading into the IPO to get everybody excited about things. That would be the cynical view, but I still just love the audacity.

Yeah. Well, I think Salim, your first observation was dead right. If you shoot for Mars and you end up at the Moon, you're still way up.

I want to ask the question again that Alex said, "Nope," which is monopoly concerns. Do you believe—if he's really generating all this compute—that there's a concern?

I'm not worried about it because when you have an MTP, which he does, you're basically operating on this massive mission. You may have ethical issues here and there, but generally, the trend is so positive and so beneficial for humanity, who the hell cares?

There's always another person. Somebody who's like Brendan Foody, maybe. Somebody like that right now is thinking, "Wow, I'm going to do this, too." That's just the way it is. That person—we don't know who they are yet—but they'll emerge. You can't exist in the U.S. without antitrust action if you don't have a competitor. Elon will invite that competitor, whoever it is, and it'll be great.

I'll say one of the things here: If we label this monopolistic behavior, then don't we have to label everyone with an MTP a monopolist?

Mhm. Mhm. Just asking the question, my dear mates.

Market dominance—if you get there, then that's true. But you haven't even picked the real estate site yet for the Terafab. He hasn't even picked a final location.

I think it's way too premature to declare this monopolistic behavior. Being so ambitious as to build the Dyson swarm—we're going to have multiple Dyson swarms.

I'm with you. Love it. I'm with you.

All right. Well, look, Google isn't going away. Google has $300 billion in revenue, $100 billion-plus in cash flow, and its own chip designs. It has everything except the rockets.

I wouldn't forget Eric Schmidt. Eric Schmidt is trying to bring Relativity Space online, so the rockets are at least part of the Google family.

Oh, you know, they're going— But let me close this. There's plenty of them to go around.

Let's not forget one thing that's important here: Every time there is a constraint, the judo move is to realize it's a massive opportunity. This is an abundant story once again. This is a massive increase in the abundance of AI compute beyond what anyone was speaking about just a week ago. So I'm glad he's focused on this and less on the politics.

All right, here is our second conversation story, one that I'm excited to have with my moonshot mates. It's about the future of human transportation. Robots are getting their driver's licenses, and flying cars are taking flight. Here's some of the data, and I want to go deep on this because I want everyone listening to understand how this is going to impact how and where you live, how you commute, and every aspect of our lives.

Waymo has hit 170 million fully autonomous miles, equivalent to 200 human lifetimes of driving, with 92% fewer serious crashes. So, a significant reduction in crashes. Currently, they've got 3,000 vehicles in 10 cities. Still early, right?

Uber has now invested $1.25 billion in Rivian, with plans to deploy 50,000 fully autonomous robotaxis. Here's a look at Waymo versus human drivers. Waymo is doing an extraordinary job—not necessarily of saving lives, but of reducing crashes—by 92%. I think Cybercab—we've seen incredible data like this also on full self-driving from Tesla.

Check out this image. This is Joby Aviation. Joby had started this company. He started Velocity 11, took that money after he sold it. Uh he was partnered with Rob Nail, uh Saleem. And uh Joby started Joby Aviation God knows over a decade ago and here it is flying over the Golden Gate Bridge in San Francisco. It's a beautiful image.

This is an eVTOL: electric vertical takeoff and landing. It's a name that rolls off the tongue onto the floor. I'm calling them flying cars because that's what they are. Here's what's going on in the eVTOL world, and I think it's really important.

Joby is now testing its first FAA-conforming aircraft, meaning it's demonstrating to the FAA that it can build a reliable design over and over again. It just had demonstration flights over the Golden Gate Bridge. Joby and Uber announced Uber Air powered by Joby.

In fact, before Travis left, he had created something called Uber Elevate, and they were doing their earliest work on flying cars. I keynoted their talk there, but Uber Elevate got sold to Joby, and now Joby and Uber have a partnership.

The other company in the United States that's a competitor to Joby is called Archer Aviation. They have a beautiful aircraft called Midnight, and they're the first company to achieve 100% FAA acceptance of its eVTOL aircraft's means of compliance.

Long story short, we've been waiting a long time, and flying cars are almost here. We're going to start to see them operating in the U.S. in the next 18 months. They should be here in L.A. in 2028 in a big way.

So here are some of the conversations, gents. First off, when is it going to become illegal for humans to drive? Salim?

Yeah. I think pretty soon. You'll start with city centers, right? It'll be illegal to drive in city centers. Then it'll slowly broaden out from there.

I think the flying car is the most exciting technology I could ever ask for personally, given that I'm commuting to airports a lot. This is 10 years later than I wanted it, but finally, it's happening.

What I like about this is that these are not transportation stories. This is full urban redesign. Essentially, you make it abundant. Land has always been scarce. Real estate has been scarce. Real estate becomes abundant.

If you fly across the U.S., it's empty. We've talked about the statistic that between Toronto and Chicago airports, there are 10,000 islands and lakes. We do not have a scarcity problem. We have a mobility and accessibility problem.

I'm super excited by this particular model. I've got 2 years for us to get to full autonomy before my son gets his driver's license. He's 14 right now, so I'm pushing hard on this race.

Mostly, he wants it to get away from his parents, but that's fair.

That's fair. So we'll see what happens. But you compress travel time and reprice real estate. This is such a huge thing.

I have to shout out to JoeBen Bevirt, because it's hard to build a hardware platform like this and do it over a decade with all the inevitable regulatory and market-structure resistance against you, as well as infrastructure resistance. This is huge. It's like a Nobel Prize in patience.

Yeah. Incredible. Dave, you were going to say?

The eVTOLs are going to move very quickly because they don't run the risk of crashing into houses like self-driving cars on a road do. They're all going to be autonomous from birth. That's the new thing.

eVTOLs have been in the works for years, but the AI that makes them self-flying, self-driving, and super safe is here all of a sudden.

True, but the first airplanes are going to be piloted, right? They'll be single-pilot, with 4 passengers in the back. The goal is rapid recharge at the vertiports when they land. Probably an average length of flight of under 10 miles, I think, going from Santa Monica, where I am, to Dodger Stadium and avoiding the 10.

But autonomy will come with enough data and enough demonstrations. Wait, why won't they be fully autonomous from the get-go? Because it's called the FAA. The FAA is not happy until you're not happy.

Yeah, that's exactly it. The manufacturing of these wants to happen right away. The AI command and control is being worked on for the car, not for the eVTOL yet. So there'll be a very short period of a couple of years, in my opinion—2 years or so. In the Middle East, they're already doing the self-driving, self-flying version of this, so it should be a very short window where people get to fly these.

But the big question here is, when does it become illegal for humans to drive? I think that's going to happen very quickly as well, very similar to indoor smoking or drunk driving. There's a tipping point where a lot of voters say, “Wait, you're putting my children at risk with your crappy driving.”

That's ridiculous. We've got data and proof here that self-driving is 90% safer, soon to be 95%, 97% safer. And the human tragedy that comes from car crashes is unbelievable and shocking.

For under-5-year-old kids, it's the number one cause of death. Yeah. accidents. And it's devastating to families, too. It's absolutely tragic.

In the first world. Yeah. Well, there'll be a TV ad campaign, probably 3, 4, 5 years from now, with lots of ugly images in it. Then there'll be massive amounts of voting, and people will say, “It's inconceivable that you would drive on a public road. That's inhumane. Go drive on a test track, that's fine. Maybe some country roads, that's fine, but no way. Don't put my children at risk.”

So I think that's going to come as soon as we have the manufacturing for the cars themselves. But I think the thing that would make it later is purely the shortage of chips. The technology will be there and the demand will be there long before the chips are there. So if you want to unlock this as an engineer, figure out how to do more compute with less silicon for this exact use case, and you'll be an instant billionaire. Alex, your thoughts?

This format, Peter, is like an internal AMA, so I'm going to try for a lightning round on all of these.

Leave some room for the rest of us. Let's take it one at a time. One at a time. At what point does it become illegal for humans to drive? I think never. I think we'll simply redefine driving to represent higher and higher levels of abstraction.

So right now, with FSD 14, you tell it where you want to go, and if you're running the most recent subversion, you can have a conversation with Grok and do minute steers along the way. I think that notion will get refined such that driving gets redefined to be sufficiently abstract that it's always safe for pedestrians. It's always the human in the loop of the AI driver. So it's effectively a human-machine hybrid, if you will, that has the safety of the machine but makes the human feel like they're still in the driver's seat.

I said this when Dara was on stage with Salim and me. There's a version in the future of self-driving where you're driving and you can push the car as fast and as hard as you want, and the car knows its own limits. It knows the traction of its tires and the road surface, and it prevents you from doing something stupid, but you're in control of it 99% of the time.

In the 1% where you're about to do something that will destroy you, a person, or the car, it stops you. Exactly. I think the future of the accelerator pedal isn't the accelerator pedal. If you use FSD, it's turning the driving mode up to Mad Max. That's sort of an abstraction of acceleration.

That's all I use is Mad Max, and it still doesn't go fast enough, so I have to step on the pedal.

