AI动态：NVIDIA创纪录营收、Elon在太空建设数据中心，以及Gemini 3的惊人性能

Just when you thought Nvidia couldn't do it again and again, we're now at 57 billion, 62% year-on-year growth.

We're in the beginning of a very long-term buildout of the fundamental infrastructure of humanity, which is computing.

Nvidia has just become the central bank for AI, and they're minting their own currency, which is compute. And everybody's got to buy their currency. There's got to be somebody who's going to challenge Nvidia.

The non-incumbents are obvious.

Saudi Arabia is positioning itself as a global AI superpower. That to me looks like we're seeing the rearchitecting from the ground up of an entirely new sovereign stack.

Since I was a teenager, I've been trying to visualize how the singularity happens and what does it look like in the last few years before the singularity. And now we're right in the middle of it and it's just I'm giddy with the excitement. Now, that's a moonshot, ladies and gentlemen.

First of all, guys, what are you doing tonight? Anybody know what they're doing tonight? I know what I'm doing tonight.

What are you doing tonight?

I'm watching “The Age of Disclosure” on Prime Video. Alex, you—

I'm already well into it.

You're okay. [laughter] That's right. It went live at midnight last night. Do you guys know what this is?

No.

See, it is a tell-all documentary about the fact that we've been covering up alien visitations and spaceships for the last 80 years. I can't wait to see it. I'm going to drag my kids along with me to watch this thing. It should be epic, Dave.

Well, now that you say it, I will. I didn't have plans otherwise, so I guess. Right. Well, the fact that, at the same time that AGI and ASI are coming online, there's increasing evidence of—I don't know. I just think, is it causation or causality? Or just—

Correlation.

Correlation is what I meant. Yeah, for sure.

Well, it's definitely tied to human events that trigger it. Every sci-fi movie knows this: when you hit a milestone, it triggers something, and then the aliens reveal themselves.

Okay.

I'll just comment as well. I mentioned in the last episode that, given enough superintelligence, any and all hidden agents become shallow. So we'll see what happens.

Yeah.

You know, Alex, I'm not big on the aliens, but I am very big on intelligences everywhere. I really love your theory on that. After we discover the nature of intelligence, we're going to find that you can compute with virtually anything, and it's happening all over the place. I just can't wait for that breakthrough, which is connected. It's not quite as—

I think it'd be interesting, once in a while, to have a couple of dedicated episodes on very specific topics. One could be, “What is intelligence?” Another could be abundance: what the hell is it, and how do you measure it? We could really go deep on a topic for a whole episode.

You know what I was thinking after the last episode, too? Peter, “The Future Is Faster Than You Think” is incredibly timely. Reprinting or rewriting it with you—there's so much that's changed, and it's only been a few years. That book is the one on my backdrop in the podcast studio in the office, and I just moved it to the front and center because I really want everybody to flip through it again. It's really, really prescient.

Thanks, buddy. The next book that Steven Kotler and I have written comes out April 14. It's called “We Are as Gods: A Survival Guide for the Age of Abundance.” It's a follow-on to the original book, “Abundance.” Anyway, it's probably the best work we've done, but I agree: “The Future Is Faster Than You Think.”

I'm not sure what the new title would be. [laughter]

Well, it's faster than you think.

Oh, no. Clearly, I'm not going to pull a Criswell on this. [laughter]

Hey, that outro today is unbelievable. I don't know—the creativity is crazy. We'll show it at the end of this episode.

Ask the audience during the episode to rebrand your books with video. Or just all of your content. If you want to flatter Peter, take any of his books or his TV shows—“For All Mankind,” for example—and redo any of that through Sora. We'll post it.

Yeah, sure.

All right, guys. Let's jump in. Our episode today has a lot that we didn't get a chance to cover in the last episode. We really went deep on hyperscalers, and it was a great pod. Today, we're going to continue the conversation. There's a lot that we didn't get a chance to cover, so we're covering it today. I'm excited for it.

We're going to start this episode today with chips and data centers. Here we go. NVIDIA beats earnings and has record revenues. Just when you thought NVIDIA couldn't do it again and again, we're now at $57 billion, with 62% year-on-year growth driven by AI compute demand. NVIDIA has announced 10 gigawatts of AI infrastructure deals, and Jensen is expecting $65 billion in revenue next quarter. Quarterly earnings are increasing at an extraordinary rate. Let's dive into this. Dave, do you want to kick it off?

Well, yeah, just to set the context for where we are: NVIDIA was a graphics card company doing polygons for video games. The AI community discovered that it was a very good fit for neural networks, and NVIDIA just caught the wave and took off. But they're only about halfway down the journey of redesigning the chips to be perfect for AI.

They can price up the chips tremendously, and there's another 2 to 10x performance gain in just the architecture and design that's still in their future. Every time they roll out an improvement like that, it's a chance for them to increase prices by 50%, 60%, or 70%, and the community will buy them no matter what. So that's what you're experiencing right now with NVIDIA. You could argue whether that's sustainable or not, but clearly there's more headroom in these ridiculously high margins that NVIDIA has.

Can I give you what I think the real story is? NVIDIA has just become the central bank for AI, right? They're minting their own currency, which is compute, and everybody's got to buy their currency. I think it's extraordinary. AWG, what are you thinking?

Yeah, I think that party can continue as long as revenue generation continues. I think 2 things need to happen for revenue generation to continue to justify the trillions of dollars of AI compute capex. Those are, first, automation of the existing service economy, and, second, creating transformative new markets through the discovery of transformative math, science, engineering, and medicine advancements. If we do those 2 things, the revenue generation continues, and the AI compute and NVIDIA capex party can continue indefinitely.

Part of the question is whether people think this is a bubble. NVIDIA is driving incredible revenues. My question is, who's going to challenge NVIDIA? There's got to be somebody who's going to start developing systems that are competitive. Any thoughts on that?

Well, I think the non-incumbents are obvious. There are TPUs from Google powering Gemini and, if you believe the reports, eventually lots of other first-class frontier labs as well. You have AMD, and you have a whole bunch of ASICs that are specialized in inference-time transformer compute. I think we are moving toward a heterogeneous ecosystem of lots of different architectures, not just CUDA.

Yeah.

Yeah. I'm not buying NVIDIA stock. I am invested in a company, Standard Kernel, which is building AI that redesigns kernels and, soon, chips to fit specific AI algorithms.

That's where the Google TPUs come in. Google has this vertically integrated approach: we innovate at the algorithm level, we roll out Gemini 3 days ago, and we immediately start designing the next-generation TPU to fit the algorithm change. That's something that Jensen doesn't have quite yet. Unless he invests very, very quickly up the stack, it's a big weakness in the competitive positioning because, right now, it's very hard to redesign a chip.

But with AI doing the redesign, that cycle time is going to come way, way, way down. It's not a manufacturing problem. You just have to crank out the masks. It's more of a debugging and simulation problem, which can be solved with AI.

Every week, my team and I study the top 10 technology meta trends that will transform industries over the decade ahead. I cover trends ranging from humanoid robotics, AGI, and quantum computing to transport, energy, longevity, and more. There's no fluff. Only the most important stuff that matters that impacts our lives, our companies, and our careers. If you want me to share these meta trends with you, I write a newsletter twice a week, sending it out as a short two-minute read via email. And if you want to discover the most important meta trends 10 years before anyone else, this report's for you. Readers include founders and CEOs from the world's most disruptive companies and entrepreneurs building the world's most disruptive tech. It's not for you if you don't want to be informed about what's coming, why it matters, and how you can benefit from it. To subscribe for free, go to dmmandis.com/tatrends to gain access to the trends 10 years before anyone else.

We've got a couple more stories on NVIDIA kicking us off here this morning. Let's dive into that. NVIDIA announced its strategic partnership with Anthropic and Microsoft. Anthropic just agreed to spend $30 billion on Microsoft Azure cloud, all powered by NVIDIA's latest chips. In return, Microsoft and NVIDIA are putting up to $15 billion in Anthropic.

This isn't a product launch; it's the formation of an AI power bloc. These alignments we're seeing week on week, right? The deck keeps getting reshuffled. Anthropic has been, I guess, under-resourced with compute, and so this is a power move by them. Dave, do you want to weigh in, or Alex?

I remember very clearly when Microsoft hit a $300 billion valuation and became the most valuable company in the world, and we were like, "Wow, that's huge." Now Anthropic is worth $300 billion, and a big company is now worth $3 trillion or $5 trillion. That wasn't that long ago.

You'll see that now Anthropic has this $15 billion plus another $30 billion. That's a lot of capability. You'll see them start to do things where you're like, "How are they getting that done? How are they getting that done?" But you really have to back up and look at the magnitude of these dollars. It's incredibly empowering, this amount of money.

