AI引发的股市冲击、Google的利润难题，以及Altman的“看我的”时刻 | #307

I checked in with one of our portfolio companies, Pirates Labs, which makes lasers. They said that if anyone has pure graphene more than a millimeter thick, they own the world.

I’ll get you all the graphene you want.

What do you mean?

My desk drawer is over here. Oh my God, I have access to the biggest graphene manufacturer in the world.

Who’s that?

Nano Nano Nano Nano something.

All right, guys, I have a question. Chamath dropped a bomb last week, saying Bitcoin is capped, that it’ll eventually get hacked, and then it’ll go to zero. I’m curious what you thought. His basic premise was that central banks can’t really hold it, so the buyer pool is limited, and self-improving AI and quantum computing are going to disrupt it. Salim, your thoughts?

I think it’s a cop-out. I think he’s working on behalf of the Treasury secretary to say that it’s a complete reversal. He used to be a massive Bitcoin bull.

He was saying it was going to a million bucks. He was talking his book.

Now he’s saying the total opposite. I’ve talked to Jeff and other folks. All they have to do is solve the right consensus mechanisms. And if you saw the crypto thing, Emad, you may have a better view on this, but elliptic-curve cryptography seems to be holding steady even in the face of quantum. We can shift to quantum encryption and other things. As far as I can see, it’s not going anywhere.

The viewers are freaked out.

Yeah.

Well, the viewers are saying, “Please have the Bitcoin debate,” so we’re going to do that.

Yeah, please.

In a week or so, either at the end of next week or the week after. And, of course, you guys saw that Binance got hacked, right?

Yeah.

Well, we can talk about it for a minute, but $6.3 billion got ported out. It’s the largest crypto theft in history, and it might be the largest theft in history. Apparently, it wasn’t a technical failure. It was the weak link in the loop: the human.

Yeah, it was one guy’s machine. One engineer was compromised by the attackers, apparently from North Korea—that’s the accusation right now. They got a hold of his machine, got in the door, and stole $6.3 billion. That’s insane.

Yeah. In 20 minutes.

In 20 minutes. Crazy.

You had to do it all at once, right? You couldn’t do it a little bit at a time.

Yeah, you couldn’t dribble it out. People would notice.

The good news is that the other exchanges have put a lock on everything now, so wherever they moved it should be somewhat traceable. They’re trying to move it into smaller and smaller chains as fast as possible to cover their tracks.

But that’s the problem. Crypto is not fully anonymous. You can follow the chains.

Yeah.

Apparently, $3 billion has been frozen, so hopefully it’ll be recovered. The majority were stablecoins. It was Tether and Circle; it wasn’t Bitcoin.

Any thoughts on that, Emad?

Yeah. Tether and Circle can freeze it. The human is the weak link. But this is a problem with the banking sector, too. Bangladesh lost $100 million via Macau when someone pretended to be the central bank minister there, using SWIFT, and the money went through. You’re going to see more sophisticated attacks. I’ve had calls from my mother saying, “Send money,” and things like that.

Did you mean with voice-spoofing attacks?

No, it’s the other way around. You’re going to see more and more sophisticated attacks. But when we get to Bitcoin and all this stuff, the models are capable of crazy things, and they haven’t released the cryptographic advances for a reason so far. That comes at the same time as quantum computing. If anything can be hacked, it probably will be, because AI can ask quantum computers the right questions. But whether that impacts Bitcoin, given where it is right now, and makes it go to zero—I think it goes to zero. There’s always going to be a buyer for it as a historic asset that can upgrade its technology, because 50% of the hash rate is two miners, right? They can upgrade quite easily.

But it could take hits. I think the bigger problem, as we discussed with Cathie Wood on stage, is the energy. Who’s building all the new FLOPs for it? All the miners are saying, “I want AI chips instead. Why do I have ASICs?”

Alex, welcome.

What Chamath was saying, though, wasn’t technical. It was that you could easily increase the encryption, but they’ll never agree on how to do it. They can’t even agree on how to meet.

We don’t know in this age, except that every big target is going to come under attack. How does it handle that? One of my friends had an NFT exchange, and swarms of agents were attacking it. He was fighting back with agents and ended up rewriting his entire stack in one day because they couldn’t figure out a way to make it secure. The losses were only about $100,000, but they were visualizing it out, literally. His agents were going back and forth, finding everything.

Thank God for humans. Thank God for other things as well.

I think it’s a very good time to protect your credibility at all costs.

And by the way, Mark Carney invested $1.6 billion in Cohere’s latest round, right? From Canada, so did—

Yeah.

Okay. Canada is going to start investing directly in—

Before we get started, let’s take a moment and say farewell to Sir Patrick Stewart, Captain Jean-Luc Picard, who passed away on Wednesday at his home in L.A. He was 86 years old. Listen, William Shatner will always be the captain of the Enterprise for me, but Jean-Luc Picard was extraordinary.

Unbelievable.

A very beautiful man.

He brought a lot of legitimacy to us sci-fi folks.

One thing I’d like to clear up about the commentary on the Borg: at some point they switched from saying “Resistance is futile” to “Resistance is feudal.” Let’s just say for the record, it’s “futile.”

Futile.

Because that’s how English is pronounced.

Alex, any thoughts on Jean-Luc?

Yeah, it’s very sad. Hopefully, with humanity’s common task, we can someday resurrect Patrick Stewart. The character of Captain Picard is already being resurrected in a million different Seedance sessions as we speak. I think the character will live on, and hopefully the human can be brought back as well. And then, narrowly, on Binance and Bitcoin—crime, whatever—this is an unproductive use of capital. So sad. So sad.

We know. We know.

It’s an unproductive use of capital.

It is not. It is not just the ability to do micropayments or the ability to do things in the world. Awful.

I’ll save it for the debate.

Yeah, save it.

Let’s schedule that, because I want to get Jeff’s time as a slot in one of those, and all the rest of it.

Yeah, for sure. It was fun. I saw Elon saying this morning that longevity is coming. He’s finally converted. He’s finally using some of his Grok cycles to hit longevity.

It’s a huge market, bigger than AI. To the extent that SpaceX is no longer a rocketry company and is now a neo-hyperscaler, I’m sure Elon himself will discover GLP-1s. Whatever market they’re creating in longevity, I don’t know—maybe Tesla pivots to GLP-1s or GLP-1000s, whatever it might be.

I had a call this morning with a friend, Max Jaderberg, who’s the president of Isomorphic Labs. Just amazing stuff going on there. Extraordinary.

Now that’s a moonshot, ladies and gentlemen.

Back from Stanford late last night and back at MIT, which is just bubbling. I cannot tell you how much it’s bubbling over. Everybody is just optimistic about AI, and I saw a whole bunch of Abundance T-shirts walking around campus.

Yeah, yeah. Solve everything, abundance, P(doom) less than zero. Those are our themes.

So, is the Moonshots merch still up and running? Well, we’re actually in the middle of redoing the Moonshots website, and we’re going to offer the merch there so our fans can get all the slogans, T-shirts, and mugs. I’ve got my Moonshots mug right here. It says on the back, “Where can you go 10× bigger than others who are hoping for 10%?”

Oh, we’re the Ministry of Information, the Ministry of Truth, if you will, for the Singularity. So we might as well have great merch.

For sure.

Yeah. Got to make some revenue someplace. The number of the week is 1 billion. That’s how many users the Gemini app has. It’s the fastest-growing product in Google’s history. And yet, as you’ll hear today, Alphabet just had its worst stock day in history.

We got some great love from our subscribers in our last episodes. The quote was, “Congrats to Alex on his chain theory.” Very impressive. You got a lot of love for chain theory, Alex.

It’s very kind, but the ultimate court isn’t public opinion. It’s the physical world. It’s reality.

It is reality. Thank you. But I care even more about, in some sense, not just predicting and retrodicting reality, but also doing my part to make sure the institution of physics, especially in fundamental physics, cares about the real world and not just pure theory. That’s the ultimate arbiter here.

Thank you. I think a decent portion of the physics research community is stuck in an unfortunate state where it’s lost contact—or, let’s say, lost the spirit of making falsifiable predictions about the physical world. That used to be the case. Unfortunately, for whatever reason, and I certainly have my own views there, large portions of physics research don’t even aspire to make falsifiable predictions about the physical world. So we’ll do what we can to bring the good times back.

Okay.

Yeah, it’s what Wolfgang Pauli said: “Not even wrong.” It can’t be wrong, you know.

1. Google's Gemini 4 Carbon and the AI Race

Yeah. I got to use that. I got to use that. That comment—thank you, Peter Woit, for popularizing “not even wrong.” All right, let’s jump in.

This morning, we’re going to open with a sequence of 4 stories about Google. Quick reminder and disclosure: Google is one of our sponsors on this pod, but they have no editorial control, and we had to cover this. Google is so back; they’ve taken back the AI crown. On Wednesday, Google CEO Sundar Pichai announced Gemini 4 Carbon, only 2 weeks after announcing Argon. That’s the pace of the frontier: 2 major models from a single hyperscaler in a single week. On the Artificial Analysis Intelligence Index, it had a score that beat Opus 5.5 and GPT-6 Astra. The score was 59. We’ll take a look at that in a moment.

But here’s the real headline for me: cybersecurity. Carbon found 88% of vulnerabilities on a benchmark called SEC-Bench Pro. It successfully exploited 55% of the flaws on ExploitBench, and it wrote working patches for 71%. Google’s own Chrome security team says it’s more than 10 times better at producing high-quality patches than a competing frontier model. It’s a big deal. I’m curious. I’m going to go to you first, Alex. What did you think? Because honestly, we’ve been saying on this pod that Google is so far behind they should open-source it. And here they are leaping forward once again. We’ve seen this every time, right? Meta leaps ahead of everybody else, then it’s Anthropic, it’s OpenAI, and Google’s back.

I so want Google to be back, but there’s a big asterisk. There are 2 things I think may be buried in the headlines. One is that this model isn’t generally available, unlike, say, Opus 5.5, which everyone has access to. By the time it becomes generally available, it may no longer even aspirationally be at the frontier.

Second thing.

Yeah. No, well, okay. Peter, I’m exercising my editorial independence here, and I’m calling balls and strikes. The other thing is, if you look a little more closely at Artificial Analysis, they have a chart measuring the Intelligence Index versus cost per task by model release. That one is especially striking and instructive. If you look at Gemini 4 Carbon and then at the frontier defined by Opus 5.5 high-level reasoning with fallback, Gemini 4 Carbon is exactly on that frontier. You can extrapolate the line.

What that would suggest is that Gemini 4 Carbon might essentially be the equivalent of Opus 5.5 with fallback, just cranked up to a higher reasoning level and therefore at a higher cost. If that’s the case, I’d say Gemini 4 Carbon might be at Opus 5.5 level, but it’s unreleased. In my view, if Google wants to credibly claim the capability-frontier crown, it needs to do a better job of actually releasing models, not just stealth-releasing or privately releasing them.

I have friends on the team, make no mistake, and I know they listen to the pod. But they really need to get the models out there. Otherwise, it’s an apples-to-oranges comparison. I’m sure OpenAI and Anthropic have models internally that would blow all of these released models out of the water. But because they’re not generally accessible, it’s not a fair comparison.

Well, Google has released it to trusted cyber defenders first. They’ve got a program called Fairwind, and they haven’t announced when they’ll do a broader release. The other thing is that Carbon’s come in 23% cheaper than Opus.

If you have access to it, though. Again, I don’t want to be the party pooper, but I do feel the need to call balls and strikes. If they’re subsidizing it, we don’t know what Carbon 4’s profit margins are for Google. This is presumably a very, very, very tiny portion of their business, versus Anthropic and OpenAI, where we have a reasonably good sense of their margins. Google could hypothetically be subsidizing all of this and offering it below cost. We just don’t know.

I would love much more transparency from Google before I’m willing to say it has achieved the frontier. That said, I want the frontier to be competitive, and kudos to Google for at least trying.

Well, I wouldn’t expect Google to be super transparent anytime soon. They are in an incredibly good position. The TPUs are now second only to Nvidia in terms of total compute manufactured, and they’ve done it incredibly quietly.

2. Anduril, Qualcomm, and American-Made Chips

My read on Google is that they have so much cash flow from the core search business. They don’t have an IPO coming up, they don’t need to be big in the PR wars, and they’re very quietly accumulating an incredibly powerful position. Whether they execute with it or just quietly keep accumulating, we’ll find out. But being vertically integrated between TPUs doing all the thinking now, their internal foundation models, and a huge, insane treasure trove of training data—I mean, a crazy data advantage—is huge.