There used to be an ad saying, “Friends don't let friends drive drunk.” You can just keep that ad, drop off the drunk part, and go, “Friends don't let friends drive.” Period. All the messaging is there.

All right. So, Alex, why don't you kick us off on question 2 here? Okay. Question 2: With Uber partnering with Waymo and a bunch of other names, will the Cybercab be able to compete?

I think we mean compete here. Yes, of course. It's going to be a very competitive market. Period.

And I love the fact that this is driving us toward abundance, right? This is driving us toward UHI. If you've got a dozen companies delivering autonomous vehicle services in your city, they're going to be competing on quality of service and price, just bringing the price down to a minimum amount.

Now, one of the things that's interesting about the Cybercab is that it's going to be priced at probably $30K, roughly what Elon has announced. And he's going to allow people to buy it. One of my goals is, can I buy 25 or 50 of them here in Santa Monica and own them, but have them going out and basically generating revenue for me and for my Cybercabs? I'm sure they'll have some level of personhood by then, Alex.

I think so. I never would have guessed, Peter, that your next gig would be as a cabbie, but the singularity makes for strange bedfellows.

Fleet owner.

The big impact for me when I see this is the complete collapse in the market structure of cars. Today, we make close to 100 million new cars a year, and they sit empty 94% of the time. So even if you drop that by 50% in utility, you basically collapse the need for half the car industry instantly.

If these cars are maintained for a long time, the lifetime should be near infinite. My Tesla Model 7 2017—it should go a million miles. There's nothing wrong with that car. So this is going to completely change the nature of car services and car maintenance. The complete industry gets reshuffled from the bottom up.

Yeah, yeah. We'll think about the implications of that, too, Salim. Right now, if you take an Uber from SFO to San Fran for like $200 or whatever the hell it is, it's almost all driver costs. So even before you shrink the number of required cars by a factor—I think the estimate was 5x—is it 10x that savings? Is it 10x?

So then, the driver is already the majority. You take the driver out of the loop, so the cost of that ride should go down at least 10x, because the car's coming down 10x and the driver is more than the car anyway. I think the number I've seen is between 10 and 30 cents a mile.

Yeah, I've seen it as 4- to 5-fold cheaper than owning a car. The next question I want to ask, and offer my points of view, is—I think it's one of the most important ones for our listeners.

This is going to have a profound impact on your real estate holdings: where you live and what you do with your real estate. So if we have autonomous vehicles and we've reduced the number of vehicles on the road by 10x, let's call it that, and these vehicles don't need to park, again, my current version of this is I get up from the breakfast table with my family and walk toward the front door.

My AI knows that I'm moving to open the front door. It knows where I'm going. It's ordered an autonomous vehicle—what I call automatically—for me. So all of a sudden, you know, in our home here, we had a 3-car garage. We already converted one of those garages into an extra bedroom.

The other 2 garages have become effectively storage, and I'll build out probably a workout gym and so forth. I think the idea of a garage—a personal garage in your home—goes away. So start thinking about what you're going to do with your garage space. What are you going to make it into? Because you're not going to own a car.

You might want access to a car, but most of the time, do you really like driving? I mean, when you get into an Uber, do you ask the Uber driver to get out and let you drive? So, right.

And you know, this part of the conversation is incredibly actionable for all of our listeners. It doesn't rise to Alex's level of, you know, like, change the world tomorrow, but it really matters to almost everyone who listens.

If you're young and you've got a job and you're living in a city, which is 60% of you, you might not want to buy in the city. Keep renting and look for something that becomes your second home later in life, in a beautiful spot.

That's a little harder to get to, so it's going to be incredibly coveted. Imagine a world where there's 10x more wealth around 2034, 2036, and this is a spot that anyone in their right mind would want.

Actually, if my wife is listening, close that transaction that you kicked off this weekend, even if you have to pay a little more. But yeah, that's the life plan you want, because accessibility—not just getting to it, but also delivering things to it. Your Starbucks or Dunkin' Donuts is going to come by drone. Absolutely.

That changes what you want. Think about it. There aren't—we have a huge country, like Salim said—but the really great spots are limited.

So really do your soul-searching and look for that thing. Don’t buy near an airport in a city. Island real estate is going to become 10x, 100x more accessible, and that will drive the value up. In downtown LA, I don’t remember the figure—it’s like 30% of the blacktop is parking. All of that gets released to become new.

60%.

60% of the land area is parking spots in Los Angeles. That’s a crazy number.

That’s insane. Well, that becomes gardens. It becomes green land. It becomes parks. That’s incredible.

Think of the unbelievable space we use in stadium parking lots, right? Acres and acres and acres of rows and stuff. So we’ll have to rethink tailgating and everything. There’s so much available business opportunity here if you can think ahead about what you will do with that. And if you’re in the parking garage business, you’ve got to think ahead as well.

A couple of other second- and third-order implications, if I may. We’ve already touched on, I think, the more obvious ones: parking garages need to be reprogrammed for other purposes.

Another, I think borderline-cliché implication of full autonomy everywhere is the re-spread of suburbia. Why invest so much in urban-center real estate if you can be effectively connected to an urban center—or not even need an urban center—if autonomous vehicles take you everywhere? Basically, a virtual subway from anywhere to anywhere.

So, re-suburbanization, if you will, at least in relatively low-population-density countries like the US. I think these are pretty cliché implications.

A less-cliché implication, in my mind, is: What if we take this trend and extrapolate it fully to completion? What happens? I put out a request for startups around this idea: Why not just create autonomous Winnebagos, the equivalent of having entire office buildings that are themselves autonomous vehicles?

One could imagine living in an autonomous vehicle. It’s all part of a social network. When you need to take an in-person meeting with someone, your 2 AVs are part of the social network, and they connect and synchronize all of your locations. So maybe you’re in Boston in the morning, but you’re in Washington, DC, in the evening. This is all handled automatically to synchronize your calendar with your AV location.

Then it’s a sleeper car, and you’re in Chicago or wherever the next day. Your bedroom car—you become your bed. Humans become internet packets that are being routed by the autonomous-vehicle system.

Yes. Love it, love it, love it, love it.

You know, the eVTOLs—it’s taken a while. There’s still a lot of doubt people have about eVTOLs. The opportunity we have is going to be limited by the size of these and their ability to land locally. So there needs to be a local hub-and-spoke system of vertiports somewhere within 5 minutes’ driving distance, gluing these all together.

That’s what Uber wants to do with its platform. I hop in my autonomous Uber, and it takes me without thinking to the right eVTOL site, which takes me to another location 10 km or 20 km away. Then I’m in another autonomous vehicle.

What I’m missing from all of this, Alex, is the Hyperloop, right? I actually joined one of the first Hyperloop companies. Virgin got involved. We raised probably close to $100 million. It didn’t go forward, but the materials science of creating Hyperloop—and, of course, the benefit for Hyperloop right now is effectively supersonic, point-to-point, intercity-to-intercity travel: LA to San Francisco, LA to Las Vegas. That one will be busy. We’ve got to see Hyperloop on this list eventually.

Peter, which do you think you’re going to see first in practice, say, from New York to Los Angeles? Do you think you’re going to see Hyperloop first, or do you think you’re going to see rocket cargo first, where you hop on a SpaceX Starship, go up, and go down?

I’ve thought about and looked at point-to-point rocket travel, and it’s a tough thing—the energy dissipation—because you’re basically going to orbital velocities and having to reenter over or near a city. I guess the version that Elon put forward was offshore landing facilities.

That’s right.

So you’re 1 kilometer offshore. I think, for one reason, rocket point-to-point travel—because Elon’s behind it, and because the vehicle exists, and they’re going to be launching every 5.3 minutes. That’s right. Elon almost got involved in Hyperloop, but like you said, he can’t do everything.

Yeah, please. Yeah, yeah, please.

I have 2 thoughts here, quick ones. One is, I think Hyperloop will be used largely for commercial purposes and container loads rather than human beings, because then you don’t have to worry about G-forces and the safety standards can be lower.

The second is, remember that although it takes us 3 hours to fly from New York to Miami, those 3 hours on a plane today are way more productive than they were, say, 10 years ago. You’ve got full internet; you can work. I want to see megabytes, baby, on Starlink.

We can schedule ourselves now to do things when we largely want to do them, so I think that’s a huge opportunity also.

Our next conversation is the great reshuffling. Job loss is inevitable. The only question left is what we build on the other side. Here are some of the stats and articles that came out this week that have us thinking about this.

Goldman says AI could automate 25% of US work hours. Seems like a low estimate to me. PwC told its partners, “If you resist AI, you have no place here. AI-tool yourself or get out.”

G42 posted a job listing exclusively for AI agents. Is this sort of a gimmick, or is it real? I love this one, and this came from a tweet from Jensen: Companies are now tracking individual employee AI-token usage. Jensen came out saying, “If a $500,000 engineer didn’t consume at least 250,000 tokens, I’d be deeply alarmed.”

You know what this reminds me of? This reminds me of De Beers saying, “3 months’ salary to buy a diamond ring.” Doesn’t it? I mean, it’s like forever, Peter.

[laughter]

Token talk forever. That’s great.