Since I was a teenager, I've been trying to visualize how the singularity happens and what it looks like in the last few years before the singularity. Now we're right in the middle of it, and I'm giddy with excitement. This is one of the ways it plays out.

If you have true AGI to work with, you can do anything. You can win a Nobel Prize in chemistry like Dennis. You have unbelievable capability. If you have the capital to invest in the teams, it's wide, wide open.

I'm predicting a merger between Anthropic and Microsoft here. This is clearly them diversifying their bets beyond just OpenAI, and enterprise is where Microsoft wants to make sure they're really interesting. Anthropic is there, so I predict a merger at some point.

Interesting. I thought there were a lot of rumors out there. One of the rumors I heard was the potential of Google acquiring Anthropic, but you're right, I think this relationship sort of quelled that. There's a close relationship between Demis and Dario.

Yeah, I mean, because they're the leaders in the field, they're the ones who actually go home at night and play with the algorithms. They write the code themselves, and they're on a different level.

If you look at Jensen, Sam, and Elon, they're very much business dealmakers. They're engineers at heart too, but they're dealmakers at heart. They're not going home and tweaking parameter files. But these two guys are.

Which two guys? Oh, yeah.

Dennis and Daario.

Yeah, for sure.

I think I'd be surprised if Dario is willing to let Anthropic get acquired by Microsoft, unless Google just says, "We're not interested at all." But I really doubt that. I think Dario and Demis really want to stay as close friends and work together on this.

They're also the two leaders in ethics. Both guys, at the core, at the absolute bottom of their hearts, are concerned about ethics. The other guys might be too, but they're really entrepreneurs.

That is so Dave. That is such an important point. I've heard Demis and Dario talk about how this is the right thing to do independent of maximizing profit. This is what we have to do for humanity. Those conversations make you feel so much safer in the world when you hear that coming from leaders like that.

Totally right. If you don't believe it, read Machines of Loving Grace that Daario took the time to write. It's epic. Then listen to Demis' Nobel Prize acceptance speech.

And also The Coming Wave. Yeah, okay.

This is very inspiring. Peter, you identified this in Abundance, right? The tech philanthropists. If you went back 100 years, the richest folks in the world had incredibly extractionary business models.

What's incredible here, starting from Google and so on, is that these guys all have a deep sense of ethics, and they all want to give back. I think that's so inspiring. That feedback loop starts to get really incredible.

I'll just say one thing, and I want to turn to you, Alex, here for your masterful analysis, as always. One of the things I still feel, despite some criticism, is that Google is always taking the long view. They're always asking, "How do we help humanity?"

Their original motto was "Don't be evil." I remember those early days at Google with Larry and Sergey. Of course, it then became a real business, a massive business. But I still think that underlying current of "What can we do to make the world a better place?" drives a lot of their decisions, especially for Demis.

Alex, this alignment, this power bloc between Nvidia, Microsoft, and Anthropic—what are your thoughts?

There are multiple power blocs here. If you look at the larger picture, Anthropic is the last of the 4 or 5 frontier labs that doesn't have its own data center and chip-architecture play. There are other stories just in the past few days that Anthropic is finally moving into the data center space, finally moving into chip-architecture design space.

Going back to the earlier discussion, I think we do move to a heterogeneous future where there are vertically integrated players, like Detroit had the Big Three—the largest car companies. I think we're going to have the Big N frontier AI companies that are vertically integrated with their own data center design, their own GPU or equivalent chip-compute architecture, their own models, and their own applications.

It seems like we're moving to this vertically integrated future.

The big tech companies are now in a position where they could literally take over any industry anytime they want. The only barrier to them completely absorbing the entire world at this point is antitrust law. They like these partnerships without acquiring each other because it achieves their goal without triggering antitrust action.

If you look at all of them, including Google right now—and Microsoft a few years ago—they were stopped dead in their tracks by antitrust legislation. It's torture when they're subpoenaing and investigating you. They give you a notice that requires you to start saving every document and email. Just that alone is torture.

Really important point, Dave. Salim, you were going to say—

I want to go back to the bubble conversation because I think we're pretty clear here that we're not in a bubble, right? Can we say that categorically?

Yeah, I think there will be a correction of some type because so much capital is being thrown without regard, but I don't think we're in a bubble. Revenues and valuations are going up in parallel.

I think it's a function of revenue generation. If revenue generation continues to scale really quickly, it's not a bubble. If revenue generation fails to scale, it looks bubbly.

I'll make a prediction. I think there'll be a slight correction as these guys do what Dave is talking about and start going after other industries, which will take a little bit of time to figure out and penetrate, and then it just goes vertical again. It's like a Gartner hype cycle.

I think this is also a hedge for Microsoft against a full dependence on OpenAI, right? This relationship they're getting with Anthropic—

All right, let's go on to our next article here. Again, an Nvidia article: Nvidia may shift from GPUs to full AI servers. Let's go to you, Alex. What's the story here?

I think the bigger story is that, for the first 60 or 70 years of electronic computers, the form factor of computers got smaller and smaller. We went from mainframes to minicomputers, if you remember those, to microcomputers, to PCs, then to smartphones and maybe wearables.

That's a pretty monotonic trend toward the form factor of standard computers getting smaller and smaller. It probably peaked sometime around the 2000s or 2010s. But now I think we're actually seeing that cosmic calendar of compute form factors reverse itself.

The key form factor of the most important computers is arguably now starting to look like these AI data centers, these coherent superclusters, and they're getting larger and larger again. I attribute that to Moore's law ending, as sad as that is, and the beginning of horizontal exponentiation.

Seeing Nvidia generalize and vertically integrate from just offering the chips to full AI servers is actually a reflection that the form factor for computing is no longer getting smaller. It's just going to get bigger and bigger and bigger.

Is there a name, Alex, for the next version of Moore's law? Is it the AI scaling laws?

If you ask Ray Kurzweil, he'll talk about the law of accelerating returns, and he'll point to a much broader trend that generalizes beyond electronic computers or CMOS.

But is there an equivalent for the GPU world that we're living in? Are we going to create Wang's law or Jensen's law?

There are dozens of experience curves. If you want to name one after yourself preemptively, you probably can. The naming rights are still open, but there are so many experience curves now, including the one that we were chatting about for the past couple of episodes about the hyperdeflation—by 40× year over year—of the cost of intelligence. There are so many new experience curves coming out of AI.

I'll make one on the spot, though. It's the law of abundance, right? Technology takes something that's scarce and makes it abundant.

But it has to be quantitative, Salim, to get the naming right.

Fair enough.

I think the short answer to Peter's question is no, and it's wide open. If somebody out there listening wants to name one, it has to be simple and quantitative, like Alex said, and it has to be accurate. It has to last a while. That's a good challenge.

All right.

For me, this was, by the way, an iPhone thing, where the iPhone succeeded by owning the whole stack—from design to experience to hardware, all the way down—or trying to get the whole ecosystem. This is NVIDIA trying to grab the whole ecosystem, create more value, and aggregate more value in one spot.

Agreed. All right. The next article here is a fun one: xAI will be the first customer for an NVIDIA-backed data center in Saudi Arabia. A lot is going on here, and I think basically Saudi Arabia is positioning itself as a global AI superpower.

This is part of their Vision 2030 play and their commitment to $100 billion. They're trying to diversify beyond oil. I've been going to Saudi at least once a year, sometimes twice a year, in my board role at FI, and the conversations are very clear. MBS, who runs the nation—the Kingdom of Saudi Arabia—has wanted this. Their goal was to be number 2 behind the U.S. in AI. Obviously, with the U.S. and China, maybe it's the number 3 position, but this has been at the top of the conversation across the leadership of the Kingdom of Saudi Arabia for years.

This is a massive commitment, and this also brings Musk what he needs most, which is massive GPU capability. We'll talk about that. Alex, let's go to you first.

Yeah, I think the story behind the story is, as I've said before, inference-time compute. This AI data center capex is just the opening act for an entire stack that's about to pop out, including AI-powered humanoid generalist robots.

When I see stories about countries positioning themselves for sovereign inference-time AI compute, that to me looks like we're seeing the rearchitecting from the ground up of an entirely new sovereign stack. It'll start with the data center inference compute, it'll run through humanoid robots, and you'll see, I think, Saudi Arabia and other countries—probably beyond just demanding sovereign inference compute—starting to demand sovereign compute for humanoid robots as humanoid robots start to be more broadly deployed.

All of societal automation, to the extent we have sovereign countries in the present paradigm, wants this entire stack to be sovereign, at least at inference time.

Well, there's a long history of the U.S. needing international partners. In biotech, we can't just try everything here; the FDA is way too slow. So you roll out in Panama, you roll out in Caribbean nations. In chemical manufacturing, India just took off long before the U.S. because regulatory barriers make things so slow here.

So, in sovereign AI, Saudi is going to move much faster. I think Bermuda also is going to move much faster. They'll learn, and everyone's like, "Well, who cares about sovereign AI?" At the rate of technical change, we need a much faster process for creating laws and rules.