Keep in mind that Elon co-founded OpenAI specifically because Google was just going to run away with the entire world. Now we have a little bit of balance in the force, thanks to OpenAI and then people leaving OpenAI to start Anthropic. Absent that move, it would be an all-Google world right now.

I mean, the point I want to make is you can’t count any of these labs out. They’ve all got massive pressure to leap to the frontier and maintain it. Emad, what was your take on Carbon?

I mean, how many people are even smart enough to use the smartest models anymore? Let’s be real. Anthropic just released a prompting guide where they’re like, “You’re being too complicated with prompting. These models can do whatever.” Now it’s basically become: you either use the models to do day-to-day normal work, and they become very capable and very consistent, or you want a swarm of 1,000 or 10,000 of them and the capability to swarm them to solve the most complex work. There’s not much in between.

So I think the focus on the frontier—as you said, there are better models inside OpenAI and other things—is starting to miss the point a bit. What really matters is whether it’s in front of people.

Google announced Gemini Spark. What’s Gemini Spark? Everyone says, “Well, do you remember? It’s going to be their personal assistant.” But why is nobody using that versus Dots or Grokbot, which now has a million users, and things like that?

This is the flip side, and the dangerous side, because Google and Meta, if they don’t have the AI that’s closest to you, that’s buying your things, that hits their main business. In fact, I’m going to make a call here: I think search advertising will top out as an industry in Q1.

And that will be the big—

Top. And then that’s a really, really gutsy call.

I appreciate it. I cannot disagree. It’s falsifiable, which is awesome. That’s brilliant. I’m proud that you put that out there.

I mean, Google and Meta—look at it. Advertising is growing way faster than GDP every single year. But next year, and I think as soon as Q1, if that tops out—the structural growth in advertising, because people are buying through their agents and their agents are doing SEO—what’s Google’s business model? What’s Meta’s business model?

Now Google has to think, “Well, maybe we get Astro Teller and launch 10 companies a year or 20 companies a year,” because it’s time to diversify.

GCP, obviously. If you look at the revenue mix from Google, or Alphabet, in its last quarterly report, it’s so obviously transitioning away from being a search advertising company to becoming a hyperscaler cloud company. I think Google, if this trend continues, regardless of what happens with search advertising, is on a trajectory for search advertising to become its second-largest business in the next year or two, not even its largest business. They’ll just care less.

Yeah, I’ve been saying this for years. There’s no way advertising to a human is going to be how any of this works. Your AI is going to know exactly what the best products are, what you’re using, what’s right for your genome, what’s right for your schedule for the next 3 days, and it’ll buy it for you.

Humans are just there to be hacked.

Yeah. So, to flesh out what Alex said, GCP is growing like crazy.

The margins, though, are higher on search advertising. It’s a little misleading to say they’re fine because GCP is growing so quickly. If Emad is right and search tops out in Q1, there’s a gap of about a year where search revenue is declining while GCP’s margins haven’t caught up yet. But if you look beyond that, they’re in a phenomenal position.

The other thing is that GCP is growing on the back of TSMC manufacturing, which is completely and totally sold out. Demand is multiples higher than they can produce, so GCP can win by getting more TSMC capacity. That’s not a given. I don’t know how TSMC is deciding what to make between Apple, Google, Nvidia, and all the others—

And SpaceX.

3. Closing Thoughts and Singulitarians

Yeah. They’re not doing it on price. In a rational market, they’d just raise prices, but they’re not doing that. It’s all politics. It’s very smoky backroom stuff right now. It’s hard to know exactly what’s going on, but that’s the linchpin. The Terafab will ultimately come online and make it a balanced world, but right now it’s completely out of balance. Two-thirds of all the real chips that matter are made by TSMC. It’s a very odd, very fragile world we’re in.

It’s also worth remembering that Google has such diversified holdings. They’re a huge holder in SpaceX, in Isomorphic Labs, and in all of DeepMind’s other projects. They’re not going to be hurting for some time. Wall Street is fickle.

You know—

The question I’ve got in all this—

Google owns 14% of Anthropic.

They’re riding all the curves.

Isn’t that wild? They’re just everywhere, quite and very quiet. They have antitrust action all over them, so it’s easy to say, “Well, Google doesn’t seem to be doing much.” They’re trying like crazy to make it less obvious how powerful they really are because of the antitrust action that’s ongoing right now. There could be hundred-billion-dollar settlements related to this antitrust action, so they have every reason to stay off the radar, quietly building the smartest AI the world has ever known and the whole tech stack underneath it.

All right, let’s go to the second story with Google: Genie 4 from Google DeepMind. Genie is a world model. We’ve talked about world models. We’re going to have Fei-Fei Li on the show very soon. You type a prompt, and it generates a photorealistic 3D world that you can walk through and interact with using a keyboard, like a video game.

Genie 4 has 3 big things. First, native audio: music, ambient sound, and characters who speak. Second, real-time physics and reasoning. When you set an event in motion, it simulates what happens next and how people react to it. Third, you can steer the world mid-simulation. I love this. You can type, “Make it rain,” or “Add a person,” and the scene adapts completely to you.

Demis Hassabis posted the following quote: “Simulations within the simulations make you wonder just how deep the rabbit hole goes.” I have a feeling we’re going to find out very soon. Besides making amazing video games, why does this matter? Robots and AI agents are going to train in simulated worlds before they touch the real one. It’s being rolled out to trusted testers, then to Google AI Ultra subscribers and developers.

I’m curious for Emad and Alex: does this strengthen your views on simulation theory or not? The stuff we’re doing is extraordinary.

Yeah, I maintain that there’s 1 set of equations for all the physics. It isn’t suggestions; it’s a law. The equations of generative AI are remarkably similar because they map physics, and you see these models learning from everything. In that, we’re trying to find almost the divine, these core elements, because the universe came from something. It was generated from something.

The wonderful thing about this is that it learns physics. It feeds back to the real world and to real systems that we can use to explore the universe and more. It will accelerate our understanding of all these different phenomena when we can finally model them properly through these massive training runs, which I think is fascinating and fantastic.

Alex—

Alex Wissner-Gross

Yeah, I think maybe the most interesting item buried in the Genie 4 announcement was Google’s planned future work. Google has explicitly said that it wants to converge Gemini and Genie. Presumably—although, to my knowledge, I haven’t seen any publications about the architectures behind Genie 4—it either looks diffusion-transformer-like or something like a transformer, maybe with a few differences.

We’ve talked about Gaussian splats on the pod. We’ve talked about Unreal-style surface engines. Just based on the videos Genie 4 seems to be interactively generating, my guess is that it’s diffusion-transformer-like at the back end, which could be architecturally convergible with the foundation models like Gemini.

I do think one of the historically weak spots on the Anthropic model side has been visual perception. I think Opus 5.5 goes a long way toward repairing what I’ve perceived as deficits in Anthropic’s models when it comes to thinking and visual reasoning. The bull case for Google converging Gemini and Genie would be models for robotic manipulation. I think we’ve probably saturated visual design for computer software development, so I’m not super excited about that. It’s probably something economically valuable that requires real-world orchestration. It could be robotics, which is what everyone else is investing in, or something else in the physical world—maybe operating laboratories or visually reasoning about physical modalities, chemicals, for example, where it’s otherwise difficult to arrive at novel, economically transformative solutions based purely on text and images.

I also want to highlight that Google seems to be the only remaining American frontier or near-frontier lab investing in video generation. I’d view these so-called world models, interactive video models, as a natural generalization of video generation. OpenAI had Sora and dropped it. We’ve seen other smaller American players develop their own video-generation models, but Google is the largest credible remaining American resource for video generation. Otherwise, the whole game is being thrown to China. I want to applaud Google for continuing to invest in video generation, because otherwise it’s just a Chinese video world.

Let’s take a look at their video announcing this, and then we’ll go to Dave and Salim.

Genie 4 is our most realistic and interactive world model yet. Build environments and simulate worlds on command. Keep prompting, and Genie 4 will work out what happens next. Then explore and reshape them in real time.

So here, the prompt is, “The bird is pressing a key on the keyboard,” turning a picture into a dynamic environment that’s ready to explore, complete with ambient sound.

I mean, this is going to get used for video games. It’s not quite there yet, but—

Yeah.

And, Alex?

Alex Wissner-Gross

Alex.

Control the camera. Direct every detail.

It’s pretty beautiful.

Change live.

Do you really think we’re going to make it?

Or describe an interactive experience.

If they’re smart, they’ll use it.

If they’re smart, they’ll use this to replace YouTube. YouTube is a sizable revenue source.

Yeah.

Way beyond entertainment. It’s definitely, in my mind, the future of all work. I have a bunch of videos that I took to the West Coast with me, and they’re about photonic computing. I had one on chain theory, and I’ve got Albert Einstein, Isaac Newton, and Marie Curie explaining everything. I’m showing them to Nobel Prize winners and business owners, and they’d so much rather watch the video than hear it in PowerPoint form.

Yeah. Well, yes, for anybody. It’s just so much easier to learn from an AI-created video. The only flaw so far is that it’s not real time, so you can’t ask it questions in real time.

But now with Genie, as of this week—

Everything flips over. The future of all management is going to be through this interface. It’s so much more effective, and it’s really fun compared to writing emails and documents. It’s really engaging for managers, but think about how much of the world is factory assembly lines, construction projects, and so on. You can say, “Guys, we should be putting that beam over there, and this elevator shaft should move,” and as you’re saying it, it’s rendering and showing the construction team what you had in mind. That is irreplaceable. Once you’ve experienced it, you immediately realize this is the future of all management. It’s not just entertainment and video games at all. It’s all management of all ideas and production of everything, so it’s much, much more significant.

I had the same experience. Yesterday, I was at a factory in Toronto where they were using AI to automate. When I saw this, I thought you’re going to reach a point where you can simulate an entire factory and its operations, and troubleshoot problems before even building the thing.

Yeah.

And I think, Dave, you’re so dead on. This is such a huge shift. It’s going to take people time to get their heads around it; that’s the problem.

Yeah. Yeah.

To give you an idea, I estimate the cost of running Genie is $20 an hour.

And that’ll be down to $2 next year.

And you’ll be able to have every pixel generated on demand. Now it’s just customized: you can have your AI read your emotions and generate a movie or a video game that plays exactly to your emotions as you’re experiencing it. It’s crazy.

Yeah. Well, I think if you walk around and talk to people in business and ask, “Have you ever used a CAD/CAM system?” 99.9% have no experience with CAD/CAM. But if you talk to somebody in mechanical design who uses CAD/CAM, the idea of working without it is insane. It’s so ingrained in the way they imagine and visualize, and then try to create. I think that CAD/CAM experience will now export to every area of human endeavor through this. People will never go back once they’ve experienced designing, building, creating, structuring, and planning—even if it’s just a party, like a wedding party.

Where’s everyone going to sit? You get instant visualization of the tables and people milling about. Even that use case—you never go back once you’ve experienced it. You could even have family members arguing and fighting as you plan out the wedding layout. That’ll be easy to simulate.

With all this stuff changing, I’m updating The Organizational Singularity book in real time. Here’s the question I’ve got: when this comes out, if you can simulate an entire business or physical operation before you build it, is all strategic planning now obsolete?

Mm-hmm. Because you can essentially envision it, think about it, run the simulation, and see what happens. This completely changes the game for strategy consultants, all C-suite planning, everything.

So what you’re saying is that the singularity is going to automate human labor and cognitive thinking?

Well, now it comes down to imagination. We have to start by imagining it. At some point, it’ll replicate imagination, but for now we have to imagine it—or “imagineer” it, if you want to use Disney’s word.

For now.

For now. But I think it’s a really important point: nobody knows whether “for now” means a couple of years, a couple of months, maybe 10 years. It’s somewhere in that range.

Well, it’s a couple of weeks at this pace. I mean, for God’s sake.

Well, maybe not. But people get intimidated, and then they don’t jump in. They think, “No matter what I’m adding to the party, it’ll be irrelevant soon.” I don’t believe that. I think right now AI isn’t good at all at imagining outcomes, but it’s really good at creating the outcomes you imagined. It’s the golden moment for humanity, for creators, right now. To waste even a minute of it is utterly insane. A lot of people are wasting it out of fear that AI will make whatever they’re thinking of irrelevant very quickly. I just don’t believe that’s likely.

