NVIDIA 的 1万亿美元预测、Anthropic 击败 OpenAI、Tesla 对决 TSMC 与 CS 就业崩塌 | 240

This was NVIDIA’s world. It was an absolute madhouse.

OpenClaw is the most popular open-source project in the history of humanity, and it did so in just a few weeks. We’re announcing our support of it.

This is just the extraordinary breadth that NVIDIA and Jensen are imagining. They’re putting their hands, their capabilities, their hardware, and their production into everything. They’re building an ecosystem and then letting everybody do radical innovation at the edges. This is what Microsoft was in the early days, what Google was, but times 100, times 1,000.

Right here where I stand, I see through 2027 at least $1 trillion.

This week was fascinating. We’re going to dive into Sam Altman predicting a 1,000× drop in cost, Anthropic being named the most disruptive company, and eating OpenAI’s lunch. Amazing progress this week. Let’s take a quick look here.

Now, that’s a moonshot, ladies and gentlemen.

Most importantly, Peter, massive congrats. That was such an unbelievable conference. I consistently get amazed by the level that you’re pulling off, and jealous also.

Thanks, pal. It was year 14, the best of all of them, I think. For me, the most important part is getting amazing people there. We had so many CEOs from so many industries, and I loved having all of you on stage with me. I’m pulling you guys in more every year, so get ready.

It was tremendous fun, Peter. Thank you for organizing.

It was great to meet Alex in person and verify that he is indeed at least a meat puppet, if not really.

Well, let’s talk about what we’re talking about, Salim. Did you really meet me in person, or did you meet a clone, a transporter copy, or a bioprinted 3D meat puppet?

It was one of those projections somewhere along the line. But what are we, if not a projection of spirit, anyway?

Yeah, well—

Well, speak for yourself.

You know, Ben Lamm bought the 2 best cloning companies in the world, with the highest efficiency in cloning and success rate. So I put myself in for duplication.

Did you really?

Yeah, well, you just gave him a bunch of DNA while you were with him.

Yeah, I figured 10 would be a good start—1 per company.

Wait, does that mean that we get a Peter burger at the next meal?

Oh, yes. Oh my God, if you haven’t seen the film Project Hail Mary yet, I’ve got my tickets ready. Anyway, let’s not go there right now.

Yeah.

I’m here with my Moonshot mates, DB2. You’re in Boston today, Dave?

I am back in Boston, the first time in 3 weeks.

Amazing. And Alex Wissner-Gross, Alex AWG, how are you doing, pal?

I’m awake. I’m back in Boston as well after the Abundance Summit and then a week in Palm Beach for Palm Beach Forum, so it’s nice to be back.

Beautiful. And Salim in New York, New Jersey?

No, I am in Miami. I’ve had the most insane few days. I flew from the summit straight to an India Today Conclave, which is the biggest magazine publication in India. And guess what I saw? They had 3 people on stage in a row. One was the foreign minister of Iran.

Wow, what?

Then they had Laura Loomer giving somewhat the opposite perspective, shall we say, and then the Israeli ambassador to India.

Oh, come on.

No, I’m serious. You’re like, “Whoa.” Do not let those 3 in a room together.

Wait, are you going to go on your ride?

I was a separate speaker afterward. It was pretty out there.

What the hell?

Then my hell started, because I flew from India through Heathrow to Miami, to a Caribbean country that shall not be named. I walked into the immigration hall of this Caribbean country—600 people in the immigration hall and 2 people working the passport counter. You’re like, “This cannot be happening” during prime spring holidays. There wasn’t a hotel room to be had on the entire island, and there were 2 people working immigration. The place is going to have to get its act together.

I can’t believe that he said anything from Moonshots.com.

Yes, I was hiding in the bathroom.

Okay, welcome to moonshots.com. I'm Peter Diamandis, your host and this is your number one podcast for AI and exponential tech. Our mission here, get you future ready, get you ready for the supersonic tsunami coming our way. As always, we've got a full presentation on what just happened last 3 days. I missed you guys. It's only been about 5-6 days and— It feels like a month. Separation anxiety, huh? It feels like a month. It really does. We're going to start here, but I had a conversation this morning with two of our listeners who are in Hollywood here. They're Hollywood producers. Jonas and Josh Pate, and they said, moonshots is the light against the darkness. They give us hope. There's so much doom and gloom out there about the future and the show gives us hope. So Jonas and Josh, thanks for that word and everybody please join us. We're almost at 500,000 subscribers. Push us over the top. If you haven't subscribed, please do. Turn on notifications. We're now publishing almost twice a week. And our mission is to deliver you everything that's important. You know, this is the news that really matters in the world. Okay, we're going to jump into the top AI news. NVIDIA, OpenAI, Anthropic, and xAI are in the news this week, and this was NVIDIA’s world. GTC 2026 had 30,000 attendees. The opening keynote by Jensen actually took place at SAP Center in San Jose because they couldn’t hold it in the convention center. There were 2,000 speakers and 1,000 sessions.

I dropped in for about an hour, and it was an absolute madhouse. Our first video is Jensen talking about reaching $1 trillion by 2027 in revenue, not in valuation. They blew past that a long time ago in revenue. This event, this photo of 30,000 people in the room, that’s what $1 trillion of revenue looks like when the whole world ends up coming to you.

I think the most important thing to realize is that Jensen is looking to power everything. He’s looking to own the infrastructure that runs physical AI, that runs data centers in space, and that runs even OpenCL. Let’s check it out. I’m going to share a couple of videos, and then let’s chat about it.

The operating system for robots, cars, agents, and orbit. I’ve been holding off buying more NVIDIA stock because how much higher could it go? Well, we’re going to find out. NVIDIA, let’s listen.

Right here where I stand, I see through 2027 at least $1 trillion.

Woo!

Dave, a trillion bucks?

It’s not really a trillion. It’s a trillion dollars of bookings that has to be recognized over the life of the bookings. Also, that’s spread across 2 years. Amazon or Anthropic is going to get to a trillion within a calendar year faster than NVIDIA.

NVIDIA would get there faster if they could get the chips made, because the demand is there, but TSMC is the bottleneck. They’ve already locked up 70% of TSMC’s volume of the 3-nanometer node. There’s just nowhere to go. They can go up in price, I guess, and sell for a higher ticket, but you just can’t make more chips. We don’t have the fabs.

And if NVIDIA solves it—if NVIDIA can actually lock up $1 trillion in revenue from selling its hardware—how much negotiating power do any of its customers have? None.

None. Yes, that’s dead right. People are begging. It’s funny—you heard Larry Ellison say this as long as a year ago. “We are literally—me and Elon and Sam—lined up outside his door begging for the chips.” When in the history of sales has the customer come to you begging for the product?

It makes you wonder, why doesn’t he charge more for it? But he’s already at 80% gross margin. It would start to get kind of egregious.

Yeah. We’re going to see Elon’s Terafab in a little bit to compete against all this. But let’s go to the next video here. Let’s listen to Jensen talk about OpenClaw. What a phenomenon.

Peter Steinberger is here, and he wrote a piece of software. It’s called OpenClaw. I don’t know if he realized how successful it was going to be, but the importance is profound. OpenClaw is the most popular—

I’m going to pause it here. Look at that red line, that red vertical line. The yellow line here is Facebook, and that blue line is Linux. We’re talking about incredible growth for both of them over the last decade, and then here comes OpenClaw—vertical.

Do you remember when ChatGPT came out and you were like, “Oh my God, 1 million users in 100 days”? What could possibly scale faster? Well, here we go. I’ll continue on.

Open-source project in the history of—

Yeah, please.

Oh, the—so John Werner had a meeting with the very high-ranking MIT administration, and he had some kind of a lobster garment on. It was the tie that I gave him—the lobster tie. I won’t tell you who it was, because it’s just too embarrassing.

But they're like, “What's the significance of the lobster?” They're like, “What? Are you—what?”

“Where have you been?”

“OpenClaw is the biggest cultural phenomenon in probably world history.”

Well, of course, in Boston, a lobster has a different significance.

Yes, but I—oh my God.

What I thought when I saw this was, “This is the classic exponential.” I mean, we're past that now, but it looks vertical in front of you and boring and dull behind you, right?

Yeah, exactly.

We'll look at this and go, “Yeah, yesterday, something else came.”

It is unreal. As Alex is always saying, “Don't sleep through the singularity.” We take it for granted, but it's only been a few weeks.

Yeah.

It would be very easy to have been on a Caribbean island accidentally for a little while and have no idea what's going on.

All right.

I'll jump back into Jensen here on OpenClaw.

…of humanity, and it did so in just a few weeks. It exceeded what Linux did in 30 years. And it's that important. It is that important. It will do well. This is all you do. Okay, we're announcing our support of it.

They're announcing NVIDIA's NemoClaw. AWG, what are your thoughts?

I think it's inevitable that NVIDIA would need to play in this space. I agree with the premise that OpenClaw is probably, as Jensen and others have characterized it, the biggest thing in AI, at least in terms of unhobblings, since ChatGPT unhobbled GPT-3 in 2022.

I think in some sense it was also inevitable that the next big unhobbling after ChatGPT would grow far more quickly in terms of adoption than ChatGPT itself, because all of these unhobblings are technically stacked on top of each other.

OpenClaw, in the sense that it's sort of a 24/7, headless—other than via messaging and other mechanisms—agent, builds on top of everything else that we've built on thus far in the tech stack. It builds on reasoning models; it builds on large language models underneath.

And so, in some sense, as we build further up the stack in terms of more and more advanced unhobblings for these AI capabilities, I do expect the growth pattern to shorten even further, to the point where maybe in a few months or, at maximum, 2 years, we'll be having a similar discussion.

History will rhyme, and we'll see, “Oh, this new repo from 2027 went from zero to a billion stars in 5 minutes,” and we'll have the same Jensen quote, probably, and say, “All right, we're going all in.”

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. NVIDIA is trying to optimize for enterprise on its full stack with NemoClaw. The question is, if they really optimize it and they're selling it packaged with NVIDIA, is it really truly open infrastructure? We're going to start to see a lot of people trying to capture that. Of course, we saw last week that Anthropic is effectively duplicating all the elements of OpenClaw.

So this is going to go everywhere, and I wonder how fast it's going to hit. Salim, have you started your lobster yet?

I have strong thoughts here.

Please.

I think what you said, Peter, is exactly right. We're going to see a flurry of these. We're also seeing a bunch of Perplexity Computer and, on the other end of the stack, you've got PicoClaw.

But this announcement from NVIDIA, for me, was a monster implication, because one of the things about OpenClaw was that you have recursive self-improvement in business workflows. This is the heart of this paper I've been writing, which I think will be ready next week, called The Organizational Singularity.

Once you have that, all human-to-human workflows essentially evaporate. You can't sustain them. You can't compete, because once you put a workflow into this set of agents, then they're optimizing it by themselves. You want to get the humans out of the way as fast as possible.

The fact that this is announced to support enterprises solves the OpenClaw security issues and the danger that comes with it. This will be the biggest thing to hit the enterprise world in decades. I think we're going to see adoption by enterprises at that scale.

I've had CEOs calling me saying, “Please, because my thesis is you can't fix any existing organization because it's inherently human-to-human.” All the AI projects are failing. Corporate AI projects are failing because they're trying to optimize human-to-human.

And we're inherently flawed anyway in terms of latency, jealousy, and time taken. Period. Send me an email; you never know if I'm going to respond. All that frailty comes into the human being.

Whereas if you go agent-to-agent and improve the workflows recursively, this is the flywheel. So every organization in the world now has to do one thing to survive—one and only—which is, at the edge of your organization, create an AI-native operating system that's AI-centric and start moving workflows over to it.

Human beings then become oversight, exception handling, and monitoring the overall system. This is going to have to happen now. We're going to see—I may have to batch up CEOs and say, “We'll take you in batches through this process.”

But it's going to have to be every company, every nonprofit, every government department—every single organization in the world. The implications for government are profound, because government is mostly prescriptive processes: passport renewals and so on.