I think Saudi will be a great test bed because they're so nimble. They can make a decision and act on it instantly. You see this with simple things like health data.

Saudi is miles ahead in health data. Trying to do anything under HIPAA in the U.S. takes way too long.

Can I tell you what I think one of the underlying stories is here that I want to hit on, and then, Salim, I want to hear from you on this? It's that xAI needs compute. xAI needs massive amounts of compute for Elon's vision to happen, and they've got Colossus 1. They're about to bring Colossus 2 online.

But if you look at the players out there, Google by far is the number 1 owner of compute on the planet, with its TPUs and GPUs. OpenAI is next. xAI is coming in third. So this relationship has got to be critical for Elon as he's standing up Colossus 1 and 2 in Memphis and Mississippi. Any thoughts on that, gentlemen?

Well, one takeaway from our trip to Saudi 2 weeks ago is that I had gone in thinking it was just going to be a massive amount of compute all over the desert—huge data centers. It's actually much more of an investment vehicle.

I think they're building big data centers to learn, and their goal is to invest $1 trillion as intelligently as possible. But I think you'll see later in this pod that the data centers are more likely to go to Texas and then space, and Saudi will be along the way. It's not going to be like Saudi is the compute capital of the world. It's more of a learning experience, a test bed, and a sovereign, controlled subset of compute.

I've got nothing. There's this sequence of these stories that just run from one to the other, each one an order of magnitude more crazy, and my brain is just blurred and fogged up with all of this.

They're all amazing and significant, but the broader implications are something that I'm more interested in. For the specifics of this, let's move on; you guys have covered it way more.

Yeah, one of my questions is: xAI is—I mean, they're all compute-starved, right? So xAI in particular needs NVIDIA chips, and he is getting them left, right, and center.

I think one of the things we're seeing is Elon getting access to NVIDIA chips very early as he's been building the Colossus clusters. I'm just wondering—and there was a rumor I saw out there about Elon potentially buying Intel, right? I mean, that would be a fascinating move by him.

Yeah, that's one that actually—well, anyone buying Intel would be a genius. But he might actually get that through the regulators because he's not Google. Google could never get away with it, and certainly NVIDIA could never get away with it.

Could you imagine the world if Elon was running Intel just to accelerate it? To literally move it at warp speed? That would be amazing.

That would be incredible. Keep in mind also that Elon's getting tons of Jensen's chips, but he also signed up for $16 to 45 billion dollars of Samsung manufacturing to make his own chips.

So that's not exactly a single path. They're going down a couple of different paths, some of which are competitive.

If you pull back one second, the multipolarity of this is the most exciting thing for me. They've got major different parties doing amazing things, and that means the whole rising tide lifts everybody. I think that's really great.

Yeah. Alex, do you want a final comment before we move on, or are you good?

I'll just say there is a race to superintelligence, and we're going to get abundance from superintelligence. Each one of these stories is, just as Salim says, a facet of this common theme.

Nice. All right. This is one of my favorite stories of this week, so I want to dive into this. I care about this deeply, and I know Alex, you do as well.

This is Elon talking about 100 gigawatts per year of data center installation in orbit. Let's listen to the video, and we'll go from there.

Like, we see a path to putting 100 gigawatts per year of solar-powered AI satellites into orbit and having this be the lowest-cost way to power and operate AI at a very large scale. For reference, the United States consumes roughly 460 GW on average per year, because the average power load in the U.S. is 460 GW.

The whole country.

All electricity from all sources in the U.S. Yes, and you're talking about 100 being added.

Well, roughly a quarter of the U.S. electricity output.

We have a plan mapped out to do that.

It gets crazy.

It does get crazy, but that's 100 GW per year, not in total.

Yeah.

Do we have a timeline for that?

Well, listen, Starship has to become operational on a regular basis, which I think we'll see in the next 12 to 18 months. You're then basically taking compute off-planet on Starship.

But what I find even crazier, and Alex, we can talk about this, is Elon's plan to mine the Moon for silicon, right? So we're going to mine it for compute and solar power.

I think his estimate of what he could do if he was launching using O'Neill-like mass drivers off the surface of the Moon is to go from 100 GW per year to 100 TW of energy per year, which is 5 times the total energy output of the Earth.

As I've said, the moon's had it coming for a long time.

Time to make use of that.

Time to pay up. I mean, just to take a step back, it's remarkable how quickly we've moved from the Dyson swarm being science fiction to, “Okay, we've deployed an H100 to orbit,” to now deploying multiple 100-gigawatt data centers in a swarm in LEO, GEO, and probably soon solar orbits.

I think the Overton window just zoomed by on the Dyson swarm. I think we're now moving to a regime where we potentially have to worry about interoperability between multiple competing Dyson swarms, maybe even thinking about an Internet. We've talked for decades about an interplanetary Internet. While there were limited efforts in that direction, I think we're about to hit a regime where we really do need—

Vint Cerf was designing that version of it. Yeah, the interplanetary Internet. Yes, for sure.

Wait, can we go back to the timeline?

Exactly. Let's come back to the timeline, because I don't think it's as long. We talk a lot about robotics and refactoring construction all over the world, but that's a pretty long, 15-year timeline. I think this is actually sooner than people might think.

I'm a huge believer. I've been studying it since—I didn't believe it a year ago. I was like, “Why on Earth would you want to put a server in space?” But now, if you think about the cost of compute coming down 1,000 or 10,000 times, is there any barrier to that? The answer is no. There's no barrier whatsoever other than power.

And power is free and abundant in space. And then the only barrier after that is radiant cooling. But they've they've you know we have our first 100 in space.

There's work to be done on cooling the systems. And just for a reference point, if you take into account dayight cycles and the you know the solar flux in orbit and other factors, it's about potentially a six-fold increase in energy density in orbit versus on the ground.

Can we just slow this down for our listeners for a second? In the first phase of this, we're talking about using Starship to launch this next generation of Starlink. Elon has already built the largest satellite-manufacturing capability on the planet. Starlink has about 10,000 satellites, and they're going to Version 3 very shortly. Those will be launched on Starship.

The next iteration is going to be putting whatever the top chips are in orbit, powering them from solar, and then the biggest challenge, Dave, you nailed it, is how you cool them.

You know, one of the things that is also underappreciated as we're moving forward with millions of autonomous cars that are imaging everything on the street, where we have millions of drones flying over your head that are imaging everything at millimeter resolution, where people are walking around with their AR glasses looking around, when we have, you know, thousands of satellites in orbit imaging at submeter resolution, we're entering a point where everything knowable on the planet is being imaged and recorded constantly. Which leads to a point at which you can ask any question and get an answer. This is a very different universe where you can know anything you want, any time you want, anywhere you want.

I remember the statistic that each Waymo car is recording a gigabyte of information per second per car. You can do radiative cooling, but there's no material, no atmosphere, and no liquid water to carry the heat away. Heat is one of the biggest issues you've got.

Space is absolute zero. I mean, it's got to be easy to create that mechanism, right?

Take it away.

Hold on. Let me ask one more question. So basically, you get a satellite up there, it unfolds a solar panel that collects the energy, you do the compute on it, and then beam the information down. Essentially, that's what we're talking about.

Yes.

Yeah, exactly.

And beaming the data back and forth is easy, easy, easy. A lot of people think, “How are you going to get the data up and down?” That's lasers, and it's trivially easy. That's not a problem at all.

Let's go to our resident super genius here. Alex, talk to us about radiative cooling.

There is this misconception that you can't operate a thermally intensive data center without conduction-based cooling. It's completely untrue. The cosmic microwave background is about 3 Kelvin. The universe, on average, is pretty cool.

The trick is that you want to make sure you're radiating in the right direction. You don't want to try to radiate heat in the direction of the sun, for example. That won't help you from a thermodynamics perspective. But as long as you can make sure that you have radiative cooling in the direction of the cosmic microwave background—which is most of the sky—and aim your radiative cooling toward the cooler directions, you're fine.

You can use radiation rather than conduction to radiatively cool. There is a bit of a navigational issue with making sure that you're pointing in the right direction and oriented correctly, but this is completely doable. To answer the timeline question, I would be mildly, if not significantly, shocked if, 10 years from now, conservatively, we don't see low hundreds, if not many hundreds, of gigawatts of AI compute not on Earth's surface.

Yeah, I think Starship is the means. To be clear, this is an attempt to move the bottleneck of AI from Earth's power grid to SpaceX's launch grid. In other words, all of a sudden, it's not power on the ground; it's how many launches to orbit we can get.

The alignment that Elon has built in his ecosystem of xAI, Tesla, and SpaceX is extraordinary. Is it love, or is it genius?

An extremist. Maybe both, in some quantity. But, in extremis, this involves taking apart the solar system, which is perhaps the most tantalizing part.