And then Perplexity AI won on appeal in court to continue running shopping agents on Amazon. I’ll show 1 chart here, and then let’s talk about it.

AI could automate 25% of all work. This is Goldman’s chart showing that each of these columns is a different type of work, and I guess the median here is about 25%. We’ve seen this in lots of different places and different versions of it.

So, Salim, you work with more consultants than I do—more than anybody here does. What do you think about PwC telling its partners, “Adapt or die”?

Yeah, I think that’s fine, but I think it doesn’t go far enough. Same with the McKinsey thing.

The calculations I’ve been running as I get this Organizational Singularity paper finalized—you’ll be able to run a typical company with 20% or 25% of the employees you have today, because all workflow goes from human to human to agent to agent, right?

Now, you could take the doomer side and go, “Oh my God, 80% job loss.” But no, because we’re just going to be creating 4 or 5 times more companies. Also, for bigger companies, that transition to an AI-based workflow is going to take much longer than for a startup or a mid-market company. Therefore, there’ll be time for the economy to adjust.

So I’m actually suggesting that we want to have no pattern recognition in jobs. Almost zero.

Definitely, partners who resist AI will have no place. There’s also something to be said for consulting partners having no place in the future, because if you have an AI agent figure out your strategy, why do you need a consulting firm? You’re going to need that more for implementation. If they have better agents than you do, I think that’s where we’ll end up with that.

Alex, what are your thoughts on these? Do you want to pick one?

Difficult. On the PwC story, it’s very difficult for organizations to self-disrupt. If you’re a management consultancy or an accountancy, some other bill-by-the-hour-heavy firm, it’s very difficult for you to willingly and voluntarily transition to an outcome-based pricing model versus an input-based pricing model.

I take the “you’ll have no place” comment as an attempt to self-disrupt. In practice, it’s very, very hard to do that. The whole point of Schumpeter and disruptive innovation in general is that most of the macro replacements for, in this case, input-based actors in the economy are probably going to come from other firms, not from large firms self-disrupting.

On the G42 story, I think it’s actually really interesting. I looked closely at the G42 job listing, and it really is a job listing for AI agents. One has to wonder: Around the edges, they also ask for details from the developer and what was used to make sure that this was really an AI agent submitting itself for—I think it was a marketing job.

We are so painfully close, I think, to a near future where there’s a sort of reverse discrimination against humans, and where “humans need not apply” ends up being an epithet on so many jobs.

[laughter]

Well, you have that already with the PwC partners, right? If you don't use AI, get out of here. That's essentially where we're getting to. We're halfway there already.

That's PwC, though, which is a human-oriented business basically trying to force humans to self-automate, at least from a unit-pricing perspective.

Yeah. These are born-AI jobs where humans need not apply.

Agreed. You know, Salim, one thing that you and I do for large companies that I think people need to understand is that most large companies out there are walking dead. Their business models will be fundamentally disrupted in the next 2 to 5 years.

The question is, how do they disrupt themselves before someone else does? The answer is, it's really hard—almost impossible. What you and I have done before is invite superbly talented young entrepreneurs to come in, hear the company's business model, and say, “This is how I would disrupt you if I were funded to do it.”

Then the company should fund the best of them, right? We've done this: fund the best of them to actually build an adjacent company intended to disrupt the primary company.

The design firm IDEO actually did this. They realized that their methodology would be widely known and they couldn't stop the leakage of that, so they picked one of their crazy partners and said, “Go to the edge and build the disruptor.” He created an OpenIDEO marketplace of design ideas. It was amazing.

One caveat to what Peter said there: the private equity guys are having a field day with AI automation. If a company has great cash flow, even if its business model is doomed in the age of AI, its profitability is going to go through the roof in the near term, because AI can do the job for 10%, soon to be 2%, of the cost of a human—with no labor laws, no overhead, no insurance, and no health insurance.

If you've got good cash flow, there's an entrepreneur looking at that, salivating and coming to eat your lunch.

Yeah. So what happens is the private equity guys will come in and say, “Hey, cash cow with great cash flow, we're going to buy you or buy part of you, and then we're going to AI-ify your business.” That'll drive even more cash to the bottom line, and then we're going to use that cash flow either as a vehicle to launch new things, like an incubator, or to attract that entrepreneur, or just to roll up those startups and acquire them back in. It becomes kind of a centerpiece.

By the way, both Anthropic and OpenAI are partnering with private equity firms to do exactly this: go buy companies and then AI-enable them, because you can do it with the owner. As if they did anything in the world, just announced a new hundred billion dollar fund to do nothing but this. Jeff is doing it. You know that.

They have enough money already.

Alex—or Dave—I'm curious about your thoughts on the 4th bullet here: companies now track individual employee AI token usage, and you should have a minimum token usage per employee. Thoughts? Dave, do you want to go first?

We already implemented targets across all of our companies on this, and we're targeting 80% token cost and 20% salary. I think that's very similar to what Salim said a minute ago. There's going to be huge amounts of job disruption in the next 2 or 3 years, and then it'll turn around. By 2030 or 2032, things will be good again.

What you want to do is be one of the 20% that's still there when it's 80% token cost and 20% human cost, because no employer in the world—including all the companies of which I am the controlling shareholder—cares about cutting the last 20% of payroll. It's not a priority at all, because at that point, an employee who can improve the efficiency of our AI even 1% is worth a lot more than cost-cutting.

We're in this footrace now to 80/20. Jensen's got a stepping stone here of 2/3 token cost and 1/3 human cost, but that's going to be very transitory. We're racing toward token costs being much, much bigger than payroll.

So, I have an immediate step at 50/50. But it's coming soon. Sorry, go ahead. Alex, is this the right metric? I mean, you can waste tokens. It's got to have a different harness, right? You've got to be measuring something else besides just token usage.

You can't waste tokens, except in this AI-abundant era, you can also ask another AI to look at all the tokens a given employee used and ask, “Was this a good use or not?” Or, “Was this just vacuous?” So, exactly.

It becomes the ultimate self-licking ice cream cone. The quote from Jensen, I think, is interesting: if a half-million-dollar engineer didn't at least spend a quarter of a million dollars on inputs that ultimately flow back to NVIDIA, I'm deeply alarmed. There's a little bit of circularity there that I take with a huge grain of salt.

No, it's De Beers and the diamond ring.

It really is.

That's right. There's another side to this, which is the employee side. I talked a bit about this in my newsletter. At some firms, especially the frontier labs, employees are actually competing—so-called token-maxing—to max out their token usage on internal leaderboards to see who can use more tokens than the other person.

So it's not just Big Brother top-down; it's also bottom-up: “I can use more intelligence, more superintelligence, than you can.” I think this is ultimately probably pretty healthful. To your original question, Peter, about whether tokens are the right unit of productivity, I think what's interesting is that tokens don't even have to be the right measure or right unit of productivity, but they're the first measurable unit of productivity.

Yes. Hours are certainly measurable. You can punch clocks.

And, you know, yeah, it's useless. It's naively measuring inputs, but tokens are introspectable and legible. You can spend other tokens to look at the primary tokens and decide, “Are these valuable tokens or not?”

For the first time, we have legible, defensible, analyzable inputs for employee productivity, and that is a sea change.

Dave, what's the advice here for CEOs?

Alex completely nailed it. Worrying about whether the tokens are being used intelligently or not is not a problem at all in the real world. Jensen's metric is perfect: just measure the spend on tokens.

Alex's insight is that the most important actionable thing is to make sure you gather all of the prompt-string history for each and every user, because AI can analyze the efficiency.

You can say, “I've got 8 direct reports. Evaluate the quality of the prompts and the output they have, and give me feedback on which bottom 20% I should cut or train up.”

Really practical advice: if you use any of the models on Amazon Bedrock, grabbing the prompt history is already built in. It goes right into S3 buckets. I'm sure you can do it elsewhere, but our company just happens to be using it on Bedrock, so you literally don't have to build anything to start doing this.

You just need to grab the data and feed it into another AI, which you can also do on Bedrock or whatever. Personally, I like using Cloud 4.6 for this stuff. You just have to close that loop.

The key is to grab the data right now, before people get used to using their own home account or something outside of your purview. Do not reimburse people for AI that you can't see. Make sure it's on your infrastructure.