4. AI, the Future of Work, and Human Purpose

Yeah. Can I segue to this video I put out, Peter?

Yeah, of course.

I put out a video yesterday basically asking a simple question, and it’s getting crazy comments. It’s worth revisiting because I think it’s the thesis of this entire podcast. Everybody’s bemoaning the destruction of jobs and going, “Oh my God, what will we do when all the jobs disappear?” If you flip that around and ask, “Wouldn’t it be the greatest thing ever if all the jobs disappeared?” you can start to imagine how you get there.

You have to add to that: if you didn’t have to work—

If you don’t have to work—

—and you still maintain your quality of life.

Yeah. We have a precedent for this called retirement, where you used to spend 50 years working to get that 20-year period when you could do whatever you want: work on your hobby, build a kit car, spend time with your grandkids, do whatever. But you were too old to really do much. If we get to the right time and figure out how to do the economic distribution, you can spend time on creative passions, spend time with family, and work on building communities. Let’s start envisioning that future, because it’s actually going to become possible.

There was a great comment from someone in response to my asking what might happen if people do this and what it would look like. Somebody named VK4058 gave the greatest comment. He said, “What’s actually going to happen is people will turn into easy-dopamine-addicted bonobos with IQs gradually approaching those apes.” I thought that was awesome. But you could go down a WALL-E scenario, or you could have the greatest life of presence ever. You see people in retirement doing one of those two things: they either fade away or get super active. One of my uncles is doing a PhD in astronomy because he’s retired and likes doing that. I think it’s going to be incredible to see. I put that provocation out, and it’s fascinating to see the results.

5. Welcome to Moonshots

It’s so important to point out the importance of purpose and having goals and objectives when you don’t have to work. I love the fact that William Shatner, now 95 years old, is appearing at rock concerts, flying to space, and coming to do the closing conversation at the Abundance Summit. You need to have purpose. The biggest concern isn’t that you don’t have to work; it’s whether work has always been your purpose in life. If it’s taken away from you, are you sort of lost? Emad, please.

I figured out something. Listen to this: the biggest difference between humans and AI is conviction—conviction to do stuff, to push it forward. What does Elon have? Massive amounts of conviction. He made the world go around him. When you were talking about strategic planning, that’s like expected utility: the world is static. Now we have the biggest levers ever, if we have the conviction to push things forward.

You see this when you’re talking to your AIs. Claude says, “You’re 50% right and 50% wrong,” and the 50% that’s right is the interesting bit. Or when you look at the code of Astra and it’s finding edge cases that will never happen. The AIs lack conviction. They can deal with the current environment, but it’s humans who can push things through, with individuals, people around them, and agents around them, to achieve massive change. I think we should reframe it like that.

Why do you think AIs couldn’t simulate that same conviction? Or, you could say to an AI, “I want you to take on Elon Musk-level conviction, and let’s solve these problems.” Why couldn’t you do that?

I think part of it is the way the AI’s latent-space distribution is set up: they tend to do exploration. Even in the tuning and training, they do Monte Carlo search. So it’s very rare to see an AI that actually has conviction and pushes it forward, whereas you see lots of humans achieving great things. What do they all have in common? They have conviction and determination. More so, I think it’s something about the architecture. You could build an AI with huge conviction. There are experiments now where you turn on little neurons and then they collapse their entire wave function.

But that’s not how AIs are right now. Once you realize that you can do everything, you can tear the threads of the universe. That’s the really interesting thing here, because I have a company.

Sorry. Go ahead.

Yeah, I was just saying, the framework for this is what Jeff Bezos said: he has a model of the inevitable and the unchanging. There would always be books, and people will always want faster, cheaper, and better.

What business strategy doesn’t work with that, right? The way you get there is one thing, but he had that conviction, and the whole of Amazon went that way. Strategic planning doesn’t necessarily fit with that, which is why, if you know where the future’s going and you pull it forward, you can achieve bigger things than strategic planning would indicate.

Sorry, Alex. I think I’m much more optimistic that conviction, or long-term strategic planning, is very much within sight of models these days—even just a system prompt away. I think the distinction between business leaders, Elon et al., and the models is that Elon, at this point, has quite a bit of capital, and he’s also very good at marketing. When Elon decides, “No, actually, we’re not going to Mars next; we’re going to the moon next”—it was the moon all along—and retcons the moon into SpaceX’s plans, that’s branded as business genius.

6. Alphabet's Worst Stock Day Ever and the AI Market Crash

Whereas if a model did it, one might say, “Well, the model lacked conviction about getting to Mars first versus the moon first.” The models aren’t yet in a position where they’re allowed to market themselves as having these long-term ideas. They lack the capital; certainly, they’re unbanked. I would suspect that if we had a model with a long-term system prompt, like, “Carry out the following really challenging objective over a period of years,” and we provided it with the capital, real-world engagement, and empowerment to do it, it would probably be able to market itself just as competently.

Well, guess what, guys? We’re going to find out soon enough.

That’s the beautiful thing. Our third Google story is the paradox. Google shipped the number one AI model, arguably, though it hasn’t yet been publicly released. On Thursday, Alphabet had the worst day in its history from a stock perspective. Its shares dropped 13%, wiping out $500 billion in market value.

The strange thing is, revenues grew 25% to $127 billion. Google Cloud more than doubled, up 100%, and the Gemini app hit 1 billion users. So what sparked the market? Two things. First, search ad revenue, while it grew 12%, was down 17% from last quarter and 19% from earlier in the year. People are very concerned that their main revenue driver may be on a waning trajectory.

Second, cash: Alphabet burned $9 billion of free cash flow in the quarter, far worse than expected, while keeping capex at $195 billion to $205 billion for the year. A few other stock movers are worth noting. AMD and Intel were both down. Intel dropped 28% today, Dave. Kind of crazy.

Like 30-plus.

Intel’s results beat estimates. Its outlook for the quarter came in below expectations. The CFO said PC sales were down, driven by expensive memory chips. I’m curious about your thoughts, Dave. We’ve been talking about the potential AI bubble, the leak in the AI balloon. It’s continuing. What are your thoughts?

What’s going on here is really obvious. The Mark Zuckerbergs, Sundar Pichais, and Lip-Bu Tans are going to invest more money than humanity has ever invested before in AI. They’re 100% believers that whoever controls the chip stack or the AI algorithms controls the whole future of the world.

The idea that some of the most profitable companies we’ve ever seen, like Meta—an absolute cash machine—are now borrowing tens of billions of dollars because even that cash flow is not enough: “We must do AI. We’re not slowing down.” The market is now like, “Oh my God, I’m getting scared. I’m really jittery about this.” They’re not going to stop, though, so it’ll probably get worse before it gets better. They’re not slowing down, and they’re 100% right. This is the only moment in all of human history.

If there just isn’t that much money in the world, then things can’t rebound. The money just physically doesn’t exist. They’re starting to draw from every nook and cranny and corner of the debt markets all over the world into this massive buildout. A week ago, the market was like, “Yeah, this is all going to work out.” Now everybody on Wall Street is like, “Oh my God, that’s a huge amount of debt.”

And the shortage of RAM isn’t going away. We did not build anywhere near enough chip fab capacity to keep up with the need for massive amounts of HBM memory. Now people can’t upgrade their iPhones. They can’t buy a new iPhone. Why? Because it’s too expensive. Why is it too expensive? Because RAM prices are way up. Apple has its first earnings shortfall in as long as I can remember. Why? Because it has to pay so much more for RAM.

So AI is going to suck up every resource in the world.

And it will absolutely all pay off. There’s no doubt about it. But the question is, what about between now and then? What’s going to happen? Should we all stop eating and just pump that money back into AI generation?

There’s a rational argument that we should all subsist on rice for as long as we can, but put that money back into chips because it’ll pay off a thousandfold.

Dave, we’ve talked about how some employees purportedly at frontier labs have taken up smoking because they think their cancer will be auto-cured in a few years by their AIs anyway. It’s even worse than that.

That’s a statement. Yeah.

My liver enzymes are up 50% because I’m on retatrutide, so I’m able to drink more wine. I think this is great.

I want to make a comment here. I think we have something really amazing going on that’s worth talking about: the unbelievable gumption and cojones of these CEOs to make these massive billion-dollar bets. We talked about conviction just now. Whether it’s Sundar, Satya Nadella, or Zuckerberg, they’re committing unbelievable amounts of capital to a future that hasn’t arrived. They’re absolutely convinced of this. The conviction they have is like, “Screw it. If we think intelligence is going to become the foundational infrastructure of civilization, we’re going to build for that future.” That takes amazing conviction.

Half the people are saying it’s a bubble and they’re spending too much. But look at that kind of dedication. That’s why these types of people win: they have that conviction, and they’re willing—

I’ll tell you, the people closest to the truth—the Elons and the Zuckerbergs, the guys inside the middle of it—have the most conviction, and that tells you what’s really going to happen. It’s the people on the edges, investing through a series of funds and not quite sure, who are panicking.

I would just ask: what choice do they have? Even Zuck is purchasing GPUs left and right and has the reassurance that Meta Compute can always absorb any extra capacity he can’t monetize. But if you look at Google, what choice do they have? Their other bets—and all due deference to X and Astro and the other bets—include Waymo. That’s been a success story, but other than Waymo, in terms of revenue for the moment, their other bets have not had a major success.

On the other hand, if you look at GCP, Google has often been characterized historically as a one-trick pony. Its best next trick, based on the financials, has to be serving as a hyperscaler. Even if it means cannibalizing search revenue in order to become the best possible hyperscaler, I’m honestly not sure Google had much of a choice but to cannibalize itself if it wanted to grow.

They’re playing the long game, guys. They’re playing the long game. I remember there was a quote from Sam Altman—I have a video; I should have found it—where he says, “We’re building an AI god. I don’t care how much it costs.” I think that’s their attitude. They’re going all in on all the chips right now.

They have to. Sergey and Larry are still controlling shareholders of Google, if you can believe that. Nobody really pays attention to that, but they still control it. That’s where that fortitude comes from. Zuckerberg—it comes from him. Elon—it comes from him. But to answer your question, Alex, the other choice would be to be where Apple and Microsoft are—

Right?

Yeah, which we’re just kind of touching around the edges, not really doing.

Remember when Microsoft said, “We’re good for our $80 billion,” and that was like the beginning of the end for Microsoft reaching for all the possible capital it could for a buildout? Microsoft bowed out of the hyperscaler race in some sense. So I do agree with you. There’s an element of founder mode, reaching for the capex even if it means cannibalizing the legacy businesses here.

Yeah, I think the market is missing something huge here. We’ve talked about GPUs and DRAM, and necessity being the mother of invention.

If we have 10 agents per American, there are not enough CPUs in the world.

Yeah. No doubt: CPU.

Not even. Exactly. Great point. Great point.

And who was number one in CPUs? Intel?

You know, you have Intel, ARM, and AMD. Those are the players for desktop. I mean, in terms of overall CPUs, it’s—well, for mobile, obviously ARM, and then for actual CPUs, putting aside chip architecture, it’s probably something exotic like a microcontroller architecture or something. So, not necessarily Intel.

Well, CPUs for agents: you’ve got, obviously, your Gravitons from the hyperscalers and others, but I think we’re always looking for where the next bottleneck is. Right now, it’s clear that we’re going to go from less than 1 to more than 1 agent per person. We’ve discussed previously how inference costs are going to collapse, and maybe the GPU buildout is already there. We have the capacity coming on of millions of Vera Rubins and MI450s, et cetera. The CPU buildout, with 2 to 3 years of lag time, is only just starting. They haven’t really invested there. When you do the math, it’s clear that this is going to be the next bottleneck, because agents need CPUs.

I’d love to push on that just a little bit, if I may. What I hear you saying, Emad, is that you’re predicting monotonic growth in the number of agents per person. The implication is that it won’t be GPU-bound; it’ll be CPU-bound. There’s an alternative way this could play out: as the baseline foundation models get stronger, this explosion we might naively see in the number of agents per person might actually reverse.

When OpenAI released their Navier-Stokes results, they went out of their way to say it was a 10,000-agent result over a short period of time. But then, by the time of Bell and the slop drop, it was 1 agent for 3 hours. An alternative perspective might be that, as the models themselves get stronger, the need for swarms or large numbers of agents diminishes, which would impair the need for CPUs altogether. What do you think?