You are dying to nail the solar system, Alex.

He's going to accelerate, right? But—

Disassembling the moon is just a milestone. If Jupiter isn't decompiled, then there's something wrong.

To our listeners, if you love looking up at the sky and seeing the moon, don't worry about it. Enjoy it while it lasts.

Okay, but let's actually talk about that part. Gerard K. O'Neill, a professor of physics at Princeton University whom I consider a mentor, passed away far too early. He wrote a number of books and papers about how to mine the moon.

One of the things that's true about the moon is that there is no atmosphere and there's a lot of solar flux. He came up with the idea of creating these mass drivers—basically electromagnetic guns—that you could put something into at one end, accelerate it to lunar escape velocity, and shoot it toward Earth.

What Elon has been talking about here is basically the idea of setting up satellite and data-center manufacturing on the moon and launching those using railguns into Earth orbit. His objective is 100 terawatts of capacity per year.

One thing that was a big shift in the last couple of weeks is that the first H100 went into space 2 weeks ago. It's operating and cooling using aluminum only as the radiator. Prior to this, everyone was thinking you needed obscure metals and whatever, but apparently it's working with aluminum only.

If you read Peter's books, the first chapter of Abundance—the greatest story ever—please read it. 7% of the Earth's crust is aluminum. We have enough aluminum oxide, so it's a great story to read. Anyway, Pliny the Elder—that's just such a good story.

Thank you for listening to that.

Well, that's a big breakthrough. If you want to make money today on space-based compute in the future, we have to get the weight down. We have to get the mass down.

That first H100 is 50 kilograms, which is way too much. Most of that is sensors and other things. It'll be very easy to get its weight down. Invest in whatever it's going to take to get the mass down so that we can launch these more cheaply. Eventually, we'll make them on the moon, but that's way out there right now.

A data point that I talk about a lot is that when we were launching space shuttles, it was, what, $600 million per space shuttle launch?

Yeah, about $1 billion per launch.

SpaceX dropped it to around $50 million to $100 million. Then Relativity Space, where Eric Schmidt is now the CEO, plans to do it for about $5 million to $7 million per launch.

I find that interesting. That's about a 100-fold drop in a domain. This is not a Silicon Valley social-media gaming play. This is real physics trying to get out of Earth's gravity well. Even there, we've seen a 100-fold drop. What can you do with 100 times more capacity? It'll go another order of magnitude.

I was in a conversation with Elon a couple of days ago when he was briefing some xAI investors. I'm not disclosing anything confidential, but one of the things he said is that in the future, if you could set up a Starship launch facility on Earth next to natural-gas production capacity, because that's what it's basically burning—methane—and you could then use solar power to pull oxygen out of the atmosphere, the fuel would effectively be free.

His estimate of the cost of transport would be that it would be cheaper to go to orbit as an individual seat ticket than it would be to fly transatlantic as an individual. This is an incredible vision he's been building.

I want to put one more figure out there for our listeners. We talked about 100 gigawatts per year of capacity being launched in the next decade or so. That's the equivalent of 100 nuclear power plants. These are typically 1 gigawatt in capacity. I mean, that's awesome.

All right.

I'll tell you what else is awesome: The people who are making these choices are math, physics, and computer-science geniuses.

If you look back in the history of business, Jack Welch—he's great—but John Chambers, these were the people running billion-, hundred-billion-dollar budgets, but they didn't have that background. And so now, when Elon talks about these things, he's almost always right. In fact, so far, he pretty much always is right because he looks at first principles, actually does the math, and says, “Look, that'll never work, but this actually will.” It sounds crazy—

But it actually will work. There's no fundamental barrier. So, it's just a different community making these choices than ever before in history, and you're seeing the things that are possible actually starting to happen.

You know what? Saturn has its fate marked as well.

Oh no. Stop taking apart our solar system.

The solar system is a dead mass right now. We've got to fix that.

Oh my God. Computronium. We're turning it all to computronium.

You can't actually disassemble the Moon. And it controls the tides.

Not only the tides, but also the molten core of the Earth. The magnetosphere requires the Moon. We can't disassemble the Moon. You can do Saturn. No problem. I have no problem.

Thank you, Dave. Thanks for defending our sisterly body here.

All right, let's jump into the topic of energy. There is—we talk about compute as energy and energy as compute. So, I put this chart here for you, Salim, principally. Here we're seeing China driving global electricity generation. On one side of this chart, if you're listening and not watching, is the change in electricity generation over the last 12 months, from September 2024 through September 2025. What we see here is China basically 4× the U.S. in terms of total energy production. We're seeing China leading the world in solar, wind, nuclear, and gas.

Another chart next to it shows the change in electricity generation just by itself. We see again that this is China, basically putting out around 325 terawatt-hours versus the U.S. putting out about 80 terawatt-hours. Salim, over to you.

There was one little piece of this that I found very exciting, which was that big red drop in coal.

We've been predicting this for a while. The cost of solar is now so cheap. I've mentioned before that the capex and opex of solar is now cheaper than just the opex of fossil fuels. What we're now seeing is that the economics mean we'll start to dismantle all of our fossil-fuel facilities unless we need incredibly dense energy generation. For most normal use, we'll be doing that and then moving fully to solar. I thought that would be hugely powerful and important for climate change and all the carbon extraction stuff that we need to do.

I just think this is a shocking difference in energy-capacity production between the U.S. and China. I mean, kudos to China for having really gone all out. I'm kind of shocked at this point that we haven't seen President Trump stand up and say, “We need Operation Warp Speed for energy.” We're seeing it in different places. We're seeing regulations change.

Yesterday, they said we're going to make all this drilling in open water available, and there's a big paradigm shift that's missing around the power of solar energy.

Well, solar and fusion and Gen 4 nuclear—all of these things. We've got to be competitive if the U.S. wants to be competitive in the long run, unless we skip this entire decade and go straight to orbit, right? That's the alternate. Alex, what are your thoughts here?

Yeah, I think we have seen an Operation Warp Speed for at least certain energy categories, including nuclear and maybe certain fossil fuels. I think the increase toward, call it, Kardashev Type II civilization-level energy production is going to happen with time. Barring some shocking, ontologically shocking discovery about the nature of our universe, it seems like we're on trend.

Whether one nation is temporarily in the lead or another is temporarily in the lead, I think the long-term trend line is fission happening, fusion is going to happen. Maybe there will be even more efficient ways to recover energy. As I like to point out, if we had microscopic black holes, we could just drop objects into them and recover their rest mass. That, in principle, would be far more efficient than fusion power.

Fusion is about 4% efficient. Fission is less than 1% efficient. We could recover nearly 100% of the energy associated with rest masses if we could just drop objects into black holes. So, the known physics of our universe allows enormous amounts of energy production. I think whether one country is ahead of another in the short term, these are just nonsecular—

Blips. But, Alex, we are electricity-limited in the U.S. We're not chip-limited. We're not real-estate-limited. We're electricity-limited.

And I'm just—

For the moment.

For the moment. I mean, this is a blip. Again, fission is in the process of coming online. Fusion is in the process of coming online.

Okay, I'm just calling it like I see it. When I see the timelines for fusion—for real production of fusion—the earliest is going to be 2030 to 2035 for the first plants, and then mass production really until 2040. That's 15 years out. That's insane. And the timelines for even bringing existing fission plants back online are like 5 years.

Minimum.

Exactly right, Peter. The gap is massive. We're fine through 2030 because we're looking for 100 gigawatts, and we can't make the chips any faster anyway. So, between now and 2030, we're fine just taking old manufacturing power and redirecting it, just like Elon did in Tennessee—redirecting it toward data centers.

But then from 2030 to 2035, we have a massive gap because the chip fabs have accelerated like crazy. Fusion kind of works, but it's not online yet. And then there's this big gap from 2030 to 2035. Keep in mind, as Alex is always pointing out, we're going to discover brand-new physics and math between now and then. Anything could happen.

Yeah. I just think solar is the easiest to scale, other than oil and coal.

I'm predicting we'll see a massive breakthrough in photonics or some of those domains over the next couple of years.

But we're in for a 5–7-year difficult period, and I think the only option is to steal all the energy from residential.

Yeah. Don't go there.

I think—

That's the only option.

There are lots of options. I would also say, don't underestimate the power of the market. If the party continues, the capex party continues, and revenue generation from killer new AI apps continues, we have the ability to reprogram the existing electricity production of our civilization to extraordinary degrees. If the market absolutely demands it—and I think we'll know the answer to that in the next few years.

You know, we should have Saul Griffith on. He probably has the most macro view of all this stuff.

Yeah, Saul is great, as is Ramez Naam in this area.

All right, I'm just saying we need to scale up solar in the U.S. Perovskite is coming. I'm super excited about perovskite as a technology for higher energy-conversion rates and lower costs.