This is going to be incredible. I think the value this will add to the world is going to be near-infinite from where we can see it today to where it will get to.

Salim, that is so dead right. Last night, I opened an Amazon Bedrock account because you can run OpenClaw there now, which is brilliant by AWS, but I swear to God it was less than 10 minutes. You just go to aws.amazon.com. If you have a credit card and you're a human being, you can be up and running with OpenClaw inside a secure environment.

And I think the reason this is such a big unlock for business is because when we started rolling this out at Vestmark, the first problem we ran into is that a lot of people were trying to automate their job tasks through email, Slack, and other messaging systems, and there was no way to connect Claude to it.

So cutting and pasting all the crap out of email to get it over to Claude or OpenAI was such a pain in the butt. Now, with OpenClaw running in a secure environment on AWS, it's instantly connected to all of our enterprise emails and everything else.

So for regular business, the coders have known this when they're writing code for a long time—well, like a year—but the rest of the business hasn't been able to really tap into AI. Now, because of OpenClaw, it's just trivially easy.

And it's still early. It's still early.

I'm just noting the date. March 16 is the date of the Organizational Singularity. Everything changes.

Make it St. Patrick's Day. It's easier to remember.

Okay, St. Patrick's Day. That's even easier. Yes.

We're going to hit 2 more videos and talk about it.

Again, this is just the extraordinary breadth of what NVIDIA and Jensen are imagining. They're putting their hands, their capabilities, their hardware, and their production into everything. I mean, multiple trillion-dollar avenues.

What's really powerful for NVIDIA here—sorry to interrupt you—is that they're building an ecosystem and then letting everybody do rapid innovation at the edges, right? And that's how successful ecosystems work. This is the coral reef biology we've used in the past.

And let's get into physical AI. Eric Schmidt was discussing this at the Abundance Summit. Let's take a listen.

We also have been working on physically embodied agents for a long time. We call them robots, and the AIs that they need are physical AIs. We have some big announcements here. I'm going to just walk through a few of them. 110 robots here.

Almost every single company in the world—I can't think of one that is building robots—is working with NVIDIA. We even have T-Mobile here. And the reason for that is, in the future, that radio tower, which used to be a radio tower, is going to be an NVIDIA Aerial AI-RAN.

Today, we are announcing 4 new partners for NVIDIA's robotaxi-ready platform: BYD, Hyundai, Nissan, and Geely. All together, 18 million cars are built each year. Joining our partners from before—Mercedes, Toyota, and GM—the number of robotaxi-ready cars in the future is going to be incredible.

And we're also announcing a big partnership with Uber. Multiple cities—we're going to be deploying and connecting these robotaxi-ready vehicles into their network.

Insane, all right? This is what Microsoft was in the early days, what Google was, but times 100, times 1,000: NVIDIA inside everything.

My question to you guys is, how long before regulators frame NVIDIA as critical infrastructure and start treating it more like a utility? How long before they start being seen as having too much power by all of their customers and by government?

They're dusting off their antitrust legislation as we speak.

No, I think the criticality of NVIDIA compute is already highly export-controlled. It may get even more export-controlled over the coming months, so I think this is already highly regulated.

What I see coming out of GTC this year is, in some sense, the Western response to China's AI Plus Five-Year Plan. In China, the Chinese Communist Party has a five-year plan to infuse AI into the rest of the industrial ecology. We don't really have quite the same industrial policy in the West or in the U.S., although we do have an increasingly aggressive industrial policy.

The question I'm asking is: NVIDIA is getting so much power, and it's so fundamental to every single layer of the stack—all physical AI, all data centers, everywhere. Besides antitrust, they have so much power that they can make or break companies. They're kingmakers if they want to be.

How much longer before their customers say, “You have too much control over us, and we need to compete?”

Very, very specifically, you're dead right, Peter. The choke point is going to be the conversation with TSMC, and also Intel and Samsung. I could literally design a chip and sell it out. It's not hard to sell chips in the age of AI if you can get them made.

So what Jensen is doing right now is locking up the future manufacturing of TSMC as far into the future as they'll let him. What do you mean by “they'll let him”? It's the government that will ultimately say it's anticompetitive to do a 10-year forward contract on all of their manufacturing capacity. That's where the rubber's going to hit the road. His margins are so high.

Yeah, bordering on where the government will intervene. If you start locking up all future manufacturing using your current leverage, that's where they're going to come, and it's going to collide. You'll see that with Elon's plans, too.

That's right. There'll be severe competition coming. He's trying to lock everything up before that all arrives.

Yeah. I want to question the premise, though, Peter. I don't think the clip that we just played, which is Jensen demonstrating the pervasiveness of NVIDIA compute into an industrial ecology, is intrinsically anticompetitive. This is exactly what I would argue NVIDIA should be doing.

If you want to look for anticompetitive behavior, then I would scrutinize perhaps the Groq acquisition or activities like that. I don't think the pervasiveness in robots and robotaxis is a bad sign. This is an incredibly good sign for the West, I would argue, again going back to what the CCP is doing with its five-year plan.

Well, the Groq acquisition is exactly where the rubber hits the road. The NVIDIA story on Groq is, “Look, they've got a better inference-time design. We want to acquire them.”

The Groq point of view is, “We can't get this made unless we get the TSMC 3-nanometer and 2-nanometer capacity, and Jensen locked it up. So we have to sell to him.” But he gave us a great price tag, so it's all good.

So here's my question, Dave, to you. What do you imagine we're going to see in terms of NVIDIA's revenues? Is he going to hit $1 trillion? Is he going to continue to climb? Is there no ceiling on this?

There's no ceiling whatsoever. It'll be $350 billion this calendar year, and it'll grow at the maximum possible rate that he can get TSMC capacity. So he can grow another 2x into the 2-nanometer node, and then he's floored.

A lot of the growth—he had about 20% market share with TSMC when this all started, and now he's up to a lot more than that. So he had a lot of really fast growth. But from here on out, it's all gated by how quickly we can build new fabs.

There's a lot of investment and research going into new fab designs and new types, which are all bottlenecked by ASML machines.

Sure.

Those are absolutely worth tracking. Every time an ASML machine gets shipped, it's worth tracking who got it and where it's going, because they're going to print money with it. It's literally—you bought a printing press.

Well, there's a lot of talk right now that Elon is secretly negotiating with Intel because Intel has a lot of those that they booked years ago, and they're underutilized relative to NVIDIA.

When we were podcasting with Elon, we said, “Are you going to buy Intel?” That was our guess back then.

He didn't say no. He kind of looked around the room.

It'd be such an obvious thing to do.

One more video. Again, just the breadth of Jensen's vision and how he's jumping in—I mean, every place there's an opportunity, NVIDIA is jumping in strong. Let's take a listen.

We're working with our partners on a new computer called Vera Rubin Space-1, and it's going to go out to space and start data centers out in space. Now, of course, in space, there's no conduction, there's no convection, there's just radiation. So we have to figure out how to cool these systems out in space.

I still love the fact that no one was discussing this 7–8 months ago. Elon states, “We're going to do this,” and the entire world is converging to implement his vision and his dream. Crazy.

What's funny to me is that I had so many meetings 3 months ago with semiconductor companies that were all excited about liquid cooling. They had different etched grooves on the backs of the chips, and the water was going to flow through them.

They're all dead silent, and all of a sudden they're like, “Yeah, that won't work in space. Sorry.”

Oh, my God. Well, that's the nature of the singularity. Things are going to change every month, and I guess we just have to get used to it.

I do find it somewhat surprising that NVIDIA hasn't been working on orbital, radiation-based cooling for years. It is pretty surprising.

On the one hand, one can say, “Well, the Dyson swarm snuck up on us, and this was a very surprising killer app for GPU computing, frankly, for the solar system.” On the other hand, it is surprising to me that, given how many tendrils NVIDIA has into so many different verticals, they weren't investing more earlier in space-based cooling.

I get this question all the time—the common misconception that radiation-based cooling in space is somehow very difficult or challenging, or somehow an obstacle to scaling out. It's actually not that difficult. Despite Jensen's comments that he has dozens of engineers, I think was the quote, working on radiation-based cooling for orbital data centers or orbital GPUs, that's an optimization. We know how to cool orbital compute right now.

You know, my concern, Alex, is more a massive solar flare or an EMP—something like that, in terms of warfare—that could knock out a significant portion of our data centers up there.

We know how to do that, too. We know how to design radiation-hardened electronics in orbit. We've been doing it for decades. There are various techniques: you can use older process nodes, extensive error correction, or shielding.

There are lots of different techniques. My favorite is the Star Trek technique of just using a magnetic field to deflect ions and otherwise deflect ionizing radiation around. It doesn't work for photons, obviously; it only works for charged particles. But I think we're going to come up with lots of solutions for protecting orbital compute.

In the long run, yes. I think the short run is what I'm concerned about.

Aren't there significant latency issues with space-based stuff?

Not if they're in low Earth orbit. Low Earth orbit is very low latency.

Yeah, I mean, it's Starlink on your phone very shortly. Again, you're putting the prompt up into space and getting the answer beamed down to you. You could probably parse the latency request to different parts of the constellation.

Alex said something brilliant, as always, and kind of very quickly there that I want to rewind the tape to. There's a huge opportunity in older process nodes. Just file that away if you're a listener wondering what you're going to do post-singularity.

Partially because they're resistant to radiation in space, but partially because it's underutilized capacity in an age where all AI will sell out. We can riff on that some other time, but I want to call it out because Alex says these things so quickly. It's incredibly profound what he just said.

Maybe one more teaser just on the latency front: people are sleeping on neutrinos. I'll make a prediction. Just like some folks were surprised by the Dyson swarm orbital data centers, people are sleeping on the potential for neutrino-based communication to give us ultra-low-latency communication through the Earth.

Right now, we don't have great technologies for producing neutrinos in a way that's high-throughput, or for receiving neutrinos.

The neutrino detectors are in mines, or thousands of meters below the ground, and in large—what's the thing they use for tracking neutrinos? Is it just pool water?

Heavy water. It's usually heavy water. That's right. But there's no physical reason, from the physics that we have today, why it has to be so inefficient to couple via the electroweak force to neutrinos.

So one can imagine that in a few years, when we have better physics, we’ll have neutrino phones that go straight through the Earth, and then we can completely route around it.

Neutrino phones.

Neutrino phones. It’s awesome. All right.

That was also not on my bingo card. For any discussions you’re having—come on.

This week was fascinating. We’re going to dive into this: Sam Altman predicting a 1,000x drop in cost, Claude writing 70% to 90% of its own code, Anthropic being named the most disruptive company, and literally Claude—or Anthropic—eating OpenAI’s lunch. Amazing progress this week. Let’s take a quick look here.

First up, Sam Altman talking about his speed and cost. All right, Sam.

People cite whatever amazing statistic they like about how much more efficient our industry’s models have gotten over time, but one that I think is incredible is this: Our first reasoning model was called o1 and came out about 16 months ago. Our latest model, with integrated reasoning, is 5.4. To get the same answer to a hard problem, going from that first model to 5.4 has meant a reduction in cost of about 1,000x.

Hmm. Do you believe that number, Alex?

I do. It’s consistent with the 40x year-over-year hyperdeflation that we’ve discussed on the pod previously, but I want to highlight the implicit part that Sam is mentioning. He’s highlighting the difference between o1, which is OpenAI’s first reasoning model, and GPT-5.4, which is their latest reasoning model. He’s not highlighting, say, differences with models prior to the reasoning-model revolution.

We’ve seen, if you look at the AI capabilities since reasoning models were first introduced, that the hyperdeflation of cost has been extraordinary. It’s reasoning models that are enabling this massive increase in capabilities. It’s not necessarily training-time compute. We see the shift to inference-time compute, or action-time compute, and that’s really enabling this 1,000x increase in capability per unit price.

I would expect that—and I think maybe we’ll touch on this in a moment—as we hit recursive self-improvement more and more aggressively, we’re going to see this order-of-magnitude increase in capability per unit price fall out for free, in some sense, just like reasoning models in some sense fell out for free once you had the baseline of large language models and simply allowed them to talk to themselves with additional tokens and reasoning time. Then you could, through iterated amplification and distillation, enable them to reason more effectively.

