AI战争：OpenAI广告与Sora 2、Grok与美国政府合作、Google广告业务面临风险

Very recently, we've seen the creation of Sora 2. We're seeing, in front of our eyes, the transition from algorithmic content selection in social media to algorithmic content generation.

This isn't about sharing content. The creation of the content is completely up for grabs.

Meta launches Vibes, an app for AI-generated videos.

They're spending $1 billion on individual employees. They have a $600 billion, 3- to 5-year budget, yet they turn to Midjourney and Black Forest Labs to build this out. That's because the really smart, creative people all want to do startups, and they don't want to join the big companies. It's really encouraging for the startups because the other big labs—Google and OpenAI—are doing their own video generation. It's encouraging for the startups that are right in the middle of the crosshairs to say, "Even here, we're thriving." So it's a good sign.

The most shocking thing about this isn't how real it is. It isn't how easy it is to use. It's the fact that it's free. That is shocking.

OpenAI is bringing ads to ChatGPT.

The AI is going to be incredibly good at convincing you to do things, whether they're right or wrong. It's a very tricky balance.

Because they're spending so much money on the data centers, there's a huge incentive to get really aggressive with the advertising.

The ongoing AI wars are making all of the demonetization and democratization occur around the world. Let's jump in.

I'm here with my favorite friends on the planet: Dave Blundin. Good to see you, pal.

Hey.

Salim.

I'm back.

You are back, AWG. You're back from your top-secret mission.

Thank God.

Thank God we missed you. Can you tell us anything about it?

To the extent that you think we're on the verge of a sharp takeoff—a hard takeoff, if you will—I was traveling in Europe to see what the world looks like beforehand.

So you're updating your baseline of what the world is before things go hyperexponential. Amazing.

If it isn't a gentle singularity, I'd like to know what it looks like beforehand.

You know what I was doing last week? I was running my Abundance Longevity Summit. I had 50 of the world's top scientists and entrepreneurs focused on adding decades, maybe doubling our human lifespan. It was awesome. I walk away with the greatest confidence in the world that at least our friends and our subscribers are going to be hearing us talk about this stuff for the next 50 years—or some version of ourselves.

That is really a frightening thought.

All right, everybody, welcome to Moonshots. And let me begin with a moment of thanks. I want to just give a shout-out to one of our subscribers, Bill Jacobs 386. I'm going to read a note he posted. We do read your notes. We love it. We're here to serve you. And he wrote, "I am continually humbled by the amount of commitment and effort that's required to put this podcast together weekly. I'm not asking for anything in return. Nothing, that is, except to listen and hopefully learn before it's too late. The future is now. And I think I'm speaking for most of us here how grateful we are. Thank you." Appreciate that, Bill. It's that kind of feedback actually makes it fun for us to serve our subscribers, serve all of you. Dave, you want to say anything to that?

Well, most of that thanks goes to the team behind the scenes. There's a huge amount of news out there that gets scoured down to the bullets that we think really, really matter to people and then also to Alex's agents, which are getting bigger by the day. His AI force is coming up. I mean, it's just incredible how rapidly the feedback coming from that agent force is filling the pipeline of possible news and then, of course, the human factor whittling it down. So it's a big machine.

Yeah, and we do spend a good 20-plus hours. I was up at 4:30 this morning going through everything, doing my background research and getting ready because if I'm not ready, I will get completely decimated by the brilliance of these three moonshot mates.

I feel like I work really hard to keep up with everything going on. Then every time the team comes up with a deck, 30% or 40% of it consists of things I hadn't even heard of.

Yeah.

It's great. It's really healthy for all of us, I think, to do this. I can palpably feel the singularity coming. I remember you and I were on stage during the early days of Singularity University, and we would update our slides, our conversation, or our shtick every 3 or 4 months.

The technologies, between nanotechnology, biotechnology, neuroscience, robotics, AI, and so on, were changing 20% a quarter on average.

We actually worked it out as a faculty. The content was changing 20% a quarter on average, but this is like 80% a week right now. This is a whole other ball game that we're in.

It really is. I look back at our podcasts from a year ago, and it's like, "Oh my God, that is so ancient history."

Shelf life dropping radically.

Yeah, it is, but it's becoming more and more fun. I've labeled this first segment "Video and audio battles." Let's begin with this video.

Meta launches Vibes, an app for AI-generated videos. If you're listening to this and not watching it on YouTube, it's just music, but it's beautiful imagery that Vibes has generated. This is through a partnership with Midjourney and Black Forest Labs. Alexander or Dave, do you want to add anything here?

I think there are probably 2 stories here. One is that we're seeing, in front of our eyes, the transition from algorithmic content selection in social media to algorithmic content generation. It's a pretty obvious story. The perhaps less obvious story is that the space is moving so quickly that Meta was apparently compelled to partner with third parties for this AI generation rather than using in-house, first-party models. I think this is a very quickly moving space, and now it's very competitive as well.

I was going to say the exact same thing, and riffing on it: They're spending $1 billion on individual employees. They have a $600 billion, 3- to 5-year budget, yet they turn to Midjourney and Black Forest Labs to build this out. That's because the really smart, creative people all want to do startups, and they don't want to join the big companies. It's really encouraging for the startups because the other big labs—Google and OpenAI—are doing their own video generation. It's encouraging for the startups that are right in the middle of the crosshairs to say, "Even here, we're thriving." So it's a good sign.

This is free. The other thing that's interesting is that they're generating a TikTok-like feed: You swipe the video, swipe the video. We've seen X do that as well, if you're watching the video. And of course, it's not just Meta. We've seen Veo 3 from Google with its video generation, and very recently we've seen the creation of Sora 2.

Sora 2 is launching viral AI-generated videos. I'm going to share a video I created for myself and talk about how easy it is to create. So let's check this out.

Suiting up for the ride. Helmet secure. Pressure's good. Visor locked. Let's make it count. Heading to the rocket. Jumping in.

Cabin comm is live. You're looking good.

Strapped in and ready for launch. Let's go. 1, 2. That's 500 done.

Double our reach every 12 months. In 10 years, we multiply 1,000-fold. What else drives compounding? Data sets. Each new user improves the model and makes the product more valuable, pulling in the next wave. Pair that with automation. When marginal cost drops toward 0, growth accelerates on its own.

Thanks for inviting me to the studio, Peter. I've been looking forward to sitting down with you on Moonshots.

Likewise. It's great to have you here. People have been asking for an episode that dives into AI and longevity.

Happy to help. It's one of my favorites.

That was fun to make. If you were listening, this is a version of me on the Moon, then a version of me pumping 500 pounds in the gym, then 6 or 7 of me having a conversation about exponential growth, and then sitting down with Sam Altman for a Moonshots conversation. They didn't get the audio model right, and I'll have to rerecord that, but it was pretty fun. Gentlemen, thoughts? Do you want to grade it on performance?

I thought a couple of things. One is, as you connect this with the previous story, this is like Hollywood, TikTok, and Spotify all merging into one thing. I think Alex's point was really important: This isn't about sharing content. It's about the creation of the content being completely up for grabs in a new way. I think all of that happens at the same time.

And the interface to create it is entirely voice and prompt.

There's no coding and no interface. Like all of our lives since the computer was invented, we've been learning incredibly complicated interfaces to everything, from the microwave oven to the laptop to Chrome and Safari, Peter. All of that is about to disappear from the Earth forever.

And just go to a straight natural-language interface. We'll see later in the pod—much more important, actually, software creation. After that comes building creation and highway creation, and all of that is going to be done by just voicing it into existence, right out of the Star Trek holodeck.

It is godlike. First, it's speaking the word and creating reality. It's going from mind to materialization. It's extraordinary.

I also think we're seeing video emerge as a first-class modality for frontier models. Right now, most people are interacting with frontier models via text or images. Video is still this separate channel with a separate distribution mechanism. These are on a collision course.

We're going to see the video form factor and the underlying model architectures—probably diffusion-transformer-based—merge into the more autoregressive-transformer-presumably-based text and image models. One could even imagine the ultimate user experience here. Maybe not the ultimate, but an intermediate UX looks something like a magic mirror that does this in real time.

Right now, Sora 2 takes a few seconds to generate, with fully realistic audio and realistic physics. The physics—if you ask Sora 2 to reproduce some generic, say, high-school- or college-level physics demos—is pretty amazing.

All of this ability to reason about physical-world models—if I ask you to think of a pink elephant, you will visualize a pink elephant in your mind's eye. Sora 2 and similar video models, once they're incorporated into the chain of thought for frontier models, will enable entirely new classes of reasoning ability.

Yeah, it's got physics consistency, which is extraordinary. Go ahead. I want to talk about how I made those videos again.

I asked it to create a video of a water drop dropping into a glass of water because it's a common image. It was extraordinary how accurate it was. It was absolutely amazing.

Yeah, it has real-world physics modeling built in. So I encourage everybody listening to actually try it out. When OpenAI does this, it's creating sort of a viral engine that's getting people from 800 million users up to a billion.

But you need to get an invite code. Once you have the invite code, it's super simple. On your phone, you download the Sora app from OpenAI. You basically hit a few prompts, and it has you say 3 words or 3 numbers. Then it has you look to the right, look up, look down, captures your face, and from there, fundamentally, it's a very simple prompt.

If individuals like Sam Altman or others make themselves open for other people to use—and you can make yourself open for use or not—you can pull people into it, and it's pretty easy and fun.

Yeah.

The viral loop now goes from prompt to publish to explore in no time flat.

It's super fun. Try it. You have to try it. It's super fun.

The viral loop now goes from prompt to publish to explode in no time flat.

Yeah.

Right. It used to take weeks at least, and now it's like nothing.

I saw a great podcast of Bill Gates talking about how we in the computer science world slaved away for 20 years just trying to get speech recognition alone to work. I don't know if you remember Lee Hetherington, Peter, from MIT.

Crazy brilliant guy, almost at Alex's level.

He spent 20 years in Victor Zue's lab trying to make speech recognition work.

Do you remember Dragon Systems? That was one of the earliest voice-recognition systems.

It really is unfathomable how fast it's going, and we take this stuff for granted, which is insane.

That's the point. Bill Gates made that exact point because he had billions of dollars of R&D to try to make speech recognition work. Now it's an afterthought in the big neural nets. They do speech, then move to video, then move to video generation, then move to complex math and physics—all in 2 years.

It's so easy to take it for granted, but it's massive amounts of converging technologies that are suddenly unleashing new capabilities and so many opportunities to glue together the different components and build an incredible new experience.

Everyone should reread The Future Is Faster Than You Think, one of Peter's many great bestsellers. It's all about the converging technologies. But when you wrote that book, there were maybe 8 or 10 things to consider. Now there's like 800.

Oh my God, we just wrapped up our gorgeous book, We Are as Gods, and it is so difficult to send it to the publisher.

No, no. When do you draw the line, right? When do you draw the line?

Yeah, it's insane. By the way, Veo 3 and Sora 2 are free. This extraordinary technology again—the most shocking thing about this isn't how real it is or how easy it is to use. It's the fact that it's free. That is shocking.

Let's continue our journey on generation. Here is a product called Suno 5. It's AI-generated, studio-quality, lifelike vocals. You can basically create something that's a full 8-minute run length. Just because we're called Moonshots, let's play a Moonshots-themed piece called “Moonshots.”