All right, let's move on to this next article here. Fascinating. “Google's investments run deep in the heart of Texas” is the concept here. Google announces a $40 billion investment in Texas through 2027 to build new cloud and AI infrastructure. The project is adding 6.2 gigawatts of new energy generation and a $30 million Energy Impact Fund.

When I see 6.2 gigawatts of energy and remember 100 gigawatts per year in orbit, these things just warp my mind. Space is beckoning and calling. But here's what the real story is for me: Google isn't just going to Texas for sunshine, because part of this investment includes renewables. They're going because AI is about to become the biggest consumer of electricity in the U.S. It's going to be bigger than steel, bigger than crypto, bigger than every industry combined. And I think Texas is the only state that can build fast enough in the energy world.

Well, and they have space.

And they have—yes, they have open area.

And a friendly regulatory environment. Texas is doing a great job of welcoming energy, data centers, and AI leadership. I would love to see other states—

Launch vehicles.

Yeah, launch vehicles. I would love to see other states provide as welcoming an environment for acceleration as Texas is doing.

They'll have to. The competitive nature of this will open this up.

I don't know. I wish that were true, but if you look at what's actually going on in California and elsewhere, it's like, come on, guys. I mean, I don't know.

Yeah—

The government's so messed up.

No, but this provides a natural competitive opportunity for many, many states that have a difficult time competing with, say, California and New York.

It's one of the best parts of the U.S., having different state legislatures. The United States of Texas is definitely pulling in some great opportunities here.

There's one problem with the U.S. right now: you have to drop the word “United” off everything.

It’s not really…

Yeah. But that’s been true since the 1700s. As long as people don’t get violent, the variety is actually very healthy, and the internal competition is very healthy as long as it doesn’t tip over to dysfunction, which it does every now and then. I agree.

We’re founded on freedom and variety, and on a weak central government and strong local governments. Trends go in the other direction all the time, but I think Texas running away should put competitive pressure on other governors, and the governors in Ohio and Wyoming have reacted already. So it does work. It’s just the pace that’s frustrating.

This episode is brought to you by Blitzy autonomous software development with infinite code context. Blitzy 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 Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy 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 Blitzy 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 blitzy.com to schedule a demo and start building with Blitzy today.

This next article is a really fun one. It’s an important one. The title is “X-energy begins construction of the first Category II nuclear fuel facility.” Kudos to Kam Ghaffarian. Kam is a friend, and he’s on my board at XPRIZE. He’s the chairman and co-founder of X-energy.

This isn’t about creating a new facility—a plant for producing energy at a nuclear plant. It’s about creating the nuclear fuel, something called TRISO, tristructural-isotropic fuel. It’s an advanced nuclear fuel, one of the safest and most robust ever created, to power the next generation of SMRs. Alex, let’s go to you on this.

Yeah. So the headline is: pebble-bed nuclear is finally happening. After decades of people hand-wringing and asking, “When are we going to get pebble-bed reactors?” we’re getting pebble-bed reactors.

So, what is a pebble-bed reactor, Alex?

What is a pebble-bed reactor? Think of a pebble-bed reactor a little bit like a gumball machine, where you have billiard-ball-sized spheres that have uranium cores, or lots of particles in them with uranium cores, surrounded by a carbon-ceramic composite. The basic idea of a pebble-bed reactor is that you have all these billiard balls. They’re being heated by radiation from the uranium cores. You pass helium gas through the gumball machine, as it were. The helium gas is heated by radioactive decay, and that powers a steam turbine, which generates electricity.

Critically, this is much safer than many other forms of nuclear reactor because the individual uranium cores are nicely encased. Over time, you could imagine this gumball machine where these spheres fall. They decay over about 3 years. The spheres are lowered into one gumball machine, then offloaded into another and recycled a few times.

But it’s these pebbles—these gumballs, if you will—and the way they’re packaged up that make this safer. This is what our country’s first Category II nuclear fuel facility is about in this story. That’s the announcement here: we’re finally manufacturing these pebbles domestically. This is, in principle, far safer than the fuel-rod-based earlier generations of nuclear reactor.

Yeah, I think one of the key points to make is that these don’t melt down. They literally cannot melt down as a nuclear reactor, which gave us Fukushima and Three Mile Island and all of those failed earlier versions. Again, kudos to Kam Ghaffarian. This is first-principles capability driving us toward a new generation of nuclear.

It’s unfortunate it’s taken us this long. My father worked on pebble-bed nuclear reactors. Really?

Yeah.

He was working on that in the ’80s, and he was going crazy, going, “Why the hell aren’t we using this for everything?”

This has been decades coming. We spoke in a previous episode about thorium reactors, which go back to the Manhattan Project. There are so many concepts for nuclear energy and energy in general that have been sitting on the shelf, in some metaphorical sense, for decades, if not almost a century, that are finally only now being put into production. Wow.

It’s not often I can see a term from some of this obscure scientific stuff that I recognize, but this one I saw.

Wow. You know, this is a really important point here. AI is the string pulling everything forward, right? AI is driving all of the technology we’ve been thinking about forever. It’s driving us into orbit, it’s driving us to go back to the Moon, and it’s driving us to build global energy infrastructure.

I prefer the analogy of a gravity well, but I’ll go with string. Riffing on that point, Peter, my modal scenario for the next 10 to 20 years is that most science-fiction concepts all happen at once. We’re going to live in a future that’s not just like a Star Trek future, not just like an Accelerando future, not just like Heinlein or Asimov. They’re all going to happen more or less at once, in the same universe—our universe—over the next 10 to 20 years.

What you said, Alex, still rings in my mind. We’re going to speed-run the Star Trek universe in the next decade, the Asimov universe and Heinlein and Charles Stross. All of these are going to happen.

Maybe we need to rename the pod to “Everything Everywhere All at Once.”

Yes. Oh my God.

I think that one’s taken.

But what an exciting time to be alive. I get it—yes, there are problems on the planet—but also, what an extraordinary time to be alive. During what period in Earth’s history would you ever prefer to be living other than now, other than perhaps tomorrow? Dave, you want to close us out on this one?

I’m just going to repeat Alex’s point: the innermost loop of humanity is what’s driving this. It’s AI at the core. It’s the innermost loop, and everything around it is accelerating because of that innermost loop. It’s just so exciting to watch. All right, let’s jump into drones and robots. A lot is going on in this universe.

I pulled this particular story out because I love this company. This is a company called Zipline that’s about to begin producing its delivery drones at 20,000 per year. This is Keller Clifton, the CEO of Zipline, an amazing entrepreneur who built this company against all the odds. Let’s listen to Keller, and then we’ll talk about it.

Everyone, it has been an insane 2 months, but I thought it’d be cool to give a 2-minute update. I’m standing in our expansion space for the manufacturing facility as we speak. We’re getting ready to build 20,000 autonomous aircraft a year, all here in South San Francisco in the United States.

Right now, we’re growing the number of deliveries we do per day at around 15% week over week. We’ve been growing that fast for about 30 weeks straight. A lot of our customers out there are placing orders 3 to 4 times per week. In fact, some customers are ordering 3 times a day. People have fundamentally changed their ordering behaviors. Some people are grocery shopping once every 1 to 2 weeks and then ordering from Zipline 3 to 4 times a week just to do fill-ins.

We’ve also been able to launch a new Walmart Supercenter every week across Dallas over the last couple of months in the US. Right now, Zipline is doing an autonomous delivery about every 30 seconds. We have one very big announcement that we’re expecting to come out in about 10 days.

All right. An incredible story behind Zipline. They were founded in 2014 and are headquartered in South San Francisco. This was a vision that Keller put forward. I did a Moonshots recording with Keller; you can go back in our library and find that podcast.

I remember him saying that when he started the company, he told his employees there was less than a 10% chance of success. They started in Rwanda and Ghana. It was a sort of regulatory arbitrage: flying drones to deliver things in the United States without the support of the FAA and DOT wasn’t going to happen, but in Rwanda and Ghana, there was a real problem. They focused on that problem, which was delivering blood and critical medical supplies to different parts of the country from a central repository.

They screwed up in the beginning. It was difficult, but they got better and better and better. Finally, they were operating at such a high level of success that they were able to bring those operations back. By the way, they operated in Rwanda and Ghana from their headquarters in South San Francisco. It’s an extraordinary story, and today they’re doing about 1 million commercial deliveries per year. They’ve flown 70 million to 100 million autonomous miles.

They’re doing this for a whole bevy of companies—Walmart, Chipotle, and others. It’s sort of a 30-minute retail-to-delivery capability. Alex, excited about this one.

Drone delivery is happening. It's happening here in the United States, and it's happening in China. I think it's an interesting thought experiment to ask what happens when this is fully realized, when this is a fully mature technology. I think it leads to, ultimately, a localization of the supply chain.