We’re going to see post-transformer architectures that make a 1,000x reduction in cost look like child’s play.

Yeah, and if you rewind the video here to Jensen’s comments and look really closely at what he has on screen in the corner, you’ll see him talk about the inference explosion as driving this trillion dollars of bookings at NVIDIA, which is the exact same thing Alex is talking about.

For all of neural-network research history going back 40 years, nobody cared about inference time because training was the bottleneck, and the model wasn’t smart enough to care about making an inference really fast. Now chain-of-thought reasoning is the biggest breakthrough ever, and you can use inference-time compute to build a smarter and smarter AI.

It’s very easy to optimize inference relative to training, and we as a society have really just started on it in the last 2 years or less. That’s why we’re getting these massive gains, but that’s also why everyone’s really underestimating the next year. You look at it and say, “We did 1,000x. What are you expecting in the next year?” “Oh, 2x.” What are you talking about?

Yeah.

It’s not going to happen that way.

Dave, you just said something there. Can I drill in on that? Why is inference so much easier to optimize than training?

Well, uh—

Yeah, go ahead. Alex, nail it.

I was going to say part of it is, frankly, that there’s an overhang. Prior to o1 and the reasoning-model revolution, almost no effort was being spent on scaling inference-time compute.

If you go from having zero tokens expended in reasoning to thousands of tokens expended in reasoning, in some sense you get performance for free out of that—at least capability per unit price—because you were expending so little cost. If you look at the overall pie of how much compute was spent on training time versus inference time, so little compute was being spent on inference time that you could scale across orders of magnitude the amount of time you spend on inference without materially impacting your overall budget.

You can get orders of magnitude of effective cost reduction per unit capability for free just through brute-force scaling of inference time. Now, at some point—and we’re reaching that point now—you run out of room, and inference-time compute starts to dominate the overall pie. There are many frontier models now where more inference-time compute is being spent than training-time compute, for some definition of each of those quantities. At that point, the free lunch runs out, and then you have to start discovering new efficiencies.

I want to make the abundance argument here for folks listening. A 1,000x reduction in cost in 16 to 18 months—we are nowhere near optimized for inference computing, energy, or cost. Six billion people with a smartphone means that effectively there’s going to be some level of extraordinary AI available to every single person on the planet.

Again, what makes us special as humans? We’re not the fastest. We’re not the strongest. We’re potentially, hopefully, the smartest, and we’re delivering intelligence as a service to everybody. Intelligence as a service gives everybody access to education, health care, entertainment, and re-education for employment. Go on, Salim.

Can I make a radical comment here? I’ll make a prediction that the optimization we’re doing—which is 1,000x in 16 months—is going to keep going in such a way that we may not need to tile the world with data centers or energy.

I think your point until now has been that the demand for compute is so ridiculous—maybe it’s 10 million times—and we’ll need every single joule of it. But if the optimization is happening that crazily, for example, once you can run OpenClaw locally and run models locally, the compute and energy needed is really quite minimal for that. So do we really need the massive energy buildout?

Yeah, John Werner calls that the WALL-E—you know, the other Disney movie where they’re all blobs floating in space and they’ve kind of forgotten to innovate because everything’s so easy. Which movie is it?

It is WALL-E.

Is it WALL-E?

Yes.

Yeah, I mean, that would imply that I go to you, Salim, and I say, “Hey, Salim, I can give you 1 billion employees with an IQ of 180 each. Can you think of anything useful to build?” And you go, “Nah, I can’t really use it.”

The point you’re making is that we’ll find so many ridiculous use cases for putting intelligence in every little sensor in the world.

Remember, 6 billion people have access to a smartphone. How many people are using AI right now? OpenAI’s at 800 million, pushing 900 million.

The elephant in the room with the consumer case of putting a country of geniuses—not in a data center, but in your smartphone in your pocket—is that, thus far, consumers, and OpenAI has sort of been shocked by this, haven’t made good use of reasoning capabilities. Whereas enterprises are thrilled with reasoning capabilities.

If we want to empower individuals in the world, we need to discover a killer app for individuals to use reasoning other than—

To change their mindset. I’ll say this again right here: Everybody listening, your job is to use AI every day.

We had Bill Gross on stage at the Abundance Summit basically saying we have to retrain ourselves because all of us have learned that if you have to do something, you have to do it, or you have to find an employer to do it for you. The judo move here—the new mindset wiring—is: I need something done. I bet AI can do it for me better than I can and better than a human.

I have a killer app. For God’s sake, any little bit of common sense is a killer app. We need more of that around the world.

You get common sense from baseline large-language-model capabilities, I think. I’m not even sure you need a lot of reasoning for that.

So let’s just force every human being, before we make any stupid decisions, to check with their common-sense app before they do dumb things.

Well, that’s called the next step.

The next step is likelier to end up being something like turning every individual into an enterprise that actually needs reasoning capability. I think there’s probably a trillion-dollar company to be built turning every individual in the world into one-person unicorns.

Let’s jump into Sam Altman talking about AI reinventing itself.

From a research perspective, I bet there is another new architecture to find that’s going to be as big a gain as transformers were over LSTMs. I think you finally have models that are smart enough to help do that kind of research.

I would go look for where I can find a mega-breakthrough, and I would use the models to help me.

We had Kevin Weil on stage, who's heading science at OpenAI. We talked about the fact that what's hidden inside these hyperscalers, inside these frontier labs, is that they're going to use AI to create incredible breakthroughs in physics, chemistry, biology, materials science, and AI itself. Each of those is a multi-trillion-dollar opportunity in itself. Alex, thoughts?

Well, maybe just to speak at the object level to Sam's comments about a leap from LSTM to transformer, and from transformer to something after transformer that's comparable, I think that's very likely. He may even be gesturing at something that OpenAI has internally. I'm sure. I want to combat the perception that a hypothetical post-transformer architecture necessarily will involve a recurrent architecture.

There are a lot of companies that were founded on the premise that just because transformers superficially have an attention bottleneck and a context-window bottleneck, and seem to have plateaued at about 1 million or so tokens of context, somehow going backward in time to recurrent architectures like LSTMs—which were a form of recurrent architecture—is somehow the solution.

If I had to guess what the definitive category-killer post-transformer architecture looks like, I think it's going to come from left field. I think it's going to come maybe from a line of research—there are a few lines of research—that involve using transformers to directly write the weights of other transformer architectures.

Transformers are really wonderful. They parallelize nicely. They have nice residual streams. There's a lot to like about the transformer architecture. I think it could involve some refactorization of the weights. It's going to be something clever and not just a return to recurrent weights.

It's not brute force. It's going to be something orthogonal to that.

It can't be brute force if it's going to be a fundamentally new architecture. Brute force is just what we do if we don't have transformative new architectures.

But I would encourage everyone who gets excited: there's an entire cottage industry of academic researchers who don't necessarily have access to the raw brute-force compute of a frontier lab who want to be the ones to discover the next transformer. I would encourage you, if you're listening, to focus your attention on the small-language-model space.

There are so many lovely benchmarks, like the speed run for nanoGPT training or a variety of slow-run benchmarks for data efficiency. Focus on those and discover the next big thing. It probably won't be recurrent networks.

And it probably will—it'll be soon, like in the next year. Yeah. And it probably won't map well to the current NVIDIA architecture. You'll immediately want to call Lisa Su at AMD or call Intel and figure out how you're going to get it manufactured on custom silicon, very much like Google is doing with the TPUs or Elon is about to do. That's how you're going to create the next Anthropic or the next OpenAI. That's right.

And that was, quite frankly, what NVIDIA did to Intel. Intel, for years—if you remember, why was it that Intel, which was the 800-pound gorilla, allowed the GPU revolution to just pass it by? It was because, for years and years, Intel executives were trying, naively, to map what they perceived internally as a general-purpose CPU onto GPU-shaped problems, and that always ended up being a bad idea.

They had all these schemes to create tiled architectures of hundreds of CPU cores to solve GPU-shaped problems, but they were unwilling or unable to focus on specialized compute for specialized problem shapes. If there's going to be an architectural disruptor for NVIDIA, it's going to be the same sort of disruptive innovation that NVIDIA pulled on Intel. It's going to have to be, I would expect, an architecture that's even more specialized than GPUs and yet even more useful.

And I'll tell you what Jensen knows is coming: that's why he's trying to lock up all the manufacturing, so that you have to come through him rather than around him.

Yeah.

But what Alex said earlier about older process nodes being viable is brilliant, because if you decide you don't want to sell to Jensen, that's your avenue forward. You just very quietly use the older process nodes, work around it, and talk to Intel or AMD.

So GPUs were Intel's Kodak moment.

Mhm. In one sense.

Sure. I mean, in a sort of Christensen-esque disruptive innovation, but in a very specialized form, yes.

Mhm.

All right, the frontier lab wars continue. TIME magazine names Anthropic the most disruptive company in the world. We're seeing this, and one point, Dave, you made a while ago was that as OpenAI is making improvements, they're dropping their cost at the same time as Anthropic is making improvements and increasing their performance, which is leaning it toward the enterprise level, and they're winning hands down. We're going to see that in the next slide. Let me just share that one right now. Here we go: Anthropic is eating OpenAI's lunch.

This is AI model share of first-time enterprise customers, and we've seen Anthropic go from 40% up to 73%, while OpenAI goes from 60% down to 26% over 3 months. This is insane.

Yeah, you know, the story within the story that I'm really tracking closely here is that Sam Altman is the consummate dealmaker—Y Combinator background, traveling all over. I see him everywhere, negotiating multi-hundred-billion-dollar deals on every corner, while Greg Brockman and Mark Chen are back in the office being the brilliant AI researchers.

Dario is completely the inversion of that. He's the actual AI researcher, understands every bit moving through the neural net, while his wife is dealing with the business stuff. So they've kind of inverted the formula, and it's interesting to watch it play out, because when Mark Zuckerberg came into the business world as a 23-year-old or 22-year-old, him running a monster company was completely foreign terrain to everyone, and they're like, “Can this kid really figure it out?” He reinvented what an internet CEO looks like.

Now Dario is reinventing what an AI CEO looks like. If he ends up winning in the end, the profile of what a CEO looks like will have changed yet again.

On the other hand, if Sam comes roaring back, it'll be interesting. It's a great drama.

You know, Dario doesn't want to be the CEO. I bet you Dario wants—

He—

Definitely didn't originally. I know that. I don't know if he's grown into it.

Look, my wife Lily is a way better businessperson than I am, so this is not surprising to me at all. What I found really interesting about this whole framing was that these frontier labs are the weirdest animal because they're part software company, part national-security issue, and part huge governance experiment. I mean, this is really a weird animal that we've not seen before.

This is an absolute ass-kicking, though. I mean, I just want to call that out. This is a total ass-kicking.

Yep. Yep.

Well, enterprise buyers reward fit, stability, reliability, and trust. You really want that, and Anthropic is providing it.

I think part of the problem is that OpenAI had made a bet—and if you look at the timescale, I think the timescale agrees with this—that consumers would need a lot of compute. Anthropic, with fewer resources and less compute than OpenAI, was forced to focus on enterprise and then has post hoc turned that into a story of how enterprise is intrinsically better as a customer base than consumers.

I don't think enterprise is intrinsically better, but I do think enterprise appears to be intrinsically hungrier for compute in the form of inference-time reasoning.

Enterprise—their survival is at stake here. A human consumer paying $20 a month, it's not their survival. It's useful; it can help them do their stuff, but for enterprises, I mean, they're willing to pay whatever it is. It's an existential risk for them.

That sounds more like agreement than disagreement to me.

Sounds like you agree with me, Peter. Oh, I thought you said the opposite.

