SpaceX 的 2万亿美元逻辑、Nvidia 惊魂抛售、美国转向反 AI、Trump 撤回 AI 行政令、债券危机？

Sacks is out today, but we're very lucky to have Gavin Baker from a Treaties Management joining us. The spicy takes must flow. Welcome back to the program, bestie Gavin.

It's been a huge week in tech. We can start with the SpaceX and OpenAI IPOs. We've got Andrej Karpathy joining Anthropic, Nvidia crushing it—so many different things to cover, but I think we'll start with Andrej Karpathy joining Anthropic.

Karpathy is only 39 years old, and he's already a legend in the tech industry. Obviously, Andrej was a founding member of OpenAI, and he led the self-driving team at Tesla. He also coined the term “vibe coding.”

He recently built AutoResearch, which is an open-source training tool that helps AI models improve themselves by running 5-minute experiments. That got over 82,000 stars on GitHub. He did that as a weekend experiment, and all these civilians started building their own recursive LLMs. It's really inspiring.

Andrej Karpathy Skills is a tool based on his set of principles for Claude Code, and somebody just released that. It's pretty crazy when you think about it. He's going to be in charge of a new pre-training team at Anthropic, with the focus obviously being recursive self-improvement. In other words, they're going to have Claude improve itself, and they've already talked a little bit about AI improving AI over at Anthropic. Chamath, what's your take on this? Is this super important in 2026? Obviously, Karpathy is super well respected. He's one of the true talents in the space, but we're in a different inning than we were, say, 10 years ago when he was at Tesla or 5 years ago when he co-founded OpenAI.

You know what's interesting? If you go back to Google, the culture of Google, which they got right, was the singular technical talents there. They were singled out and called Google Fellows. I don't know if you guys remember this: Amit Singhal, Sridhar Ramaswamy, Jeff Dean. These guys are stars.

What's interesting is, if you track what folks—particularly Jeff Dean, I guess, because the other 2 aren't there anymore—did inside of Google, wave upon wave, they were at the foot of those waves. What's interesting about Andrej is that he's been at wave upon wave of AI. He was probably the first person who really commercialized Richard Sutton's “The Bitter Lesson” essay when he was leading FSD at Tesla, which was really about brute-force computation.

I remember him telling me the story—I don't know if he said this publicly or not—about how he spent a portion of his time, I want to say a quarter of his time, labeling data. Could you imagine, in 2016 or 2017, hand-labeling video data from Teslas? So he did that, then he was a co-founder of OpenAI. He's a star, and he's an exceptional human being. He's super curious.

What he's done as a kind of free agent is also quite impressive. So I think this is a really important deal. I think he's one of these really curious people who can be sent off, and they'll just go and invent new things. I think this idea of recursive self-learning puts these models on a combination of overdrive and autopilot.

If you put those 2 things together, I think you can potentially live out this idea that there's an order-of-magnitude improvement on a yearly basis—a new form of Moore's law. The model quality just goes absolutely parabolic.

Just like this. Straight up. I think the high-order bit there is that you throw a bunch of compute at the problem and these things learn really quickly. Gavin, what's your take on Anthropic's recent success and their massive hiring binge?

The success is extraordinary. It's undeniable. I think the fact that they were even positive, per The Wall Street Journal, in the most recent quarter is a really important fact for the whole AI narrative.

You could talk about circular funding, you could talk about ROI, and we could go look at the ROIC of the hyperscalers. But if OpenAI and Anthropic are at, call it, $100 billion of ARR now, with 80%-ish gross margins on inference, the returns are there.

They're growing really fast. If we add in Gemini, Cursor, xAI, and open source, it's not hard to see $200 billion, $300 billion, or $400 billion of ARR at the end of this year at high margins.

Across all of the language models? And you're talking specifically about the private language model companies, maybe not Google, which is obviously public. You're including Google. Okay.

But I was excluding a lot of the returns to this GPU spend that have come from better recommender systems at Facebook, Google, and Amazon; better ad targeting; and better ad measurement.

So, excluding that and just narrowing it to LLMs—which I think is the possible definition—it seems like there's going to be a really strong ROI this year, even excluding what are still some of the most economically important and profitable use cases for GPUs and AI infrastructure.