We talked earlier in this episode about sovereign stacks. If you extrapolate the ability to do drone delivery of supplies and supply chains, I think ultimately we're going to find ourselves in a regime with hyperlocal manufacturing. I think this is very exciting. We asked for flying cars. I think that this is, in microcosm, the first version of flying cars, where in the near-term future, optimistically, we find that our skies are filled with delivery drones.

A couple of points here. One is that this is such a big thing. This is one of the Gutenberg moments that changes everything, right? AI changes everything, but drones really do change everything. The last number I saw was that we're saving about 1,000 people a year using drones because we can use thermal imaging to find people in earthquakes and so on. We're dropping in lifesaving stuff to them.

It's going to grow exponentially as we propagate the drones. This is going to be very, very big over time, but it's already had a big impact. I remember in 2010, one of the Singularity University teams launched this idea of drone delivery, and then Amazon, a few months later, copied that and announced they were going to do delivery by drone. That has inspired a whole vector of this. It's really exciting to see.

Of course, Google Wing—Wing, which is part of Alphabet. I had the CEO of Wing on my stage at the Abundance Summit last year, I think it was. Again, Zipline—just a beautiful design. If you go back to the podcast I did with Keller about 18 months ago or so, you can see how they operate.

They can set up an automated launch location in your store, where the drone comes back, picks up the next order, and delivers it. There are a lot of drones flying in the air, and these drones, of course, have cameras and sensors. These drones are going to help create this imaging of the surface of the Earth at millimeter resolution when there are millions of drones flying in the air.

For many years, they were doubling every 9 months. The price-performance was doubling, so that's a hell of a curve to be riding on. Let's also acknowledge that China is already doing delivery of coffee by drones and all sorts of stuff. They're living in this universe already.

Yeah, big time. It's that last point Peter made that I really was hoping Peter would do a relaunch of The Future Is Faster Than You Think, one of his best-selling books, because it talks about this topic, and it's all about converging technologies. But that was before neural-net-based vision and feedback-control systems were perfected, which was really just in the last year or two. That's a total game changer in that whole thesis.

Yeah, I talk about him in my next book. Yeah, for sure.

Oh, do you? Okay. Well, he's doing Tertill now, which is basically a gardening and farming robot. I'm sure he'll expand it out from there. I had dinner with Rodney Brooks, and he was really down on the supply chain for robotics. He was like, "What happens is we invent it, it's great, and then the Chinese come and clone it, and then they undercut your pricing." I thought, well, okay.

That's what happened to iRobot. I get it.

But that was before the vision and feedback-control systems. We also have a much more protectionist economy now, which is another story. Your ability to launch a robot that does something very specific is a completely different world today than it was 2 years ago in terms of its dexterity and its vision.

It's not just vision; it's any kind of sensor. Now you can train a neural net very, very quickly, and once you have the data—you know, this came up with 1X when we were talking to Bernt Ørn—the data that comes back from this deployed set of robots gives you a huge flywheel effect and a competitive advantage.

If you get your robot to market in any given use case, just like Zipline is doing, then you can use that to retrain the neural net every single night, and it just gets better and better and better. I think people are underappreciating the degree of dexterity and the degree of capability that's in these new generations of robots, and it should absolutely take off.

I remember the statistic that each Waymo car is recording a gigabyte of information per second per car.

Yeah, it's crazy. Alex, can you give us some wisdom on that idea?

I think this is the planet waking up. I speak about decomposing the solar system to build a Dyson swarm. I think a baby step in that direction is deploying small masses in Earth's atmosphere to transport other masses around.

If we're on this trajectory—and I think the jury is still out as to whether we are—but if we're on the trajectory toward a Dyson swarm, it's very natural that, in the intermediate term, we'll see drone delivery and all of these masses start to move around in Earth's atmosphere.

Can I make a couple of quick points?

Sure.

There's the negative side of it, where people go, "Oh my God, loss of privacy, et cetera, et cetera," right? But there's also the unbelievable positive side, where there's an unbelievable amount of illegal fishing going on, for example. We're literally scraping the life out of the oceans, and now we can track that and police it a lot better because we just know. I think there are so many positive use cases that they outweigh the negative by a long margin.

This is a point I've made. People act differently when they know they're being watched. This brings an entire conversation around police states in China and so forth. But I supported the Lindbergh Foundation—this is Charles Lindbergh, now Eric Lindbergh, his grandson—and they were funding drones that would fly over herds of elephants and rhinoceroses to keep the poachers away.

Because the poachers knew they were being imaged, they stayed away from illegal activities. Anyway, there are a few more robot stories I'd love us to chat about.

Can I talk about that topic, just for the poaching topic? Yeah.

There was an idea a few years ago, which I think is now starting to be implemented, where if you use synthetic biology to create rhino horn powder, you flood the market with super-cheap rhino horn powder, and then you take away the economic incentive for poaching.

Yep. Awesome stuff. Okay, this is Unitree's G1 learning to do humanlike chores. I'm showing this video just because I want people to start to imagine what it's going to be like to have these robots living in your home with you. Again, this is a $16,000 robot. Maybe this is a little more advanced version.

Vacuuming the bed?

No, I think it's ironing.

Ironing.

Ironing the bed. Yes.

Ironing the sheets. Okay.

That's a service I don't really need.

Don't knock it until you've tried it, I guess.

Yeah. Well, that's a good point. Actually, all kinds of things I don't feel like I need are going to be so cheap.

I just don't want to be in the bed when the robot tries to iron it.

Unitree is about to go public. It's probably the leading manufacturer of robots on the planet in terms of volume. Again, I would never bet against Elon. He's convinced that what he's building with Optimus Gen 3 will be the single most useful machine on the planet.

By the way, guys, I was texting with Brett Adcock last night, the CEO of Figure.

Cool.

Brett has agreed to come back on the podcast, so we'll get an update on Figure 3. I'm excited about that. We have to decide whether we go to his facility or do it digitally here.

I just stuck this in because one of the things I do is put out a humanoid robot report every year. It's going to be coming out in a couple of weeks, but the number of humanoid robot companies is exploding. I've never seen anything like it. It's just crazy. This is hardware. This is real hardware being manufactured.

This is a company called Sunday Robotics. What I found fascinating—and you'll see this in the video—is that they've created an army of 500 people they call memory developers. In this photo, you see a human wearing these gloves, and these gloves are basically the hand manipulators of Sunday Robotics. The human goes about their normal daily chores with these gloves.

It records every action, every motion. So they’ve claimed at Sunday Robotics that they’ve created the single largest robotic data set for dexterity. Let’s take a look at this video, and then we’ll chat about it.

He’s opening a dishwasher and putting these wine glasses in the dishwasher with very high precision, and folding socks.

Nice. Hey, NEO Gamma, would you please bring me a cappuccino? I’d appreciate it. I think by next year, hopefully, that will happen, and the cappuccino will actually materialize magically via robot. Dave, what are you thinking here?

Well, a couple of things. The programming by actual action—you just do it, either through vision or you just move, and that programs the robot—is a huge unlock versus sitting down and writing the Python code. It could even be Python, because it’s too slow, so you have to write assembly or C code, which takes forever. That’s completely gone now. You just train it, and that’s a huge, huge unlock.

You know, we tend to visualize the humanoids working in the home because that’s what everybody loves. But these things work in nuclear reactors, where no human can go. They work inside pipelines. There are huge amounts of use cases for this in areas where there’s no other option. Build a data center in space? Well, it’s not going to be people with spacesuits. It’s going to be robots in space.

100%.

And so, lots of unlocks here. You almost can’t go wrong right now. I will be curious, when we’re talking to Brett, about whether he’s using NVIDIA chips, because you said Elon almost never loses, but Elon’s making his own Dojo chips for these robots. Brett, I don’t know what Brett is using. I know that 1X was using 2 NVIDIA chips per head of each robot.

So that’s a big constraint. We’ll see what Brett’s plans are.

Yeah, Alex, we know algorithmically how to solve robotics. It’s vision-language-action models—foundation models that are generalizations of LLMs like ChatGPT. The hard part is the training data. There are so many different approaches emerging for training data for these VLA models.

I think, to Sunday’s credit, what they’re doing is, as alluded to in the video, using special gloves that human operators wear that are exactly matched to the size and form factor of the hands of the robots. This is one approach, and I think it’s a very promising approach. To their credit, it requires, as they say, zero teleoperated data for training purposes.

But there are lots of other approaches. We’ve seen Figure announce cameras on the palms of its hands. We see a lot of so-called sim-to-real approaches that generate lots of synthetic data and attempt to translate that over to reality.

I’m optimistic that one or more of all of these approaches being tried for data-set generation and data-set curation will solve robotics imminently. I’d like to point out that somebody needs to figure out how to program the swarm, too, because when you walk around the research labs, a huge fraction of people are working on grippers. It’s like, “Okay, I picked this up. I’m squeezing from both sides.”