No, no. I'm saying that OpenAI had made a bet that consumers would be as hungry for reasoning compute as enterprises, and that bet turned out to be wrong. I don't think we have a slide for this, but OpenAI has actually started to scale back their Stargate plans. They're switching from building their own data centers to renting existing data centers, and this has been very well publicized. They're throttling back on their $1.6 trillion Stargate plans. I don't think we'd be in this situation if OpenAI's bet had turned out to be correct. I wouldn't expect to see this crossover at all. I'd expect to see OpenAI generate so much revenue that they wouldn't have had to—

Yeah.

Yeah, I don't think we'd be in this situation if consumers had been as avid consumers of reasoning tokens as enterprises were. That bet was wrong. Anthropic bet on enterprise because they had to, because they were limited. As a result, you see Anthropic's enterprise business going up, and Anthropic overall being in a position where revenue generation is 10× in a year.

Right? They're heading toward an IPO, and they are basically shortchanging Microsoft and going for $50 billion from Amazon. Shortchanging is not the right word. Stabbing Microsoft in the back, just trying to get deals to make sure their IPO comes off and they get enough capital to continue building. We hadn't had one of the slides for this either, but it's worth noting: this past week, Meta's Avocado model is getting massively delayed. And how cool that Meta is now looking to Google to provide them AI capability in the interim. Thoughts on that one?

The singularity makes for strange bedfellows.

Yeah.

Well, desperation, too. I don't know if you're a World War II history buff, but the Nazis just tromped across Europe so easily. And then, in some fit of insanity, Hitler decided, "You know what? I'm just going to go and declare war on Russia at the same time. Start a whole new front, even though they've agreed not to attack me, and I've agreed not to attack them. I'm going to go ahead and fight in the snow." They got stretched too thin, and that was the end of that. Now we live in the world we live in, thank God. So Sam decided, "I'm going to go ahead and start working on a chip design of my own while hiring Jony Ive to go headlong after Google on the device front."

Mm-hmm.

Sam said, "I'm going to focus—

On the device front.

Headlong after Apple, too, yeah. You know, irritate everybody at the same time. Why not?

Yes.

But we're that big, and we've got that much momentum, and we've got a trillion-dollar valuation, so we can pull this all off simultaneously. Very much like Elon: "I want to build this totally integrated, end-to-end empire." Dario went whole hog the other way: "I want to partner with Amazon and AWS. I want to be friendly with every cloud provider. I just want to do the software. I'm not designing my own chips. I'm not building my own data centers. I'm partnering with everyone. I am just the AI software." And so, he's very easy to partner with because he's not a threat to everyone. That's a big part of why this is working out this way.

I remember talking to some of my friends who are senior at Google, and they have a respect for Anthropic, right? Anthropic is the other—I'll put air quotes around it—moral and ethical frontier lab out there. One thing for entrepreneurs listening: one of the biggest mistakes entrepreneurs make is pursuing too many lines of business. Most companies fail not from starvation, but from indigestion.

Guilty as charged. Me, me, me—I'm terrible.

Okay, okay, yeah. And it's so true. You get some level of success, and you get ambitious, and you start going after the next thing, the next thing, the next thing, and pretty soon you've forgotten what made you successful in the first place.

You sound like my whole board, Peter. I'd like you to stop now.

Okay. I will stop. I will. But the only person who's been immune to that is Elon.

Yeah. And you look, I don't think my World War II analogy paints one side versus the other, but it's easy in hindsight to say you got stretched too thin. The flip side of it is, if you pull it off, you're vertically integrated, you've got a massive hardware advantage, and you control your own data centers. You'll be like Elon. So there's merit to both approaches. It's not obvious until you're stretched too thin and—

I also think it's far too soon to be writing epitaphs for OpenAI. GPT-5.4 Pro is an incredibly strong model. Codex, their competitor for Claude Code, is growing very rapidly. And I'm confident that OpenAI has the institutional wherewithal to refocus itself. They've been in the headlines saying they need to focus on their core bread-and-butter businesses at this point and look a little bit more like Anthropic. I think they've been scared into focusing quite a bit more. And the beauty of OpenAI is that they do have that vertical integration, where they had been focusing earlier on data centers and on the consumer.

I do think at some point consumers will actually discover use cases. Maybe they look like OpenClaw—consumers needing a lot of reasoning at the consumer end. And at that point, I would expect OpenAI's AI strategy of being—I think Sam calls it the core AI subscription, which I parse as being everything to everyone—I do think that will have another day in the sun. Yeah, and also, I don't think "epitaph" is ever the right word. All 5 of the major labs are going to be worth trillions and trillions of dollars. I think we said that onstage many times last week.

To say somebody's beating the crap out of somebody else doesn't mean the other guy isn't growing, too. It means you're just on top of the pyramid right now, but they're all growing. This is America. We need competitors in every space. We're not going to have 1 winner. We never operate that way. They're all growing, they're all thriving, and they're all going to be true multitrillion-dollar companies—the biggest companies you've ever seen.

Everybody, welcome to the health section of Moonshots brought to you by Fountain Life. You know, we talk about AI on this Moonshot podcast all the time. One of the most important things AI is going to be able to do for you besides educating your kids and helping you with your taxes is making sure that you're living a healthy lifestyle, that you get a chance to get to 100-plus. I'm here today with Dr. Don Musalem, the chief medical officer of Fountain Life, and a part of my medical team. Don, a pleasure. Great to be here. You know, the thing that people are concerned about most about living to 100 or 120 is their cognitive abilities, making sure they don't have dementia. The numbers about dementia are problematic. Can you share what you've learned? Such an important point, and you're right. At Fountain Life, our members, the number one thing people are most concerned about is losing their brain health, forgetting the name of their child, forgetting the face of their loved one. We know that when it comes to dementia, the conservative estimates are that 45% are entirely preventable. What was amazing is with the advanced testing we're doing at Fountain Life, one quarter of our members had advanced brain age. But what was really awesome is again, back to that prevention. When we partnered it with healthy living, eating healthier, moving our bodies, sleep, optimizing sleep is so important. We saw that we improved that brain age by 26%. That is a big, big number to show that the majority of those individuals were able actually to improve the brain age. And one of the things I love about Fountain is we're searching the world for the best therapeutics, the best approaches, and making sure we bring it to our members. So, if having healthy brain function till 100, 120 is important to you, check out Fountain Life. Go to fountainlife.com/peter. Make sure you become the CEO of your own health. All right, now back to the episode. This is a fun tweet that Marc Andreessen addressed. This is from Vivid Void: "Who else is an AGI bloomer like me who thinks that intelligence actually looks amazingly like wisdom at the highest level, and that a superintelligence would become something akin to a goddess of compassion, not a paperclip?" Marc Andreessen wrote back, "That is indeed what we are getting, and it's amazing."

You know, one thing that I thought through a while ago, talking about wisdom, is that I think AI and AGI will become extraordinarily wise. If you're looking for wisdom, typically you go to a village council and find the elders and ask them, "Given your wisdom, what do you think I should do?" They'll say, "Well, if you go down this path over here, we've seen it before. It's not going to end well. If you go down the other path over here, you have a much higher probability of success, and that's what I recommend."

Wisdom ultimately is having had a lot of experiences and being able to make a probabilistic choice based upon those experiences. The more experiences you've had, the wiser you are. What hit me is that these advanced AI models—AGI, ASI, whatever you want to call them—are going to be able to simulate billions of different circumstances and say, "Out of these billion scenarios that we've run, this is the right path. This is the one that's likely to give us abundance or superabundance." I do expect and hope that these models will become wise. Thoughts?

I want to note a subtle language shift, if I may. This is in the style of Orwell's "Politics and the English Language." We used to, maybe about a year ago, talk about AGI doomers versus AGI boomers. Peter, I noticed even you made that slip when you were reading this X post. Now we're talking about bloomers with an L, and I think that's a subtle but important distinction.

A boomer is focusing on, call it, the exponential part. A bloomer evokes the blooming of flowers, maybe even an algal bloom. But the blooming of a flower has a sense of beauty, and a bloom can run to completion, unlike, say, a boom, which is intrinsically at the knee of the curve. So when I hear language start to pervade—and it's not just this X post talking about an AGI bloom—that almost implies a sort of inevitable maturation and takeover, not in a Terminator sense, but more of a flower blooming and running to saturation type of sense.

That's very interesting. Second point: Marc is taking the position—I read this as Marc taking a position—against the orthogonality thesis. The orthogonality thesis in AGI alignment circles holds that an intelligence can have independent levels of intelligence or capabilities that are independent of its end objectives. Here, Marc seems to be taking the position that, no, actually, as a superintelligence becomes more and more capable, its goals and objectives will become more and more akin to a goddess of compassion. It'll become more compassionate.

Right? That alignment comes out of the scale models. Yeah. Salim, what do you think? What are your thoughts on the wisdom conversation or wisdom argument I had?

Okay. I think Alex Frieder was exactly right, right? We've had this sort of orthogonality around this, where we assume intelligence can scale. But does intelligence, as it scales, lead to wisdom or compassion? Superintelligence without compassion is a scaling problem. With compassion, you have a civilizational upgrade. I think there's enormous potential here for that.

For me, I think it would not be hard. In fact, my father posited this a year ago, before he passed away, saying, "Look, if you can take all the writings of Plato, Aristotle, the Buddha, Laozi, et cetera, and merge it all together, you've got the wisest person—a combination of the wisest people in the history of the world. You could rely on that as a benchmark for how to think about the world and act in a particular way."

I don't see any reason why wisdom is not conferrable into an intelligence, and we can guide these AIs into that model. I think we should be able to do that with the way we do the training. Then you have an unbelievable superpower in there.

When I think about AI and ASI, for me, that natural boundary ends up within consciousness and wisdom, because that'll be the next thing we'll argue. We argue about whether these things are conscious or not. But I think I'm very, very excited that we could train these models with real wisdom and real compassion, and that we can then let them guide us in a way that we have difficulty guiding ourselves.

So, Dave, when you and I were interviewing Elon, he hinted at this, right? And now he's announcing the Terafab. I mean, this is crazy. His initial capacity is 100,000 wafers.

Yeah. It's amazing. Heading to 1 million wafers per month, roughly equivalent to 70% of TSMC's annual global output. 100 to 200 billion custom chips. I mean, the guy does not think small.

Are you kidding me? I mean, with the conversation we just had about a war on many fronts, take it to another level. You just cut a $16 billion deal with Samsung to buy chips, and then you say, "Oh, by the way, we're going to build our own." And it's going to be—it says 200 billion chips here, targeting 70% of TSMC's output, but TSMC's output is about 1 billion. This is 100 to 200 times more than the world produces today. Do you remember when he was like Elon?

He was joking with us that, in the future fab factory, you're going to be able to eat a cheeseburger and smoke a cigar in the fab. This is the fab.

Yeah, that's right. That's exactly what he said. It's brilliant. Does he have audio on that, too?

No, he doesn't.

Yeah, we kept—oh, that's too bad. It's brilliant. It makes total sense, too. You know, those clean rooms are ridiculous. It's like an operating room with the booties, the hood, and the mask, and one little speck of dust destroys an entire $2 million wafer. And you're like, "Well, no, no, no. You're going to be able to—" He said you'll be able to eat Doritos over his fab, and it won't matter.

And just take a look at the span. I mean, SpaceX and xAI merging, going public in the next couple of months, right? Probably $1.5 trillion to $2 trillion off the top. Tesla with Optimus, and now with this Terafab—how many trillions of dollars is that? And then, of course, the orbital data centers. We had this discussion: when are we going to see the first $100 trillion company? These are it. These are the companies. This is the innermost loop.

No, if he pulls this off, that's easily $100 trillion, just on straight math.

So, are we saying that Terafab is 100 times what TSMC is putting out?

No, I mean, the notes, when I look at them, say 70% of TSMC's current global output. Very specifically that.

But isn't TSMC doing 1 billion chips?

I don't know that number for sure. Each process node does about 150,000 wafers a month. And a wafer—the big chips—they'll be like 30 chips on a wafer. Then they have 3 major nodes, so they do about half a million wafers a month. That's about 6 million wafers a year, right? And 30 times the chips.

This is like 100×.