I do think what Karpathy is working on, recursive self-improvement, is really important. Unlocking that and continual learning may be the 2 final frontiers for AI.

The idea of recursive self-improvement is that the model, while it is training during a forward pass, has input into its training, or another model has input into the training. I think that could be really powerful, and Chamath's statistics of 10Xing every year might seem conservative if that comes to pass.

Then, of course, continual learning is the holy grail, where the model learns from experiences the way humans do. We haven't unlocked that yet. Those 2 things combined might pull the future forward in a very real way.

Right now, Anthropic has a decent lead on everybody else, whether it's 3 months or 6 months. Obviously, they're probably 6–12 months ahead of open source. Maybe they're 3–6–9 months ahead of their contemporaries, but they have a lead.

You put Karpathy in there, Friedberg, and now you have Karpathy doing recursive self-improvement. At some point—and it may have even occurred at Anthropic—the AI is going to be improving the language model more than the humans in the loop are doing it. Obviously, they're orchestrating it, Friedberg, but at what point do we think this, let's call it, super-recursiveness occurs? When will we cross the recursive valley and AI is doing more to build a language model than humans are?

I'm not sure when this idea that you feed the whole model into a context window to train itself and build a new model is going to happen, but I think there are probably a lot of different architectural paths that could be walked here.

One of them is this idea that you could make much smaller models and then create networks of smaller models that work together, where you ultimately have less energy or less cost per token produced out of an aggregation of models than you did with 1 single large model.

I've said this probably 3 or 4 times now: there's a lot of work and a lot of opportunity ahead in rearchitecting models and rearchitecting how models work together to solve problems. My guess is that there's a lot of leadership that he can bring to exploring those paths.

All it takes is a minor breakthrough, and your cost per token drops in half. That's a tremendous efficiency gain that seems very much on the horizon, because some of the early papers—I think I shared 1 from MIT a few weeks ago—indicate that there's a lot of room to run here.

In terms of rearchitecting models and deploying models, these very small models—small language models—and then verticalized ones are the future. We've got a company, Abacus.AI, that's doing it for corporations. Crushing it. Everybody's got an interest in doing this.

I don't know if you saw the news this week, Chamath. It happened about 2 weeks ago, very quietly. Google included the Gemini Nano model in its Chrome browser without telling anybody. It's 4 GB on your computer, and that's the one that does proofreading, spelling, autocomplete, and all that. So now we have Google covertly installing this on everybody's operating system.

“Covertly” is a strong word, so let's not use that word.

Without telling people, without giving people a heads-up. Let's say it in a way that we can both agree on.

We're in the phase now where I think breathlessly talking about every model improvement is a waste of time. There's no ROI in it. We are on a path of accelerated learning, and we're going to start to see end-user achievements that were heretofore impossible. That should be the focus.

For example, we were able to solve—I'm just collectively saying “we”; in this case, it was specifically OpenAI—

By the use of a human—and this is important—we were able to solve a math problem that had remained unsolved for decades and decades. I can tell you, in a different example, there are drug candidates that are about to enter clinical trials and INDs that were sitting on the shelf, and people didn’t think were very viable at all.

So, we’re at the phase now where these things are front and center. They’re useful to people, and they’re increasingly valuable. I think what we should do now is focus on these end-user use cases, because the way that you say it, in my opinion, is part of the problem. It starts to create this bogeyman, us-versus-them thing, and I’m not saying you’re doing it on purpose, but I’m saying this is exactly why I think so many people are becoming—when you mention AI, it’s sort of like a four-letter word now.

It’s presented as this thing, and I think we have to present the other side of it, at least so that people have the data. I don’t think Google is in the business of doing shady things. I don’t think they’re that company. There are other companies that would.

Meta. [Laughter]

You’re referring to your alma mater?

I’m not going to say which ones, okay?

But Google is not that company. I think the reason they did it was probably because there’s user utility. My point is, we should focus on the user utility, because I think that’s the story we’re telling from now on.

I think we, collectively—the 4 of us—can responsibly tell both sides of the story in a well-balanced way. Nobody wins if we become Luddites and go back in time.

Yeah.

And I think that if we’re not careful with our words, that’s what will happen.