You can do that with a gripper, but you can also do it with 2 independent drones pushing against each other. You can’t program a swarm of 50 drones by showing it what to do with your hands. So someone has to crack the code on how you use that same exact approach for the swarm version.

Someone’s going to have to start strapping wearable cameras to swarms of birds or something like that to gather the good training data. There’s a startup there for someone in the audience.

So, I’ve got a positive and negative perspective. Positive: a few years ago, when Baxter first showed up, you could train it. Instead of saying, “Lift object, turn 90 degrees, move over here, turn back, put it down,” and explicitly programming the code, you could move its arms back and adaptively show it what to do, and it would learn that. I think we’re seeing the fruition of that vector, which is very, very powerful.

On the negative side, I think we’re a long ways away from picking up salad plates. I think Dave’s point of dull, dirty, dangerous—the DDD thing—is where we’ll go for a long way before we start doing this as general-purpose stuff at home. But that’s still my viewpoint. Let’s see what happens. I may be wrong.

All right. I’m going to move us along here. This is a company called Clone Robotics. They’re going to be one of the robot companies at the Abundance Summit this year. We have 4 robot companies on stage with their robots. Let’s take a look at Clone here.

What you’re seeing is basically a humanlike setup. What’s different here is that instead of electromagnetic motors, these are hydraulic systems moving the tendons and muscles to get dexterity. It’s sort of Westworld robotics, and I just show this to show the variety of different approaches that are going on today.

In brief, we’re seeing the $100,000 robot from Boston Dynamics, in particular this robot dog, entering police work and safety work. Do you want to add a point here, Alex?

Yeah, I’ll point out that it’s really interesting. The ratio between humans and dogs on Earth is 9:1. The ratio between humans and domesticated four-legged mammals on Earth is approximately 3:2.

I think the elephant, no pun intended, in the room is: will we see robot dogs and robot quadrupeds scale in proportion to humanoid robots? Yes or no? I don’t know the answer to this either, but one can imagine lots of scenarios where you don’t want humanoid robots, where you want something for difficult terrain or other more difficult circumstances. You want lots of other animal, nonhuman-animal form factors. Maybe you want snakes. Maybe you want insect-type form factors. I can imagine—

Microdrones, flies. I think we’re going to see all of these, and this is one data point. Yeah, you’re dead right. You’re feeding Salim, which is sad, but more nightmares for you, Salim.

No, no, no. I’m actually, again, for policing and dangerous work, absolutely right. But if you want a dog, give me a male dog and a female dog, and I’ll get you a dog.

I mean, oh my God. See, I put this chart in here for you. We’re going to wrap very shortly, but EVs are set to soar as gas-car efficiency stalls. Do you want to speak to this slide?

Oh my God.

Give us a rant, buddy. Give us a rant.

No, I’m going to keep it short. This is the International Energy Agency again, making its predictions. I’ll go back in history and say that back in 2012, it put out a prediction that it would take until 2040 before we had 1 million electric cars out there. By the time they finished their report, Elon had 1 million Teslas on the roads already.

In this case, there’s one data point here that is completely wrong, which is the number of electric cars in India. As you see, they’re predicting that it’ll be near zero until 2035. That’s complete horseshit. It’s going to be vertical like all the others. So what the hell are they doing? Europe and the US, I can just about understand, but that’s even going to go away just because of the sheer economy of it.

Well, autonomous electric cars—the robotaxis, the Cybercabs—are going to be displacing gasoline cars.

Yeah. So what are they talking about here? What are they missing? How do they miss this year after year after year? The number of moving parts in a combustion-engine car is 2,000 moving parts in the drivetrain, and a Tesla is 17. You can’t compete with that in terms of design, reliability, maintenance, and so on.

Yeah. The rant is over. Okay. All right. I want to close us out today with the notion that science and technology are creating an increasing world of abundance.

The first story here is “Epigenetic reprogramming trials are close.” We’ve got an image here of David Sinclair. David is the founder of Life Biosciences. Life is a company we just had at the Abundance Longevity Summit, which I do every fall, and a number of my Abundance Community members are investors in Life Biosciences.

What is it? Life Biosciences is commercializing the work that David has done with partial epigenetic reprogramming. He’s using a modified adeno-associated virus to put the OSK genes into cells. What does this all mean? It means that he’s demonstrated age reversal in mice and monkeys, and for the first time this age-reversal technology is entering human trials in the 1st quarter of 2026. It’s a big deal and has the potential to really transform things.

He’s going into the eye to deal with NAION, which is strokes in the eye, and glaucoma. But if it’s a true age-reversal technology, it’s going to affect the entire body. So he’ll go from the eye first to the liver and other organs.

You have to understand: when you were young, you didn’t have a specific disease. As you got older and your epigenome shifted, this disease materialized. So if you can reverse your epigenome to an earlier state, the disease should go away. There’s a lot of work here, so I’m super excited about that.

Let me link it with this story here: Anthropic is hiring life-science researchers. Again, we’ve heard Dario Amodei talk about whether we could double the human lifespan on the back of progress with AI. Anthropic has been going very hard and heavy here. Alex, what are your thoughts, please?

Yeah, Dario Amodei has stated publicly that he expects disease—or, let’s say, disease biology and medicine—to be solved by the end of the decade, in the next 5 years.

And I think Anthropic's doing an amazing job of pushing forward AI for science in general. But I think this is what solving biology looks like. It looks like starting to apply dedicated efforts, hiring biologists, and building out facilities to solve biology by the end of the decade. I think they will succeed.

Amazing. Part of what we're talking about solving as soon as possible is the cost of sick care. Again, reminding everybody, we do not have a health care system. We have a sick care system. The system takes care of you after you're sick. A health care system would keep you healthy, right?

AI is going to be the biggest impactor here. We saw a study out of Stanford and Harvard Medical School that definitively shows that an AI diagnostician looking at your data will diagnose you far better than any human or even a human with AI, because humans introduce bias into the results.

This is the progression of AI as a radiologist. In this chart, we're seeing basically how Gemini 3 is performing versus radiologists. Gemini 3 has outperformed radiological trainees, and it's on its way to taking over the role from board-certified radiologists. It's not there yet, but I think in a year we'll be there. Alex, your thoughts?

First of all I love benchmarks. I I've made no secret that I love benchmarks. The benchmark in this case is named radiologyy's last exam. It it's composed of I think 50 radiological images spanning multiple modalities, multiple body systems. And the basic task is handing an image or imagery to an AI of various sorts and asking it to to specify a final diagnosis. And as you say, Gemini 3 Pro, which just launched, beats radiological residences. And on its present trajectory, if you just plot a a straight line through progress from GPT5 thinking to Gemini 3 Pro, I I think we're going to see radiology get solved in the next year.

And my brother-in-law Tim is a radiologist, and surprisingly, he cannot wait for AI to outperform him, which he says is basically today. He got into the business to save lives, not to have a job, and he is just super excited about it saving lives.

Yeah, I think you're going to find people who are going to actually say, “No, no, no. We can't let AI do this yet. They're not good enough.” And you're right: there's not enough radiologists. I know we hire radiologists at Fountain. I hope we're going to transition to our AI radiologist soon enough.

Talking about creating abundance in the world, this was a fascinating article. I think you found it, Alex, so thank you for that: “Rainmaker's Bold Plan to Refill the Great Salt Lake.” Alex, you want to hit this one?

There's a scene in the Johnny Depp movie Transcendence where nanobots are released throughout the biosphere, remediating it. I do think we're going to find ourselves in a world soon where some facsimile of that scenario is the case.

In this case, these are drones for weather modification that are ultimately being used to refill the Great Salt Lake. I think this is an early preview of that scenario, where many of the most, in principle, labor-intensive, capital-intensive, energy-intensive environmental issues that we face can be addressed. In the case of the Great Salt Lake, it's losing water, which is resulting in arsenic that was already in the salt bed being aerosolized, and people are breathing it. It's not a great situation.

All of these environmental scenarios can ultimately, I think, be remediated at scale with enough automation and AI. I think this is just a sliver, just a preview of what we're going to see biosphere-wide.

Yeah. You know, we talk about water scarcity in the world, right? Water is one of the most important things: clean drinking water. If you've got that, you can eliminate half of the disease burden on this planet.

We have to realize that we have an incredible supply of water on this planet. The problem is that 97.5% is salt, 2% is in the ice caps, and we fight over a half a percent of the fresh water in the lakes and rivers.

It turns out there is another source of water. It's quadrillions of liters of water in the atmosphere. If you can access that—we've had a couple of XPRIZEs on that topic—you can move the water to where you want it and need it. This is about creating water abundance on the planet, which I think is so cool.