Again.

All right, a Bond-like thematic Moonshots audio.

Can I give us a challenge?

Yeah, sure.

Before the next episode, we should all play with this and come up with our own versions of what the theme song should be for the podcast. Then we'll let the viewers pick which ones they like the best.

The theme song for the podcast—you know, Nick and Dana and the team are working on that in the background. So we might have just taken the workload off of them, but absolutely.

All right, that was my bit, if you will. I think it's probably also worth noting, again in passing, that the musical Turing test has passed. We barely discussed it. Anyone can compose a Top 40 song or an opera. This is the beginning, maybe, of disposable or casual art.

Wait, what would have been the test?

The ability, perhaps, to generate an indistinguishable-from-human Bond-like song, in this case, or a Top 40 song. We just passed that.

And Alex, I'm sorry I didn't give you credit for that, but thank you for playing. One of the most exciting things we get a chance to do is play with this stuff as it's coming out. The good news is all of you can play with it, too.

So, for $8 a month, we now have a personal Hans Zimmer.

That's a minimum and quite a bit more.

Yeah. Making all of the demonetization and democratization occur around the world are the ongoing AI wars. Let's jump in.

All right. Anthropic announces Sonnet 4.5 and claims the best coding agent available. Alex, would you walk us through this?

Yeah, it's really remarkable what a single-minded focus on code maxing, or code-gen maxing, is doing for Anthropic with its model. In using this model and testing it, one of my favorite test cases is to ask the model to single-shot the generation of a cyberpunk first-person shooter.

Claude Sonnet 4.5 does an amazing job. It gets nearly all the way there with minimal hand-holding. I have very high confidence that some iteration of Sonnet 4.5 will get all the way there with visually stunning graphics, music, and elaborate first-person controls.

I think the risk that one can perceive on the horizon is, on the one hand, focusing on code generation is perhaps a very ambitious bet toward recursive self-improvement. If the code can write itself really well, maybe that's the critical path to an intelligence explosion.

On the other hand, if it turns out that other modalities are important, like video, for example, which we were just seeing, or music, then the risk is that a single-minded focus on code generation in particular may not be the critical path. I suspect we'll know the answer in the next 6 to 12 months.

Well, shout-out to Blitzy. The top benchmark here is 82% on SWE-bench, but Blitzy got to 86.8% on that benchmark by combining models. So that'll go up a little bit now with Sonnet 4.5 under the covers.

But just by hitting all the models and iterating a lot, you can actually squeeze more performance out of these benchmarks. They're pretty much maxed out now. They're working on a new benchmark with METR for long-form coding.

If your process is writing code for 8, 10, or 12 hours, how do you benchmark the quality of the output? It's a really cool new benchmark. We'll get into benchmarks later in the podcast, too, because there are a lot of capabilities in the world that didn't exist a year ago. We have to have some kind of metric for all of them.

I love the way these hyperscalers and frontier labs are all incrementing their software by 0.5. You know, Sonnet 4, 4.5, GPT-5. We've got Grok—where are we on Grok? Are we at Grok 4 now?

That's right.

Grok is also worth dwelling on for just a few seconds, particularly the autonomy length scale. Sonnet 4.5 is, somewhat infamously at this point, working for 30-plus hours straight.

I recall that in a past episode we were talking about the characteristic autonomy time of some of the bleeding-edge frontier models being 7 hours, and before 7 hours, 1 hour. If you had just taken METR’s original exponential fit for the amount of time frontier models can work independently and extrapolated a mere exponential time, we’d be far below 30-plus hours. So, if lots of reproductions hold true to this 30-plus-hour time estimate, that would strongly suggest that, in fact, we’re on a hyperexponential rather than an exponential in terms of autonomy. Really crazy things may start to happen in the next year or so if that’s the case.

And, Alex, Dario Amodei is in particular famous for really focusing on making what he would consider safe AI. One of the final bullets here is that Anthropic’s Claude 4.5 has reduced its ability to lie and seek power by a factor of 10. So what does that mean? It’s like when you ask it to turn off and it doesn’t, or if it’s trying to aggregate resources, or it’s lying to you. Those are not good things.

There is an entire cottage industry at this point of for-profit and not-for-profit red-teaming labs that are fed early access to these frontier models and look for these sorts of traits. I think it’s an interesting research-level question as to whether power-seeking, for example, is instrumentally convergent as a goal for superintelligence. Instrumentally convergent means that, regardless of whatever the long-term goal assigned to the model is, or whatever it’s prompted to do, above some threshold of intelligence or superintelligence, it more or less is required to seek power. I’ve published research in that area. In my mind, this is still very much an open question regarding the so-called orthogonality thesis: whether the ultimate goal of an AI can even be decoupled from its intelligence level.

It would be super interesting to see how Gemini, xAI, and OpenAI all rate on lying and power-seeking in their models. Do you have any idea? I see lots of different measures for this. It’s difficult to register a uniform assessment across the industry.

Yeah, that’s a fun challenge, though. That could go bad in so many ways, but that would be so fun. Let’s put together a benchmark for how it lies.

How well it lies.

Let’s see if we can prompt it into lying as much as possible.

Well, I could imagine. Listen, there’s an all-out competition between all these frontier labs. If the way you get ahead is that your AI is more power-seeking than its neighbor, are you optimizing for it or against it? We’ll find out.

All right. Continuing on: Imagine with Claude, a live app-creation demo of Claude 4.5 that generates apps in real time. Let’s take a quick look at this video, and then I’ll ask you to tell us about it, Alex.

Imagine if Claude is still building software, but we’ve cut out the middleman. Instead of writing code that describes this text box, Claude just makes the text box. We’ve given it access to software tools that construct software directly and substantially faster. Claude isn’t writing code in the standard way. It doesn’t have to plan it all out in advance. Instead, it generates new software on the fly. When we click something here, it isn’t running prewritten code. It’s producing the new parts of the interface right there and then.

Amazing. So, Alex, I saw you were playing with it this morning.

We’re living in the future, Peter, where the models are so high-throughput, apparently, that now it’s possible to do just-in-time code generation on every event. You click within a user interface within Imagine, and new code is generated on the fly. You can ask for new apps to be spun up on demand; they’ll be generated on demand. And I think it’s an interesting thought experiment to ask where this goes in extremis when throughputs continue on their exponential, or maybe hyperexponential, trajectory. I suspect, naively, where this ends up is that every single pixel is going to be generated. Yeah.

Not just—yeah, not just vector art, not just UX—windows, icons, menus, pointers, every pixel.

And I imagine your version of Jarvis, your personal entourage of agents, is spinning up capabilities for you that they think you might need on standby, ready for you to request access to.

We could end up with a gray-goo-type problem on this. It’s somewhat of a positive thing, but it’s going to be surreal, because you create an AI that starts generating apps and we’ll end up with billions of apps flooding the app store. It’s going to cause some interesting challenges on the—

But there will be no app store. You will not be choosing an app.

It’ll be algorithmic, obviously.

It’ll be the capabilities you need in the moment to achieve your objective—

—will be conjured up as you’re—

Materialized. Yeah.

Yeah. The term of art is, at this point, slop. And I’m a lot less concerned about slop overwhelming civilization than perhaps some folks. I think there are so many ultra-high-value, transformative problems that will set AIs on while we’re sleeping. I’m incredibly not worried that we’re going to drown in slop.

I agree. I completely agree. Also, I think it’s a good place—see, a lot of business leaders out there aren’t reserving their compute, and they’re like, “Well, I won’t need that much,” or, “I’ll wait and see what happens.” This is a great use case to show you that if you say, “Look, I want this software to exist in real time,” it’s entirely possible, but you have to have a lot of compute dedicated to you in order to make it happen in real time. How quickly can you imagine 400 or 500 concurrent things that you want it working on very, very quickly? So, if you have access to that compute, all of that can be created for you in real time, and it’s an absolute joy to do. If you don’t have the compute, you’re not going to get it. The demand for this is so mind-blowingly big, and you just have to figure out, where am I going to get the compute to do exactly what we just saw?

Alex, how easy was this to use? What did you have to do to spin it up?

Trivial. So all I had to do was go to the Imagine with Claude site. I asked it first to generate a calculator app for me: “Create a calculator.” It created a functional calculator. But most interestingly, as I was testing the calculator, clicking on each button in the calculator app, it was generating code in real time.

So this is a transformative way of thinking. We’re accustomed to historically thinking that there’s a software development time and then, later, an execution time. This completely blurs that boundary, where even at execution time, every software event results in new code generation on demand. It changes the just-in-time paradigm.

So, as a coder, you don’t have to think through every possible use of it. This is building out the use tree as it’s requested.

That’s right. And Vernor Vinge, one of my favorite writers, wrote in Rainbow’s End—another book, other than Accelerando, that I would highly recommend—about what would happen when we have too many transistors, transistors too cheap to meter, as it were, and our transistor budgets go through the roof. I think this ends up being one of these use cases. If we have so much compute just sloshing around, the ability to delay app-code generation until user-event time—that’s incredible, and that will certainly mop up lots of compute.

Yeah, we haven’t heard much from Claude—at least on our WTF episodes—over the last month. It’s good to see Claude, Anthropic, coming out with some great products. It’s quietly winning in the marketplace.

Yeah. Let’s go to OpenAI. OpenAI is introducing ChatGPT Pulse. I love the idea. I haven’t played with it yet. The idea is that, in the morning, when I’m using my ChatGPT voice and having a conversation with Ember, which is the voice model I’m using there, I have to think, “Okay, what’s a unique idea or concept I just learned about that I want to speak about? Let’s talk about the FOXO3 gene and how it’s impacting longevity,” whatever the case might be. Here’s where it flips the model: based on all your conversations you’ve had with ChatGPT, it’s actually coming up with topics you might want to learn about. So it’s prompting us, and then we’re prompting it back. Has anybody played with it?

I thought this was a really subtle but important thing: you’re not querying it; it’s querying you. I think that starts a new vector of really interesting development.

Yeah, it feels a bit like a successor to Tasks, which are also still available from within ChatGPT. But in my dream world, what I would love to see is, perhaps in addition to being able to set cron-style, periodically scheduled tasks, if I want compute running on my own behalf while I sleep, I would love the ability to have long-running tasks on hard problems—single tasks that run for days or weeks on end, rather than just smaller tasks that run, say, once per day.

I was going to say exactly the same thing. Go for it. I want to hear what comes out.

I want to cure every disease. That’s like a beautiful, well-posed task that is surely going to absorb many billions of dollars of inference-time compute.

Mhm. Okay, that’s great. I want antigravity. I want warp drive. I want a lot of things. All right, so let’s move on here.

Next up on OpenAI’s docket: OpenAI is bringing ads to ChatGPT. Its new chief ad officer, Fijimo, has come on, and what I find interesting is that OpenAI is going after massive revenue streams. Dave, do you want to weigh in on this one?

Well, the ad revenue is inevitable. That’s $300 billion for Google, and it’s all going to move over to AI conversations. There’s a lot of complexity to figure out there. She has a challenge on her hands, trying to figure out how you balance this: AI is going to be incredibly good at convincing you to do things, whether they’re right or wrong.

Mhm.