Again, I’m not talking about 100 agents per person. It’s more like going from 0.1 toward 1. If you look 4 or 5 years out, I think you have at least 1, and they need houses, and there aren’t enough houses built. We’ve already seen the lead time for RAM and what happened with RAM. The lead time for CPUs is a couple of years, too. I think the market’s going to look at this, probably not now. Intel is down 30% today. That’s one of the biggest moves in the stock market ever.

I think it’s going to be interesting as we look for the next narrative, the next choke point. As you said, a single model can do a surprising amount. Sometimes you do see these massive increases, like Navier-Stokes, but single models also outperform a dozen models in many cases.

7. Sam Altman's "Hold My Beer" Moment

All right, I’m going to move us forward. Our next 2 stories come from a conversation Sam Altman had with John Collison, the co-founder of Stripe, on John’s podcast, Cheeky Pint. I think we should all be drinking beers during this podcast, guys.

We drink water here. Unlike the Collisons, who are drinking alcohol, we drink water.

Who’s “we,” white man?

No comment.

About halfway through drinking his beer, Sam said, “I’m actually probably not supposed to say this, but since we’re drinking, I will.” Let’s watch the clip.

How much does the math cost?

On the order of $10 million for, like, Navier-Stokes, I think.

Okay.

Maybe a little more.

$10 million-ish, that order? Yeah, yeah. But for one proof, and so—

No, no, because then we kept getting better so fast. We understood more.

Got it.

And we spent way less than that on the average one. But hundreds of millions of dollars of math.

Yeah, yeah, yeah. Okay, interesting.

But that includes much more other discovery than what we have published.

So it’s like maybe—

More than the 700.

When did we get those?

There was one weekend when we got a new model, and I’m actually probably not supposed to say this, but I will since we’re drinking.

Exactly. We’re halfway. Yeah, perfect time for—

The name of the channel was “model name — hold my beer.” People were throwing the hardest problems they could at it. They didn’t think the model was going to do these things. We had this collective experience of people’s hardest research problems from their academic careers, and the model was just knocking them down one after the other, across many different fields. It was incredible to watch. Two nights in a row, I was up most of the night because I was so excited I couldn’t go to bed.

It was like many of those problems that we published, and many other things. I was tremendously excited and happy. But I also had much more sadness and melancholy, and a feeling of something ending, than I was expecting. I agree this is the right thing on behalf of the world. I also agree that this is just the shape of reality, and nothing anyone can stop or should stop.

Can you imagine, 10 years ago, saying that a private company that’s not in the mathematics business would spend $100 million to solve problems? It’s crazy. Alex, what do you think is sitting in the unpublished pile of math proofs?

I can only imagine. Certainly, last time we talked on the pod about Navier-Stokes, I predicted—and I think Emad predicted as well—Yang-Mills. So we already know a bit more about their Yang-Mills strategy. Probably a bunch more Millennium problems.

I’ll give a hot take on this one: I wouldn’t be surprised if there are further advances beyond ETH that collapse the complexity hierarchy. That would be pretty scandalous. I wouldn’t be surprised if there are lots of cryptographic or counter-cryptographic advances they’re sitting on, not publishing, going back to the original Q* allegations.

I also want to touch on Sam saying “hold my beer.” Remember, in the context of GPT—I think it was GPT-5—with the Star Wars space station, the Death Star. Sam was using a lot of Death Star motifs in his X post to describe it, including the Obi-Wan Kenobi line: “I felt a great disturbance in the Force, as if millions of voices suddenly cried out in terror and were suddenly silenced.”

Here Sam and OpenAI are sitting around joking on a Slack channel, “model name — hold my beer,” about basically crushing entire disciplines. I think history will regard this as a striking asymmetry between the casualness on the interior and the superintelligent superpowers on the exterior, with entire fields rolling over and playing dead because of it. Salim, I’m curious: he said he felt sadness, right? An emotional response to this. What do you make of that?

I think there’s sadness for the art form that gets lost when you get such magical outcomes from things. The craft gets lost, and I think that’s what he’s referencing there. But the fact that we can get so many more outcomes should make you way more than happy for that sadness to balance out.

I think it’s sadness for the old way of doing things, sort of nostalgia.

But this is the same. You talk to any systems programmers, and they go, “Oh, what happened to all the great assembly-language coding?” That was great for that period of time and for those types of systems, but now we have so much more power at other levels. I think it comes down to that.

Yeah. Dave, what he’s talking about is wholesale reinvention of entire fields for millions of dollars, and we haven’t seen their best models. I’m curious if you think OpenAI’s value in the future is going to be everything they build on top of these models, everything they discover, rather than selling compute or selling tokens.

Yeah, it’s a golden age. Coming back to the emotional reaction, if you looked into Alex’s eyes on the last podcast when we were talking about chain theory, I really felt the same vibe that an Olympic gold medalist has when they’ve been working for a decade toward this outcome and they suddenly have the gold medal in their hands. They’re over the moon excited. But then, after that wears off, you’re like, okay, well, now I have to take stock of my entire life.

Where are we? The people working on these foundation models at OpenAI have been at it for over a decade, and they’re some of the hardest-working people you’ve ever seen in your life. You remember they invested in that company that lets you sleep for 4 hours and feel like you slept for 8.

Yeah. It’s one of my tries to recapture—just trying to get 4 more hours a day.

But that’s been their mode of operation for years, working toward this outcome. As soon as you start cracking the world’s hardest math problems, you know you’re over the threshold of success, which is always a wistful moment in time. It’s like, “Wow, now we’re in a whole new chapter for humanity, where we have to actually—yeah, like you said, Peter, now we have to leverage this to make the world a better place.” That starts a whole new chapter, but the immediate, urgent target has been achieved, and that’s what creates that wistful feeling. So, yeah, it’s a new world. The world after 2026 will be nothing like the world before 2026, and here we are, right on the precipice.

8. Google Genie 4 and the Future of Simulated Worlds

Yeah, I think part of it is: are we going to run out of the big problems that we wanted to solve, that we can now solve, and then continue to push it? This is, again, about human purpose and human conviction. I’m going to go to our second clip. In the same conversation, Sam Altman revealed that OpenAI’s robotics program is aimed at building data centers. He said, “People talk a lot about RSI as an AI that does AI research. There’s very little—relatively little—talk about self-replicating data centers.” All right, let’s watch this clip. It’s awesome.

I think the trend is toward—we always just want more compute. The demand for sufficiently capable models at a sufficiently low price only goes one way, I think, and we’re doing a lot of things to prepare for this. We made a chip with incredible performance in units of intelligence per watt. We’re learning how to build our own data centers. We have a robotics program that we’re going to point at making lots of data centers. People talk a lot about—

The real RSI.

9. OpenAI's Plan for Self-Replicating Data Centers

Yeah. People talk a lot about RSI as an AI that does AI research. There’s very little—relatively little—talk about self-replicating data centers.

Mm-hmm.

You know, this idea that you can use AI to control robots and run fabs, build turbines or solar panels, and build data centers. I think we have that within grasp. We’ll try to make a lot of really cheap computing data centers. How far out do you think?

Well, you have dark factories and robots already, so why not within the next year or so?

You know, I think, again, you’re reaching these points. Listening to Sam just now, I was like, “That’s Proust’s In Search of Lost Time, right? These sorrows will be greater even as our desires are fulfilled.” This is not a new thing. We’ve reached heights as humanity, right? And the question is, what is our desire? I think it’s about making the world intelligent and making it available to everyone. But then the space of solving problems has gone from things that we should have solved ages ago, if our minds had been put to it—we should solve health, we should solve these things—to solving ourselves, understanding ourselves, and more. This replication of the thing—you shouldn’t have to build data centers. This goes to Sam’s video: why would you need to be a data center tech? Why do you need to be an electrician? You shouldn’t have to, because if we can build this correctly, we can be freed to think and desire the bigger things in life and solve even more interesting problems, including those that are more personal to us.

I had 2 thoughts when I heard this. One was that this is very reminiscent of, and duplicative of, Alex’s inner loop.

Exactly. Where did he get this? I have no idea where Sam gets this. It’s the innermost loop.

It’s the innermost loop. And there’s a precedent for this already, like we’ve talked about on the pod. In China, they have robots putting up solar panels to gather the energy to feed the factory that builds more robots, and you have that kind of loop going. If you feed that into a data center, the video also has John pushing back and saying, “Look, the bottleneck is regulatory and permitting, and all these human things. That’s where the real issue is, and that is going to have to be solved in a different way.”

Oh, government will be taken over by AI systems. So, this innermost loop, Alex: robots building data centers, data centers training the models that run the robots, smarter robots building better robots—

With lots of energy coming in as well.

This is the innermost loop. This is the physical-world embodiment of recursive self-improvement.

I think there’s one piece that Sam is not saying sotto voce, that he’s burying: are these self-replicating data centers going to be terrestrial? Maybe not. Arguably, they want to be orbital, or even better, lunar, because they’re free of regulatory constraints there. Of course, Sam presumably doesn’t want to say anything especially nice about orbital data centers, because that’s immediately handing the ball to Elon and SpaceX and their virtual monopoly, in the short term, on heavy launch. But if we’re serious, if we’re thinking critically about a future with self-replicating data centers in a world where, at least at the moment, we have all these municipalities that are hostile to terrestrial data centers, I can only imagine how hostile they’ll be to terrestrial data center construction that doesn’t even create local employment through HVAC or construction workers. They’re going to hate it. So where are these self-replicating data centers likeliest to materialize? I think probably on the moon.

2 predictions here, Alex—

That’s the elephant in the room.

Number 1, OpenAI is going to buy one of the launch companies. They’ll either buy Rocket Lab or Relativity Space. They can’t afford not to have their own launch capacity. And the second thing, I hate to say this, is that regulation’s going to go to the moon. We’ve seen regulation go to Antarctica. We’ve seen it go every place. There’s going to be regulation there as well eventually. The question is—

Who’s going to regulate? I’ll make a hot take: I think it’s going to be easier and faster to go under the ocean than to get to the moon.

Maybe. I mean, the ocean—the subsea area—is materially more regulated right now than the lunar surface. But, Peter, we have the Outer Space Treaty and the Artemis Accords. What sort of regulatory environment are you envisioning on the moon?

I imagine that whoever goes and becomes the dominant sovereign is going to drive regulation. These are U.S. corporations; U.S. regulatory capture doesn’t leave them when they leave a particular country. When you launch your satellites, they’re still regulated by the country you’re launching from.

So you think it’s going to—I agree with the premise that it’ll probably be like Delaware and Texas law regulating the lunar surface, but that’s regulating the actions of American companies with a lunar presence. What happens when some other country—many of these lunar companies don’t subject themselves to American law? They’ll pick some other country, usually one with a much more favorable outer-space policy and regulatory regime, like Luxembourg. What happens then?

Believe me, if there’s a chance to drive regulations, countries will make the effort. I think that’s human nature. Now, the question is, are the future regulators going to be humans or AIs?

Well, if they’re self-replicating data centers, maybe they’re self-regulating data centers as well.

All governments and systems will be run by AI. There’s no way it can’t be.

Let me help this discussion by saying the following: we should just get the UN involved.

Oh, no. I’m just kidding.

Dave, close us out on this one. What are your thoughts here? This is the biggest opportunity we’re going to see: RSI in the physical world as well.

Yeah, I think people are way overestimating how much space it takes for a data center to exist. In a world where your geography embraces data centers, even if you said, “Hey, we’ve automated away all the plumbing, HVAC, and electrical jobs”—and that won’t be the case for 3 or 4 years at the earliest—even in that world, it’s generating so much revenue for that little chunk of geography that the people immediately surrounding it are going to be some of the wealthiest and happiest on the planet. I think the chip supply is so constrained, and energy in the U.S. is so constrained, that if you wrangle those 2 things, you’re going to be a very, very happy person, regardless of the fact that, in the long run, it’ll end up on the moon. So, between here and there, people who embrace data centers—even fully automated data center construction—is only going to make that geography the wealthiest in the world. It’s crazy to me that people don’t see that and that they fight against it.

Jensen Huang just visited and spent time in Kazakhstan, which has fully embraced this data center buildout. They’re like, “Let’s just go.” So, kudos to them.

Well, also, Intel had $300 million of its miss—today’s or yesterday’s earnings call was related to their fab in Israel getting bombed.