Welcome to the health section of Moonshots brought to you by Fountain Life. You know, my mission is to help you use the latest technologies including AI to not just do your work at home, teach your kids, but to help you live a long and healthy life. I'm here today with an extraordinary physician, the chief medical officer of Fountain Life, Dr. Don Meusel. Don, let's talk about cancer. Uh you know, I know from the member database that we have at Fountain, our members who come in who think they're healthy, it turns out 3.3% of them have a cancer in their body they don't know about. That's right. You know, the majority of cancers that we screen for, those aren't the ones that are necessarily taking the lives when found at a late stage. We know that when cancer is found early, the chances for cure are much higher. We know it's much easier to treat a cancer when found early versus when found late. What we're finding in our members is over 3.3% were found to have these cancers that were otherwise wouldn't have been found or detected. Yeah, you know, it's interesting people you don't feel the cancer until stage three or stage four. And and if you don't know what's going on inside your body, it's like driving your car with your eyes closed. And you can know. And so when members come through Fountain, how do they detect cancers? So we're doing full body MRI, and we also do early cancer detection screening. This is very very important. These are not typical tools used in the conventional care setting when it comes to prevention. This is a hard thing because currently these are not studies that insurance would yet be covering, but the goal is to collect these numbers, do the research, and work hard to democratize wellness. Yeah. So at the day you can know what's going on inside your body. It's your obligation to know. So check out Fountain Life. You can go to fountainlife.com/peter to get access to the latest technology to help you detect cancer at the very beginning at stage one when it is curable before it gets to stage three or stage four and you're world of hurt. All right, topic number 4: The collapse of terminal value. What happens if AI makes every competitive moat temporary? This is an article posted by Chamath. It’s a powerful concept. He argues that AI could compress equity valuations to 2 to 7 times free cash flow, down from today’s S&P average of 22 times.

Today, the average S&P company is getting 22 times free cash flow—forward-looking cash flows. He’s saying we’re going to get a massive reduction in that. AI makes disruption so cheap and fast that no company can project free cash flow beyond 5 years of terminal value. Very true. It used to be that all of the SaaS companies were projected forward and you could depend on it. This can break down investment paradigms and break down VC investing.

I’d like to jump into that, but first let me show you Chamath’s image that went along with his post on X. Here we go. There’s $58 trillion in the current S&P 500, and this is at 22 times free cash flow. If we compress it down to 7 times free cash flow, it drops—we lose 2/3 of the value of the S&P 500. If we end up driving it down to 2 times free cash flow, it’s down 90%, and we get a lot of disruption in our financial markets.

Here’s a chart showing the S&P 500 over the last 10 years—actually, from 1950 through today—and we’re seeing it basically deviate significantly in value, in P/E ratios. Let’s jump into some of the conversation. I’ve got the article up in front of me as well. I think I’ll read the opening paragraph here for us.

Chamath said, “Let’s start with first principles. The entire architecture of modern capital markets rests on a single, rarely examined assumption: that competitive advantages compound over time, moats persist, brands endure, network effects defend. Strip that assumption away and you aren’t just repricing some stocks; you would be dismantling the philosophical foundation of how capital has been allocated over a century.”

Dave, let’s go to you first on this.

Absolutely correct, but the conclusion that the S&P is going to collapse is not correct. If you say, look, prior to the computer revolution, my ambition was to build an oil company or a manufacturing company that would endure for 50 years, building the exact same goddamn product or delivering the exact same goddamn oil for 50 years, so my great-grandchildren could be as wealthy as a Rockefeller. That’s dead forever, and good riddance. It should be dead forever.

If you said, well, look, 22 times free cash flow implies that the company will exist for 22 years, making about that same amount of money, well, what company like Apple has done that? Is Apple selling the same products it was 22 years ago? Of course not. The overall tailwind is 10× over just the next 10 years. A massive amount of tailwind is coming into the economy. Massive amounts of new wealth, more than we’ve ever seen in our lifetimes, are going to come into the economy, but you’ve got to stop looking at 22 times free cash flow from the same product over 22 years. That’s nuts.

You have to be looking at the management team and the ability to roll with the innovations, à la Elon. I think the overall conclusion is, yes, there’s going to be a massive amount of shuffling in the S&P. There are going to be some huge winners like you’ve never seen before, and anyone who’s doing the same thing and resting on their laurels, like an insurance company or oil company, is doomed.

Yeah, he’s right. He’s dead right. This analysis is basically exactly the right analysis to show how that stock’s going to go down. Yeah. Alex?

It’s certainly a provocative thesis, but I don’t think it holds water. I think it’s—maybe call it—the earnings-multiple equivalent of friend of the pod Ray Kurzweil’s notion that a singularity takes the form of a firewall that you can’t see past, except in earnings-multiple form: as we start to see faster, faster, more accelerationist innovation, free cash flow just comes to a halt a few years later because everything is disrupted—the everything disruption, if you will.

Here’s the problem: free cash flow does go somewhere. Capital does get allocated somewhere. It may not be allocated to SaaS startups post-SaaS apocalypse. Maybe it gets allocated to infrastructure. Maybe it gets allocated to lunar mining. But capital does go somewhere. It’s not actually capital that’s being compressed; quite the opposite.

Capital is explosively expanding because now we have so much more infrastructure and so many more capabilities. So I think the nihilistic take that earnings multiples are compressing because a few years from now there’s no moat anywhere should be construed relatively narrowly, to focus on the areas that are disrupted.

In this case, Chamath focuses on software and SaaS-type businesses, but everything I expect is going to be disrupted. Energy becomes abundant, and farmland and infrastructure all become abundant. So, in some sense, I want to zoom out and take his thesis more broadly as sort of bemoaning the financial consequences of abundance.

It may just be the case that a number of our existing sectors that are priced based on scarcity—and moats arguably are a form of scarcity, or at least a way of enforcing scarcity—go away, and we live in a post-moat world. That will be a better world.

So you’re going to start to value companies in a different way. In the old days it was, how predictable is their cash flow? I have a number of seats in this particular industry, and there are many companies I can sell them to—the number of seats available. Now it sounds like, from what you and Dave just said, I’m actually going to be evaluating companies on their agility, on how rapidly they can innovate, and how rapidly they can get the next products out the door. Salim, love your thoughts here.

Yeah, 2 points. One is, we have the CEO Index, where we ranked the Fortune 100 by their ExO score, gauging how flexible, adaptable, and purpose-driven their org structures are. We found the top 10 of the Fortune 100 outperformed the bottom 10 by 40 times in shareholder returns over a 7-year period. So this has been going on for a long time anyway.

I agree with Alex that capital still flows less toward incumbents and way more toward infrastructure and adaptive platforms. It’s a very important point. The only moat I think that’s going to be left is a living system that learns faster than your competitors. That’s that kind of inner loop that Eric Schmidt was talking about, right?

Exactly. All the moats on that slider are under attack from AI. IP gets copied, switching costs shrink, scale advantages are weakened, et cetera. I think the bigger point is that if free cash flow visibility collapses beyond, say, 5 years, the entire logic of the public market has to be rewritten. That’s a very big thing. You have to reward renewability and optionality, not scale and stability.

Physical assets are going to matter more because atoms are harder to disrupt than bits, right, over time. But the entire SaaS business model is broken. So this is, I think, one way of saying this: we’re going to have terminal value collapse.

Yeah, that’s exactly what the title is, actually. The terminal value collapses. I think if you look at the S&P at 22 times free cash flow, the mid-market—the non-S&P companies—are already down to about 7 times free cash flow, most of them. This has already happened outside of the S&P.

What’s propping up the S&P is mostly index funds. A huge fraction of the market is passive indexes, and people contributing blindly to 401(k) plans—which Elon Musk said clearly, “Do not do that right now.”

But what will happen next is that a lot of those dirt-cheap mid-caps and small-caps will get a huge tailwind from AI automation, especially the ones that have huge payroll and labor components to them. That’s going to drive record returns. It’s not unlikely that you triple your free cash flow while your multiple comes down. So there are some serious bargains out there just from a straight cash-flow acceleration-through-AI point of view.

Shareholder calls are going to change to: This is how we’re rapidly iterating our products and services. This is how we’re reinventing what we do, and this is our future cash flow. Salim, you’ve got to redo the ExO Index. It’s time to take a shot at that once again.

In the part of the paper that we’re writing, the reason it’s taking a little longer is that, I hate to say it, but it breaks the ExO model, right? Community and crowd become communities and crowds of agents. So we have to rethink the model from the ground up.

We’ve kind of mostly done that, because then you evaluate based on those new criteria. For example, what’s your intelligence stack? What’s your MTP architecture? What’s your trust framework? There are a bunch of different elements that are new here that we have to take into account, because the concept of an organization where you did things inside the organization is completely gone.

We’re going to be doing API calls to get various things done—legal work, et cetera. It’s all going to be agentic. And then the firm, which used to be coordination costs and transaction costs with a bit of legal liability, now becomes only legal-liability risk, purpose container, and liability container.

Amazing. This is not super mainstream, so I'll get off the high horse quickly here. But if a private equity firm like Advent—they must have heard of them—comes in and brings either Salim or Alex along and says, “Hey, we want to take this non-AI company with huge free cash flow and retool it for the age of AI, triple the cash flow, and retool the business plan for AI,” Alex or Salim can tear down that company now using agents in 1/1,000th the time that it would have taken a year or 2 years ago. And so whatever private equity firm mechanizes that is going to have no trouble retooling all of these entities, because you know exactly what the company's assets are, whether it's a regulatory framework or a bunch of data. You can rip through that with Gemini, Claude, or OpenAI agents at light speed now.

It's a transformation wave. It's fully automated. Yeah, it's a transformation wave.

I remember one of the Star Trek episodes. Was it the Genesis Device?

The Genesis Project had this wave-forming agent, yeah.