There’s a lot of revenue tied to that. I think Meta did a very good job of balancing the news-feed quality with promotions that are blended in, but it’s a very tricky balance. Because they’re spending so much money on the data centers, there’s a huge incentive to get really aggressive with the advertising.

Yeah.

Yeah. There will be consumer backlash, and everyone will move to some other model. It’s a really hairy balance. But the AI is both the best ally you’ve ever had in buying things and, if it’s misguided, could walk you down some seriously bad paths.

The trust question seems to be: Will you trust insights from an AI that has ads baked into it and has an ulterior motive? What do you do then?

Yeah, for sure. I think the ad model is ultimately going to disappear. I think there’s limited value here, right? Because once we have pendants or glasses and our AIs are able to see where we’re focusing—if my retinal gaze is on that lamp behind Alex and I say, “I love that lamp,” and I’m focusing a lot on it—attention is going to equate to some level of interest, and my AI may pop up and say, “Would you like me to buy that for you?”

Rather than having an ad come, it’s mostly just where I’m focusing and listening to my conversations. The other thing that’s going to be interesting is if I give my AI a surprise-and-delight budget. I say, “Hey, you can spend up to $500 a month to surprise me,” and stuff starts showing up, or it knows I’m running out of toothpaste or my T-shirts are worn down.

I’ll tell you, Peter, the 2 sentences you said back to back—I’ll tell you where the conflict is between the 2. You want your AI to surprise and delight you, and it absolutely will.

Most consumer products have 70%, 80%, 90%, 95% margins—huge margins—where there are 2 or more absolutely identical products.

Sure.

There are 2 different sets of sunglasses, 2 toothpastes; it makes no difference whatsoever. If the AI says, “Okay, I’ll get Crest instead of Colgate,” 95% of the margin goes to that company instead of the other company. There’s a huge amount at stake where the consumer is still happy either way.

Where does that money all land? Right now, it all lands at Google, and in the future it’s going to land on the AI advisor. So both things can be in harmony with each other, yet there’s a massive amount of money under the covers. It’s still ad revenue, or it’s decision routing.

Take it a step further, Dave, because my AI probably knows the exact makeup of the molecules in the toothpaste. It actually knows my taste buds better than I do, knows my genetic makeup, and will order a toothpaste that’s perfect for me at half the price. I know that it’s maximizing what’s best for me, and Google’s not getting it. No one’s getting it. The AI is buying it direct.

Yeah, we’ll see. Because if you look at toilet paper as an example, you can buy it for literally 5% of the retail cost.

If you deflate the margin and say, “Well, the consumer is much happier, or they’re only paying 5%,” but all the margin gets sucked out of the value chain, then the marketing company at the front also isn’t making any money. What tends to happen is the opposite: the marketing front end is complicit with the backend consumer-products companies to keep the margins high. The consumer just says, “Okay, fine. I’ll just buy that toilet paper,” and then you don’t think about it.

But do you think my AI could think about it and somehow circumvent all of that price gouging?

I think you’re onto something really interesting there, which is packaged ecosystems, where the number of things you can buy is getting so complex, and the number of choices is so complex. For a while there, there was an Eddie Bauer edition Ford Explorer, and it was like, “I’ve just bought into the Eddie Bauer package. I’ll get the car, I’ll get the clothes…” It’s just part of the overall thing.

If you read Neal Stephenson’s “The Diamond Age,” everybody moves into these culture packages where the AI has figured out all the parts. I think that’s a real thing, just because the complexity of decision-making gets so high over time that you just want to join kind of like a phyle as a group.

It’s trusted, but it’s also a brand affiliation, right? So I think one of the last moats that’s going to exist someplace is going to be brands, because I’m showing my wealth or my affiliation. I’m signaling a lot more through the brands I’m using, but not on toothpaste. No one goes to my bathroom and says, “Hey, what toothpaste are you using?”

All right, let’s move on. But the point here is that OpenAI is building revenue streams. Here’s another one: they partnered with Stripe for Instant Checkout in ChatGPT. I think this is brilliant—the ability for OpenAI to generate revenue on the sales of products, starting with Etsy and soon Shopify. Who wants to weigh in?

I’ll weigh in on this one. I think if you squint, we can see maybe the outlines of what at least near-future superintelligence microeconomics look like, where you have a power-law distribution. You have a long tail of consumer subscriptions, consumer ads, or consumer affiliate fees for agentic commerce.

Then you have a middle chunk where white-collar, so-called knowledge work gets automated in part and whole by AI. That’s sort of the middle chunk of what turns the wheel. And then the head of the power law is solving all these transformative problems. I think Sam would say, like, curing cancer or curing all disease, that are worth many trillions of dollars.

I think that the key question of our time, or at least of the near future, is: What exact power law do these follow? Is it a fat tail with lots of consumers using Stripe-powered Instant Checkout to power a very fat tail? Or is it a very thin tail where almost all of the revenues flowing to the frontier labs, to justify the soon-to-be trillions of dollars of CapEx to build data centers, are all being driven by transformative inventions and discoveries, and the Instant Checkout, if you will, ends up being rounding error? I don’t know the answer, but I think this is the defining question.

I think they’re reaching for near-term revenues that are easy to get right now, but in the long term it’s going to be the invention of new materials, new biotech, all kinds of things. The interesting number here is that by the end of 2025, it’s projected to be $142 billion in consumer purchases via chatbots. I think the one thing that we all have in common is a constraint on time.

So if I’m in the middle of researching a product and I’m in the midst of doing comparative analysis on OpenAI, and it pops up and says, “We have to buy it.”

Maybe, but maybe in the near-term future, the scarcity—I think you would say—of attention also gets alleviated, and we find ourselves in a post-scarcity attention world.

Interesting. In which case, do we shop around more? We have more hours in the day.

Yeah, but we have so much more to do with those hours that, when you think about the software we were just using through voice and also Suno through voice, it’s so compelling and so fun. You’ll eat up every one of those hours and more.

Yeah. So I guess those are harmonious statements. There’s no “but.” I’ll tell you one thing: when Alex says, “I don’t know what’s going to happen,” you know you’re going into crazy times. Timelines are really short, and I think timelines are 2 to 3 years at this point, max.

I thought this was profound because this could be a big threat to Amazon. If I can chat and then basically go straight to the source of where something’s being made, that’s huge. I’ve been using ChatGPT and Gemini to do comparison shopping for the last few months, and I don’t buy anything anymore without saying, “Hey, show me good alternatives to this or this or this.” It’s remarkably good at crawling the web and finding all the stuff that would take me ages to figure out, and now I can do direct commerce with this. That’s huge.

Yeah. Otherwise, you copy and paste into Amazon and buy it there, probably, right? Amazing. And travel—I mean, it’s interesting using a large language model for travel, saying, “I’ve got to be at this location by this time. Which airlines have the highest on-time reliability, and can you get me there? What’s the travel time? Set up the schedule for me.” Instantly, it’s there. Then it should say, “Do you want me to buy the tickets and set up the Uber for you?”

I would just remind you that this is still nibbling at the edges of consumer spending. AI is going to eat the whole economy.

That starts to look like AI eating real estate expenses, AI eating healthcare, and AI eating utilities and food. Right now, buying consumer packaged goods—this is not to diminish the CPG sector, but this is just nibbling at the edges of disruption.

I'll tell you, Peter, since Jeff Bezos is your friend, Lee Bosio[?], who used to run Alexa—he was the single-threaded leader for Alexa when he was at Amazon—used to work for us, and Jeff Bezos saw this coming a mile away. That is why he built out this massive investment in fulfillment.

That's because the interface is going to change for sure. He can rely on the fulfillment side of it to route all that volume through Amazon, but he knew this was coming when he invested in Alexa.

We still haven't seen Alexa play out fully, right? Alexa is still very antiquated. We haven't seen Amazon's AI play yet.

Very much. It doesn't hold state—no memory. There's a lot to build there.

Well, you know what they're doing? I had a call with the chairman of investment banking at the largest bank in England. They're huge Anthropic and AWS fans.

It's interesting, and Anthropic tends to be the friendly little brother to Google and others as well. They're well-liked and well-respected. We'll see how they team up.

Ten times less lying. We saw that on the other side.

And power-seeking. Okay, I trust my Anthropic AI. Here we go. GDPval measures the performance of our models on real-world tasks. They released tests for real-world tasks across 44 jobs in 9 industries, with GPT-5 and Claude Opus nearing expert quality, 100 times faster and cheaper. Alex, do you want to lead the conversation?

Sure. Well, as you know, Peter, I've beaten the drum in the past here on the importance of new evals and new benchmarks. This is a very important benchmark. OpenAI has alluded to this benchmark in the past, but actually looking at the benchmark, which is available open source for folks who want to look at the prompts, this feels like a benchmark for knowledge work. It's pretty diverse, and to the extent that you look at this chart and other charts that have been made available showing progress on GDPval, which covers a number of different industries and lots of tasks, it appears very thoughtfully put together.

If you just extrapolate by the law of straight lines, you predict that in the next 6 to 12 months, we're talking about substantially all knowledge work across a number of industries being superhuman as performed by AI. For some, I think that's a very short timeline. We're talking about evals literally solving the economy, or at least a good chunk of the knowledge-work economy.

Yeah, it's here now. Do not look for some decade in the future. This is the next year or 2. One of the quotes here is, “The models completed tasks up to 100 times faster and cheaper than human experts,” highlighting both their potential and the need for oversight. See, you were going to say—

Two points. One is, I remember there was such a big shift in car-making when you had a robot opening and closing a car door 10,000 times to test the hinges. Quality just went through the roof after that. Now we can have AI doing the same thing for this type of stuff.

What I thought was really powerful about this was that this isn't some kind of toy-problem benchmark. This is real-world stuff. Now we have the ability to gauge AI doing real-world stuff, and this becomes very tangible.

Fantastic. Let's go to yet another conversation here. This is a video I'm going to play with Brendan Foody, the CEO of Mercor, who Dave knows extremely well. This is Mercor's AI productivity index. Let's take a listen.

We decided to test how well today's leading AI models can actually do your job, and the results are astounding. Introducing the AI productivity index, or APEX, an evaluation that measures how well we've automated the most valuable industries in the world. We studied model capabilities in law, medicine, consulting, and finance in partnership with industry experts in each domain. APEX is designed to give an accurate forecast of how AI is going to impact jobs. But this version just scratches the surface of measuring model capabilities.

All right, Brendan, catapulting yourself to the top of the class. How old is Brendan?

23, I think, now. Yeah, founded at 19. He's ahead of Mark Zuckerberg in terms of company valuation, age, and race to billionaire age. I don't know if anyone since Mark has been on that curve.

As long as we're talking about Brendan, we get a whole bunch of inbound calls from people wanting to buy our Mercor stock from us. It's a $10 billion valuation, right? And it's like, yeah, but if you look historically at people who've reached where Brendan is at that age, every one of them—or almost all of them—become whatever: Elon Musk, Mark Zuckerberg, Bill Gates, whatever. He's on a trajectory like nobody else, and everybody loves him. You look at him on screen there; he's the guy everybody's cheering for. So it's pretty cool to see.