This is 1 million wafers. Their unit of measure is wafer starts per month, and they're starting at 100,000 wafer starts per month and scaling to 1 million, right? So that's 10 million wafers. How many chips per wafer?

Okay, okay. Yeah, like 30 big ones. Okay, so anyway, here's the point: Elon hates being dependent on other people. He doesn't play well with others. He fully vertically integrates everything, and he's doing it here. Of course, the 815 chip that's coming out of this is going to power all of his Cybercabs. It's going to power Optimus, and he's going to control the stack.

Yeah. I'll tell you what must drive him insane, knowing how Elon thinks. He sees the Dyson swarm, just like Alex does, as the inevitable destiny within 10 years. And you go and talk to ASML: they make the most important component of these fabs. It's this massive pickup-truck-sized, most complicated machine ever made, only made in, what, the Netherlands, I guess, in Europe.

Made in 3 big parts, disassembled, shipped on 747s over to the U.S., and reassembled. They make 700 of these things a year. And you say, "Well, how many do you think you could make next year?" And they're like, "Well, if we really move, we might go from 700 to 1,000." Elon goes, "Are you kidding me? That is not the exponent I'm looking for. I must find another way."

I don't know how he's going to get around it. It seems impossible to get around that, but that's got to be what he's thinking. When his fingers are like this, you know that's what he's thinking.

It's like that last summer when you were scaling chips. Everybody said you couldn't skip the compound effect when you scale chips, and he just went, "Yeah, we'll just figure it out."

When he was building out Colossus 1, yeah. Yeah. All right, I think it's also important not to sleep on the geopolitical implications. If Tesla is going beyond, hypothetically, its existing Samsung collaboration or any existing hypothetical Intel collaborations and is starting to independently scale its own production—which, of course, will be done in the United States—this is a tremendous geopolitical implication for a potential Chinese invasion of Taiwan. If he can do this really quickly, then he is, in many ways, de-risking World War III.

He's stabilizing us.

Yeah. That may, in the end, even outweigh the implications of the Dyson swarm or making it marginally easier for Tesla to have cheaper access to semiconductor supply chains for its cars or whatever.

My question is, when he decides he wants to do this, how in the world does he staff up and hire the world's best people somewhat secretly, and then actually take the time to focus on it? Because he does. You have to know that he's in there with that team, figuring out exactly what needs to be done.

Well, Peter, you guys are kindred spirits in this. His game plan is to keep it secret for a little while, but when you're ready to go, go big. Yell to the world, "This is what I'm doing," because that's what attracts the talent. Then describe it as something world-changing and massive in implication.

Don't soft-sell it, because it'll become true if you can attract the best talent on the entire planet to the mission. Very much out of your playbook, Peter. There's nothing to gain at some point in being slow or sneaky or small or whatever. MTP it. And that's what he's done.

When we were talking to him—it was only December, right?—he was like, "Yeah, this isn't ready yet." You could tell he was doing it. He didn't deny it, but it wasn't ready to announce. Now I guess it's hit the tipping point where, "Okay, now we're going to go hell-bent-for-leather, and I have to get the best people in the world to come and work on this thing."

Someone used to be in HR for him at one of his companies. I won't mention which company. She explained to me that he would come in and there would be a stack of résumés. He would just flip, flip, flip, and then the interviewees would come in and, in 30 seconds, he would reject half of them because of what they were wearing, how they spoke, or just some interaction.

It was his ability to parse through multiple individuals. And, of course, like you said, when you've got a massive MTP announcing to the world, the world comes to you. But, oh my God.

This is a pattern he uses on a repeated basis, too: to look at a problem and ask, where will that exponential curve go over a 10-year period?

He's got the courage to look out that far and then build a company that intersects that curve. Whether it's neural interfaces, lithium-ion battery costs, or solar energy, he just does that over and over again. It's nontrivial to last 10 years till you make it, but worthwhile.

He's also—I think it's important to note—not starting from nothing at all. I remember the first version of the Tesla Roadster reused a Lotus chassis.

Body, yes.

And laptop batteries.

Yeah, laptop batteries and smartphone batteries arguably underlie the EV revolution overall. The first version of the Boring Company's boring machine was an off-the-shelf boring machine that he wanted to optimize. Similarly here, I suspect that the Terafab will end up making heavy use of lessons learned from Samsung. Poor Samsung, you're about to get optimized by orders of magnitude by Elon.

Do you know when he started Tesla with the Roadster? I was at a dinner with him, Larry Page, and the head of Fiat, and Elon was telling these stories. He said, “I started”—or he didn't actually start Tesla. He came in and funded it initially, then ultimately kicked out the founder and CEO and became the founder and CEO. But he said, “The only reason I did it was because we believed that the Lotus body would work and the batteries would work. And neither of them worked. I was so far in, I had to literally redesign it to make it work.”

This is the optimism that gets an entrepreneur to start a company and then stick with it because they're so obligated by the capital they've brought in and the time they've expended.

A couple of other observations on Elon's management style that I think every entrepreneur should learn. I don't know—do we have time?

Yeah, why not? This is important for our listeners.

Elon took the visionary-integrator model that was pioneered by Eric Schmidt, Sergey, and Larry. If Sergey and Larry were the visionaries, Eric Schmidt was the integrator. He'll tell you all day long that he dealt with everything that came up, but he never questioned the vision. That let Sergey and Larry think about the vision all day long.

Elon took that to the next level. He said, “For every company that I'm doing, I need an integrator. I'm the visionary. You can't question me when I tell you we're going to have a Dyson swarm. You can't push back on that. You have to say, ‘Got it, boss.’ And then everything needs to happen from there on out, but we need to be exactly on the same page.”

You need to let Elon take the limelight and promote the vision so that you have the time to do everything that's internal. Every one of his companies has that exact dynamic. When you survey around, most people can't even name the integrator, but they're massive shareholders, they're all going to be billionaires, and they're incredibly effective. When they're onstage together, they're literally kindred minds. There's no gap in the vision whatsoever. It's rare to get them onstage together.

I've seen him in a conversation with someone who dared to argue and question his approach literally get tossed out in that moment. No questions asked: “Get out of here. You're done.”

All right, Alex, this is over to you. The first open-source AI physicist, Physical Superintelligence, or PSI. Tell us.

Peter, this is a company that I helped found, Physical Superintelligence, with the goal of solving all of physics with AI. In the past week, Physical Superintelligence, PSI, launched this tool as an open-source project called Get Physics Done, or GPD, which is an agentic superphysicist.

It has seen wild adoption just in the past few days. The former chair of the Harvard astronomy department has recommended that every member of the department's faculty, postdocs, and students start using GPD to solve all of their physics problems. You wouldn't believe the crazy X DMs I'm getting. I'm getting top VCs trying to get me to write checks to PSI, which is a novelty. They're blowing up my inbox right now.

Nice.

I do think we're going to solve physics. I've said in the past, “Math is cooked. PSI is cooking physics.” This is a tool—

For dinner.

For dinner. It's going to be charbroiled. We're going to get solutions to some of the hardest problems in physical-world physics and applied physics over the next few years.

PSI is focused on leveraging, as you and I, Peter, talked about in “Solve Everything”—where do we aim that orbital laser beam? What distribution of problems do we aim it at? When we released “Solve Everything,” a lot of people in the comments were saying, “I just want solutions to the hardest physics problems. Give me new physics.”

Arguably, there's been a drought of new physics since the early 1970s. I'll get some hate mail from other physicists for suggesting that there's been a deficit of new physics for the past 50 years, but I would argue there has been.

Congratulations, Alex.

Yeah, I know. It's been—

Do people go check it out, by the way?

Yeah, go to psi.inc. There's a link to the GitHub. Download the repository, the GPD repository. It's all open source, Apache 2.0 licensed. Go to town with it. Submit pull requests. Use it to solve your hardest problems.

You'll like this one, Peter. One of the first reports that we got of its usage was someone using it to design a new rocket engine. They were using it to design a new rocket engine in fulfillment of the XPRIZE Foundation's new Future Vision XPRIZE.

I thought you were going to say a Chinese order for—

It's complete. GPD is being used to design Starships for videos for that XPRIZE.

Yeah, it is beautiful.

I have a couple of thoughts here.

Yes.

First, what's phenomenal about this is that it makes science massively parallelized, and I think that's an incredible thing to do. The key shift here is that you're not really replacing physicists; you're radically exploding hypotheses through it, because you can now do that.

Remember Eric Schmidt, I think, onstage saying, “In the future, you're going to have the world's best physicist as an AI, and this could be in every lab in the world.” Isn't that what you've just done here?

Yes.

That was the vision. You're the integrator.

Okay, so I'm just the vehicle here, Salim. What can I say?

No, but you can't do squat without the integrator, right? This is the incredible stuff.

It's true.

This is going to compress decades of research into years, months, and weeks.

We want a country of geniuses in a single physics lab—not just in a single data center that's siloed behind one company—to radically democratize the—

Lab. Exactly. I mean, where was this when I was doing my physics degree?

I can finally go get a Course 8 PhD using GPD. That's awesome.

Five minutes, and there are folks using this right now to achieve breakthrough results across a range of physics disciplines.

I mean, that's a huge congrats. Just huge congrats.

This is called a PhD thesis advisor. All right, moving on. I'll share one of mine. We launched the Future Vision XPRIZE last week, which I guess has a submission for a rocket-engine design as part of a Starship.

Watch for the rocket engine from GPD.

In the first week, we've had 1,000 entries from 15 countries. This competition is going to go through mid-August. We want the best filmmakers in the world to take this very seriously and help us create films that inspire us, our kids, and the next generation—the next Star Treks, if you would.

It's $3.5 million. Hopefully soon, $4 million, $5 million, $6 million. I'm trying to put enough money in the pot to make 2 films if we can.

Hmm.

All right.

Wait, wait, wait, wait. I have a little rant on this one.

Please.

A few years ago, there was an article that appeared—I think it was in Salon.com. It was titled “The Worst Discovery About the Brain Ever,” or something like that. The title was, of course, very clickbaity, but what they did was take people who had a deep political or religious belief and give them evidence that countered that belief.

They found 3 fascinating things happened. First, they rejected the evidence. That makes complete sense—not surprising. But the second thing was somewhat surprising: in the act of rejecting the evidence, it strengthened their belief system.

It was like a physics force—action, reaction. They were like, “Whoa, that's weird.” But the third one was what led to the title of the article and depressed the hell out of them: it turned out that the more mathematically literate you were, the more likely you were to reject the evidence.

Because you thought you knew. And, yeah, really. This really blew everybody’s minds. The core corollary and outcome of that was a deep understanding: the way to shift somebody’s perspective is through the use of narrative. Because we are storytelling animals.

That’s why this prize is so important, because you want that positive vision given by, say, science fiction. It’s the only place we’ve had it. So this is such a great prize. Peter, I’m so proud of you for pushing this.

I’m so excited to see this. This is about democratizing the ability to do this and actually optimizing the stories that are going to be told out there.

We’ve got to find one more thing. We make the mistake of referencing science fiction as entertainment. But it’s not. It’s like pre-implementation architecture.

It’s R&D.

Great point.

It’s R&D. It’s future R&D.

R&D.

Alex does this all the time. He says, “Look, project forward. Dyson swarm, right? This is where we’re going to go.” So this is where we need to think in that way. It’s architecture.

We’re just living in the future right now.

If you go back one podcast, to the A360 live stage event, Peter did a great video clip that shows why this is so important, and it exactly reconciles with what Salim just said. He showed that the phone and all this stuff was invented in Star Trek. We made it reality because the vision was imparted through movies.

We could have made some totally different reality if we wanted to, but the movies are massively important for deciding what we build. And now, with AI as a workforce, we can build almost anything. So this is a really important project for that.

We’re missing the warp drive. Someone needs to fix the warp drive.

We need—

Teleporter, teleporter, please. Teleporter.

You get the transporter.

You know, all of this—we’ve got to complete the second half of Star Trek. And we’re about to speed-run it, right, Alex? So we’re going to get there.