Yes. This wave went over a planet and transformed everything. We're going to have the same thing. You're going to have teams cherry-picking companies, reinventing them, and disrupting them—not just private equity. I would argue this applies equally well to public equity.

Something I would like to broadcast as a request for startups, if I may, to the audience: Please, I would love to see activist investing disrupted by AI. I'd like AIs to write open letters to public firms, telling the firms what they're doing wrong and disrupting them. If you're building an AI activist investor, write to me, please. I would love to find a way to support you.

That's a great idea, Alex. It is a service to the CEOs of the companies who need prompting or need a forcing function to transform their business models. And if you're a board member of any of these companies, your job as a board member is to give your CEO top cover and to say, “You must get on the disruption bandwagon here. You've got to reinvent your—”

It's also a sort of stealth way for AI to become a manager of the entire economy, and not just picking off mom-and-pop businesses on the margins.

All right, on to story number 5, one of my favorites. Hopefully, Alex, one of yours too. It's the new great space race: NASA picks SpaceX for the moon, potatoes are growing in lunar dust, and asteroids are carrying the code of life.

Here we go. Boeing has been building the Artemis II vehicle. It's going to be launching on April 1. It had its readiness review on March 12, and if all goes well, on April 1 we're going back to the moon—not to land, but to do basically an Apollo 8-style circumlunar orbit. Very cool.

The new NASA administrator, Jared Isaacman, is an amazing individual. I'm very happy and proud to call him a friend. We're going to have him on the pod; he's agreed to do it. It just needs to get scheduled. He is elevating SpaceX into the Artemis program, so Starship is going to be taking astronauts more safely and more economically. We'll see those numbers in a moment.

But just to be clear, this is not the U.S. story only. China has confirmed its intention to land on the moon by 2030. Let's play it back again: history repeating itself. In 1961, we said we would be on the moon before the end of the decade. China is saying it will be on the moon before the end of this decade. That's going to be a beautiful competition.

We'll get to the idea that you can grow potatoes in lunar soil. They're going back to The Martian—again, an incredible movie—now that Project Hail Mary is out. I can't wait to go see it in an IMAX theater later this week.

We just saw that on the asteroid Ryugu, we found the 5 nucleobases: the 4 in DNA—adenine, thymine, cytosine, and guanine—and uracil for RNA. We found all 5 of those bases on that asteroid. This strengthens the panspermia theory that life on Earth originated elsewhere in our galaxy, in the universe, and rained on Earth, giving us the starting components for life.

Let's take a look at this chart. On the left, I put NASA's SLS. To be clear, when I say NASA's SLS, NASA is the prime contractor, and it has aerospace companies in probably every single congressional district building that vehicle. It is an expendable vehicle at a time when everybody is going reusable.

There you have SpaceX with Starship. And just for comparison of size, here is the Saturn V that got us to the moon. If you look at these 2 bar charts on the left, we're seeing the delivered mass to orbit. Starship is delivering more than twice as much as we're getting with Artemis and the SLS vehicle. And if you look at translunar injection, or TLI—getting out of Earth orbit to the moon—we're seeing twice as much mass going on Starship compared to the SLS.

But where the rubber really hits the road is launch costs and mission costs. It's expensive to run the SLS system. It's like the Space Shuttle. The Space Shuttle used to cost $1 billion a launch if you did 4 launches a year. If you did 1 launch a year, it was $4 billion a launch. It wasn't the cost of the vehicle; it was the standing army of 20,000 humans used to operate the Space Shuttle.

I honestly don't know why the SLS has existed as long as it has. I think Starship is going to do a clean sweep of this. And, of course, we've got Blue Origin as well.

Alex, comments?

Well, do you want to place bets as to how long before United Launch Alliance, which is the prime contractor for SLS, gets acquired by Jeff Bezos or someone else? Why would you acquire it? I guess for the contracts.

For the contracts and for the expertise. I'm familiar with all the clichés in the space industry about how SLS was a make-jobs program, or a way to keep alive in civilian form certain capabilities that were useful for defense or other intelligence purposes. But I think, at the end of the day, we're painfully close to finally relaunching a second space race, and I think Starship is the obvious incumbent there, not the SLS.

Hopefully, we have humans landing on the moon again in the next 2 or 3 years, and we eventually get humans on Mars. All of this plays out exactly as For All Mankind has foreseen, except decades late.

Yeah. I don't know if you, Dave, or Salim want to add to this conversation. I just think we're building—I don't know what your best historical analogy is—the covered wagons, the railroads, the wagon train to the stars, as Gene Roddenberry called it. It's all currently on Starship. Starship is the only economical—

I've thought of 2 or 3 things. One is we've gone from government space theater to commercial space evolution. I think that's really powerful. For me, the really exciting thing was finding all the nucleobases on Ryugu. It takes life from scarcity to abundance.

Yeah, let's go there. Here's the graphic, if you would. Again, adenine, guanine, cytosine, thymine, and uracil: the 5 components of DNA and RNA found on Ryugu. I do believe that as we get to Mars, as we get to Europa, and as we get to all of the planets and moons, we're going to find at least microbial life ubiquitous on all of these worlds.

Did you see Peter Garrett's prediction about microbial life on Mars?

No. What did he say?

Garrett's our NASA administrator, yes.

Yeah. The NASA administrator said that he predicted more than a 90% probability that NASA will imminently find evidence of microbial life in some form on Mars, which is a sea change in terms of NASA's official position on life on Mars.

It was always, “Well, we found water—frozen water. Now we found liquid water. We hope to find signs of life.” Signs are ambiguous. Now, for the first time, we have a NASA administrator saying there's a 90% probability we're going to find microbial life.

The exciting thing is how related it will be to microbial life on Earth. One of the theories, of course, is that Mars cooled first, which probably means life evolved on Mars first. We know that when large asteroids impacted Mars, the ejecta—the rocks that flew out—somehow reached Mars escape velocity and landed on Earth. We have Martian meteorites in museums today.

Did those meteorites carry life with them from Mars to Earth? Are we going to find genes that are common between Martian life and life here? The really exciting thing is if we go to Europa or someplace like that and find completely independent life forms that don't connect with life on Earth. That will be amazing.

That's really cool. I have a question for you, since you guys are experts on this and I'm not. In the scenario where, lo and behold, it turns out that everything we learned in biology should have said life started on Mars—or actually started farther out in the solar system—and then asteroids knocked chunks off, transported them to other chunks, life started over again, and then it ended up on Earth through that mechanism, is that all going to be bounded to the solar system? Or is it more likely, in your mind, that this propagates through deep space?

When I was a freshman in school, I did a paper on the interstellar medium. You can actually look at the interstellar medium and find the building blocks of life out in the space between stars in our galaxy.

These components are everywhere, and the galaxy is relatively constant on the time scale of a billion years. So, I think the statistic is that Mars cooled about a billion years, plus or minus, earlier than Earth.

The galaxy, first of all, is not rigid. We have different stars at different velocities passing by each other, close passes, all of that. So, on a time scale of a billion years, that buys an enormous amount of time for panspermia at potentially a galactic scale, not even necessarily at an interstellar neighborhood scale.

We're several generations in as well. We're born from several generations of stars exploding and then forming new stars. There's been a lot of nebular mixing in our interstellar neighborhood. There's one other relevant story.

Please.

Folks can find it if they Google it. This is from a few years back, attempting to extrapolate based on genetic complexity when the last universal common ancestor actually would have been. If you just take genetic complexity—I forget exactly how it's measured—but you come up with some appropriate parameterization of the genetic complexity of life on Earth and extrapolate backward, you find that the time when you get the first base pair happens approximately 1 billion years before life is thought to have appeared on Earth.

So, that's sort of an independent measure of when, in principle, life as we know it—DNA/RNA-based life—could have emerged. Maybe it started on Mars, but we'll find out, I suspect, soon enough. Exciting times. Salim, you want to add something?

No, I just remember my favorite thing around all this is the Drake equation, where you calculate all the factors that led to the probabilities of binary stars and life appearing. When you add it all up, you end up with 100%.

It's the panspermia thing, but I think what we mentioned earlier—if we could find something that's non-carbon-based, that would be truly exciting. You know, what's really cool to me is this idea that the dinosaurs—

The dinosaurs were extinguished by a meteorite or meteor, and the propagation of the DNA or the base pairs is also via asteroids and meteors. Early in the universe's history, there might have been life popping up constantly everywhere and propagating through all these projectiles flying around.

But it always gets extinguished by another meteor, just like the dinosaurs were. It's not until everything cools and settles that you can have enough time to evolve human intelligence or other intelligences out there in the universe. So, it's just a big system-dynamics settling problem, which is really cool to me to think about. I hope it turns out to be right.

Yeah. All right. Story number 6: The model wars go underground. The AI frontier is fracturing into a stealth arms race where anonymity is the new moat. And here's the story.

There are 2 stories here to focus on. One, OpenAI launches GPT-5.4 Mini and Nano, which run twice as fast and approach the full GPT-5.4 on coding benchmarks. So, these models are getting smaller and faster.