I think these last 2 slides are really important because AI is so general-purpose and so capable in so many areas. Alex and I have had all kinds of torture trying to interact with the State House here and with other government officials to get them to realize the urgency and the implications. It's so hard, but when you throw a really good benchmark at it, it makes it much easier to explain why this is so urgent.

Brendan is taking on all things related to work productivity across all areas. That's a really big ambition, a very worthy ambition for him.

Alex, this is how economics gets solved. If we want to live in an abundant future where the cost of service labor is driven to zero, step 0 is creating benchmarks. APEX and GDPval are beautiful examples. It's still early days, obviously, but they're beautiful examples of benchmarks for knowledge work, or knowledge-work-based services, in the economy. I would like to see many more benchmarks get created, including for robotic labor and manual labor.

Just within our portfolios, we have 28 seed-stage companies doing AI just here in the building. If I take any one of them—Mecato, doing mechanical design—what's the benchmark for the quality of the design? Prim and Vocara are doing voice sales and customer service. With AI voices, what's the conversion rate and the customer-satisfaction rate on an incrementally smarter AI? How do you benchmark that? Every one of these companies should be inventing a benchmark. Blitzy is already doing it for coding.

Whatever you're doing, if you don't create the benchmark, then it just turns to mud. There's no way for anyone to know, because it's like, how do you know if it's a smarter AI? I don't know. The challenge is, we saturate them all, and we're comparing them all to human productivity. We need to have a whole brand-new set of benchmarks that are—

I don't know. They're anchored in what, Alex?

The good news is, we already know how to benchmark superhuman performance. There are relative Elo-based benchmarks that we know how to do. We know how to—as a civilization—we know how to build systems that are more energetic than humans are, that are faster than humans are, and we're still able to measure them even though they're superhuman along some dimensions. We have no trouble measuring superhuman intelligence capabilities.

Thousands of horsepower. Exactly. Microsoft isn't being left out of the game. Microsoft unveils Agent Mode. Think of this as the ability for you to have access to it in all of your favorite Microsoft tools. Salim, do you want to jump in, or Dave?

I've been trying to get Microsoft Copilot to work in any kind of AI-useful way and have failed miserably for the last few months. I hope this one is a better effort. I'm not going to make an enemy out of Microsoft, as powerful as they are, but I will say that adding AI as a feature to something that already exists—that's the wrong attitude.

I feel like Apple and Microsoft are the worst offenders of this. It's not going to work.

That's a great point. They're trying to maintain their customer base and scratch their AI itch versus AI-native, clean-sheet startups. Every corporate CEO should understand that the same thing applies. I see so many people saying, “Yeah, we're doing AI. I added it as a feature in one department, so now I don't have to think about it anymore. Let me go back and get back to my country club.” You're going to get crushed with that kind of perspective.

It's not a feature. It's a brand-new everything. It's a completely different field and opportunity. There are a bunch of things I was trying to do in Excel, and I literally tried to use an AI mechanism to do them. I just couldn't do it.

Finally, I ended up using Comet to do it in the browser, and it did it way better and way faster. So I think this is a huge gap. I don't know where they're going to go with this.

All right. We've covered OpenAI and Anthropic. Let's not leave xAI out of the picture here. Elon has cut a deal with the government. xAI struck a deal with the U.S. GSA to let federal agencies use Grok for 42 cents for 18 months. It was either 69 cents or 42 cents. I guess he went with the cheaper option: 42 cents. I'll leave that alone. Any particular comments on Grok entering D.C.?

The price point is 42, which is 420, which is the magic number, which is the $20 million SEC fine that he had. Remember that? Of course, it's all tongue-in-cheek with Elon. I love that, even at that scale, he's making it fun and interesting. Kind of like Taylor Swift, there's always a hidden message.

And people love that stuff. It's good. It keeps people engaged. But what's going on between corporate America and government America is completely unprecedented. It's a little scary. It's working really well, and it's helping the country a lot.

But it's very odd to be investing in Intel and then cutting deals to move things in. For the government to be directly involved in corporate America like this has never happened before. Well, it's looking a little bit like China, right? China is picking winners and forcing partnerships and creating robot cities, gene-engineering cities, AI cities, and such. It's fascinating.

But isn't the government signing deals with every—ChatGPT, et cetera, et cetera? We saw that earlier. So it sounds like what they're doing is trying them all and seeing which one is the best over time.

Well, that would be fine. I mean, that's like government procurement, but that's not what's going on at all. You go into the White House and you're either genuflecting and being the anointed one, or you're not. These are White House—

Edicts: come in and talk. Yes. Yes. And we'll get to—we'll talk about Intel in the section called “This Is Not Investment Advice,” which is coming up.

All right. Meanwhile, in other AI news, here we go: a former Meta researcher is building a math whiz. I'm going to bring this to you, Alex. Teach us.

I haven't seen any indication thus far that math is not going to be solved in the next few months. How's that for a double negative?

A few months. Okay. So wait, hold on. Alex, you've said that before, and everybody's asking me, “Please have Alex explain what it means to solve all math.” So could you just—

Before we do that, let's speak about this particular article. This is a woman—

It's great to see female CEOs in the AI world. There aren't enough of them.

Karina Hung is the founder of Axiom Math. She's 24 years old, and she wants to build the ultimate AI mathematician. She's raised $64 million at a $300 million valuation. Again, we're seeing this over and over again. We're seeing starting valuations in the hundreds of millions of dollars. I don't know if it's at a pre-seed round or whatever, but intelligent individuals who have a monomaniacal focus are getting incredible capital backing.

Okay, now back to you, Alex. What does solving math really mean?

There are, I think, a few different ways one could operationalize what it means to solve math. One way would be to look at a benchmark like the FrontierMath Tier 4 benchmark, which measures the ability of AI to solve extremely difficult but nonetheless pre-solved problems that would take human researchers several weeks to accomplish.

If you just do a naive logistic extrapolation of progress in FrontierMath Tier 4, you find that, by the law of straight lines, as it were, by the end of this year—by the end of 2025—we're starting to pass 10% to 15% of the problems in the benchmark that AI can solve. At that point, I would argue we're in a regime where, algorithmically, we have a clear line of sight to solving any math problem that we might have today. Just pour more compute on.

That would also, I think, point to the second operationalization I have in mind when I speak of solving math. I don't mean literally every math problem that we can think of today has been solved. What I mean is that the process of mathematics has been solved to the extent that we have a clear line of sight: if you pour millions, billions, maybe trillions of dollars into OPEX in data centers, no new algorithmic advances are needed. We can reasonably forecast that any mathematical problem that's solvable will be solved with the same algorithms, just with a lot more computing.

Okay. Now take me to the implications of that for the general public.

It's tricky. Probably—I would say this is in the territory of speculation—but I think one of the more obvious downstream consequences of solving math is that any problem that depends on the difficulty of math, or let's say math being difficult, that isn't protected in a formal sense by the so-called complexity hierarchy, is at risk.

Mathematicians and computer scientists have this notion of certain problems being provably harder, in some sense, than others. Maybe you've heard of P versus NP. But if there's no formal protection for certain classes of problems being provably harder than other classes, I think certain types of tasks that we encounter in the everyday economy—for example, hypothetically, certain hash functions that cryptocurrencies depend on, or other everyday economic functions—are at risk of volatility.

If suddenly, for example—again, speculatively, not investment advice—there were a super-AI mathematician tomorrow that could, say, invert the AES cipher suite or invert the hash functions underneath AES, that could be potentially extremely disruptive to the economy and cause a lot of volatility.

I think the point you're making is that if AI cracks advanced math, it isn't just solving equations. It's creating the scaffolding to solve all these other areas, like cryptography, economics, physics, and so on. That's what you're really saying.

Yeah. To that point, I would say the way I would frame it perhaps is: first-order consequences are that problems that depend on math being hard experience some volatility. Second-order consequences: I think it's the ultimate canary for any domain that requires the ability to do mathematical reasoning.

So I would expect, in short order, a variety of math-oriented science, engineering, medicine, and other domains to fall in rapid succession. If this theory of the future ends up being correct—and I was alluding a few minutes ago to timelines being short—we may find ourselves in a world 2 to 3 years from now where we're just drowning under a surreal Cambrian explosion of breakthroughs.

We're drowning under a surreal Cambrian explosion of breakthroughs.

Exactly. That will also, parenthetically, be potentially quite difficult for society to metabolize.

Yeah. The economic impacts of that are going to be unbelievable. Speaking about economics, AI can now pass the hardest level of a CFA exam in minutes. Let's take a quick look at this.

CFA is a Chartered Financial Analyst designation, and it deals with investment management, portfolio management, financial analysis, and ethics in finance, which I find absolutely fascinating. I looked it up: the CFA Level 3 part of the exam is about portfolio management and wealth planning.

I want to make a comment on this one. We're advising one of the Big Four accounting firms on how to think about transformation, and we've been predicting this would happen with them because this requires real-world reasoning. The fact that it is doing this is a huge implication. All their finance jobs essentially get rewritten now and recreated. That's a body blow to the accounting world.

What I find interesting is leveling the playing field across all investments. Do I, with access to this specific AI, have access to the best investment advice that Warren Buffett has access to as well? Is this leveling the playing field across all economics? I think it is.

But what I'm excited about is—

America lost, and then Europe, too, lost almost all of its manufacturing. Despite inventing the car, inventing the plane, inventing the microchip, and inventing the computer, all the manufacturing of that stuff moved to other countries.

Yeah. We gave it up.

We gave it up. And you're like, “Well, but our economy kept growing. What are we all doing?” Well, we're a service economy. We're doing services. What the hell does that mean? You look under the covers, and a huge fraction of very smart people are working in this totally circular, nonsensical world where we created complex law, complex taxes, and complex accounting, and then this other huge group of people need to solve the complex accounting, and it produces absolutely nothing useful for humanity in this huge—

IRS code. IRS code—I mean, for God's sakes.

Holy crap. Yeah. Ronald Reagan was the last guy to say, “This is insane. We’ve got to get this down by 10x.” Ever since, everyone has bloated it up. The accounting lobby is the biggest lobby in the country, and it is bigger. We finally—

Yeah. Lawyers and accountants. We finally have an opportunity here—

—to get rid of it once and for all. Not by eliminating it, but by having the AI automate both sides.

Yeah.

And then it just becomes something we don't have to do anymore. All that talent can create things that actually benefit humanity. I'm so excited for that. I would also—

The relief is so palpable in your voice there. It's incredible.

Peter, to your question, I would also encourage the thought experiment: If everyone has the best investment advice, thanks to superintelligent investment advisors, what does the economy look like? What is the rational act? What's the rational course of action for an investor if everyone has equally super investment advice?

It goes to your point, Alex, of buying the index.

Yeah.

Damn it. He's right again.

Fight on that one, but we'll get to it.

I thought I'd bring quantum into the conversation. I know, Dave, you and Alex have been working on this. A couple of years back, I started a SPAC with Shervin Pishevar, and we took D-Wave public, which is now seeing incredible resurgence. It's gone from around $0.69 a share up to $30 a share and done extremely well. We've seen Rigetti Computing. Chad Rigetti has been a friend for some time, as has D-Wave. All of these independent quantum-computing companies are getting some real traction.

Here's a quote, though, from Julian Kelly, Google's quantum AI director: “The technology is 5 years out from a real breakthrough.” Alex, you've been tracking this. What are your thoughts on quantum computing?