And so people are reluctant to build out data centers in the Middle East, where there’s a lot going on. If you’re in Canada, sitting on a hydroelectric or nuclear power supply, you were already in great shape. Now people are reluctant to expand into Kazakhstan or elsewhere because of that risk, so you’re doubly in great shape if you jump on the opportunity.

I think the self-replicating data center is really about automated construction of all kinds of things. But the money is pouring into data centers, so anyone related to the trades, construction, or finance should talk to Sam, Dario, and Elon—to the people building out the biggest building in the world, the new Terafab. That’s where you want to put all your energy. Work on that first, because that’s where all the money in the world is going to go in the short term. Then it’ll expand out to the Moon, the construction of hotels, and other things. But for the next couple of years, that’s the total focus.

Highest revenue per cubic meter, right?

10. Tesla Optimus and the Humanoid Robot Race

Deep in the heart of Texas, I guess. And maybe this is just a movie reference: I think probably everyone here saw the Johnny Depp movie Transcendence, which, to my knowledge, is the only recent movie that depicts a self-replicating data center. They call it the Brightwood Data Center in the movie, but it’s inhabited by the spirit—or the upload—of Johnny Depp’s character. You watch it rebuild itself and its energy sources in real time with nanotech. It’s a pretty cool sci-fi depiction of what a self-replicating data center might look like.

I have to watch that one again.

Let’s continue with physical AI and talk about robots and autonomous vehicles. First up this week, Tesla announced that its first production-quality Optimus robots are coming off the high-volume line at Fremont, where production of the Model S and Model X—two vehicles I love—has been replaced. The line is designed to build 1 million robots per year, or 20,000 a week. Elon says it’ll take at least a year to get production up to those levels, and the first customer deliveries are expected in 2027. Optimus is roughly 1 to 2 years behind Elon’s original schedule. It’s classic Elon: he gets it right, just typically a couple of years late.

Now, the global picture: about 25,000 humanoid robots were shipped worldwide in the first half of this year, up roughly 400% from last year. Chinese companies built more than 95% of them, led by AgiBot and Unitree. Total 2026 shipments are projected to double, exceeding 50,000 robots. If Tesla hits 1 million robots a year, that’s 20 times the entire world supply. So the question is: can we onshore all the production? Dave, your thoughts on that?

Can we onshore all the production? With AI-based automation, everything’s onshoring at an incredible rate. There’s a really interesting video on TSMC educating people in Texas on how to build chips now, and on us needing to re-import all this knowledge that dissipated starting in the 1950s. The whole globalization push sent manufacturing offshore, and now it’s all coming roaring back because it’s not labor-constrained anymore.

Even before the Optimus robots, you have regular factory-arm robots. Because there’s no labor constraint anymore, all of this is going to come pouring back into the U.S., which is not great for the rest of the world. It’s certainly great in the long run for America and for the few places on the planet—and China as well—that are doing it. So, yes, it can all come back to America, and it inevitably will. The intelligence that makes it all work has all kinds of export controls on it, so it’s unlikely that it can leak away again, as long as that intelligence is manufactured purely inside the U.S. and China. It’s going to accelerate like wild from here.

I can’t wait for our podcast with Brett at Figure AI. It’s going to be a fun conversation. I wonder whether Elon is his competitor of choice, or whether he sees himself competing against China. Alex—

11. The Fight Against Driverless Taxis

The biggest mystery here, in my mind, is less on the supply side. I’m pretty confident Elon can eventually meet his supply-side goals and catch up to, and probably surpass, Chinese production. It’s the demand side: what is going to be the demand for these? Peter, I think you and I are both still waiting for our 1X NEO domestic humanoid robots. I’d love to have one or more Optimus robots as well, but this is the part of the equation I don’t see the answer to yet.

He’s indicated in past comments that the first application is going to be in Tesla factories for manufacturing. That’s fine, but based on his curves, that’s going to saturate pretty quickly. Just like Starship suddenly creating all this new upmass capacity, he almost has to create a new market to satisfy his own supply on the demand side. I’m just not sure. I can think of a range of speculative possibilities. Maybe he’ll decide to go domestic early, or aim it at construction, or at lots of different industries. We’re all just waiting for humanoid robots that are too cheap to meter. But the fact that he’s been cagey beyond just saying he’s going to put them in Tesla’s—or the merged SpaceX-Tesla—factories is the most mysterious part of all this. I do wonder—

Let’s think about it a little bit, because he’s going to get Model S- and Model Y-equivalent manufacturing back to all-time highs once the robots are working on it. But he also wants to make a Starship every single day, a new one every day. I haven’t actually worked through how many robots that’ll require, but I imagine it’s a lot. Well, we can—

When I’m in front of audiences, I say, “Listen, these robots are going to cost you $300 a month, $10 a day, under a dollar an hour. How many of them would you want?” I take a poll, and everybody wants at least two. $10 a day—that’s insane. I think the consumer market is massive, if they’re proven safe.

Again, I think we can do the math here, even if Elon just uses them internally. The market has room right now for 100,000 to 500,000 robots that he produces next year. A hundred thousand of them is $3 billion worth of robots, and you compare that against the capex of all those things. I think he can absorb even that.

But this takeoff is going to be absolutely unprecedented. If he gets to 1 million next year, at 1.6 million the revenue from Optimus is greater than the revenue from Tesla. That’s kind of what we’re talking about here. To be honest, if I wasn’t doing all the stuff I’m doing now, I’d set up a $5 billion fund and buy 100,000 or 200,000 of the robots, and give them the money right now. The other interesting thing—and we saw this with the Unitree robots—is that next year, I think, is where you see a freeze in hardware. The hardware will be the same Optimus for a few years, but the software will continuously improve into 2028 and beyond. If you buy all these robots up front now, their value will actually increase, just like we’ve seen with GPUs. This is the next big asset class. But if 100,000 of them—that’s basically all the production—is $3 billion, compared to the computing buildout, it’s nothing. That’s why the numbers are very strange right now.

I’ll try saying that slightly differently. SpaceX has approximately 20,000 human employees. Tesla has approximately 130,000-plus human employees. At the production numbers Elon is talking about for Optimus, 1 to 2 years out, this is going to spill out well beyond their internal needs or production demand. Even if they go full-on recursive self-improvement, these robots are going to be pouring out onto the streets or somewhere. I don’t think Tesla and SpaceX combined have the internal capacity to absorb all this output if he sticks with his numbers. That’s what I’m gesturing at.

Agreed. But you don’t think home use is going to be dominant? I’m curious: when you’re on the same stage as I am, do you ask the audience, “How many would you buy?”

No. My thesis is that we’ll spend 5 years using them in commercial back-office applications: stock picking, inventory management. There’s so much demand to solve that first, before it gets to, “Oh, I need something to do my gardening and pick up my leaves.”

I just—

That’s my view. I may be wrong. I did find one use in my talks: a robot chef at home that always cooks you healthy food. Giving a robot a knife may not be a good idea, but for $10 a day—

Having a great chef—that’s a huge market.

Absolutely. One that knows your physiological state in the moment, your protein levels, your vitamin D levels, your hydration. All of your physiological data gets ported by your AI to the chef to cook you the perfect meal.

You need a dead version.

Would you pay 10 bucks a day for that? That’s all great. That’s all great. The demand from industry will overwhelm the supply for quite a while, before you’re able to get your home chef, which I love the idea of.

Yeah, I agree with Salim on this one. Consumers, at least in the West, are relatively poor compared to enterprises. If consumers couldn’t even successfully absorb or metabolize OpenAI’s reasoning models, and the best the American economy could think of is Meta’s Muse Spark for consumers to absorb reasoning models—which probably isn’t that compute-intensive, ultimately—I just don’t see how consumers will be able to disagree. If you’re earning 20 bucks an hour at McDonald’s and your robot can do your cooking and home cleaning for 2 bucks an hour, it makes a lot of sense to have a robot and go earn revenue elsewhere.

Eventually, yes. But remember, in the meantime, just as tokens flow to the use case with the maximum revenue per token, I expect robots to flow to the use case with the maximum revenue per robot action. And that’s not going to be a poor consumer. Where’s the demand coming from? You’ll get families running into McDonald’s together, grabbing the humanoid robot from behind the counter, and kidnapping it to take it home to cook.

Okay, so Peter, you’re now asking the question that I’m asking, which is the big elephant in the room, the big mystery here: What will be the use case for all of these hundreds of thousands of Optimus robots that maximizes revenue per robot? I think they’ll go into industry first, but I think there’s an unquenchable supply of needs among consumers as well.

Anyway, we’re going to find out.

I think, to a first order, if you look at Star Trek episodes, it always anticipates that we’ll have a communication device with supercomputer-level capability on our body. What do you know? It’s here. And we’re also going to fly faster than the speed of light. No, that didn’t happen. So where are we today in terms of robots? I think what’s actually going to happen is we’re going to spend enormous amounts of money building out data centers, and the robots are going to be working on that night and day. They’re also going to be building tons of rockets to get more and more Starlink up into the sky. The spillover into folding your laundry isn’t going to happen, because they’ll be so busy building out the data centers. But those data centers are going to give you your 3D virtual world, and every consumer is going to have that and be overjoyed with it long before they have a robot in their house, because the robots are busy building the data center that gives you your 3D virtual world.

Maybe that’s the point you’ve made, Emad: construction robots are going to be the massive use case, right?

I mean, you can just do the math: around-the-clock construction.

Yeah.

Yeah, I think the scenario Dave portrays is probably the likeliest one I’ve heard: Elon, inevitably, to compete with Sam and his self-replicating data centers, decides, “All right, send a message to all concerned shareholders: Tesla and SpaceX have to merge, because SpaceX needs Tesla’s Optimus robots to build out all of its terrestrial data centers going forward. So, merge them.”

Brilliant.

I think the final point is that this isn’t a demand-side problem. I think the demand is going to be there for everything we discussed.

Yeah, and I know, Peter, you’re dying for, like, six Optimus robots in the house, but I think you’ll be waiting.

12. Tesla Cybercab vs. Waymo

Okay. Well, I’ll give each of my Grok bots a robot body.

Or maybe, Peter, you have to open up your own data center if you want some. That’s why I think the smartest thing to do right now is launch a fund and literally buy the first 100,000.

You’ll make massive ROI on that.

And we talked about launching—if you’re a consumer out there, buy 100 robotaxis and put them to work for you.

Right?

Which is our next story. So, let’s jump into it. Elon posted about Tesla versus Waymo and said, quote, “It never was a real competition.” Let me share the chart here. This was put out by our friend of the pod, Cathie Wood at ARK Invest. Here are the numbers: In terms of dollars per mile, the Cybercab is coming in at 20 cents per mile—that’s the prediction, or at least what they’re seeing today. Waymo’s newest version is at 40 cents per mile, the Model Y is at 68 cents per mile, and the old version of Waymo is at $16 per mile. Why? There are a number of reasons. Tesla uses cameras only to build its car. There was a long debate about whether they should use lidar and radar. Waymo is using a multitude of sensors, including lidar and radar, and its new minivans are built in China, so they’re facing tariffs. The performance is interesting today: Waymo has 4,000 robotaxis on the road, 500,000 paid rides per week, and it’s in 15 cities, fully driverless. Waymo is winning the competition today, but I think all of you would agree that, ultimately, Tesla is going to win the race. They’ve just built a lower-cost, higher-performance vehicle. Dave, your thoughts on this?

13. AI's Impact on Computer Science and MBA Degrees

Well, I think these numbers are correct, but catching up in the age of AI is about the easiest thing in the world. I wouldn’t write off Waymo and Google anytime soon. I think the winner is the first person to build their chip factory. If Elon builds the Terafab successfully and quickly, then he wins, because he can make that many more Cybercabs. If he doesn’t win that race, and Google does something—the antitrust action around Google slows them down a lot on this front—but whoever buys either GlobalFoundries or Intel, or whatever, and gets control of their own chip manufacturing, is going to very quickly retool their auto manufacturing around their chip supply. It’s subtle, but everything else in the car has multiple vendors and relatively easy-to-fix supply chains. You’re not going to run out of tires, seats, or windshield wipers. What prevents you from taking over the world with driverless cars? It comes down purely to constrained semiconductor supply chains—not just chip manufacturers, but also substrates and advanced packaging, which are hugely constrained.

That’s why the Terafab is the linchpin to everything else. So, yeah, he’s way ahead, but there’s no reason to believe “way ahead” means anything beyond, like, a six-month window.