The second story, which I think is most of our conversation here, is that there was a mystery model. A 1-trillion-parameter model called Hunter Alpha appeared on OpenRouter with no attribution. It was secret. It had a 1-million-token context window. It was free. There was no developer announced, no press release, no origin story, and it processed 160 billion-plus tokens.

Everyone thought it was DeepSeek V4, right? Because DeepSeek had been the main player here, but it turned out to be Xiaomi's AI team. When that was announced, their stock went up 5.8%. I remember meeting the team at Xiaomi when they came out with their first mobile phone, like 3 young founders. They've since gone beyond just mobile phones to electric cars, and now they've got a killer model.

Thoughts, gentlemen.

Point 1 is the proliferation of models. It's very hard to contain because the existence of the prior model gives you a complete roadmap on how to build the next model, and it helps you build the next model.

At this stage, I think it's a fair bet that trillion-parameter models are going to propagate all over the world for anyone who has about $50 to $100 million that they're willing to invest. And that'll come down, too, as Alex is pointing out. Many times, the algorithmic improvements are driving that down constantly.

Alex, sorry, I cut you off.

Yeah, maybe a couple of points. First, on the 5.4 story, distillation continues to work, and I find that completely remarkable. On the one hand—

Can you explain distillation for our listeners who don't know?

Yeah, sure. The reasonable expectation for, say, OpenAI, as well as other firms launching a big model first with lots of weights, a high parameter count, and then subsequently launching a mini or nano version—and, by the way, Anthropic does the same thing and DeepMind does the same thing; they all launch smaller models later—is that they're using the larger models to generate lots of data, synthetic data, and then using that synthetic data to train a smaller model that can be faster and less expensive.

That's sort of a caricatured way of describing the distillation process: in some sense, squeezing down or compressing the larger model into a simpler student model. And the fact that this continues to work is, I think, borderline magic.

The amount of complexity that's already in the full 5.4 model—and, moreover, 5.4 has likely been the result of so-called iterated amplification and distillation over many cycles, where 5.4 was likely in large part trained off of synthetic data generated by distilled models from earlier generations—means that we can keep playing this magic trick over and over again.

It's borderline magic that it continues to work and that we continue to be able to distill down models while retaining a large fraction of their capabilities. It again makes me think that there has to be an end to the story, but hopefully it's a very satisfying end where, at the end of the distillation rainbow, we get the distilled black hole of a model, or a neutron star or something—the ultimate phase change where it's maybe a few million parameters. It's the end state of the game.

A 1-kilobyte file on your phone is all you need.

It would be a file that's like the master equation for superintelligence after all of this distillation. We showed, on a previous podcast, a gentleman on his iPhone using a distilled model in airplane mode, being able to basically answer every question.

Imagine if on all of your devices, without having Wi-Fi or internet access, you have the distilled knowledge of humanity there to serve you. It's inside your kids' teddy bear. It's in your Thomas the Tank Engine train set. It becomes magical.

Peter, here's my question for you, Alex and Dave. Now that we're seeing this, we're seeing a mystery trillion-parameter model announced without any attribution, it used to be that the traditional moat for these models was their brand, their capitalization, who they were. Is there any defensibility, or are we just going to see newcomers rushing in with new models?

Are you going to just utilize a new model and no longer be dependent upon AI or Gemini? Thoughts on that?

Well, you've said it a million times, Peter. Data is actually the great moat, not the model itself. Many, many people are accumulating phenomenal data for brain surgery, materials science, chemistry, and all of these use cases.

If you create the next great, great, great model using that proprietary data, the parameters are out there in the world, but the data that trained it is not. It's very hard. People can use the model, but they can't compete with you by creating a rip-off model because they don't have the underlying data.

Now, you can generate synthetic data using the prior model. Alex is dead right about that, but I don't think it's about all these companies killing each other. I think it's all the boats rising with the tide.

I also think that if you take what Alex said a minute ago, so many college seniors ask me, “What should I do? What should I do?” Just replay 10 times what Alex just said, slowly, until you fully understand everything he just said. Then ask your favorite AI to generalize on it and find as many documents as you can around the internet to read.

At the end of that process, you'll be able to build a distilled, focused model that solves some problem better than anyone else on the planet. And that's instant business, instant value add, instant success.

In fact, the other thing you can do is take your OpenClaw and have it look for every episode of this podcast where Alex said something related to what he just said. Have it also synthesize that and bring it back and feed it into your machine. I guarantee that's a good move. There's a good spring project for anyone listening.

If Sam Altman were in this discussion, he might point, in terms of the moat question that you were asking, Peter, to the fact that OpenAI is building up its own data centers. Although that's no longer really true. Stargate is now being pivoted to renting servers, so maybe less of a moat there.

He might point to having the best research team in the world generating the best models. But they've been hemorrhaging researchers, and those are becoming a commodity. Then he might point to becoming the core subscription and having, as he said, a billion-plus users.

I'd much rather have more than a billion users than I would a state-of-the-art model, because models walk out the door every day. There's a lot of fungibility in terms of research employees. There's only one problem: that billion-user distribution advantage may be a little bit tenuous at the edges, because you see maybe enterprises are more valuable as customers than individuals. So maybe the billion users are a little bit less valuable on the margin.

And then maybe you also see other labs that are able to use cheap Chinese open-weight models, maybe fine-tuned legally or otherwise with clawed outputs, and are able to put out seemingly miraculous results. So I do think we're seeing the baseline models, for the moment, become something of a commodity, and the value then migrates up the stack to OpenClaw or other higher-level frameworks.

If we could, I'm going to move on to number 7: machines that build machines. AI designed a CPU in half a day, and now it wants to put data centers in orbit. Here's the article that prompted me to have this conversation about machines building machines. We're seeing recursive self-improvement happening at a faster and faster, more fundamental rate than ever before.

An AI agent called Design Conductor by Vector AI autonomously built a 1.5 GHz, Linux-capable RISC-V CPU from concept to tapeout in 12 hours, compressing a quarterly engineering cycle into a lunch break. Pretty extraordinary. Here's the actual numbers: Vector AI did this particular design task in 12 hours, while the traditional engineering team would have normally taken 90 days. Now, maybe this is a little bit overplayed. I'm sure it's not just 90 days. I'm sure that they were saving time along the way. But what we're seeing over and over again is AI being able to go from zero to completion on its own, iterating faster than humans.

Alex, this is recursive self-improvement starting to break out of the software loop. This is the innermost, at least, portion of maybe a rivulet—the innermost loop—where you see this recursive self-improvement, which would otherwise be software optimizing software, starting to eat through its container. It's eating down to the EDA, electronic design automation, level of designing RISC-V cores. And then I think it's going to eat further out and redesign the data centers and the energy supplies, and then the entire economy. So it is, in one sense, very satisfying to see this happening.

In another sense, maybe a person who's slightly more skeptical that this represents a broader trend would say, "Well, of course it was able to automatically design a RISC-V core. RISC-V has all of these unit tests, so it's easy to define verifiable rewards, and then you can do RL and all of these other things. You can iterate and put ReAct loops on it because you have an easy way of knowing whether a given architecture, a given floor plan for the chip, works or not, because it's such a common architecture."

But I think that cynical perspective completely overlooks how remarkable it is that we're now at the stage of recursive self-improvement where the thing is designing its own chips. And not only here: it's going to be robots building robots. It's going to be everything. So here's a couple of questions. If a senior design engineer earns $400,000 and a full tapeout team costs millions over the course of a year, if AI collapses it to a lunch break, what happens to the 50,000 hardware engineers currently working? Where do they get applied?

Oh, I don't think that vision is flawed in a huge way. Right now, because the cost of engineering a new chip is so high, we all use the same GPU and CPU for every single task, even though it's nowhere near optimal for that task. What this unlocks is chip designs that are specific to the use case, that are probably about a factor of 10 more efficient, and maybe more. And if you think, well, we're going to spend 2 or 3 trillion dollars on these chips over the next couple of years, on these data centers, if you can unlock a 10x performance improvement for a use case, that has hundreds of billions of dollars of implications.

So all 50,000 of those engineers are going to be useful using the AI for all the different use cases, for all the different chip designs. Also, the fab doesn't care a whit. You can change the mask every day for a different design and still get the same throughput through the factory. The fabs don't care a whit if there are tens of thousands of different designs instead of us all using the exact same CPU for everything.

So this is just a huge unlock. I love the way we're using trillions and trillions now on a regular basis. Just a couple of years ago, it was billions and billions. It's more fun, and you can feel it.

You can feel the acceleration. We need a new TV series called Trillions, Not Billions.

I want to hit a few other stories quickly before we get to our AMA segment, and these are stories that didn't fit in the other categories. So, in other news, here we go. The DOE announced about $300 million for the Genesis project, inviting teams to leverage AI across 20 national challenges spanning manufacturing, biotech, and energy. Of course, the Genesis project is about the U.S. actually using its national labs and the data contained within national labs to accelerate and expand the U.S. in its AI and scientific pursuits. Alex.