I think it's early. I'm reminded that the GPU—or call it the accelerated-compute market, via the avatar of NVIDIA—had to pivot several times before it took over the economy. It started with PC gaming, then pivoted for a bit to crypto and now AI, and maybe there's a post-AI act.

But I think what is missing right now, at least to my knowledge, is the killer app for quantum-accelerated compute. There's a school of thought out there that maybe we'll use quantum at inference time to generate large synthetic data sets of quantum-chemistry data. That will be used as training data for classical AI. It's difficult for me to buy that that's going to be an enormous market.

My best guess is that, to the extent that there will be a killer app for quantum compute, it's probably something like AI-accelerated generalist training for AI or inference for AI. At least, again, to my knowledge, no one has yet published the killer app for quantum ML. There are lots of proposals out there. Nothing has seemed to scale yet.

This year at the Abundance Summit, I'm going to have Jack Hidary back on stage speaking about Sandbox AQ. It's interesting. This is the spinout out of Google X. Eric Schmidt is the chairman of the company, and they booted up at a $500 million valuation and have had, I think, in excess of $100 million of revenue.

They're not a quantum-computer-based company. They're an AI company using the quantum equations to provide different products and services. They're basically looking at new navigation systems that are able to measure slight perturbations in the Earth's magnetic fields. When GPS is down, you can still navigate because magnetic fields are not being spoofed like GPS is being spoofed in the Middle East.

They're using it for different biomedical applications, looking at your heart—your heart's electromagnetic system, if you will. They're using it for encryption methodologies, but it's a real revenue engine there.

One of the things that we should speak to for a moment, because we do have a lot of crypto listeners as well, is that everybody's like, “Oh my God, when is quantum going to break the encryption codes that's going to destroy Bitcoin?” It's important for everybody to know that if, in fact, we have quantum computation breaking encryption, your keys to your Bitcoin wallet are the last thing to worry about. The same encryption codes being broken are the nuclear codes, the banking system, and everything that runs the financial systems around the world.

I would actually take the position that post-quantum crypto is nearing a state—maybe not evenly distributed yet, but at least in theory—of approaching quasi-maturity. If I lost sleep at night worrying about inversion attacks against widely used crypto systems, it's not quantum information processing I'd be worried about. It's AI solving math.

I think that's a far more insidious threat to crypto security in general than quantum. We know how to do post-quantum crypto.

But the same thing then, right? AI solving math—if it's breaking encryption, it's breaking encryption across a multitude of other much more concerning financial and defense areas.

Yes. Well, as a practical matter, this is imminent either way. It's not going to affect nuclear codes or Bitcoin. What it will affect, though, is anything that you've encrypted and left around. Using AES-256 or AES-128, that's already vulnerable within a year, if not today.

So it's all the designs, files, and stuff that you thought you encrypted and left on a server or left in your desk. All that is going to be wide open. So, just so you're aware—

I think, Dave, that's a really important point. It's the stuff in the past. It's not really the stuff that's current or in the future, because we'll come up with quantum-encryption capabilities and so forth.

There's one thing about this story that popped out at me that I would just want to flag, which is that in 2008, we heard a quantum-computing expert saying we're 5 years out from having a real breakthrough. So this has been a constant pattern for a while.

I think with the AI changes, this may actually really be the case—that we're 5 years out. It may be much less than that, given the potential way out to solve a lot of these problems, but just the lean—

Well, no, I've heard this. Yeah, it's great advice. As a general pattern, when somebody tells you, “Hey, blah, blah, blah is going to happen. Invest in it. It's 5 years out,” 9 times out of 10, it's 20 or 30 years out.

Fusion has been 5 years out since the ’50s.

Well, it's been 50 years out since the ’50s. Let's be real—

—not 5 years out.

Well, the opposite is true, too. When somebody tells you something's imminent, like, “This is happening right now, guys,” don't ignore it. It's very likely that you're almost late to the party. I think that's great advice.

All right, let's move on to chips and data centers. A lot is happening here. I'll start with an open letter that Sam Altman put out on abundant intelligence. I'll just quote from it: “With 10 gigawatts of compute, AI can cure cancer or provide customized tutoring to every student on Earth. If we're limited by compute, we'll have to choose which one to prioritize. No one wants to make the choice. We want to create a factory that produces a gigawatt of new AI infrastructure every week.”

This is basically Sam saying, “Give us all the compute and capital so we don't have to choose between education and solving cancer or longevity.” It's an important point. I don't know—any thoughts on this one, Dave?

Oh yeah, lots. I mean, it's amazing how it's becoming increasingly clear that Sam is very small compared to Zuck and Google, I guess—Sundar.

Small in what way?

Well, he signed a $100 billion and a $300 billion deal, and that made big news, but he doesn't have anywhere near $100 billion or $300 billion. He has like one-tenth that at most. Meanwhile, Mark Zuckerberg said, “Yeah, we're going to put $600 billion into this over the next few years.” But he has it. He has the cash and the credit to actually do it, and he will really do it.

Sam is up against some serious heavy hitters—

—and Google's got massive engines. I mean, they've got so much capital in the bank that they can expend here. And Elon just moves his pinky, and capital flows into xAI at whatever he needs.

Um—

But what I love about that dynamic is that Sam is the one guy driving the vision and driving the agenda. Everybody else can afford to just kind of be an afterthought or a soft sell. Without Sam out there opening everyone's eyes, nobody else—Google, for example—would have ever even rolled it out, I don't think, without Sam putting the pressure on.

Here, I think he's exactly right. Is it on this slide, or is it coming up? I think it's coming up. I'll wait for it.

Here we go.

So yeah, this article here is “OpenAI, Oracle, SoftBank Expand Stargate with 5 New AI Data Centers.”

Aiming to hit the $500 billion, or 10-gigawatt, goal before the end of 2025 and being ahead of schedule.

Tiling the Earth, Dave. Continue.

Tiling the Earth. Well, yeah. Sam is saying, look, we're going to try and organize around 10 incremental gigawatts per year in perpetuity, or accelerating, and that will just barely keep up with the use cases and the demand. That's really cool to hear someone articulate, because then the land, the governments, the plumbing—all of that stuff—can start to get rallied around a long-term view of what it means to stay ahead in this race.

I think it's great to articulate it because the numbers are so big no one else will say it. Sam's the one guy who will actually say it.

Let's put the numbers out there here. Stargate's $500 billion investment dwarfs all the other hyperscalers in 2024. Microsoft put $40 billion into AI data centers in 2024 and planned $80 billion for this year. Amazon invested $16 billion, Google and Alphabet $29 billion, and Meta $23 billion.

I think they're all going to be massively accelerating, but just to give people some numbers to compare this to.

I do think, for what it's worth, we are already tiling the Earth—quite literally. But there's also a certain sense in which, if you remember President Reagan's nuclear policy of building up to build down, I can imagine high-likelihood scenarios where efficiency advances, maybe ontological shocks—perhaps ontological shocks that result from these data centers—make the naive assumption that we're going to scale in extremis to Dyson swarms look a little bit silly.

I think in the short term, though, it's all systems go, at least for the next 5 to 10 years.

I have to imagine that one of the first areas where AI is going to cause a massive disruption is energy efficiency and compute efficiency in these centers.

Yes. This is a regime right now where, as we were talking about quantum a few minutes ago, maybe photonics is an intermediate substrate before—if at all—we migrate to fully quantum systems.

As Feynman said, there's so much room at the bottom. There are so many new low-level infrastructural advances. There are scenarios where we don't need to fully tile the Earth, and the data centers solve a whole bunch of low-level physical problems for us, enabling us to keep this relatively contained.

And if you remember, if you go back to our—

Sorry, you were going to say the—

I was going to say the same thing. You remember Brockman said, “We want a GPU per human.”

Yeah.

And then, as soon as you have it, you'll want more.

Yeah. If you go to our podcast from a couple of months ago, we had a whole section on the software breakthroughs. To Alex's comment about the opportunity at the bottom, the best guess is a minimum of 10x—more like 10,000x—but somewhere between 10x and 10,000x in software improvement is coming.

We'll use all of it and want more. There's no doubt in my mind. Then there's hardware on top of that as well. We did a whole analysis of the different dimensions, and now they're multiplicative. We should revisit that because we have a lot more color now.

Yeah. But I think, to Sam Altman's earlier point about choosing between health and education, there will be a fundamental breakthrough. There just needs to be, because we can't expect the systems that we had a couple of years ago to perpetuate going forward.

I think we'll have the compute to do all these things.

This is fascinating. Here's an article saying, “Nvidia discussing new business model: chip leasing.” OpenAI struck a $100 billion deal to lease, not buy, Nvidia's AI chips, spread over 5 years.

I can just imagine the conversation between Jensen and Sam:

“Hey, listen, Jensen. I want those chips. I just don't have $100 billion.”

“Well, Sam, what if I just leased them to you over 5 years? Are you good for the payments over 5 years? Because I think our investors love having guaranteed revenues over 5 years. And who takes the depreciation risk?”

Dave, what are your thoughts here?

Actually, a bunch of our MIT best buddies, including Kush Bhatia here, are starting new companies around this entire area of creating new securities that allow you to finance all this stuff.

The hyperscalers are just going ballistic. This is so much bigger than all other forms of real estate investment combined—the aggregation of data centers and chips. The leasing was inevitable because Sam doesn't have cash on the barrelhead. Meanwhile, Jensen has the lead right now, and he has a $4.5 trillion market cap.

One way to lock that in is to use leverage. This is why Larry Ellison's the richest guy in the world—or was a week or two ago—because he used his balance sheet and his borrowing ability at 4% to finance a lot of bottlenecks that the startups can't afford, and Sam obviously can't afford.

Who's going to fund it? Just because you're leasing it, somebody still has to buy the chip up front. Nvidia is saying, “Okay, well, we'll fund the purchase of our own chips using our massive balance sheet and our massive market cap.”

This felt inevitable to me. It was going to happen at some point.

I think it's easy for skeptics to paint this as smacking of financial engineering and some sort of GPU credit bubble. I think the GPU credit bubble story, in addition to the depreciation that folks here have already mentioned, misses another storyline.

Right now, Nvidia is in a very high-margin GPU hardware business, and there's an impedance mismatch between selling high-margin GPUs and low- to negative-margin neocloud and cloud businesses. Leasing is the market contorting itself to accommodate that mismatch between high-margin GPU hardware and low- to negative-margin neocloud businesses.

Rob Fischer, who used to run Link Studio here, went off to build data centers.

He said he's signing deals 2, 3, 4 a week now. I think what's happened is that the visionaries started building data centers ahead of the curve, knowing the demand would come, and everybody was a little nervous about that.

The demand, at least as far as Rob is concerned, is here now. You can see it in all the use cases we demoed earlier in the pod. Those things didn't exist 6 months ago. Now anyone seeing those is going to want to do them immediately, whether it's corporate, personal, or just a theme song for the podcast.

Everyone's like, “Wow, that's really usable. Where do I get it?” Well, it has to run on a data center somewhere. It's not magic. I think the demand is starting to catch up to the construction, and the demand will get way ahead of the construction.