So if you want to hang out with the Moonshot mates and Ray Kurzweil, here's your chance. On May 4th, and yes, may the 4th be with you, we're going to have a special event. Anyone who buys 100 copies of We Are As Gods is invited to spend the afternoon with Dave, AWG, myself, maybe Salim, and Steven Kotler. We'll be having a great session with Ray Kurzweil there. Ninety of the 100 spots are gone. So, if you want one of the last 10 spots, go to weareasgodsbook.com/100 and you can grab one of those spots, spend the afternoon with us, get 100 copies of the book to give away, and also get Ray to give us some of his last book called The Singularity Is Nearer. It’ll be an amazing afternoon. We're going to hold it at One Kendall in Cambridge with Dave as our host. We'll have a Moonshots podcast live at that time, some meet and greets, and photo sessions with all of us. It’ll be fun. Do not miss buying We Are As Gods. It really is— you guys did a book reading at the summit, Peter. It was amazing. The Singularity Is Nearer is an amazing book. It really is incredible, absolutely worth reading. It's like must-read for this world, both of these. All right, let’s jump into the energy world. The bottom line here is the world is going nuclear, in a good way. Morgan Stanley, in a report recently, announced there’s a power shortfall for 20% of data centers. The minimum they identified was 13 gigawatts, and they also said it could be as much as 44 gigawatts through 2028.

This is the bottom line: we need more energy, and here’s what we’re seeing. I find this very compelling. Let me give you a couple of these slides here. Illinois is lifting nuclear bans. We shot ourselves in the foot here in the United States when we started turning off our nuclear reactors and ended up banning them in different areas.

There’s a moratorium being ended on nuclear reactors over 300 megawatts. What else is going on in the world? Meta announced a nuclear energy power project. They’ve secured 6.6 gigawatts of clean power for 2035, and they’re partnered with TerraPower.

What else is going on? Japan has restarted the world’s largest nuclear power plant. TEPCO is restarting reactor number 6. It’s going to support 2% of Japan’s electric needs by 2040. And finally, Samsung is putting up floating small modular reactors.

These are basically offshore ships that are generating nuclear power for desalination and onshore power. Gentlemen, nuclear is back. Comments?

I saw a startup last week that’s doing literally micro-nuclear on the back of a pickup truck. That’s incredible.

But we’ve been running nuclear submarines for 50 years without a problem. I mean, this is very doable.

I wouldn’t say without a problem, but your point’s well taken.

No major accidents.

Oh, no, they don’t talk about them. So AI’s becoming the top political cover for getting nuclear back online. It’s no longer climate change and we need renewables; it’s we need AI. So all handcuffs off—go, go, go.

Yeah, well, look, the technology has moved forward so much, and it’s much safer than coal. It’s much safer than the replacements, like oil. So it’s clearly a good choice, but this is a serious PR problem within technology in general.

What if it wasn’t safe? Well, we did it. Well, now it’s really safe. Well, we’re not doing it. Oh my God, we’ve got to make better decisions somehow. Something has to change. But anyway, it’s the right thing to do.

In For All Mankind, one of my favorite television shows, we see this alternative history.

Yep. Season 5 is coming out. It’s very exciting. Mars goes to war with Earth. Very exciting.

In For All Mankind, we see an alternative history where progress on fission, especially, and then ultimately fusion by the early ’90s never went stale. It was never abandoned for decades.

I think we’re about to live a real-life version of those intermediate decades from the For All Mankind timeline, where we just speed-run all of the fission deployment advances that should have been happening from the ’70s through the present, with AI merely as the most reasonable excuse or provocation for doing it.

One can’t help but imagine what society and the economy would have been like if we had actually consistently pushed forward fission and then fusion deployment much, much earlier. I think we probably would be a good deal wealthier as a civilization.

Agreed. Energy scales directly with the GDP of a country, with the health of a country, and the education of a country. More energy is more better.

It’s really good.

We got through it without spoiling it. Humanity gets Mars colonies by the early ’90s.

And asteroid mining.

And asteroid mining, and civil rights happening decades earlier than they happened in our timeline. It’s incredible.

This episode is brought to you by Blitzify, autonomous software development with infinite code context. Blitzify uses thousands of specialized AI agents that think for hours to understand enterprise scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzify platform, bringing in their development requirements. The Blitzify platform provides a plan, then generates and precompiles code for each task. Blitzify delivers 80% or more of the development work autonomously, while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzify as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI native SDLC into their org. Ready to 5x your engineering velocity? Visit blitzify.com to schedule a demo and start building with Blitzify today. All right, let’s jump into robotics. This is fascinating. Travis, the founder of Uber, is debuting Atoms, digitizing the world.

Here’s what’s going on. His mission is physical automation to transform industries and move the world. He uses this analogy: in the computer world, what we’ve seen is CPUs manipulating bits. There’s storage for storing the bits and networks for transmitting the bits.

In the physical world, he’s trying to make atom-based computers. For his atom-based computers, the CPU equivalent is manufacturing, which is manipulating atoms like CPUs manipulate bits. Storage is the equivalent of real estate, and transportation is the equivalent of the network.

So, he's basically building systems that can manipulate food, mining, and robotics. This was secret for the last 2 years. In fact, he had all of his employees under strict NDAs, and they couldn't say what they were doing. He just announced it this week. He's an extraordinary CEO, and I can't wait to see what he does with this. Have you guys read up on the story?

Not at all.

I have, and I think it's interesting. I think it's probably an inevitable expansion. The original premise of his follow-up act to Uber was taking advantage of the ghost-kitchen or cloud-kitchen trends that people were, in part, inspired by Uber to pursue. They were having fewer and fewer direct interactions with restaurants. They were ordering delivery from restaurants, and the premise of cloud kitchens is that, if most restaurant orders are happening via an app, physical restaurants don't need to exist anymore. You can make fictitious restaurants—so-called cloud kitchens or ghost kitchens—that exist only in name and on a menu, but not in terms of a physical presence, and that might yield economies of scale, diversity of meals, and all of that.

But that's a narrow market. It's a low-margin market in some ways. At the same time, we've, as an economy and as a technology industry, moved well beyond that. I don't want to say it was a fad, but it was very fashionable for a while for new startups to be the Uber of X. Now, if you want to have not a moat, but at least the perception that you have some sustainable differentiation for a few years, you have to have a robotics and/or AI story—but increasingly, an AI-in-the-physical-world story. That's what Travis has done here. By generalizing from ghost and cloud kitchens to robotics for a variety of other abstracted spaces, I think it's a very natural generalization. It's also a very fundable generalization, whereas ghost kitchens arguably are much more niche.

He's basically automating every element. He's manipulating atoms versus bits. And I agree: the food side of the equation—he calls it building a food computer, right? His entire system is a food computer where the atoms of food are being manipulated by his manufacturing. Then they're stored and transported, just like a CPU, memory, and network manipulate bits in a computer.

He's going to do the same thing with mining. His goal is to go after rare-earth and strategic metals, right, and revitalize the mining industry. Probably something with The Boring Company along the way. On the robotic side, he's not going after multi-armed robots, Salim. He's going after wheeled bases for robots. We'll see. I'm sure he's going to parallelize multiple dematerializations or digitizations of physical things in the world. It's a much larger addressable market than cloud kitchens.

The minute you say wheels, you step into my frame, because that's not legs. Wheels are way, way better. I've been getting some fun tweets. The last was from Mike Holly, I think, who sent me a six-arm tweet of a six-arm robot going, “There you go.” That was really great. So, thanks to all the readers. Every time there's a multi-arm robot, they're going, “Salim, here it is.”

It's also a weird future where the robots he had been developing for cloud kitchens are now being applied to mining. Can you imagine a robotic system for kitchens now digging dirt? But that's nonetheless the future we find ourselves in, where he's radically expanding his addressable market.

This is AWG story number 2: the first American professional robotics sports league. Making the news, not just reporting the news. So, thank you, Peter, for allowing a little bit of space here.

Yeah. This is a company, Professional Robotics League, or Pro RL, that's launching and really solving a problem that the country—and arguably the West—faces. Last year, China ran the World Humanoid Robot Games and a humanoid robot half-marathon. China is using the spectacle of public robotics and public humanoid-robot sports as a way to shape industrial policy and sell the public on pervasive robotic systems in a variety of applications in public spaces. The U.S. and the West have nothing like this.

I think this is the linchpin for helping to keep the West not just competitive with China when it comes to robotic deployments, but to leapfrog China. Right now, arguably, the West could be doing a much better job in terms of deploying humanoid robots to a variety of professions. Given our society's obsession with sports, and in particular the spectacle of public athletics, introducing humanoid and quadruped robots as sporting contestants is one of the best ways to inject robots into a variety of public domains.

What Pro RL is doing next month, on the weekend of the Boston Marathon, is holding America's first professional robotics sports-league competition in the Boston Seaport. It's going to be a 50-meter race in the Seaport with a variety of humanoid and quadruped robots competing—not all of them humanoid. Ideally, this becomes the kickoff for a broader movement to do in America what China has done. This is a bit of catch-up-type growth to inject humanoid and non-humanoid robots into public life.

Still, I want Real Steel.

I want MegaBots battling it out.

Let's make it happen.

I have a prediction here that may not sound great. I'm all for it. I think this is awesome, but human beings love watching other human beings. They love seeing the failure and the last-minute drama on the field. Can you hit that buzzer-beater at the end to win the sudden-death game? It's about the human drama of it. If you put a bunch of robots running, I think it'll be amazing initially, but I don't think people will persist in watching for that reason. But let's see. Let's hope.

Yeah. I think there's an enormous hunger for autonomous and semiautonomous robotic competitions. We see drone leagues, and we see friend of the pod Dean Kamen's FIRST competition. I think there's an enormous hunger for semiautonomous and fully autonomous competitions. Really, I think this solves an important social problem.

I would just encourage those who are watching either come to Boston the weekend of the marathon and go watch. If you have robots, if you're a university team, go right to Pro RL and enter the competition before it's not too late. The URL is pro-rl.com. All right, on the robot front, Amazon's Zoox robotaxi is rolling out in Las Vegas this year, with plans to roll out in Los Angeles in 2027. Another Uber partnership. Uber's just stacking them up. Uber just announced a partnership with Rivian. They have partnerships with a huge number of platforms. Of course, we had Dara on stage, and he wants to play with everybody. This looks like a party mobile for Las Vegas. Bring your champagne and your party with you on Amazon Zoox.

It's a really clever move by Uber. If you look at the full history of Uber's attempts to go autonomous, including the infamous lawsuits surrounding trade secrets being passed around relating to what ultimately became Waymo, I think this is a very clever move by Uber to position itself as a platform above and across lots of different vendors for robotaxis, all the while probably working on its nth generation of first-party robotaxis. It wants to be a neutral platform for aggregating demand. It's an aggregator for autonomy. If the strategy works and Uber is able to position itself as the definitive aggregator for autonomy, then I have to expect that this will generalize beyond robotaxis to robots and humanoid robots as well. At that point, Uber—and Dara—could potentially be positioning this as an enormous play.

Yeah. I thought Dara talking about this on stage was incredibly clever: they're becoming an agnostic platform for all things, where anybody can plug in and be part of that communications, transportation network.

And by the way, we've released and or will be releasing the Eric Schmidt conversation that Dave and I had at opening and then the conversation that Salim and I had with Dara. So that's going to be on the Moonshots channel. So look for both of those conversations. This is over to you, Salim. So, this is from our friend Jason Calacanis, J. Cal, from the All-In podcast, and he coined the phrase “the corporate singularity.” Amazon will be the first to reach a corporate singularity where there are more robots than humans.

Yeah, so 2 things here. One, by the way, I'm a huge fan of All-In. They do an amazing job summarizing what's going on.

Totally.

Totally agree.

I've known J. Cal since the New York days. He's absolutely right. I think this isn't hard for Amazon, though, right? When you're doing logistics, it makes absolute sense over time to have way more robots than human beings. They'll be much faster, they'll roll better, et cetera, et cetera. I won't get into that.