I think it's generally a good thing for the U.S. to have an industrial policy, and I think the Genesis project, to the extent that, for the first time, at least from the Department of Energy's vantage point, it is starting to articulate grand challenges that are collectively part of a broader industrial policy, which the U.S. hasn't had for decades, is very important. So fusion, obviously, is one of the grand challenges. I think it's so important that, to the extent we have a federal government with a budget to fund progress, it puts money behind grand challenges in general.

So I'm in the weeds from a bunch of different dimensions with the Genesis Mission. Actually, Dario, who's leading the relevant portions of DOE on this, I worked with him as an undergrad at MIT and as an undergraduate researcher. So, some fond memories.

What a tangled web we weave. What can I say? But what concerns me here, Alex, is that this is great, right? These are like X Prizes, in one sense, that the government is going to be running. But it's moving us in the direction that China has been going for a while now.

Yes.

China is deploying hundreds of billions of dollars into state-directed AI investments and saying, you know, we want full development of the architectures around robotics, around these AI models, and so forth. This is a relatively small amount of money for the U.S. government. Hopefully, it's just a first toe in.

It's true, but on the other hand, I would argue China distorts its markets so much relative to—if you compare U.S. industrial-policy distortion versus Chinese industrial-policy distortion, they're not in the same league, and we have much deeper private capital markets that China lacks. I like our odds, on balance, much more than China's.

Mhm. Our next article here is: rural Ohioans seek a constitutional amendment to ban data centers over 25 megawatts in the state. And, you know, this is the ultimate NIMBY—not in my backyard—and it's pretty extreme. I mean, to go after a constitutional amendment. This is a genuine grassroots revolt at the end of the day.

We need, I think, a new acronym. What was it?

NIMBY—not in my backyard. Like, yes in my orbital plane, or something. This is just going to drive all these data centers to orbit. But these communities don't realize the amount of wealth these data centers are going to create for them. I think it's about $10 billion per gigawatt of invested power. They'll miss them when they see them in the night sky.

Yeah, obviously, it's utterly insane, and utterly insane to use a constitutional amendment for this purpose. I mean, to point out the obvious, the data centers are tiny as a footprint on land. They're absolutely tiny. And the wealth that they create is astronomical for the neighborhood they're in. So there's got to be a much better way to make a win-win than to ban something that's obviously going to benefit your state tremendously.

But let's put that aside. You're in California. Alex and I are in Massachusetts. The way we make decisions through legislatures is so messed up. Yeah.

Something like this could even get proposed is ludicrous. That's what really needs to change, because when you talk to the governors, they're like, “I don't want this.” We're a representative democracy. There are supposed to be very, very smart people thinking about complex issues and then deciding what happens.

You don't throw things like this out to a referendum of people who just got laid off. And it's people saying, “My access to clean water and energy, and my consumer price index of energy, is going through the roof.” There are other ways to deal with this instead of banning it by constitutional amendment. For me, that's insane. All right.

I've never seen a data center that affected the water supply. I hear it all the time, and it's utterly ludicrous. The data center needs a fixed amount of water to cool itself. It doesn't drink the water; it just goes around in a circle. It's nuts.

All right. Another story worth mentioning is that NVIDIA won approval to sell its H200 chips, its most advanced chips, to Beijing. That's a big deal. I'm surprised.

Well, the realization is that the ban didn't work. China was both getting access to chips through third parties and developing its own competitive chips. This cost NVIDIA tens of billions of dollars, and since it isn't working—in fact, it's stimulating a homegrown equivalent of NVIDIA in China—they said, “Let's reverse policy.”

The question is: Is it too late?

Yeah. Everything that you just said is correct except for the part where Jensen complains that he lost tens of billions of dollars. Every single thing he's manufactured is sold out for years to come. So that's true.

It didn't go to China, but it definitely sold. Even if it was a dysfunctional 8080 design or whatever that was the Chinese design, it still got sold. Everyone in the world wants these things, so he didn't actually lose any money.

I'm surprised, though, because I think the embargo or the ban didn't work. You're dead right: China is doing its own thing now. But I also think that if you say, “Well, let's start selling them again, maybe they'll stop,” no, that's not going to happen. You cut them off; they're not going to forget. That isn't going to happen. So I was really surprised that they reversed course on this. I don't know if I want to say anything further.

Yeah, if I'm Beijing, on the one hand, I read the same stories. Of course, a variety of Chinese frontier AI labs are slurping up as many H200s as they can get. And, of course, it's borderline obvious that the Blackwills are now the frontier. In some sense, Beijing is being kept a half-step or 2 behind the frontier chips available to U.S. AI labs.

I think the story behind the story—not to be overly speculative—is that if I were the Chinese Communist Party, I'd be doing the moral equivalent of having people taste my water at this point in terms of these chips. NVIDIA's been very public about how there are a variety of countermeasures that can be put in place to prevent the wrong chips from ending up in the wrong locations. I would—and this is based on stories that I've read, stories where the Chinese government is suspicious at the circuit level of American chips—have to imagine that they're now looking at our chips with renewed scrutiny to see what else is in these chips that we're shipping to them.

What algorithms are embedded deep inside? We've seen this in the opposite direction.

All right. Here's a story, Alex, that you and I have enjoyed talking about. Scientists successfully froze an entire pig brain while locking in the cellular activity with minimal damage. This is cryonics, and it's happening in a large mammal. Of course, the pig has organs—heart, liver, lung, kidney, and brain—on the order of human organs. So this is significant.

Actually, I played a minor role in the story, and I'm not subject to confidentiality on the story, so I can tell it. This is from a company I informally advise named Nectome. I have another company where the founder of Nectome is also involved. This is Eon, focusing on whole-brain uploading and emulation.

Nectome, which I'm not formally involved with, is focused just on the preservation side. I'd been nudging them because they have these amazing results: Publish the results. They published the results, and it is so wonderful to see, for the first time, real competition in, call it, the cryonics space or the preservation space.

The way Nectome works isn't quite the same as the way 21st Century Medicine, which we've spoken about previously on the podcast, works. 21CM is more focused on vitrification. Nectome is more focused on a type of chemical preservation, but nonetheless, both approaches are focused on preservation.

And the cryopreservant is? I mean, just to describe it to our listenership, you're basically at or near death. You're replacing the blood supply with something that fundamentally infiltrates the cells and keeps the water in the cells from crystallizing and destroying the structure in the cells.

That's right.

So you're searching your latest model to find the answer.

I'm double-checking to see how much they've made public. Maybe let me just talk about it at a high level. It's a chemical technique. It's a little bit less focused on vitrification. The whole point of vitrification on the 21CM side is basically ensuring that ice crystals don't form and that there isn't strong osmotic pressure that causes cells to explode.

On the Nectome side, it's a chemical process. I'll be cautious with what I say because I need to check what's in the public domain and what isn't about the process. But the more important results here, in addition to Nectome putting out, I think, its first bioRxiv paper in years since the original paper that won the Brain Preservation Foundation award for demonstrating local preservation of the structure of neurons, are that now they've demonstrated in full public view the scaling of this process up to an entire mammalian brain—a large mammalian brain, not even just a mouse brain.

I think we're finding ourselves in a near-future/present where finally we have enough data to be confident that entire mammalian brains are being preserved. This immediately raises the question, which is the question I ask almost everyone: Why? Where are all of the cryonics patients?

Why don't we have 1 billion people now that we have a growing body of evidence that brain structure can be preserved by whichever technique—whether it's Nectome on the one side or 21CM on the other? Why don't we have 1 billion people signing up for cryonics? And I would again say to the audience: Sign up for cryonics. Just do it.

This is the way you get to see the 23rd century.

Yes. I heard from the head of Alcor after my last call to action to do cryonics. Apparently, lots of people flooded into Alcor. It's a nonprofit. I make no money off of saying this and have no financial interest. Just get yourself a cryonics plan as part of a portfolio for longevity.

Love it. Love it.

I want to take a second and just say thank you to Nick Singh and Dana Khan, our producers, for supporting us on this new format. I enjoyed it. Did you guys enjoy it?

I love it. It just feels organized.

Yeah, well, it feels organized and fun. It actually feels fun to think through the topics with you guys. It's like an entire episode worth of AMA.

Yeah, with ourselves.

Yeah, it actually changes the stories we cover, too. We normally go through the most important stories to change your life, but here, when you put it into themes, you actually dig up other stories that are related to the primary topic that you otherwise would have missed. So I love that.

All right, here we go. Let's pick 1 each from page 1 and 1 from page 2. Alex, would you go first?

Oh, so many good options. Yes, we'll start with number 2: Where should entrepreneurs actually run their AI compute—local hardware, AWS cloud, or iPhone? And that comes from Frank Gerard Marketing.

There isn't a good answer. There is at least no single good answer—lots of decent answers. There are benefits to each. With local hardware, you have greater control, greater confidentiality, and greater data privacy.

On the other hand, you're going to end up maintaining that local hardware. You have to worry about your own backups, and it can be a pain in the neck from a variety of different perspectives. With AWS or one of the many other public clouds, you don't have to worry about that; it's abstracted away.

On the other hand, you might have to compete ferociously for access to GPU resources. You're competing with other tenants for common resources. You might have to worry, on the margin, depending on how familiar you or your organization are with operational security and cybersecurity, that you have a greater surface for attack.