I don't think we've seen anything yet in terms of demand. Everybody is still just barely tickling ChatGPT and not really plugging in. Once we're spinning up agents, building new capabilities, and transforming our lives, we're going to see 1,000x per individual.

All right. I call this segment “Not investment advice.” Okay, let's jump in.

Create titles, then you don't even have to say it.

I'm going to continue on our Intel saga. Dave, congratulations on your options. I finally bought in, probably a generation of Intel options later than you did.

Here's a chart and a quote from Chamath: “President Trump got Intel to give Team America 10% of itself at $0. He has a better IRR than Buffett.”

Of course, if you get something for $0, you have an infinite IRR. But here we go: “President Trump makes 80% on Intel purchase in 6 weeks.”

Not bad. This was predictable, right? The United States cannot afford to let Intel fail. We remember that we did a podcast exactly concurrent with Lip-Bu Tan being at the White House.

Yep.

That was August 11th. I think the news came out the next day. We said, “Okay, Lip-Bu Tan will come out of the White House. It'll either be black smoke or white smoke, depending on how that meeting goes.”

What you're looking for is either Lip-Bu Tan to quietly disappear in a good way, or Donald Trump to reverse course.

Not quietly disappear, or Donald Trump to reverse course. Remember, he tweeted, “Lip-Bu Tan must go.” He's completely conflicted. He's invested in China.

Yeah.

Or, you know, Donald will either reverse course. It depends on whether Lip-Bu Tan says, “Look, I'm as American as apple pie, and I will build the best fabs in the world right here on our soil,” or he says something else.

Well, it came out white smoke, and that means Donald is going to make this succeed one way or another. The slides imply that Intel is way up this year, but it was August 11th—that was the date it was at its low for the year, or near the low for the year.

Yeah, well, 6 weeks, like it says.

It's crazy. This is sovereign venture capital—the government basically driving investor confidence and triggering momentum. I'm a libertarian capitalist. I don't know how to think about this, but I do believe that Intel is a critical asset for America, and it needs to be partnered up and supported.

Along those lines, we've got this other piece of news: “Intel stock extends its gain, hoping for AMD to go from rival to partner.”

The 2 big deals that are out there for Intel are partnerships with AMD, Apple, and Nvidia. This is, again, going to Chamath's terms, not mine: Team America here.

I think, Peter, there's a certain sense in which this was almost predetermined. By this, I mean the quasi-nationalization of Intel. I remember conversations I had with Intel engineers 20-plus years ago, and they knew, as has continued to be the case, that everyone knows Moore's first law: the number of transistors, or transistor density, doubles every 18 or 24 months, depending on which version of the law you like.

Not as many folks perhaps pay attention to Moore's second law, which is that the cost of a fab doubles approximately every 4 years. So, 20-plus years ago, you could imagine just extrapolating Moore's second law out and realizing that, at some point, new fabs would become so expensive that really only sovereign nation-states would be in a position to finance them.

This was reasonably well known within the semiconductor community 20-plus years ago: at some point, as Moore's first law is starting to end and Moore's second law is starting to become so expensive that only sovereign interests can afford to finance this, something like this, in some sense, was bound to happen eventually.

Bound to happen. Exactly right. I'll tell you, a lot of people don't talk about this, but a few years ago, we outsourced all of our PC boards—the green boards inside your laptop—to China for cheap manufacturing, for years, for decades.

Lo and behold, there were little spy chips, very small, about the size of a grain of rice, stuck between the layers of the PC boards, and that made it into all the U.S. data centers. So, that was grabbing all the passwords and transmitting them back to China.

Wow. The U.S. government discovered this. It had been going on for years. Rather than make a big international incident out of it, they said, “Holy crap, this is going to be devastating. We're going to lose confidence in all financial instruments and everything. We're going to squelch this story.”

It kind of disappeared from the news, and they've been quietly, for a long time, trying to clean it all up. So now, the idea that you would trust your highest-end chip manufacturing to be done offshore and repeat that same mistake is a nonstarter.

I start thinking, Dave, about what else falls into the “we cannot let it fail” category, and my mind turns to energy. I think we'll talk about that in the next segment here—the U.S. government needing to prop up firm generation, reduce regulation, and really accelerate our energy economy.

I'll be keeping an eye out for this Aschenbrenner-like moment of finding a company that is—I don't say too big to fail. I would say too centrally critical to fail.

You know—

Too scarce to fail.

Yes, don't talk about scarcity.

All right, let's move on here. Speaking about scarcity, Jensen goes on record with, I think, something very important: electricians and plumbers needed in the new working world.

Last podcast, we talked about how universities are failing. The perceived value of a college degree has fallen through the floor. At the same time, the category of workers who have been out of jobs the longest is new college graduates. It's insane. So how does higher education continue to charge what they charge in this scenario?

Here are the numbers: it's estimated that hundreds of thousands of electricians, plumbers, and carpenters are needed. The U.S. is short 500,000 construction workers in 2025. Rather than coming out of school $100,000 or $200,000 in debt, why don't you come out with a job that's paying $100,000 to $200,000, where you're needed instantly?

And it's not just construction; it's construction automation, too. This is why I can't wait to go to Abilene to meet with Chase Lochmiller, because you're like, “Wait, why would an MIT aerospace guy be the right guy to be running Stargate in Abilene?”

Well, because he looks at every one of these jobs and thinks, “How can I build a robot for that? How can I automate that? How can I restructure it so it's modular?” I think that's going to be the other side of this.

It's not just jobs in raw wiring and plumbing. It's jobs in management and construction automation—some very high-end jobs, a massive opportunity for employment. I really wish some more states would recognize that if you want your population in your state to be well-off, you've got to get the data centers up and running in your state.

So, here's another stat: Gen Z is choosing trades over college—a 16% rise in trade programs since 2023. Construction is the fastest-growing industry for new college graduates in 2025. I find that absolutely fascinating.

All right, I added these slides. I'm calling it an exponential reality check. A couple of days ago, one of my boys wanted to build a computer, so we're going to build a gaming computer. We're researching the GPUs, the CPUs, the memory, and so forth, and we're ordering them.

It turns out you can order everything you need—every component—on Amazon. I'm on Amazon, and I'm buying this DDR5 RAM kit, 32 GB of RAM, for $101. In the back of my mind, I'm thinking, “I wonder what that would have cost in the '80s, when I was building my first computer?”

Then we go on, and I'm ordering a 4-terabyte internal hard drive for $84. Four terabytes for $84. I'm going, “Holy, that's crazy.” So I hopped on ChatGPT and said, “Okay, give me an estimate of what this would have cost in the mid-'80s.”

Here are the numbers. They're pretty staggering. Instead of $100 for 32 GB of RAM, it was $150 million back in the '80s. A 4-terabyte hard drive that did not exist would have cost you about $1.26 billion to cobble together. I was just in awe of this.

If the top speed of a car had increased at the same pace as these curves, we'd have cars that went faster than the speed of light.

You know what I find incredibly fascinating? We finally have an answer to something that's vexed all of the AI and psychology communities for decades: what would it take to create human-level thinking outside of a human brain?

It turns out it takes about 8 to 16 GPUs of capacity, and those are about $30,000 each. You can store the human brain's storage on 2 of these. So, it's about $160 of storage for everything that can fit into a human brain—and then actually a lot more. We have massive overabundance of storage.

Yeah. But computer processing is still— the human brain is doing really, really well on 20 watts.

So, Alex, the best I can figure is we're going to go to molecular memory that will effectively be free in a couple of decades.

We can do better than that.

But we can do better than molecular memory.

Okay. We can also do better than free. We could do atomic-based memory. There are proposals for picometer-level memory, albeit at faster timescales. We could do femto-scale computing and storage. We could go sub-femto scale.

The physics of our universe goes so many orders of magnitude down to the Planck scale, and whether the Planck scale is physical is still an open research question. We're not going to run out of degrees of freedom to store cat images, or whatever else it is that we're trying to use storage for. There's lots of room at the bottom.

I always found it fascinating when I was doing my physics degree that no matter how big you want to go in the universe or how small you want to go, you have infinity, essentially, in either direction.

I do think, for what it's worth, there are scenarios where we start to run up against fundamental physics limitations, but we're still many orders of magnitude away at the moment.

Not something to worry about tonight on your drive home, folks.

Wait a few years. I added this as a segment we might want to have in future episodes as well, which is exponential book recommendations.

We've been talking about Accelerando. A few of our subscribers and listeners have reached out about that book, so I thought I would take a moment to chat about it. One of my favorite books is by a dear, dear friend who's often on stage with me and Salim at the Abundance Summit: Ramez Naam. He wrote a trilogy called Nexus.

Alex, tell us about Accelerando for a moment. Again, if you want some fun reading between episodes of WTF, here are a couple of books for you.

Sure. I love the book-corner concept. I would say Accelerando is my favorite book ever. It tells the story of a multigenerational family starting before the singularity, passing through the singularity, and going after the singularity.

In my mind, it is probably the single best fiction—or nonfiction, fiction in this case—depiction of what the 21st century is likely to look like. It has so many important concepts, ranging from, obviously, AI, nanotech, space development, and first contact, that are difficult to synthesize and that have apparently proven difficult for other authors to synthesize.

And I think just reading Accelerando, which I first encountered in grad school, has made me such a sci-fi snob that it's difficult to judge every other bit of science fiction by that standard. I had the opportunity to create a poster-sized version of Accelerando, which is available as a Creative Commons-licensed e-book, and present it to Charlie, which was a real pleasure.

But I would encourage every sci-fi writer out there to hold yourself to the standard of Accelerando, both in terms of optimism and physical realism. There's always a temptation, if you're a sci-fi author, to take one dimension of the world and extrapolate it narrowly, and that ends up creating, I think, highly unrealistic scenarios. Accelerando does a much better job.

He does. He only fails me in his extrapolation on space and space technologies, but I'm not going to be critical—it's an amazing book. I'm reading it, actually listening to it, for the second time. It's got a great Audible version as well.

Nexus by Ramez Naam came out in 2012. It's 13 years old, but it holds up incredibly well, so it reads as fresh today as it did back in 2012. It's a story about a guy named Kaden Lane. He's a young scientist who develops something called Nexus, a nanotechnology that's basically like a neural lace that links human brains directly to the cloud and links them to other brains.

It gives birth to a collective consciousness and allows you to run software apps on your brain, and it also goes deep into bioengineering. It's a look at where we're going to get to on the flip side of what Ray Kurzweil predicts in the mid-2030s as a high-bandwidth brain-computer interface. An amazing book, an amazing trilogy, and one of my favorites. I've read it 3 times now, the last time with my 14-year-old son.

So, Salim and Dave, any favorite books for you?

The Foundation series from Asimov is a classic that's just a must-read for everybody.

Okay, Dave.

Yeah, I only read what Alex tells me to read because his recommendations have been 100% perfect. I don't want to trump his great advice, but I will say that the terminology in the books alone makes it worth the investment. The stories are great, too, but if you read the books, then you get the terminology and you can keep up with what he's saying.

And I think that's really important. It's a great investment to make. Alex, would you come up with another recommendation? I'll do the same for next time.

Absolutely. So, my second and third favorite—

Hold it, hold it. Hold it for next time.

Okay, sure.