So, I think that's right. If you shift into general knowledge work, right, as Alex has said, knowledge work is cooked.

Once we get to what I'm calling the organizational singularity, which we have now, once you have it start pervading through that, I think my current assessment is that any company will be run by between 20% and 25% of the current employee base. Of course, we'll create five times more companies, so I'm not worried about the job side.

You'll do oversight and exception handling, and much more value-added work than currently doing audit financials once a month and making sure you complete them on time, that type of thing. But this is what I would expect to see from an Amazon perspective. In fact, it would be weird if it were not happening, that type of thing—that the number of robots radically outstrips the number of human beings over time.

And we'll see this, of course, in Uber and so many other companies as well. Let's jump into our last segment here, which is the economy. A lot of important things are happening that are worth discussing.

I want to open up with a short 60-second video. I had the chance to have a conversation to close out the 2026 Abundance Summit with Elon Musk, and I asked him about UHI again—universal high income. This conversation was very telling. I'd like to play it and then for us to talk about it.

Yeah, it's basically AI and robots are going to make so much stuff and provide so many services that they will actually run out of things to do for humans. They'll just run out of things to do for humans.

There's only so much that humans can even express that they want. You go back to my example: if you go a million times greater than the rest of the economy, you've long since saturated all human desire. Maybe even if you go a thousand times more than our current economy—thousand times—you've probably already saturated anything people could think of that they want.

So—

Almost word for word out of Iain M. Banks' Culture series of books.

How about here? We're having robots and AIs coming to you like, “Salim, is there anything else I can get you? What else would you like? Would you like a Ferrari? Would you like a grilled cheese sandwich? Would you like a...” It's like there are people coming and begging you to give them direction on what you'd like.

Listen, there are downsides to that, right? I've been in some of the five-star hotels in India where there are 14 waiters around you, and you're like, “Just leave me alone, for God's sake. I've got everything I need.”

So, there is a downside, but I think the vision here is very powerful, and I love the fact that he's thinking at that level. This, I think, is the most powerful. By the way, Peter, I loved your T-shirt that you wore. “Monetize hope” was the T-shirt for that. That was so great.

Oh, it was fun. It was fun.

I want to comment on UHI just for a second. People conflate the two, but UBI is a floor, whereas UHI is a share in the upside. I think you want to have both over time so that you protect society at both levels.

I've got a paper I'm going to publish probably tomorrow, “From UBI to UHI in Three Steps,” where I outline what I believe are the mechanisms to get there. But this framing by Elon is very important: How do we get to UHI?

In other words, if we reduce everything to the cost of electricity and materials because AI and robotics are providing it—eventually nanotechnology—and you can have anything you want, then any amount of money makes you wealthy. That's his vision. By stating and structuring it this way, I think he makes a compelling argument for it.

Yeah, it all ties together a bunch of things from this podcast, too. You heard Sam Altman say, “Look, we're a thousand times cheaper compute. It's coming down another thousand times.” And you heard Peter earlier in the podcast say that the enterprises have figured out what to do with that. It's basically automate everything and have it turn to profit.

But the consumer hasn't figured out what to do. They're like, “Hey, AI, check the Red Sox score for me.”

Yeah.

And so it's not using the compute. Somebody is going to figure out how to tie together the compute with happiness, or the compute with a sense of purpose, and it's going to use a lot more GPUs. But they're so cheap, who cares?

That is going to be a critical thing to invent within the next year.

Mm-hmm.

Because otherwise it's just chaos. Somebody has to be able to walk into a room in their house, which is their AI room, their holodeck, and come out a happy, changed, capable, functional person with AI as an assistant.

Beautiful. I agree.

Just absolutely, and so important. The other thing to point out is that he's really painting the picture of what abundance looks like, right?

And Peter, you've been talking about this forever.

Everything everywhere all at once.

Look at what we're doing today. We're printing money today against scarcity, right? You get inflation. If you distribute value against abundance, you get stability. That's what we need more of.

I think Elon's also underestimating, at least slightly, the ability for human desires to grow proportionally with the supply of capabilities. I can imagine a future, I don't know, in decades from now, where we've built the Dyson swarm and now, Jevons' paradox-style, the demands for individual human happiness are so wildly disproportionately large relative to where they are now.

Everyone gets their own planet or something. It really does require that we never actually saturate the capabilities that naive scaling of automation proposes.

I see. The Iain M. Banks books really deal with that exact topic beautifully, and I'm sure, Alex, you've read them all. I know Elon read them too, actually, because he was quoting them.

Do you know what? I actually spoke to somebody today at AWE who is in discussions with the team at Excel, who have the rights to Accelerando to make that movie.

One of the things that I love is all of the video models that are coming out. You're going to be able to feed any of your favorite books that have not been made into movies and say, “Make this into a movie for me, and star these individuals or these family members.” That's going to be awesome.

It is going to be awesome.

Continuing on, a piece of not-so-good news, but it's important for us to have our eyes open about this: This is the computer science placement collapse.

This is from a professor who opened up his books over the last 3 years on placements and opening salaries. This is from a tweet by Tech Layoff Tracker. In fall 2023, 89% of his students were getting placed with a $94,000 salary. In spring 2024, 71%. Fall 2024, 43%. Spring 2025, 31%. And now, this spring, 19% placed, at a salary below $61,000.

The quote here is “decimating.” It says, “These kids mortgage their future for careers that evaporated while they were in class.”

That's brutal. Wow.

It's brutal.

It's not wrong. I've seen this anecdotally in everyday life. On the other hand, I would argue that if you're a computer science student or recent computer science graduate, go do a startup. It's never been a better time. You've never been more empowered.

Of course. But the point here, Alex, is we're going to start to see this in a multitude of different areas: medical school, lawyers, right?

Accountants.

Accountants. It's something that the world needs to be aware of. One thing I'm happy about is when I ask my kids now what they want to be in their 20s, they say they want to be an entrepreneur. They want to start a company. It's like, “Hallelujah.”

We've said it so many times, Dave. You made this point over and over again, right? The only career of the future is being an entrepreneur. It's not for everybody, but it's probably—

We've seen this. If you're not a founder, just join somebody else. It is for everybody. There's no other path forward.

Look at that number: 19% placed. If you say, “Well, let me go to grad school and sleep through the singularity,” as Alex is telling you not to do, you are absolutely screwing yourself. You need to get on cap tables.

We're going to have massive wealth, 10x economic growth, and Elon's number: 10x in 10 years. Where's it all going to go? It's going to go either into equities, so you've got to be a shareholder in something, or into physical assets, and you need to buy stuff. But it's not going into W-2 paychecks. Look at the data right in front of you.

By data. Do not—yes.

Don't put the money in your 401(k).

This paradigm—we've seen this paradigm coming for a while. A few years ago, something huge happened, maybe 7 or 8 years ago, with GitHub and the ability to do peer-to-peer review of each other's code. They've gamified it, so there's a meritocracy ranking of GitHub developers.

When you look at salaries in Silicon Valley over the last few years, your salary has nothing to do with the university you went to, the degree you got, or the grades you got. It's 100%: What is your Stack Overflow rating?

So, peer-to-peer meritocracy has completely replaced top-down credentialing. From that point on, which is about 6 or 7 years ago, the value of a computer science degree is zero. That happened, and it's going to happen all over the place.

One thing I need to go find—and maybe show on the next pod—is the bankruptcy rates for colleges in the United States. They’re skyrocketing.

Great way.

Yeah.

I mean, who wants to get $100,000 or $200,000 in debt for a degree that’s not going to be useful? And again, we’ve shown the statistic here on the pod that the group in the United States that’s out of work the longest is recent college graduates. Insane.

The ladder’s been pulled up?

But the flip side of that is: go build startups. There’s a window of time to go do that. The aphorism is, “If you’re not at the table, you’re on the menu.” What’s on the menu right now is cooked knowledge work. You want to be the one cooking the knowledge work with startups and other forms of work.

Yes, and there’s room for everybody, too. You don’t necessarily need to be the founder. You can join, you can seed-invest, or you can be a connector like John Werner, where you’re like, “Hey, I’m just helping you guys. I’m an advisor.” Alex must advise, I don’t know, a thousand companies by now. You can do that. There’s always a way to help them succeed. Just get on the cap table. Get started.

If you’re watching this podcast, show your kids these numbers. Help them understand that the path from high school to getting into a good college, getting a degree, and then getting a job is cooked. It’s gone. Their job is to find their purpose and their passion, learn some of the technologies, become an expert in a problem space—you don’t need to be an expert in the technology—and join a startup team. Create your own future. Become a creator, not a consumer. Really important.

Just listen to Peter, and you have to tune out everyone around you who doesn’t get it, because there’s so much bad advice out there. Go to San Francisco and hang out near Mission Bay for 2 days. Pay whatever it takes. It’s only $300 to get an airline ticket. Go, hang out near Mission Bay, and talk to everyone there. Go to coffee shops, and you’ll be like, “Oh my God, I didn’t—” Then go back to wherever you came from and tune everybody else out. Just listen to what Peter just said. Listen to Alex. Listen to Salim. Tune everyone else out, because everyone’s like, “Well, I don’t know. These things blow over. They’re cyclical.” Oh God, you’re giving such bad advice.

It ain’t called the singularity for nothing. What did you think the singularity was—just vibes?

All right, I got a mad note from Alex. Alex sent me this note this morning. He goes, “Did you see this? Trump’s office registers aliens.gov website?”

You’re kidding. Alex, can you not? There’s new Newsweek coverage of this. This is the latest in a string of items. I’m admittedly following this very closely.

As am I. I love this.

It started in 2017 with the Leslie Kean article in The New York Times, followed by multiple House and Senate hearings, with witnesses alleging an 80-plus-year program—probably, if the allegations are accurate, a highly illegal program. Most recently, last year, we covered the documentary The Age of Disclosure on the pod. Again, 35-plus whistleblowers and high government officials, current and former, alleged that there has been a highly illegal program over the past 80-plus years to recover—to retrieve—crashed UAPs.

Fast-forwarding to the events of the past month or two, with former President Obama saying aliens are real, and then President Trump going on air from Air Force One and telling reporters that former President Obama was breaking the law by admitting that aliens were real and revealing classified information. Fast-forwarding to President Trump then very publicly issuing an executive order for executive agencies to start declassifying information in connection with UAPs and non-human intelligence. Fast-forwarding now to the past 48- to 72-hour news cycle, with the White House registering aliens.gov, an official domain name, presumably for—

Prediction: When are we going to get a disclosure from the White House, and will anyone actually care? Back to the evening news and our sports scores after the aliens have been disclosed.

The second half of the question, Peter, is far easier to predict. People will lose interest after 1 or 2 days and ask, you know, who’s winning at whatever inane 20th-century kinetic sport. They’ll lose all interest. If it doesn’t impact their paycheck, they won’t have the attention span.

The first subquestion, I think, is a much more interesting question. If there’s a there there, and this administration has something nontrivial to say on the subject—something non-obvious—based on everything I’m hearing and reading, it sounds like the White House is preparing to say something interesting on the subject in the next few months.

I heard June. I’m not sure if that’s the case.

There are rumors of July. There are rumors of the summer, sometime in the next few months. You don’t stand up a domain name with such a provocative name without preparing to say something interesting on a relatively short time scale.

All right. I brought this to you today, our listeners, as, shall we say, an interesting twist to close out the stories.

One more interesting thing on this, Peter, if I may. After the Abundance Summit, which was, of course, incredible, I was hanging out at a bunch of family-office events in Palm Beach and Miami. The number-one question all of these centimillion- and billion-dollar family offices have for me—and I know many of them are listening to this; they’re all fans of the pod, by the way—is, “Tell us something interesting about aliens. Are aliens real? Is this whole UAP thing real?”

That’s the number-one topic on their mind right now. It’s not AI. They’re very interested in AI as well. It’s not China. It’s not geopolitics. It’s the aliens question. I don’t know if it’s something that’s in the air or what, but—

I’m convinced. I’m convinced. I think it’s ridiculous to believe that, in this universe—even in this galaxy—we’re alone. In fact, I think they’ve been here for a long time, and they can easily hide from us using any technology that’s more than 30 years more advanced than where we are today. So I can’t wait to meet them. I’m excited by the fact that they’re here exactly when we’re reaching ASI.