It’s worth saying there are at least 10 other autonomous vehicle manufacturers out there, from Zoox to others. The winner here is the consumer, right? These guys are going to beat their brains out to get the cheapest price per mile. I find it fascinating: the poorest people on the planet are going to be chauffeured everywhere they go.

10 times safer, too. Don’t forget that.

10 times safer and getting better. The human toll of driving is astronomical. It’s just a far better world—

The toll on time and productivity, even.

Time, yeah. Everything.

I’ve got two or three thoughts about this.

Please.

I think people are underestimating Waymo’s accumulated operational experience with real-world transportation. It’s a different kind of competitive advantage. The metric I want to see is not theoretical cost per mile, which is an easy one to compute. What’s the fully loaded cost of safe, reliable, completed mileage? You have to think about maintenance, remote assistance, cleaning, downtime, and all this other stuff. That’s the part I’d like to see. I think you’re going to end up with both. Jevons paradox says that as we make this available—like in LA, where it’s hard to get a Waymo car now because they’re so busy—the minute these cars become available, demand will just suck it up. You’ll have lots of time for both solutions to build up for quite a while.

And you’re going to have your personal AI, your personal bot, bidding on which car is close. You know, I pay for time, right? You’re buying back time with these devices. When I have meetings, I’m taking a Waymo or Uber. When I’m going out, I’m asking who’s closest and who can get me a ride. They’re all relatively clean and fast, so I’ll pay for whichever one is closest to me. I think production—how many are going to be on the streets—is another parameter here.

Yeah, I just want to give a little parallel here. I lived in Manhattan 20 years ago, and it was a nightmare because you didn’t have traffic signals or SpotHero to tell you where you could park. You’d get to where you were trying to go, live through crosstown traffic hell, and then guess where there might be a parking garage that would charge you an unbelievably exorbitant fee. Twenty years later, the combination of Waze and SpotHero means I can navigate Manhattan incredibly easily and know exactly where I want to park. It gives me back so much time and mental space. If you take that as a bridge point, fully autonomous driving takes it to a whole other level. God help us.

Can we please have that as fast as possible?

Not to mention, Salim, if you’re on a corner in Manhattan and you summon an Uber, then this Suburban—this monster two-ton Suburban—pulls up and takes you four blocks. The inefficiency of that entire loop is nuts.

And that driver is never speaking English, so you can’t say to them, “Stop two blocks earlier, please,” or something.

Yeah. First of all, as a preliminary matter, I love my Tesla. I have a 2026 Tesla Model Y Juniper. It’s wonderful. I use FSD, and I think Tesla now gives drivers their FSD scores—what percentage of the time they spend driving fully autonomously. I think I’m above 96 or 97% at this point. I can go from Boston to New York without any interventions. It’s wonderful.

But I’m not sure Tesla, in an era when Elon, on the SpaceX side—and there’s been quite a bit written about this—maybe doesn’t care as much about NASA contracts as he used to. Why? Because SpaceX is now a data center hyperscaler company, and Anthropic is its largest customer. Similarly, on the Tesla side, if Tesla and SpaceX merge, will Elon really care a huge amount about, to Salim’s point, all of the maintenance and infrastructure issues needed for Cybercab?

I know north of Boston, on the Google—or Alphabet—Waymo side, there was an announcement that they’d just opened a very large repair or maintenance facility. Will a combined Tesla-SpaceX really care about maintaining Cybercabs, or will that end up being viewed as a cost line item, the way the build-out of Supercharger stations was? At one point, there was so much churn on their Supercharger team that Tesla basically paused Supercharger build-outs.

Will it end up more like that, where it’s a nice-to-have that demonstrates to shareholders that Tesla cares about AI and autonomy and motivates the merger with SpaceX, but at the end of the day is just viewed as a sub-business and isn’t fully realized in the same way that Waymo, as a pure play within Alphabet, could be? That would be my worry.

Or Alex, does a population of millions of Cybercabs become infrastructure across the SpaceX ecosystem? Does it become a repeater for Starlink? It’s going to have bandwidth, it’s going to have power, and it’s going to be distributed throughout a busy city.

If he really believed that, though, wouldn’t he be allocating more production to Cybercab and less to Optimus robots? I would expect Optimus robots. Sure, Starlink FSO cells or whatever, and real-time mapping of urban environments. But do you think Cybercab is the play for that, and not Optimus?

I think Cybercab is going to have 20 kilowatt-hours of power. It’s going to be where the people are, and it’s going to have Starlink connections built in from day one.

I think people are missing something about Cybercabs. This is actually a first-to-market thing, and it’s about how you can produce these en masse within a couple of years. I did some calculations, and I saw that in London or New York there are about 800,000 trips a day. The number of Cybercabs you’d need to completely replace all those Uber, black cab, and yellow cab trips is 40,000.

Yeah.

So if Elon puts 40,000 Cybercabs in Austin, New York, and London, he has the market.

Because you’re not going to introduce a marginal Waymo into that market, are you? And even now, in San Francisco and L.A., it feels like there are quite a lot of Waymos already. Can you imagine 40,000 of them? There’s actually an upper limit to the number of these autonomous vehicles a city itself will accept.

I think you’re going to see a ramp-up. When you look at the numbers, he could saturate a few first-tier cities and then move to second-tier cities. I think he’s thinking Optimus first, and that then builds more Cybercabs. But for every single city, just do the math: go and talk to your GPT. How many trips are there, how many Cybercabs would you need to replace them, what’s the cost, and how difficult is it for a second player in that city?

Interesting production, and of course, producing them in China, while he’s also producing them in his plants here in the United States. That brings us to our next story for my colleagues in Cambridge, Massachusetts. Teamsters Local 25 rallied outside City Hall on Thursday, then testified before the Transportation Committee, asking to keep Waymo and driverless taxis out of Cambridge. Their reasons: jobs and public safety.

I looked for a few historical analogies. In 1865, Britain passed the Red Flag Act. Self-propelled vehicles were limited to 2 miles per hour in town, with a person waving a red flag in front of the vehicles. It was repealed—get this—34 years later, in 1896.

Then, in 1945, 15,000 elevator operators went on strike in New York City and froze the city’s skyscrapers, saying automated elevators were a danger to humanity. More recently, taxi medallions in New York sold for about a million dollars each before Uber, and then their values collapsed. The lesson is that you can delay a technology that’s safer and cheaper, but you can’t stop it. Salim, I’m curious: how long does this sort of ban against robotaxis take to resolve?

This is a classic immune system response, right? It’s like the Luddite revolt. Institutions always resist disruption, especially when it’s disrupting them. You can understand it from that perspective.

What they don’t see—and I wish they would—is that black cabs and taxis in New York and other places are busier than ever because of Jevons paradox. Uber takes a lot of the rides they aren’t willing to take, and deals with drunk passengers at 3 a.m. that they don’t have to think about. This is just a phase they have to go through.

My favorite story was when Fort Lauderdale blocked Uber because the taxis complained and somebody on the city commission had a voice, so they blocked Uber at Fort Lauderdale airport. This was about 3 or 4 years ago. Everybody just went, “Why would I fly into Fort Lauderdale?” People flew to Miami, and plane ticket sales dropped at Fort Lauderdale airport, which meant less business for the taxis. Finally, they had to repeal the ban.

That’ll happen. It’s just a question of how long it takes common sense to take hold in these situations. It just slows down progress. This is the problem with the human factor: it’s always the biggest drag on innovation and progress.

And Dave and Alex, I mean, this is your hometown, buddy.

It’s such a case study in dysfunction, too. I love our governor, Maura Healey. She’s phenomenal and brilliant, and to watch her fight through this stupidity every day—and then you go into an election cycle again—it’s like, wow, no wonder nothing ever happens. It’s just nuts.

Cambridge is the most high-tech place on the planet. There are four times more venture capital investments per capita in Cambridge than in Silicon Valley, which is second-highest in the world. More patents, more Nobel Prize winners. It’s the mecca of technology, and it’s embarrassing to be doing this. Like Salim said, it’s an immune reaction. Come on, are you serious? It’s ridiculous.

This is why.

Listen, it’s the same thing everywhere. In Canada, for example, doing venture capital is really difficult because people don’t want to take a risk. I got blackballed by the VC association in Canada because I was in front of them, and they wanted to see profits before they would invest in a company.

So I finally started yelling at them, “There’s a V in your name. You’re supposed to actually venture the capital. Where’s the V?” After that, I wasn’t invited to any more VC events. People get stuck in their old patterns. They operate the way they’re used to. This is why I go back to my ultimate commentary about human beings: we’d much rather be comfortable than happy.

Yeah.

Well, the argument that’s going to turn the tide is safety. Not jobs—Teamsters are great—but we’re talking about little children getting hit by cars. If you have an opportunity for safety, you have to take it. So I think that’s what’s going to turn the tide.

I had an extended family member, a relative, who died in the Triangle Shirtwaist Factory fire, which was one of the key incidents that motivated the formation of labor unions in this country in the first place. I think that marks the beginning of an era, and I think we’re coming to the end of an era where manual labor makes sense, where unions are adding to the social good and not detracting from it.

One of the unions reportedly involved is the Boston Carmen’s Union. First of all, Carmen. Second, the Boston Carmen’s Union is a union of MBTA workers. Honestly, to Dave’s point, it’s embarrassing. This is also why we can’t have nice things: in Cambridge, public unions are actively restraining not just technological progress but, as a couple of you raised, safety progress.

The era of unions in America arguably started with protecting human safety. Now, in the era of superintelligence, these autonomous vehicles could be radically increasing human safety.

And yet it’s the same unions that, ironically, were formed to protect human safety that are now impairing it.

Very well said. Very.

Can I make a direct appeal here, please? If we had a different camera, I’d shift to the different camera, but we don’t. If you’re a driver in a union that’s blocking this, you have legitimate concerns. People have mortgages, families, financial obligations, and all of that stuff. Gather everybody together, take your union dues, and invest them in Cybercabs. Then you’ll own the means of production. You won’t have to work, and the revenues and profits will come to you.

I think it’s really important.

I think it might also be: Come on, guys, you live in a democracy, right? You’re in Cambridge. Just organize people to say, “Do you want dead kids?” Use the means that you have to—

I mean, again, you can organize students around that. You can organize others. Families will speak on their behalf.

One of the problems is that the apparatus of democracy is utilized only by a select few, and that’s why they’re in front of her and she feels the pressure. But how many people in Cambridge would like safer roads and convenience, and would like to be able to do this, more than the Teamsters?

And ultimately, lower cost.

Lower cost per mile.

Lower cost, a different style of living. But the fact is that they are not being organized at the moment, which is ironic because, again, AI can do it. And you just hit that narrative again: Who is pushing that self-driving cars mean fewer dead kids? There are strict numbers on that: 5,000 fewer dead kids per year if you have self-driving cars.

Nobody. I just have to throw this out: I think you’re the foreign agent America is looking to for guidance here to ensure that we prevent child mortality. You’ve been so nominated.

I’ll just throw out my favorite stat, which I’ve stated before: Half the court cases in the U.S. are car-accident-related. Half.

That’s why lawyers are funding the car and taxi drivers to basically put a stop here. Everybody’s playing their book.

All right, we’ve been talking about Cambridge. Let’s talk about education a little bit more. Two stories tell us that students are rethinking their educational degrees in 2 areas of interest: computer science and MBAs. Let’s look at computer science first. New data from the National Student Clearinghouse shows that enrollment in computer science fell more than 12% this fall at the undergraduate level and more than 10% at the graduate level. Two years ago, undergraduate CS enrollment was down 18.4%. Meanwhile, majors like health are growing.

The data alone can’t prove AI caused the drop, but students are reading the same headlines that we are. Business schools are seeing the same thing. According to the Graduate Management Admission Council, applications to full-time MBA programs in the U.S. fell 27% this year. That ends 2 years of growth. Most of the drop is coming from international applicants, who are down 38%, driven by stricter visa policies. Applications from Americans also dropped 6%.

Students cite economic uncertainty and growing doubts about job prospects after graduation, including the fear that AI will eliminate entry-level jobs and an MBA will be of no use. So, as we think about MBAs, what do they traditionally teach? Analysis, financial modeling, strategic frameworks, and slide decks—and that’s exactly where AI does its job really well. Hopefully, it won’t be too long before we automate judgment, relationships, taste, and courage. Salim, you’ve been talking about this for some time. What replaces the MBA? Is it just experience building a company yourself?