I have 2 points to make here. One is that this is climate engineering at full operational scale, which is amazing. The downside that people say is, “Oh, my God, we shouldn't be geoengineering the world.” The counterpoint to that is that we have been geoengineering the world, throwing up a ton of carbon into the atmosphere for decades now.

We have to use technology. I've been watching the outcomes of the COPs—a farce going on every year where we get all the folks together. We are not going to find a political solution to climate change. Nation-states cannot solve climate change. You need a technological approach. I think this is the start of a whole array of those.

Yeah. In terms of re-engineering and gaining access to resources, one of the biggest conversations is that rare-earth metals are rare. How do we get access to them for our electronic supply chain? This is a company called Vulcan. The title here is “ReElement and U.S. Government Launch $1.4 Billion to Build Domestic Rare-Earth Supply Chain,” so critical for so much of what we're building here.

Of course, for the last couple of decades, this has been a supply chain controlled in China. Alex, do you want to give us some 101 on this? Or Dave, either of you, please.

I'd like to comment again, in extremis, where I think all of this ends up. I think it's relatively easy to imagine a future where so-called reindustrialization reaches an end state where local supply chains become hyperlocal.

For rare earths, for other key elements, and for other key feedstocks for the supply chain of the innermost loop, if you will, I think ultimately that takes us to nanotech. It takes us to Drexlerian nanosystems. It takes us to robots that are able to scavenge raw resources from the immediate environment, and immediately you get an immediate supply chain that's packaged up in some sort of self-contained way to produce finished products.

I think this, again, is just the start of maybe a 10- to 15-year journey toward those types of nanosystems. Dave.

Yeah, we have 2 investments in companies now that use AI vision systems and sensor data to scour through recycling and trash, looking for rare earths and looking for things that are extremely valuable.

Everybody loves that because the supply-chain people love it, and the recycling people and the people who want to get the garbage out of the streets love it. It's just a pure good enabled by AI and robotics, and it's a really, really good theme and a very, very profitable business, too.

There's also a book called The End of the World Is Just the Beginning that Thomas Pedy gave to me.

I love that title.

It's incredibly packed with data and statistics. It's a really good read. Amazingly, the U.S. is one of 2 countries in the world that has everything—literally everything is here.

Our rare earths come from China, but that's only because we didn't bother to mine them out of the earth here. We have them. We just didn't put together our own mining operations. We had rare-earth-metal operations. The problem is that China would undercut the marketplace and put our companies out of business. But that's a different—

Industrialization. The whole push now is, “Hey, make sure that everything is able to be done here, doesn't get undercut, and reindustrialize America.”

Yeah, and France is the other one that has pretty much everything, surprisingly, if they can just access it.

Including grain.

It's a great book.

I put this in here, Dave, because Link Ventures is a Boston-based company. So here it is, the data: “Massachusetts Leads in VC-Backed IPO Success.” Massachusetts leads with a 4.1% probability of going public within 10 years. California companies are at 2.3%, and New York companies are at 1%. Any comments, Dave?

Yeah. So, no mystery here. Massachusetts is also the healthiest state in the country and the best place to raise a family in the country. It's not magic; it just tracks university density. If you do the exact same chart on university density, you come up with the same curve.

So, as a fraction, if you walk out on the streets of Boston and touch a random person, there's a 25% chance they're a student, and some other chance they're a professor or working in a startup. It's just an incredible high density of very, very smart, upwardly mobile people. There's not a lot of a homeless problem. All of that really feeds this machine.

Yeah. Yeah.

But people come to Boston to study and to learn and to build, and then a big fraction of them stay.

Another big fraction goes to California. So this is tracking where they started.

A lot of things actually start in Boston. Many of them do migrate to California, which is why people perceive that as being the epicenter. But if you want to be there the day that they're founded, or if you want to recruit the talent, there's 20 times more engineers in Boston than there are in Silicon Valley—20 times—because of MIT, Harvard, BC, BU, Northeastern, and Tufts. It's all within walking distance. It's just a really unique place on Earth.

You couple that with the statistic that you showed me: of the unicorn venture-backed companies, the highest rate of giving birth to a venture-backed unicorn company comes out of MIT. MIT is number 1. It turns out USC is number 2, and then Stanford is number 3, which sort of shocks all my Stanford friends when I show them that figure. But it's why Link Ventures is based out of MIT.

I got to say one thing about this.

Yeah, please.

It's a wonderful statistic, but California has 10 times more IPOs than Massachusetts and only 6 times the population. So I think that speaks to Dave's comment about people moving over there.

You've got to take into account that the weather is cold, but I think when I remember building a company, the best engineers and systems engineers we could ever find were in the Boston area.

The best engineers and systems engineers we could ever find were in the Boston area.

The game plan that has worked for so many of my friends is you start in Boston, you hire your first 10, 15, 20 people in Boston, you get revenue, you get traction, and then a West Coast VC offers you a hundred billion dollar valuation. [laughter] So you move your headquarters out to Silicon Valley. You grow, grow, grow, and then you have your first kid and you move right back to Boston because the school systems are the best in the world.

It's also very community-oriented. There's a little town center with a white-steeple church, and the police and the firemen and everybody all interact with everyone. It's very, very social. You rake leaves, you shovel snow, you grow up, and the cycle starts all over again. So, just one thought on a possible life plan. It certainly worked for a lot of my friends.

Any closing thoughts here? Let's go around the horn. Salim, how did you enjoy this episode today? Any other closing thoughts? Any other rumors you're hearing? Any other fun things?

Just an epic steamroll towards the era of abundance. We have collapsing costs. We have access into multiple industries. We've got new industries forming at the seams of everything. Very quickly, you're going to have a personal AI that's a doctor, a lawyer, a tutor, a mentor, a coach, and it's all going to be free.

So when people talk about abundance, we're kind of getting there so fast. I think all these stories that we're talking about show us how quickly we're going to get there. It's incredibly, just unbelievably exciting.

Amazing. Dave, your thoughts, please.

Well, 2 things I saw in the comments. One was, “Dave, can you stop wearing checkered shirts every single time?” So I agree. [laughter] I'll try and develop a look. I don't have a look, but I'll try.

The other one is really cool. I like to listen to it at 1.5× speed, but then when Alexander speaks, I need to go to 1× and listen to it twice.

Yeah, you've got to slow it down by 2 times.

We could try to automate that. That's a very easy AI problem. We'll create a little overlay that can automate the process for you.

Alexander, please, your closing thoughts here.

Two comments. First, we didn't get a chance to talk about the new Nano Banana Pro model, which is just incredible. I encourage everyone to play with it. It has transformative new multimodal capabilities. I was very impressed. Kudos to the team.

Second, I spend all of my time thinking about solving the hardest problems on Earth with AI. As I've mentioned in the past, if you have really hard problems that you're working on solving, I would love to connect with you. I think we're entering the age where the hardest problems on Earth are solvable with AI.

And Alex will be at Nurups next week, so if you're at Nurup's, the AI conference, definitely reach out to him. We'd love to connect in person.

And plug from my side, December 17th meeting of Life Session online.

We had put out a call to see if you guys wanted to get together with Moonshot Mates at a Moonshot gathering in the fall of next year, fall of 2026, probably in LA. We've had about 700 of the thousand write back saying that they're interested in joining us. So if you're interested in a Moonshot gathering, send an email to moonshots@diamandis.com so we can hear your vote. Our goal is to get to a thousand people who say they're interested. And if we get enough interest, we'll pull this together in the fall.

If we disassemble the moon per Alex's prediction, we're going to have to rename this podcast.

No, that's why it's called Moonshots. We're all about shooting the moon, disassembling it, and building the computronium cloud.

All right. Our outro music here, Epic Fantasy Edition.

So good. Check it out.

If you are listening and not watching, it's worth going to YouTube to watch this.

Thank you, John. That may have been my favorite one of all of them.

You got some real talent there, John. I love AWG as an elf and Dave as some version of Robin Hood. And Salim, you show that to your wife. You're the sexiest man on the planet in that one.

I look like a troll. That's great.

As always, gentlemen, love you greatly. Thank you for your wisdom and your passion and your commitment here, everybody. That's a wrap on Moonshot. See you guys again next week.

Every week my team and I study the top 10 technology meta trends that will transform industries over the decade ahead. I cover trends ranging from humanoid robotics, AGI and quantum computing to transport, energy, longevity and more. There's no fluff, only the most important stuff that matters that impacts our lives, our companies and our careers. If you want me to share these meta trends with you, I write a newsletter twice a week, sending it out as a short two-minute read via email. And if you want to discover the most important meta trends 10 years before anyone else, this report's for you. Readers include founders and CEOs from the world's most disruptive companies and entrepreneurs building the world's most disruptive tech. It's not for you if you don't want to be informed about what's coming, why it matters, and how you can benefit from it. To subscribe for free, go to dmmandis.com/metatrends to gain access to the trends 10 years before anyone else.