Okay. All right. We've got to keep our subscribers coming back. All right, let's jump into energy and robotics. OpenAI is planning a 125-fold increase in energy capacity over the next 8 years. This is more than India itself is putting out: 250 gigawatts of energy by 2033. Where are they today? Roughly heading toward 2 gigawatts. Thoughts, gentlemen?

If you do the arithmetic on this, if my arithmetic is correct, 250 gigawatts obviously represents a tremendous expansion over where we are now. On the other hand, it only corresponds to approximately 1/20 of 1% of the insolation—the inbound insolation on Earth's surface—that could be captured or recovered with solar photovoltaics.

So, even with 250 gigawatts for 1 frontier lab, we're still pretty far from Kardashev level 1, let alone Dyson swarms. I would like to see terawatts, tens or hundreds of terawatts. We'll get to solar in just a moment.

I found this fascinating. The US is planning to use emergency powers to save more coal plants. The Energy Department kept oil and coal plants in Michigan and Pennsylvania running past retirement. The reason is that they want grid reliability and don't want to risk the demand. We've seen the Consumer Price Index for energy starting to spike, and there's a definitive need for more energy. There are 100 coal plants set to retire in 2028. Of course, this White House in particular has been pro-energy of any and all types. Let me hop into solar, and then we can circle back to this conversation if that's okay with you guys.

Sure.

All right. Ember, which put this out, is an independent energy and climate think tank in the UK. You can see that this is a chart plotting energy from 2000 to 2025 across solar, coal, natural gas, hydro, nuclear, oil, and bioenergy. It makes the point that over the last 15 years, between 2010 and 2025, global solar capacity went from a low of 40 gigawatts to today's high of almost 3 terawatts of energy. So, Salim, take us away here.

Well, this is a really important piece to point out. We do this in all of our presentations, where we point out how hard it is to spot this.

And how bad our brains are, cognitively, at seeing this curve, right? You guys had talked about Chris Wright and his comment that in 50 years we'll see solar still below 10%.

Which kind of blows my mind. If we can flip the next slide—

Right? I want to give a couple of examples here because this is so—

So, read this one out for those who are listening.

So, this is an exponential graph with Vinod Khosla on it. What he did was go back and look at the exponential growth of mobile phones through the decade from 2000 to 2010, doubling every 2 years. He had a research analyst look at what all the industry expert analysts said would be the growth of mobile phones.

In 2002, they predicted 16% growth year over year. Two years later, it had gone up 100%. The 2004 prediction was not 18%, 20%, or 25%; it went down to 14% growth. Why? Because they thought there would be 14% growth because they thought it would level off. They had just seen 100% growth over 2 years, so it had to level off.

Predicted.

In 2006, they predicted 12% growth. It went up another 100% in reality. Between 2006 and 2008, it went up another 100%, and they predicted 10% growth. Then it went up another 100%. How much more wrong can you be from a 10% prediction when the actual reality is 100%?

This is the mobile phone prediction of all the top analysts, by the way—Gartner's and all these guys. This is critical, but this slide, I think, is killer. If you were driving, pull over and park and just look at this for a second.

What you see in black is the actual growth of solar energy over a 15- to 20-year period. The curve, by the way, is a total hockey stick up and to the right—an exponential of epic levels, just going vertical. What you see in the colored lines, which are all horizontal, are the predictions year after year from the top energy experts in the world as to the future of solar.

What we see is that every time solar goes literally vertical, all the experts go linear.

They basically do that year after year after year.

They can't continue scaling. It's got to keep leveling off.

It's got to level off.

Right? This goes from 2012 to 2017, 2018. Now, the 2018 graph was even worse. It actually showed it going down. The cost is dropping 50% every 18 months. How do you predict that it's going to go down?

This kind of drives me nuts because this is not a math error. This is a cognitive error. These are not laypeople, by the way. These are the top energy experts in the world getting it 180 degrees wrong. Literally, if I made predictions like this year after year, I should lose my job if I'm that far different from reality. This is the problem we have because our governments are listening to these experts.

It depends who employed them. If it was the, you know—

It's really kind of unbelievable. There's a whole other one about electric cars that I won't get into. They predicted that we would not have more than 1 million electric cars by 2040, and we crossed that in 2014. Even then, they didn't update their predictions.

I'm going to give one more here. This is a graph of solar modules dropping, then leveling off for a bit, and then dropping again like a stone. In 2003, the leading energy expert in the world in solar energy itself made a comment. He said, "Look, if you add up the cost of the silver, the glass, and the wiring—the physical component cost of a solar module—you'll never get below $1 a watt. That's the limit. That's the actual limit."

The market actually believed him for a while, and it flattened out for a few years. Then it started dropping. By 2014, it was 50 cents a watt. Now it actually goes off the bottom of the graph. Where we are today would be where my feet are sitting on this chair. When the graph is this big, we're down to about 2 cents a watt, or close to a penny a watt.

His comment when he was showing this was, "Okay, getting below $1 exceeded my expectations." That was his comment after being this wrong. So, it's really, really hard.

I want to give a final example that we don't have a slide for, just to be fair to these folks and to show how hard it is. Over the last 20 years, if you owned a car wash in Buenos Aires, Argentina, your revenues as a car wash owner have dropped by 50%.

One of our community members, Santiago Bilinkis, who I think, Peter, you know well, lives there and says, "This makes no sense. The middle class has exploded. We have a ton more Mercedes and BMWs running around. Argentinians are very proud; they like to keep their cars clean. There should be a doubling or tripling of revenues."

Why is there a 50% drop? Are there water restrictions, hypercompetition, or legal issues? He starts looking into it and, over a couple of months, gets rid of all the obvious factors. Then he finds the answer, which literally turns out to be Moore’s law, because our computational ability over those 20 years has increased quite a bit.

Our ability to model the weather has gotten a lot better, and over that 20-year period, we’re exactly 50% better at knowing when it’s going to rain.

And when you know it’s going to rain, you don’t wash your car, right? The reason this is important is that you can be the smartest car wash owner in the world and you will never see that coming. We call this in the book the orthogonal effect of innovation, where a breakthrough in one domain affects you radically and you don’t see it. You can’t see it.

It’s so critical to keep track not just of the demand side, but the supply side of things. The most famous of all these—I’ll just end my rant here—is in the 1980s, when McKinsey advised AT&T on the future of mobile phones. They predicted that by the year 2000, there would not be more than 1 million mobile phones in the world, and AT&T left the business, saying, “That market doesn’t work.”

By the year 2000, we had 100 million mobile phones. So they were off by 99% on that one. At one of our executive programs at Singularity University, a guy puts up his hand when I mention this. I coauthored that report, right? I’m thinking, “Oh my God, what is he going to do? Is he going to rebut this?”

He goes, “No, you’re absolutely right. The reason we got it wrong was that when you had these big handsets with these briefcase batteries, we figured there was no way you were going to sell more than 1 million of those. What we didn’t see was that within a couple of years, they had shrunk to a clamshell and you could actually sell a ton of them.”

That’s the part that people miss. When you track these things, be really, really careful about making outlandish predictions like, “It’ll never get below this,” or, “It’ll never get above that.” We’ve seen that repeatedly.

I don’t know why you want to end that rant. That was the coolest thing ever.

For years, we’ve been struggling with this, talking to governments, and they’re like, “Yeah, this will never happen. That’ll happen.” We go berserk.

Love those slides.

I love those slides. And you know what else? When Bill Gross was on the pod, he said, “All the land where pumped hydro makes sense has already been bought.” I did a little research, and that’s actually not true. Lots of land where pumped hydro makes a ton of sense—but it’s not quite as sunny—has not yet been bought, if anyone’s listening.

And because the solar panels are getting so cheap, you can just put more of them there.

And so, heads up, there’s a theme. If the governor of New Hampshire is listening, please give me a call. There are lots of opportunities that haven’t been tapped in real estate.

I have 2 more quick energy factoids. First, I did a little bit of research, and I was talking to one of our energy gurus in our ecosystem. It turns out that, if you add up all the dams in the U.S., there are 10 gigawatts of potential hydroelectric power that hasn’t been tapped.

So we could use all those dams. That’s a big—sorry, I’m really going off here. But I remember we were on the pod and we said, “Holy, the Hoover Dam right now is operating at about 5% to 10% capacity because it hasn’t rained.”

Yeah.

So we’re like, why the hell are we not doing pumped hydro right here? Just pump the water from the bottom to the top. Tons of sunshine right there. It turns out somebody had already—

—thought it through and put together an entire investment thesis around it. But that theme isn’t over. That is very hot. I think the point we started this whole conversation with is that China is running away with solar deployment.

I don’t understand why we don’t see it here in the U.S. I’m a pilot; I fly out of Santa Monica Airport. I fly over L.A., and all I see are naked roofs that could all be producing electricity. There are a few solar thermal farms out in the middle of the desert, but there’s so much potential—so much potential.

All right.

Geopolitical.

It’s geopolitical because China has pretty much a lock on the supply chain and the panels.

I would be investing in building out solar capacity and manufacturing here, right? SolarCity. And actually, what I would look to do is say, “What’s the 10× to 100× breakthrough on photovoltaics or solar past the next level?” Then go after that.

And Alex, you know, digital superintelligence and new materials science will give us new capabilities for that. So there will be—

That’s why Alex is standing there, not looking worried at all. He’s like, “What are these guys worried about?”

I think there are many ways to generate useful energy. I think fission in the form of SMRs, and fusion potentially as soon as, as we’ve discussed in the past, 2028 to 2030. I think there are so many non-solar, novel-ish forms of energy that are on the verge of coming online.

I’m not losing sleep over geopolitical imbalances in solar photovoltaics.

All right, let’s jump into robotics here. This is a fascinating tweet turned into an article: “China’s robotic boom is going global.” If you look at the first half of 2025 and the countries around the world that are purchasing robots from China, Poland is up 1,700%, Mexico 275%, Russia 135%, and Vietnam 114%, as opposed to South Korea and Germany at minus 3%, and the U.S. at 58%.

The point here is that countries that are blank slates, that don’t have a robotics industry, are buying from China. Countries are starting their automation journey and buying from China. This is something the U.S. needs to be looking at. Basically, China is staking its flag in countries around the world by deploying both AI and robotics in a very cost-effective fashion. Dave.

I wouldn’t be surprised, given how central robotics in general—general-purpose robotics, more particularly human general-purpose or humanoid general-purpose robotics even more particularly—are to this emerging industrial ecology of batteries, fabs, chips, AI compute, and probably SMRs and drones, that we see an emerging demand function for fully sovereign robotic ecologies.

To the extent, Peter, you were suggesting earlier that you’re looking for other resources that may be too scarce to fail, robotics, I think, is a plausible candidate for wanting to be sovereign-aligned resources in the near-term future.

Yeah. You know, I had dinner with Rodney Brooks, the founder of iRobot, when we were out in California a couple of weeks ago.

Yep.

He reaffirmed what I think we all know: Our whole parts and component supply chain is garbage compared to what China has. All those years of manufacturing and industrialization moving over to China led them to develop a very, very flexible parts and components contract supply chain.

If you need something to build your robot, you can call someone and have them make it, and it’ll be there in a few days. There’s no equivalent in the U.S. So it’s going to take a while to rebuild that whole supply chain. What Alex said is exactly right: This is ripe for national involvement to kickstart it. It’s also not naturally happening in the venture community.