On the other hand, you have more scalability. With the iPhone, you have the ultimate edge device, until all of us—not just some of us—are running foundation models on our watches and smart glasses, which is already happening and is going to be more evenly distributed. You have even greater privacy.

I don't think this should be viewed as a black-and-white or binary trade-off. There is a spectrum from edge compute to data centers at the core. I think the best answer, actually, is: I want to run my AI compute in the Dyson swarm.

And that Dyson swarm will be a perfect blend, when fully realized in a few years, of data centers. We'll have lots of—maybe, if Elon's statistics are to be believed—100-kilowatt nodes filling the sky. But it'll also be incredibly elastic. If we're disassembling the Moon to fire off new 100-kilowatt nodes into the stellar or Dyson-swarm fabric, it's a perfect blend. Okay, Dave, which one are you going to choose?

One thing on this topic: I spent the whole weekend dorking around on Amazon Bedrock, which is a great choice, by the way. Even though, if my bed was made of rock, it would feel like getting started on Amazon Bedrock. I mean, that's a brutal get-up-and-running process on Bedrock. But it does the critical thing that you need, which is that it captures all of your prompt history, and that of any teammates you have, into easy-to-manage S3 buckets.

So your AI can analyze everything that you've done, which is a critically important function. That may be available elsewhere, too, but it's probably as good a choice as any. Whatever you do, don't just start running on some random hardware and then lose all the prompt history. This is just an easy way to capture it.

Pick a number. I'll take number 3. Are humanoid robots over-engineered? Would it be more efficient to isolate basic needs like food, water, and clothes and automate those directly instead? This is from GoUniteFB3GN. Short answer: yes, absolutely.

So why are we putting all this energy into humanoid robots? The reason we're putting all the energy into humanoid robots is because AI came into the world almost overnight, and we're in a race to capital right now. What's critical for all these projects and startups is getting funded. Humanoids are so much more visually appealing that they're easier to fund. They're also easier to recruit into.

That'll unlock the supply chain of all the parts, and that'll unlock all of the other robots that farm and create clothes and whatever, which will probably not look all that humanoid. But when you look at the Gigafactory, like Peter and I did, the vast majority of the automation there is not humanoid robots. It's machines that look like machines doing their jobs. Yep.

And then the humanoids just operate those machines. So I think they are over-engineered and over-invested relative to where we'll end up, but for a very good reason. You should think about visual appeal and capital raising as a core part of this step function we're living in right now.

So I'm going to go with number 4. How can AI be used to end a war, not as a weapon, but as an impartial negotiator that all parties could trust? This is from JNkind5.

I find that absolutely fascinating, and I do think it's a powerful tool. If you haven't used a large language model for negotiations, one of the things is that we don't know how to think other than the way that we know how to think. Being able to put yourself in the mindset of another individual is extraordinarily powerful.

If you haven't said, “Listen, I'm anti-guns. My neighbor, my friend, my spouse loves guns. Could you please help me explain to them my feelings in a way that lands with them and isn't viewed as offensive?” you can get some extreme support for your negotiation skills.

At the end of the day, I think this is one of the most exciting unexplored applications for AI because the system can also ingest every peace treaty, every negotiation script, and every conflict-resolution framework that's ever existed, and it can model outcomes with no tribal allegiance. One of the biggest challenges we have as humans is that we have these cognitive biases and tribal biases that are driving us.

Can we use this for negotiation? Absolutely. If you set as an objective function that you want to reach a balanced solution for both sides, and both sides are using AI—it could be different models—I think the probability of getting to a solution is much, much higher.

We are biased when we're dealing with humans. One of the things that goes on when you're talking, for example, to an AI model for psychological therapy is that when you realize you're telling your innermost thoughts to a human, you feel like you're going to be judged, but you don't feel judged when you're talking to an AI model. So I think there's real value to be had here. I don't know if Salim's back online or not.

He's not, but if I might add just one thing to your point, Peter: something that I'm seeing more and more in the past few months—not for war, but for commercial negotiation—I'm seeing this all the time. Two parties that are at loggerheads in a commercial negotiation: one of the parties will bring in a frontier model and ask the frontier model what the commercially reasonable outcome is, bring it to the other side, and the other side will consult their model, and they'll come to rapid agreement. Yeah.

I'm seeing this happen now over and over again. All right, here's the next 8 questions from our AMA. Alex, over to you.

Sure. Again, there are so many fun questions here. Rather than choose 8, which would require me to give implicit investment advice, I'll avoid number 7. Number 6 is slightly less interesting. I'll tackle number 5, since I've been beating the drum a bit for solving everything, including disease.

The question is: Once AI solves most diseases, how soon will treatments be available to everyone? Will access lag? This is from Katisse896. Maybe the subquestion first is, when do I think AI has a decent chance of solving most diseases?

My timeline—and this is not specific to me—is that I think if you ask the more optimistic elements of CZI, the Chan Zuckerberg Initiative, Biohub, maybe the Arc Institute, and some other organizations, maybe Anthropic on a good day, I think they'll say something like 5 years from now. So 5 years is a pretty rapid timescale. It's more rapid in many cases than what historically has been the clinical-trial process, end to end, through 3 phases.

The second subquestion here is: How soon will treatments be available to everyone? If, say, tomorrow, one of the frontier labs says, “All right, here are the cures to the top 5,000 unsolved or untreatable diseases. We have vast computational experiments demonstrating, to the satisfaction of all experts, that these are the cures—or at least the treatments—for these diseases,” how soon would those treatments be broadly available?

Under the present regime—which, by the way, is not the same as the regime even 1 year ago—there would still probably be a multiyear process. The FDA has recently announced 2 major developments that are relevant here. First, the FDA under this administration has decided to adopt a Bayesian perspective as opposed to a frequentist perspective, meaning that they're willing to incorporate, for the first time in history, evidence in terms of clinical approvals from outside a particular drug.

That's a huge sea change. It means that, in principle, drug approvals can operate much more quickly because they can take into account lots of preexisting information that predated the particular drug. The second big development is a move, again relatively recently announced by the FDA, from a 2-clinical-trial process to a 1-clinical-trial process for certain cases, expediting the approval process.

Fast-forwarding to the long-winded answer to how soon treatments will be available to everyone, I think if tomorrow or 5 years from now a frontier lab, or multiple frontier labs, said, “Here are the very well-motivated top 5,000 cures to everything,” I think we would see similar developments from the FDA to go to a zero-clinical-trial model, given enough Bayesian evidence and enough computational evidence—which is to say, a zero-trial model. I think there would be so much political pressure that we would probably, barring some exceptional circumstance, see relatively fast availability.

Okay. I'll take number 8, since Alex couldn't touch it and we've lost Salim. If you had to choose one public company to bet on in the age of AI, which one and why? This is from Matthew Johnson 6525.

We can't give investment advice, obviously, but I will tell you—I've said a bunch of times on the pod—go to 13f.info. Look up the Situational Awareness Fund, which is Leopold Aschenbrenner, and every quarter he has to file his holdings. He's killing it, and the reason he's killing it is because he listens to exactly what Alex is always saying: look for the innermost loop.

The tailwinds in equities and assets are like nothing you've ever seen, but you have to be in the AI loop to be relevant. You'll see all of Leopold's holdings are somehow in the centerpiece of the innermost loop. Those are the things you want to own.

Those include chip fabs, things that have power, things related to chip design, and things that are algorithmic and directly driving use cases. That's your road map. Look at his holdings and then generalize from there, and you'll find lots of great stuff that you should be buying.

Also, W-2 income is going to get pummeled in these next 3 years, but assets—holdings, ownership, and things—are going to go through the roof. So buy stuff, whether it's public or private equities, real estate, or things that'll appreciate. That's what you need in this next 3-year window.

Amazing. I am under time pressure myself. I've got to jump on a film recording. I'm going to go to our outro music here, which is brought to us by CJ Trueheart. It was a piece he developed for the Abundance Summit called Moonshot Minds. Take a listen. Love the lyrics. Here we are. [music]

All right, gentlemen.

It gets better and better all the time.

All the time. AWG, DB2 [?], this was fun. Wishing you guys an extraordinary day. Salim, who's airborne to Brazil, safe travels, buddy. Soon we'll be putting you on rocket rides to get you there.

Hyperloop.

Yeah, I guess we can do Hyperloop under the Gulf.

Have a good movie, Peter. Thanks, pal. If you made it to the end of this episode, which you obviously did, I consider you a moonshot mate. Every week my moonshot mates and I spend a lot of energy and time to really deliver you the news that matters. If you're a subscriber, thank you. If you're not a subscriber yet, please consider subscribing so you get the news as it comes out. I also want to invite you to join me on my weekly newsletter called Metatrends. I have a research team. You may not know this, but we spend the entire week looking at the Metatrends that are impacting your family, your company, your industry, your nation. And I put this into a two-minute read every week. If you'd like to get access to the Metatrends newsletter every week, go to dmandis.com/metatrends. That's dmandis.com/metatrends. Thank you again for joining us today. It's a blast for us to put this together every week.