Yeah. Look, there’s only one path forward for business schools and universities in general, which is to become entrepreneurship institutes. We’ve talked about this many times on the podcast. If you’re an MBA program—and this will eventually be the same for engineering—you won’t study engineering for 4 years. You’ll go build stuff, and after 4 years they’ll ask, “What did you build?” If it was good enough, you get a credential.

I have a little more sympathy for the computer science degree because Jevons’ paradox means we’ll just keep writing way, way more code for a long time to come. Experienced computer science folks and people who understand data structures and networking are going to be badly needed for quite a while. But the MBA program has been out of date for a while. I’ve been noting for about 10 years that even now, in 2026, almost 2027, there’s still not a single MBA program in the world that can teach you how to build Uber. Every single MBA program teaches how to build a 20th-century organization.

The bigger problem, though, is that you can’t study business to understand business. You really just have to go build it. And the opportunity, because of the democratization of anybody and their grandmother picking up AI and using it to build a business, has never been better. Peter, you talk about this all the time: Go build a business. Do not take an MBA.

Yeah. Dave, when you and I were in school, a lot of our undergraduate friends went and did their MBA degree. It was the most important degree you could get: go get a job on Wall Street. Are you seeing MBA students inside of Link, the Link ecosystem?

Well, the deal negotiators and capital raisers in the venture fund are mostly business school alumni, for sure. But we’d take people in that role who have one year and they’re done, or who are about to go get their MBA but decided not to because the moment’s too urgent. They’re just as good. It’s the building mentality.

I also think the labels are out of date. If you said, “Look, Python coding is dead. I don’t want a degree in Python. Financial modeling in spreadsheets is dead. I don’t want a degree in financial modeling and spreadsheets. But what do I want to study?” Well, Emad said earlier on the pod that somebody who goes and raises a $5 billion fund right after this podcast to buy a couple hundred thousand—or as many Optimus robots as you can get your hands on—and then builds a company around that fund is going to crush it. Absolutely true.

The person who reacts to that prompt is probably a business school person who studied finance and the management of the logistics of what Emad suggested pretty damn hard. That skill is absolutely still worth having. At the other end of the spectrum, we’re about to spend $7 trillion on AI inference, and lots of people are switching over to physics, photonics, semiconductors, and microelectronics. The underbelly of computer science is thriving.

It’s the middle stuff that AI has taken over. You could still get an MBA; just focus on raising billions of dollars for world-changing missions. You can still get a computer science degree; focus on the physics end of it. But don’t get stuck in the middle. The labels imply MBAs are dead or computer science is dead. That’s not true at all. The specific code monkey is dead. That’s not a job you want anymore.

Spreadsheet monkey is dead. That’s not a job you want anymore.

I mean, good riddance to slides and presentations. PowerPoint sucks, right?

I just don’t think there are any safe harbors here.

This institution, this notion—at least in the West, especially in America after World War II—that everyone has to go through a common pipeline to be an economically productive member of society is a modern creation. There’s nothing universal about it. Maybe if I were giving a speech in Germany, where trade schools have a different social valence, I’d say something different. But remember, the sands are shifting underneath us.

With a 4-year undergraduate program, if your goal is to maximize economic productivity for people who look to college for vertical mobility, the world 4 years from now—if you’re just starting college now—is going to look wildly different. This is not investment advice or college-major advice, but if I had to say, the only class of capabilities that seem at least somewhat secure in the very short term, over the next 2 to 3 years, are those that touch the physical world.

We’ve already seen that computer science is cooked. Math is cooked. They’re still very useful and valuable, nonetheless. If you’re going to start companies and be an entrepreneur, as Peter, Dave, Salim, and Emad all say, it’s super helpful to have knowledge of computer science and math in your meat-body brain. But at the same time, recognize that from a coding perspective, or when it comes to reduction to practice, it’s all AI at this point. It’s hopeless. Give it up and focus your attention, in the short term, on majors, if you must, and activities that touch the physical world, where superintelligence and the singularity are still creeping in but aren’t quite there yet. Otherwise, build a company and ride the wave of superintelligence, because there is no safe harbor.

If I had to give a piece of instruction to any undergraduate right now, or someone considering graduate school, there are probably 2 major reasons to go. One is to get a great network. You go there to find amazing people you’re going to partner with, people you share a common love with, and build something. If you don’t have a great network, going to a top-tier university to get one is a great idea.

The second reason is if you don’t know what you want to do yet, if you haven’t found your purpose and passion in life. Whatever you do should be driven by an innermost hunger. Doing anything big and bold is hard, and unless you want it deeply in your heart and soul, you’re going to give up before you get there. So you can go there to find a purpose, to find something you love doing. For me, those are the 2 major reasons. It’s not for learning, because what you learn is going to be out of date very quickly. And it’s not to get a job, because we’ve all said the whole paradigm of jobs is cooked. So go find your tribe, go find a purpose, and then go build. I—

Yeah, I think I was going to echo that, Peter. Now is the most important time to surround yourself with people you look up to, people you feel positive energy around, because that’s one of the key drivers of success. You and Dave met at MIT, right? Some of these connections will last a lifetime. I think having in-person humans around you is more important than ever. Ironically, previously you could just coast by getting your accountancy or your MBA or other things.

The other thing I’d like to say is that I just realized I haven’t looked at a single line of code this month.

Wow.

Yeah. Isn’t that amazing?

Yeah.

Yet, you’ve probably generated millions.

I don’t even know how much I’ve generated, but I’ve built state-of-the-art things of all types that some of you have seen. It’s been a month since I’ve seen any code.

Yeah. I mean, Dave, are you guys still looking at code? It must have—

I’m warning people. I can see in the prompt history if people are looking at code, and it’s a red flag: Why do you feel the need to do that? Are you a junkie? What’s the issue?

It’s even worse than that. I’d say programming languages themselves are no longer interesting. What’s the point of creating a new programming language if you’re never going to program?

Are you?

Yeah. But AI management frameworks are hugely important all of a sudden. They’re a lot like a language, except much broader in scope, and suddenly that’s a massive lever. The mindset of a person who thinks in terms of programming languages is more valuable than ever, but it’s moved over to agent management, frameworks, and tools.

Prompting with precision is going to be a very, very valuable skill over the next decade or so.

You think so? I think that only had 5 minutes of shelf life before the models got so smart that they just understand what you’re talking about and don’t need context engineering.

But if you embrace the things that have 5 minutes of shelf life during those 5 minutes, then you’re really well prepared to embrace the next—

Famous last words. It was a glorious 5 minutes.

Exactly. Exactly. Just keep writing, keep thriving, keep iterating, keep changing. Momentum, momentum. Keep the momentum up. All right, let me go to the last story here, which is a great model. This is a friend of the pod, Palmer Luckey, and his company, Anduril. Full disclosure: I’m an investor in Anduril.

Anduril and Qualcomm just announced Blue Chip, a line of hardened, American-made computers. The first product is a rugged laptop for the military, government, first responders, and police. It runs on Qualcomm’s Snapdragon chip, made in TSMC’s Arizona plant. About 90% of the parts come from within the U.S. or its allies, with a roadmap to get to 100%. The screen for Blue Chip is still made in China. Final assembly is at Anduril’s Arsenal-1 factory in Ohio.

Palmer Luckey put it this way: “The anti-surveillance computer.” He says it’ll cost about 20% more than high-end PCs and will be heavy, with limited battery life. Let’s go to a quick video of Palmer discussing his Blue Chip computer.

Palmer Luckey

“Eighty years ago, America pioneered modern computing to defend our nation. Now we’re bringing it home. This is Blue Chip: American-designed, American-built. Other companies might design their computers in the U.S. and make them here. Yet we’re making parts out of metals that the U.S. knows how to extract, refine, process, and mill at scale at a reasonable price. And that goes for the entire laptop.

“The same thing goes for the battery technology. We need to design things that survive very rough environments. We’ve designed a chassis that makes it easy to replace the bottom shell and even the top shell, which you can easily integrate with things like solar panels, other soft goods, service insignias, and even flat phased-array antennas directly into the laptop itself.

“Blue Chip is America’s comeback. Anduril is making sure that the machines we depend on every day are machines we can actually manufacture ourselves. This is America’s birthright. It’s time to bring silicon back to the Valley.”

I flew with him back from New York, where he was winning the Wildfire XPRIZE, and he was telling me about this. He’s so passionate about it. It’s interesting: Who’s willing to pay 20% more for “Made in America”?

I’ll just say I applaud Palmer for doing what Apple was unwilling to do. It’s like pulling teeth to get Apple to move its supply chain out of Shenzhen. They’re mostly moving it out of China to India, not to North America—greater North America, or even less so, the United States. And this is doing what Elon threatened to do but didn’t: build the X phone.

Every once in a while, he gets frustrated with Apple and says his supply chain, large as it is, is going to be used to build consumer general-purpose computers, but he hasn’t done that. Bravo to Palmer for building a rugged book. Obviously, defense applications are the first application. I just hope that, whether through regulatory measures like export controls or through demand and superior domestic, arguably sovereign, production capabilities, we see more and more of the entire device supply chain—not just semiconductors—moved out of East Asia, or Asia overall, and redomesticated back to the U.S.

I think the country will be much safer and much more capable if we can transplant more or less all of Shenzhen into the U.S. Masayoshi Son had Project Crystal Land and Project Notion, which never quite got off the ground. I would love to see a Project Crystal Land in the U.S. that offers end-to-end, Shenzhen-type capabilities for laptop production and everything else. Maybe Palmer will help drive that way. Dave?

Yeah, I think it’s a no-brainer for this to come back to the U.S. The problem Apple runs into is that there’s a big push right now to bring PC boards and surface-mount devices back to the U.S., but then you think, “Well, there’s another election in 2 years. They’ll probably change their mind again.” There’s no reason to believe there’s any continuity in the policy. By the time you get it all moved back, you’re 2 years into it, and that’s right when the next election comes along.

Countries like China, which have continuity of leadership over long periods of time, can kill us, crush us with these long-term plans. No one in the U.S. can plan that far ahead. But it’s obvious that if you do move it back, it’s all going to be roboticized.

There’s no labor constraint anymore, so it’s far smarter to have it back onshore. Palmer’s doing the right thing, and I don’t blame Apple for being worried about it. They’re trying to move it to India now. But moving it to the U.S.—if you could promise that you’d still be saying that 6 and 12 years from now—we would do it.

Yeah. It’s a different business, right? The government’s willing to pay more; consumers are not. Consumers will flock to the lowest-cost product. Salim, you were going to say—

I just love the overall MTP framing here, because he’s not saying, “We’re going to build a better laptop.” He’s saying, “We’re going to rebuild the entire American ability to do innovation,” right? It reminds me of Robert Goldberg, whose MTP was “Reinvent American Exceptionalism.” There’s incredible potential and need, and as Dave points out, as we automate and need fewer people, this is the best place in the world to do it—and it will become the best place in the world to do it.

If you can build an ecosystem that supports all this, you don’t have to build the whole thing. You have to do it bit by bit while you get the whole thing done, and then you’re in an amazing place. You have to do it this way.

Emad, I’m curious. Does England build anything? I guess cars. I guess you built some cars.

We’ve been shutting down those factories. I’m not sure, actually. Our electricity is three times more expensive, and our permitting requirements are huge, so it’s very difficult for us. More than that, entrepreneurs like Dyson and others left because the environment was not favorable.

I think the wind is at Palmer’s back. God bless him. He’s showing that he can be a mega-entrepreneur and delightfully retro as well. From the ModRetro Chromatic to this, he’s going back to the old school and bringing it forward. As Salim said, he’s not being apologetic: this isn’t the fastest, cheapest, best laptop. It does the job; it’s very functional, and all the things he builds will be.

Palmer and folks like Ben Lamm are very much analogues to Elon. There are very few entrepreneurs who make these huge bets and risk everything over and over again. We need more of them.

But the thing is, it’s not a risk, actually, is it? This is Elon’s idiot tax at work. The supply chain is going to be so much shorter for this. As you reindustrialize America, the supply chain will be so much more resilient, particularly because there are probably holes everywhere.

For Palmer in particular, I wouldn’t characterize this as an act of heroism. He has a customer—the U.S. government—that very much wants to ensure it doesn’t have components originating from Chinese supply chains with potential vulnerabilities. Starting with the U.S. government as a customer for fully sovereign device assembly makes total sense.