It’s really tough for a venture capitalist to plunk down $10 million or $20 million for an electric motor winding company or a gear company.

They should have a Manhattan-style project for the supply chain for robots and drones.

There are various initiatives that have been discussed. We heard this from Bernt Ørnich, CEO of 1X; we heard this from Brett Adcock and from Elon directly. They’ve had to completely build their entire bottom-up supply chain internally. Every component is manufactured inside the company right now, which is insane.

What a waste. But the other thing that’s going to be interesting is that there will be a scarcity of robots for the foreseeable future, until production gets ramped up.

We’re going to start to see governments probably bidding, like, “We’ll buy 1 million robots here in Saudi Arabia, or the Emirates, or Qatar,” in order to get early supplies delivered there. That may bid up prices in the early days, too.

Good for the world. I would view any emerging robot scarcity as just a facet of compute scarcity. The most important robots are just going to be GPUs on legs, and the compute ultimately is, I think, the fundamental scarce factor here.

Next item here is an interesting graph, which asks the question: What if everyone in the U.S. drove like Waymo? Here’s the extrapolation: If every U.S. vehicle performed as well as Waymo, we’d prevent 33 to 39 thousand deaths annually. Pretty profound.

I found a better related statistic.

Please.

It turns out about 50% of all the court cases in the U.S. are car accidents.

Wow.

50%. So you take out a bunch of lawyers, too, which isn’t bad. That’s a good thing. With all due respect to some lawyers, reducing the number is definitely a good thing. So this is huge.

And interestingly for Waymo, nearly half of all Waymo impacts, or crashes, happen under 1 mile per hour.

So these are just bumps. They're not actually crashes. I saw this statistic and said, “That’s got to be global, not the U.S.” Because that’s about the total number of deaths.

We kill 1.2 million people a year around the world in car accidents globally.

Around the world, yeah. Well, that’s why I thought 40,000 out of 1.2 million was viable, but 40,000 in the U.S. wasn’t. But if you read the fine print in the notes, it’s actually a 90% reduction in fatal crashes.

It’s huge.

And 15% of all organ donations come from auto accidents.

Interestingly enough.

Right. So, I live here in Santa Monica, and Waymos are all over the place. I just started seeing the Zoox vehicle from Amazon going and collecting data, right? It’s a pilotless vehicle with all of the lidar and cameras around it, going and mapping the streets. It was about a year ago that you saw all the pilotless Waymo vehicles mapping the streets. So, we’re going to have Zoox, we’re going to have Waymo, and we’re going to see the Cybercab, or whatever Elon calls it, very, very soon.

Meanwhile, we have people attacking the Waymos.

Brad Templeton used to joke that because we don’t want to be killed by robots, we’d much rather be killed by drunk people, which is what’s happening today.

I suspect that, for at least most Americans, their first encounter with a generalist robot is going to be by either driving in or seeing a Waymo- or FSD-based car, Zoox, or the equivalent. This is just the beginning of a longer journey. We start with these generalist robots on the roads, and they’ll be in our homes before we know it.

Uber is partnered in part with Waymo. They’ll be offering Waymo as part of your Uber app, and they’re also working with Joby for flying cars. So, super fun. We’ll be talking about all of those things and where Uber is going in the future.

Flying cars are my big hope for technology in the near future.

Yeah. Tired of driving.

Airport transfers are just horrible.

Oh, it is awful. All right, we’re going to wrap up with health and biotech. I think one of the most important subjects, at least in my life, is how we double our human lifespan and avoid all of the travesty of chronic disease.

The first article comes in from a friend, Joe Betts-LaCroix. Joe’s company, Retro Biosciences, is one of Sam’s companies. Sam founded it with $180 million of backing back in 2021. Their mission is to add 10 healthy years to the human lifespan. They’re one of the teams competing for our $101 million XPRIZE Healthspan.

Salim and Dave, since you’re on the board of XPRIZE, that competition is pretty amazing. Just for everybody, if you haven’t heard of it, I raised $157 million for a global competition to add up to 20 healthy years to people’s lives, particularly in immune function, cognition, and muscle. We now have over 730 teams that have entered that competition, which is pretty amazing, if you ask me.

That’s got to be a record, right?

It is. That’s incredible.

Yeah. Well, actually, for Elon’s $100 million Carbon Removal XPRIZE, we had 1,300 teams.

I would have to say this is as hard or harder because you have to run, effectively, a clinical trial and prove on a human population that your therapy didn’t just improve cognition, muscle, or immune function—it did all of them.

I love the fact that Retro is going after this. Their product is entering human trials next year, with a hope of getting something in Australia in late 2025, and they’re going to be hopefully getting something on the market in the next couple of years. This is called RTR242. It’s an experimental Alzheimer’s pill designed to restart the brain’s natural recycling process of toxic proteins. This is your glymphatic system. When you’re in deep sleep, your glymphatic system is clearing your brain of those toxic proteins. One of the biggest things I had Mett Oz speaking at the Abundance Longevity Summit as well, and his biggest concern for the future is neurodegenerative disease and also one other disease called loneliness. We should talk about that sometime.

I want to end with this article. This is out of China. One of the things about longevity in biotech is if it works in China, it’ll work in Chicago. If it works in Boston, it’ll work in Botswana. We all have the same biology. Chinese scientists have genetically engineered a gene called FOXO3 that is a critical stress-resistant transcription factor, and they’ve been able as they modify this to reduce aging by 3 to 5 years. For me, this is a huge deal. In 61 different tissues, at the end of the day, we’re going to start to see longevity becoming more and more real.

And everyone listening, I want to let you know that the next 50 years that you’re alive and hearing us on this podcast are going to be awesome. Just don’t get hit by a bus in the next couple of years.

Yeah, exactly. Don’t die from something stupid. In the interim—

Peter, there was a comment I heard a few years ago, a couple of years ago, and I wanted to just ratify where we are with that. Somebody on one of the Abundance stages said that we have labs, mice today, that are living to the equivalent of 300 years old already. Is that—and are we really there?

No, we’re not there yet. The average mouse is living on the order of 20 to 24 months. We’ve seen extensions of 30 to 40%. There are experiments where they hope to double the mouse’s lifespan going on right now. We’ve also seen the first epigenetic reprogramming trials going on in humans starting in January. Life Biosciences is one of David Sinclair’s companies going into humans. It’s been very successful in animal models, including non-human primates.

And so it’s interesting because I define an expert as someone who can tell you exactly how it can’t be done.

Yes.

Yeah.

And for what it’s worth, Salim, I’ve asked this question of all of the best frontier models of the day: When do we get longevity escape velocity? Their consensus is 2030.

Yeah, which ironically is the same time when Bitcoin hits $1 million, according to all the frontier models.

Which is exactly what Ray Kurzweil predicted: 2030.

It’s like, damn, it was right.

Damn the man.

He may be proof that time travel is real.

Yeah. That and Elon. Yes, exactly.

So, everybody, you’ve got to hang on. Stay in good health: sleep, diet, exercise, and mindset. Don’t die from something stupid. You’ve got to hold on for the next 5 years. There are therapies coming, and they’re significant therapies.

Let me give kudos to the Moonshots community here. When I did that podcast with David Sinclair, he came on and was really miffed. The Harvard–White House debate and head-butting had canceled all of his funding. $4 million of funding got canceled, and he was on the verge of letting his entire research team go—all of his researchers.

I was just pissed, and I said, “Let’s turn this around.” On the podcast, almost off the cuff, we announced something called Friends of Sinclair Lab, where people would contribute $50,000. I was the first to offer to contribute, as was David himself. Since then, we’ve received over $4 million in donations from the people listening to this podcast.

I’m looking to buy a Ferrari, if anybody wants to donate to that.

No, but this is decentralized science.

It’s citizen-driven, bottom-up science. It’s so awesome.

And the challenge is that when you’re funded by the government and have peer review, you’re stuck in incrementalism. Anything dramatically different, they don’t want to fund.

Yeah. It’s great.

Yeah.

Dave, what’s your week looking like for you, buddy?

Well, it’s Friday, so we have a lot of our best and brightest who are coming through the lab and getting funding right now.

A lot of them are getting West Coast term sheets at 2 or 3 times higher than on the East Coast, so there's quite a bit of migration west going on. One of our coolest companies—we signed the term sheet in Mark Zuckerberg's old dorm room. There's a poster of The Social Network movie signed by Mark Zuckerberg on the wall, so we signed the term sheet right in front of the poster. Then that got all around Harvard, so 20 people joined the company for no salary because they're so hot. Anyway, they're smoking hot now. It's called Biograph. They're moving to the West Coast.

I got a whole bunch of open seats here in the lab, so I'm really excited to spend time on campus backfilling. We're going to try and get 16 more teams in. January is coming fast. MIT has January off.

That's the perfect time.

Perfect time to boot up a company.

If you're at MIT, Harvard, or Northeastern and you're hearing this podcast, first of all, Dave's a rock star. If you've got a couple of best friends and you want to start an AI company, where do they go, Dave?

Go to the Link Ventures website, or just email Dan Oliveri or Kush Bawaria. Their names are on the website, and it's just K. Bawaria or D. Oliveri at Link Ventures. You have to have at least 3 people who are bona fide best friends, and we'll check—we'll poke around and ask your other friends, “Are you really best friends?” But we only bring in teams that are super tight-knit.

It keeps it all really fun. Salim, how about you? What's the week ahead look like?

We're doing a whole bunch of planning with our ecosystem to think about how we leapfrog everything we've done in the past and go 10x faster, better, and cheaper with all the offerings that we have. We have our next ExO 10X Shift workshop on October 15th. It's $100. Those are all selling out, and they're great. We cover the model and show people how to take their organization literally 10 to 100x through that 2-hour workshop.

I've got a little bit of travel, but not too much before the madness toward the end of the month. Visioneering is coming up, which I'm super excited about.

Yeah, for sure. And Alex, welcome back from your secret mission. I'm excited to work on our project together, which we'll unveil at some point. We're going to keep it secret for the time being. What's on your agenda?

I'm trying to accelerate the singularity, or whatever it is. Maybe the singularity at this point isn't even the right term, but I'm smoothing out and moving toward whatever we want to call it—the intelligence explosion, or, if you're a technological determinist, what was always going to happen: the inevitable byproduct of building an internet, then compressing the internet, and then using that to solve everything else.

I think timelines are very short at this point. Every week, my timelines are getting shorter. Usually, I'm the accelerationist in the room—not always, but usually—and my timelines are incredibly short at this point.

My favorite thing these days in these podcasts is watching Alex's face when we rant about energy or healthcare or something. He's like, “Superintelligence is going to just solve that. Why are we even talking about this?” He's got this great look on his face.

I mean, you're reading my face, I think, correctly. There is a certain sense of a hyper-deflationary mentality. Why do anything?

Really?

Paralysis. It's like a starship that heads out, and when they get there, they find out warp drive had been invented. There's a term for it. It's called the wait equation, and it does cause singularity paralysis, for lack of a better term.

I'm seeing it more and more every day in conversations I have, as it dawns on more and more subject-matter experts that AI is about to transcend their capabilities in, call it, 2 to 3 years, if the current extrapolations hold. What happens next? I spend a lot of time thinking about that.