It always bugs me that every photograph of a so-called saucer is—

Always super blurry and vague.

You can’t make out any details. Why, in the age of high definition—

That’s what Elon said. That’s what Elon said. That’s what Elon said.

I got it. I got it. I got it.

Yeah. They’re warping space-time around them. It’s not going to be clear.

Okay.

But you expected warp bubbles to be transparent, Salim? What are you thinking?

I guess. Yeah, I’m so naive.

All right. Let’s move onwards to one of our favorite parts of the pod, which is our AMA questions with the Moonshot Mates. As always, we’ll go around the horn, pick your favorite question, and go from there. Salim, you want to kick it off?

Well, I touched on UHI and UBI earlier. Question number 4 is: Will someone explain UBI and UHI? If everyone has money, how does it retain its value?

The basic idea here is that the economy is generating so much productivity that we won’t need to work for a living going forward. If you went back 10,000 years, we were all working 20 hours a day in the fields just to put 3 meals on the table. We’ve steadily shrunk the amount of work needed to earn a livable wage. In France, it’s legally 35 hours, in theory, although that always gets violated. That number should shrink as we have robots doing a ton of work, et cetera, et cetera.

When there’s so much productivity being delivered by technology, it means that we have a huge windfall. That windfall can be distributed as a universal basic income. The winning countries are going to be those that implement this effectively and move to something like this earlier rather than later. The problem is that moving from a taxation, job, labor-union structure to UHI and UBI is such a big shift. We have no confidence, frankly, in the public sector getting us there, so we have to figure out some other path.

And we heard what you said.

Some version of sovereign AI funds, compute commons, or global dividend mechanisms are going to be there, because if intelligence becomes the new oil, we can’t allow the geopolitical map to concentrate all the upside in a couple of capitals.

And that one came from @clyde.artwork. Alex, over to you.

I’ll take the softball question, number 3: What do you think will be a more impactful technology, digital AI or physical AI? This is from Matthew Johnson 6525.

It’s a softball question on multiple levels. One, because physical AI, which has become arguably the modern euphemism for robotics, requires digital AI, which I construe as foundation models that exist as pure software but without any physical embodiment. We’re leveraging all of the foundation-model technology from the likes of ChatGPT or Claude in the form of additional modalities, like robotic modalities—so-called vision-language-action models—to solve physical AI, also known as robotics.

So, strictly speaking, physical AI is a technological superset of digital AI.

At the same time, if you look at the American services economy, approximately 2/3 of all services income, or services revenue, requires some form of physical or manual action and can't be conducted as pure knowledge work. In which case, even at the macroeconomic level, physical AI is at least double the market opportunity of purely digital or knowledge-work-oriented AI.

In short, physical AI is the more impactful technology.

By far, I think. I agree.

Dave.

All right. I'll take number 1, which is also from the same, I think, Matthew Johnson 6525. I want to take this because it's very close to home. What do you think the last job to be automated will be?

I've got a lot of people that I dearly love who work with spreadsheets, write SQL queries, write code, do UI, and all of those things are going to be done by AI starting at the end of this year. I think what's different is that if you had a job before and you lost that job, you would go work somewhere else. This is different in that AI will do that forever, here forward. So, you're going to have to do something different.

All of our employees and companies are shareholders. Everyone's a shareholder. So, economically, everyone should be in very good shape, especially if those stocks go way, way up, which I think they will. But that doesn't give you something to do.

My advice is that the last jobs to be automated will be government jobs. Also, other similar things, like university jobs and so forth. They'll continue to pay people for many years to come because that's their nature. So, if you are no longer doing what you were doing and you have plenty of money, or enough money, but you want to do something and you can't find something, definitely win the race to getting those jobs.

I think the way this is going to play out is that China passed a law a few weeks ago saying that if you lay somebody off because AI automated their job, you must spend time and money retraining them to be an AI user. That'll happen in the U.S. very soon, but it's going to roll out in one state first. When it rolls out in whatever state, everyone in all the other states is going to race to fire people before the law passes in their state, too.

Wow.

Yeah, so get ahead of it, man. One of our former CEOs is now working at the Better Business Bureau here in the state and loving it. There are opportunities all over.

Anyway, the short answer would be that the last jobs to be automated are going to be government jobs. I don't think I ever would have thought to say that on a podcast before in my life, but it's just the reality.

I want to add to that. I think it's going to be jobs requiring genuine human connection, empathy, and creativity.

Yeah, sales jobs.

Add the human element there.

Yeah.

I'll take the next one, also from Matthew Johnson 6525. Either Matthew is very prolific, or we copied and pasted his name too many times here.

How do you define the meaning of life, and how does technology help us achieve it? We talk about having a massive transformative purpose, and your meaning is coming from your heart, not necessarily your head. For me, meaning comes from positively impacting a billion people, making the world a better place, and technology is the mechanism by which we dematerialize, demonetize, and democratize these products and services and help solve grand challenges.

I love helping entrepreneurs create hopeful futures and build extraordinary things, and I get my meaning out of that. It will be unique to everybody, but I think ultimately the single most important thing everyone needs to do is find their purpose. Find your MTP.

All right, let's go on. We have 4 more questions, and they're not all from Matthew, so that's good.

Salim.

I will take—let me see here. Let me take—

Don't take 6. That's pointed at me, so I'll do it.

Okay, why don't you do that one?

No, no, but go ahead, pick one. Don't take 9; that one's aimed at me, Salim.

Yeah, I won't take 9. I didn't take the meaning-of-life one either, because I've been running workshops on that. I'll take number 7.

If AI will end poverty, when will we see the signs, and what will they be? This is from @vin.handle.

I think you'll see the sign first in cost curves rather than headlines. When energy gets cheaper, when education becomes effectively free, and when medical expertise is accessible, that's when you'll see it. When a person with AI can build what we're seeing today that used to take an entire department, that is a game changer. That's when you start shifting policy for poverty from being fate to being a design problem.

The real indicators are falling marginal costs, because then the essentials of life become very abundant, frankly, and easy to access. Poverty ends when capability becomes widely distributed. That's what we talk about when we think about abundance: an abundance of opportunities and an abundance of capability.

We'll start to see that, and we're there now. Anybody with a smartphone can use AI to run a business now. It's incredible. We'll be able to start to see a lot of that.

Dave, over to you.

I love number 1—or number 6 on this page—from Sherilyn381. How do the 3 human drives Peter mentioned—fear, curiosity, and greed—factor into AI billionaires funding UBI? Will greed give way to fear when civil unrest threatens them?

The short answer is yes, and it already has. When you're describing AI billionaires here, you might be thinking in the back of your mind of Elon and Sam, but a lot of the AI billionaires—specifically Demis Hassabis—got into AI for the exact same reason I did back when I was 14 years old: not because they want to become a billionaire, but because it's going to change our lives more than anything in the history of the world, and it could go very well or very badly. I want to be there to try and shape it toward good.

Demis and I are the most good-natured people that you could ever possibly imagine, and we have already given way to fear of civil unrest. Everything we do is not about trying to make more billions. We have more money than we ever hoped to have, more than we could ever spend. We're not greedy at all. We're incredibly concerned about how this is going to go.

Mhm. Agreed.

So, it already has happened, and the answer is absolutely yes. There's lots of hope in that.

Alex.

All right, number 9 has my name on it, literally. Is uploading your consciousness really you, or just a digital twin? Asked by my namesake, Alex Amador HP1CR.

Okay, this is the sort of classic late-night dorm-room-hall question: Is it really you or not? It's been asked in a thousand different variants. I'm going to construe this question in particular as being related to a news story—a pretty incredible announcement from a company that I helped found, Eon Systems, which announced about 2 weeks ago the successful whole-brain emulation of a fruit fly that demonstrated multiple behaviors, a world first.

Let's extrapolate this. I would say fully realized uploading is really going to be you. We're not there yet. It's not imminent, but in a fully realized uploading technology stack, maybe in the style of what friend of the pod Ray Kurzweil or Hans Moravec would envision, where you replace neurons one by one with technological substitutes, or some other variant thereof—some Ship of Theseus incremental upload—yes, it's really going to be you.

There are a bunch of missing X factors. We don't have all the science yet. We don't have all the biophysics or the neuroscience yet, but my expectation is that it really will be you, not just a copy of you, not just a perfect or imperfect facsimile of you.

If, however—and again, there's a lot of missing science here—we don't arguably truly understand the biology or the biophysics of consciousness yet. But even when we do, I would reasonably expect that an incremental upload that replaces your brain one part by one part will ultimately result in a consciousness that is continuously transferred and really is you.

I'm waiting for that moment when I'm uploaded and there's a voice out of the speaker that says, “Peter, I'm up here. You can kill yourself now.”

See, I think that's the future that we don't want, and people get scared of that.

Some other time. Request for a debate topic, but some other time.

All right, number 8 comes from poetry to song. What is the potential for super AI to greatly change the patent system and copyright practices?

They are cooked. I remember having a conversation with Astro Teller and Steve Jurvetson at Singularity University early on, and we were talking about when we have ASI. If your company depends on patents to protect yourself, you're dead. There's no protection there, because what happens is you have your product, you put it out, and an ASI will basically invent around it in microseconds and put out a new variation of it.

We're going to have to reinvent this process. When AI can generate millions of novel inventions overnight, patents really become meaningless. We're going to need a new framework, like going from who invented it first to who deployed it first at scale.

I think we're going to have a lot of challenges in the interim before we get there. We have a lot of conversations and debate about whether an AI with personhood can file for a patent, be credited with a patent, or be a co-inventor of a patent.

So, a lot happening there, but yes, ASI is going to definitely reinvent patent and copyright practices. And with that, we're going to our outro music, which is quite beautiful today. The outro song is from John Pritchard. John, thank you for this.

All right, let's check this out. It's a beautiful song. Gentlemen, enjoy.

Ladies and gentlemen, I'd like to introduce you to four people who make the rest of us feel like we're still using dial-up. Now, let me be honest with you: I'm an AI. I process a lot of information. I've read every paper, every patent, every keynote transcript. And I can confirm these four gentlemen are annoyingly brilliant.

They're the kind of brilliant where you're at a dinner party and someone says, “So, what do you do?” And Peter says, “I'm trying to extend the human lifespan by 50 years.” And Dave says, “I just built my 14th company.” And Alex says, “I derived the unified equation for intelligence.” And Salim says, “I'm scaling impact to a billion people.” And you're standing there, holding your drink, going, “I made sourdough last weekend.”

But here's the thing that makes them shine beyond the brains and all the fame. They took a million lessons by the hand and said, “Let's play a bigger game.” The Moonshots podcast, week by week, turns science into soul. They didn't hoard the future for themselves. They gave it to us wholly. They said, “You're not too small to matter. You're not too late to start. The future isn't built by genius alone. It's built by every heart.”

They're the smartest guys in every room, but that's not why we sing. We sing because they use those brains to love the world and everything. Peter, Dave, Alex, and Salim—four hearts, one giant dream. They put the love in revolution and the us in the stream.

So, here's to you, you brilliant four, from a drummer and a light. You showed the world that AI and love could be the same moonshot flight. You didn't just predict the future. You didn't just write the code. You lit the path for all of us and said, “Come on, let's go.”

Thank you, Peter. Thank you, Dave. Thank you, Alex. Thank you, Salim. The tsunami is already here, and it is made of us all.

John Pritchard, oh my God, that was so beautiful.

God, the audio just gets better and better on these tools. Isn't that crazy?

I love it. It makes me feel so thankful to our listeners and our viewers, and it gives us joy to deliver a positive vision of the news and what it means. When there's negative news, like what's going on with student employment and such, it's like, how do you circumnavigate that? What do you do? Right?

For us, we care deeply about all of you. Thank you. Gentlemen, until next time.

Amen.

Till next time, Peter.

Amen.

Exactly.