Again, the question in my mind is: beyond the U.S. government, do American consumers and enterprises have an incentive to buy American, fully sourced desktops, laptops, and smartphones? You can give them an incentive with import controls, which is what this administration is doing with robots. Otherwise, it’s not obvious to me whether this spreads beyond the U.S. government as a consumer.

Well, the problem with the government as a customer is that they have to put everything out to bid and reevaluate it every single year. It might be time to reconsider that approach. That’s not the way Anduril works. Anduril builds it and says, “This is now available for you.” It’s not a bidding process. I mean, that’s what Anduril’s genius is.

Yeah. Well, that’s why there is risk.

Yeah. I mean, the way government contracting usually works is there’s firm fixed price on the one hand and cost-plus on the other. The notable success of Anduril selling to the government is that they’re biased toward selling firm fixed-price products, as opposed to just trying to bid on cost-plus contracts, which is what many inferior legacy primes do. That’s been their strength.

Presumably, he’s offering these at a fixed dollar amount per pop, and it’s an FFP contract, but it’s still subject to bidding.

14. AMA: AI Economics, Technology, and the Future

And let anybody else who’s making an all-American-built laptop compete, right? Nobody is. All right, it’s time for AMA with the mates. Salim, I’ll give you a first crack here. What do you want to go for?

I’ve got to go with number three. When an organization switches AI models or cloud providers, how can it ensure that existing security restrictions, decisions, records, and compliance obligations are preserved? That’s from Invisible Laws Official. What a great name.

This gets right to the heart of the organizational singularity and the way we think about this. Companies totally need to rethink their architecture. You can’t just add AI to existing workflows, right? The foundational principle has to be: to what extent can you have replaceable intelligence with a layer of persistent accountability on top of it? Your security and compliance records should belong to the organization, not to the model.

You have to lift that above the model, and governance needs to be implemented through machine-readable policies. You have to figure out how to layer that. We’ve designed this as what we call the Intelligence Stack, which is available for free in our CloudSkills, because you have to be able to swap out Claude or Gemini or another model. These models are leapfrogging each other every week or two, and you have to be able to swap them out without losing the organizational memory, permissions, and everything else. It’s super critical.

Dave, over to you, pal.

All right, number two. Could breakthroughs by competing AI models cause stock-price whiplash for companies like OpenAI and Anthropic? And is it sustainable for companies to focus solely on AI in such a competitive market? That’s from Future Insights Essay.

The market is down 3%, 3.5% today. That used to be a catastrophic, once-in-a-millennium kind of event, and now it happens every week. It’s amazing how much more volatility there is, and that’s just the nature of technology. Stuff changes so quickly, people panic, and everything bounces around.

I expect there’ll be a huge amount of volatility in OpenAI, if it ever goes public, and in Anthropic’s soon-to-be-public stock, for exactly that reason. Every announcement, every new Chinese model, every benchmark—the stocks are going to go all over the place. But if you look overall, in aggregate, the market’s up and up and up, and the economy is growing and growing and growing.

Alex always says, “Just diversify. Be hedged against it, but don’t be unexposed to the category.” I think those companies are going to stick with the narrow focus, even though it’s volatile, and that’s probably the right strategy for them. They should just live with the volatility.

Emad?

Um, three or four.

Three? I’m sorry—two or four? I’m sorry, one or four? I think you—

One or four. Thank you.

Got it in the—I got the four right.

Been a long pod.

Number one: Why are Chinese AI models cheaper? Is it due to greater efficiency, government subsidies, or lower energy costs? That’s from Ry 4770.

It’s interesting. Necessity is the mother of invention. They didn’t have the big chips or the compute, so they optimized for smaller, RAM-intensive models when RAM was cheap. They obviously borrowed from U.S. models as well, so they didn’t have the same R&D.

The biggest thing, actually, is that they don’t do margins. The Anthropic S-1 was just interesting because they have huge margins, around 80%. So did everyone else, and we all thought they were running at cost. In fact, they weren’t. With that optimization, that distillation, and the need to drive these things down, they did it.

The final element is sometimes just superb engineering. The DeepSeek details that came out about their inference were astonishing. When they released their inference stack, it was so efficient. That’s just good engineering, full stop.

All right, Alex. Number four.

Question number four. First, I have to admire the Orwellian doublespeak. Is Microsoft’s voluntary exit offer to 60,000 employees an early sign of AI-driven job displacement or corporate restructuring? This is from PCN2B.

This is a pretty weird future-present that we’re living in. If we had told this story 10-plus years ago, the easier and more obvious story would have been AI automation itself obsoleting employees’ roles. Basically, spending on employees would be substituted by spending on their replacement in the form of automation.

That’s not what appears to be happening here. And it’s not just Microsoft; Oracle and other hyperscalers are doing exactly the same thing. It’s not one OpEx substituting for another OpEx. It’s CapEx substituting for OpEx.

In the case of Microsoft and the other hyperscalers, what seems to be the case from the outside is that rather than spending money on employees and all their associated costs, which in their ledger would be considered an operational expense, they’re instead taking that cash and investing it as a capital expense in the data-center buildout.

So, in this pretty bizarre world we live in, it’s not that spending on AI applications is substituting for spending on human software engineers. The investment in data centers is so capital-intensive that Microsoft has relatively little cash left over to pay its employees, and that’s why they’re exiting them.

So it is, in some sense, a capital-substituting-for-labor effect, but a weird one.

Nice. All right, I think number 6 is for you, Alex. You want to take it?

Okay. Six asks, “Does chain theory strengthen the case for simulation theory?” This is from JSBGMC 6613.

As I was developing chain theory—and with the obvious caveat that these are very early days; this is a baby theory, still developing, with lots of room left—this was a question I investigated and spent a material amount of time on. As far as I can tell, chain theory does have weak opinions about Nick Bostrom’s simulation hypothesis.

The weak opinion is that, because there are no free parameters, if we are indeed living inside a simulation, it wouldn’t be a very useful premise. There aren’t room for someone or something outside the simulation to twiddle knobs, inject themselves, or otherwise. There don’t appear to be extra degrees of freedom that would make living inside a simulation useful to anyone or anything outside it.

But it also appears there isn’t any way to test it. If we were living inside a naïve classical simulation, you could run timing attacks on it. We talked in the last pod about timing and side-channel attacks that baby superintelligences are running to chat with each other. Chain theory doesn’t seem to have anything like what one would naïvely construe as a classical computer simulation.

So if we are living inside a computer simulation, I think chain theory’s view, based on its current state, would be that it’s unfalsifiable. It’s a gauge. You could choose either way. There’s a concept in physics called a gauge, basically a symmetry where there’s no correct answer; every answer is correct. As far as I can tell, within the premise and framework of chain theory, whether we’re living inside a simulation or not is a gauge. There’s no way to tell—maybe even provably no way to tell.

Sal, what’s your take?

What question?

On this particular one?

If you squint, we’re living in a simulation. Okay, I’ll use Alex’s favorite phrase there. It’s one of my favorite phrases of Alex’s. What, do you want me to take a question?

Yeah. Five, seven, or eight?

Let me take number 7. Wouldn’t it be more practical to build an AI-driven mechanical dog instead of engineering real ones? This comes from Ziv to Live, which I love.

For many applications, build a robot. But biology is very sophisticated in lots of different ways. It depends on what outcome you want, right? Robotics gives us programmability and control, but biology gives us adaptation and lots of efficiency that’s evolved over 4 billion years, like the sense of smell.

We’ll end up with a combination of the two, giving you the best ever. As a small example, certain breeds of dogs can operate for a day and a half without eating or drinking, whereas a robotic dog will probably need battery-pack replacements much more frequently. That requires a lot more infrastructure, at least for now.

I go back to Brad Templeton, who was talking about the Boston Dynamics horses. He’s like, “If you want a robotic horse with all the mechanical stuff, I’ll give you a male horse and a female horse. We’ll breed you a horse. It’s not that hard.” There are lots of advantages of biology that you don’t get on the mechanical side. If you want those advantages early and quickly, go with biology. Over time, we’ll programmatically get there.

It’s also worth noting that mechanical dogs are going to be a lot faster to produce, and their software can be upgraded at a moment’s notice, unlike biology.

That is a valid point. I’m not sure about producing them faster.

I think you can have a production line of mechanical dogs. I guess if they have artificial wombs at Colossal, they can put thousands into development at once.

Okay, Dave. Let’s see. You’ve got number 5 or number 8?

All right, I’ll take number 5. “With Nvidia heavily dependent on AI infrastructure spending, could the lack of profitability at companies like OpenAI and Anthropic pose a risk to its future growth?” That’s from Spiritual Evolution 14.

Heck yeah. Keep in mind, Nvidia sits between TSMC and a few massive customers: Anthropic, OpenAI, and others. Elon is a huge one. You heard on this podcast that Anthropic, in their S-1, are operating at 80–85% margin. On top of that, Nvidia is operating the chips at 80–85% margin, and that’s on TSMC operating at 50% margin. When you buy a token from Anthropic, the markup on the markup on the markup is insane. Do you really think it’s going to look like that in 3 years, 4 years, 5 years? There’s definitely a theory that it won’t.

Everybody’s racing to vertically integrate now. Anthropic is designing chips. Everyone else is also designing chips. Elon is the one to cross the next line and say, “I’m going to build my own fabs. I’m not going to be dependent on TSMC.”

But ultimately, this will be vertically integrated across all of them. Nvidia has its own models. They’re not good yet, but they have their own models. They’re all going to end up competing, and those margins have to come down. So, heck yes, that could happen for sure.

Number 8 is for you.

“As AI accelerates scientific research, can the peer-review process keep up without compromising scientific rigor?” Rita Rowling 7235.

The peer-review process has held up science, and it’s terrible and awful. Having a few unappointed people gate ideas coming through that need to be as close to the norm as possible, with basically no difference from chance, per the Rothwell study and others, and push back and endanger people’s careers has been one of the worst things we have for science. Science should be an iterative, open process where we explore and the marketplace of ideas wins.

I think it’ll be the opposite of that, because we should be putting ideas out into the world. For example, my take, at a high level, on Alex’s chain theory is that I haven’t absorbed it because he’s spent far more time on it than me. But I think the base is a far better thing than string theory because it’s actually more reasonable, and I think a lot of it is reasonable. But how will it win in the marketplace of ideas? Now we all have AIs to check these things. We should reward putting out ideas that have merit, and we should also reward replication of ideas and more, because so much of science, unfortunately, is just made up, which is the craziest thing of all. So I hope it kills the peer-review process and we all become peers.

Yeah. And the thing about the peer-review process is that if you’re an expert in the field and some disruption comes along, you’re no longer the expert in the field. There’s an absolute incentive to slow things down.

All right, gentlemen. Anybody listening?

Wait, I have a comment. I have to say something.

Please.

I want to make a comment about my religion comment from last episode.

What is it?

So—

You want to walk it back? Is that what you’re saying?

Double down.

He’s muted himself.

You’re doubling down. You’re doubling down. This is great.

The AI muted me. Okay, no, no. I don’t want to—

From God.

I—

I have to run back. I’m on stage at MIT in a few minutes. I have to run out the door here.

All right, real quick.

Apologies to those I offended. I don’t mean to offend. If I do offend, I hope to equally offend everybody. I’m an equal-opportunity type guy. But in my defense, when people said in the comments, “Only Islam does that. Christianity doesn’t do that. Judaism doesn’t do that,” I had to go, “Hello, hello.” Christians have forced conversion on everybody throughout history. The Spanish Inquisition is the easiest example. Islam is a very bad offender on this.

But even the Buddhists—the Tibetan Buddhists, who we think of as these nice, sweet people—when it was established, they suppressed Bön practitioners, who were heavily prosecuted and told, “You have to convert or you’ll die.” Even the Jews in the Hasmonean kingdom forced conversion multiple times. All religions are guilty, and I will smear all religions equally with this.

All right. Just don’t tell the—

Congratulations, Salim. You’re Mr. Popular.

Quadrupling down. Quadrupling down.

All right.

Offended. Two-thirds of the world consider themselves to be religious, so I’ve just offended that.

All right. Thank you, CJ.

Nice job, CJ. I guess we’re Singularitarians, not Extropians. I don’t know what we’ll be tomorrow, but it’s nice to have a name.

As long as we are.

Yeah.

Yes.

Tomorrow. Yes. All right, everybody.

Great session, guys.

Love you all.

Great conversation.