Yeah, and by the way, I’m obviously not a Luddite, but this is what has been reported by a lot of folks: People were surprised—shocked—when they saw the size of the model being done in the background, and it triggered some people to look at it around privacy. I do agree that Google is not a bad actor in the space, so probably a PR error more than anything, I would say.

May I just add 2 things? There’s attorney maxing, and then there’s attorney maxed. Google is probably attorney maxed. [Laughter]

Yes. Yeah.

It happened for a long time. The second thing I would say is, I do think it’s incumbent on all of us as Americans who are involved in the technology industry in 1 way or another to be advocates for the positive, optimistic possibilities that AI introduces to everyone in this world.

It is starting to feel or seem like there may be a CCP-funded campaign against AI and data centers in America. That’s very logical for China, but it is not good for America. I just think we all have a responsibility. That’s what I would say.

Who do you think is doing a poor job of that and is responsible for this? Is it Dario, with his constant, “Hey, everybody’s going to lose their job”? Who’s responsible for this? Is it the CEOs blaming AI for their layoffs? What’s your take on this, Gavin?

Look, hold on a second. Everybody is trading their own book. It makes enormous sense for Dario to try to create the boundary conditions for a regulatory moat, because he will be inside of the tent, pissing out. He’s big enough now.

If you notice, a lot of the breathlessness has ramped up. Jason, we’ve talked about this. You can annotate successive rounds of fundraising and successive scale with the volume. I think it’s a reasonable business strategy, and I think that he’s quite clever.

If you actually just have a chatbot, a chat agent inside of Claude, and you ask it what it would do, it would come up with this strategy. Meanwhile, there are other versions of other counterstrategies and counter-exploitative strategies.

The point is that each CEO has a clear incentive. They’re operating at such a level of scale that they’re just reading their own book. It’s up to us to take a step back and actually see the forest from the trees.

I think—Nick, can you find this? There was a clip of Shyam Sankar, friend of the pod, fabulous guy, the CTO of Palantir. I think he was on Fox News, and he said, “Stop breathlessly asking these model makers what they think. Go to the end user and ask the person in the factory who’s using the model what he or she thinks. Ask what the doctor thinks. Ask what the scientists think, and start to tell those stories. That’s what we should be talking about.”

Yeah, and Gavin, I was sort of asking you your opinion on what’s causing this and then what’s the solution. Do you have folks you think in the industry who are representing it particularly well, where we can point out, “Hey, Elon has said we’re going to move to a world of incredible abundance and working will be optional”?

I think that’s a little scary for people to hear on the margin, because they hear “no job.” But he does say, “Hey, universal basic income is probably going to have to come into place,” and he’s said that multiple times.

You have Chamath, according to Chamath, talking his own book, scaring the bejesus out of people in order to get regulatory capture. What do you think is going on here, and how can we do better as an industry?

I think Chamath outlined a very viable and positive path forward. We’re just real people who are not at the tip of the spear. These are the positive impacts AI has had on my life.

I was at an event maybe 10 days ago, and someone who runs a hedge fund—his daughter was born with a very rare genetic mutation that would normally have condemned her to a life devoid of joy, meaning, everything. The neurons in her brain were not firing, so who knows what her life expectancy would have been or what the quality of her life would have been. It’s a tragic disease.

He said he didn’t accept that as an answer. He did an enormous amount of research with LLMs and found an existing, safe drug on the market that they thought would have a meaningful impact on his daughter’s condition. And it did.

I think the percentage of times that the neurons were firing was 30% or 40%, and it took it up to 80% or 90%. That means that she can live a normal life. She may not be as smart as she would have been, but she will live a normal life.

He’s now figured out how to use AI to further tailor that drug. There have been all sorts of advances in protein design, et cetera, et cetera. He’s reasonably confident he’s going to have a drug in months that is a complete cure.

That’s just 1 person, 1 dad, who was unwilling to accept defeat for his daughter and who changed her life and the life of everyone else with that disease. We should tell those stories.

I think Elon’s doing a good job. A future where work is optional sounds great to some people and scary to others. A 4-day workweek is probably something that sounds good to a lot of people.

I think Jensen is doing a good job of being an effective advocate. I just think we need to stay focused on the positives. That’s what I’d say.

Yeah. Friedberg, what’s your take here on the AI PR crisis, if we’ll call it that? We had 3 different commencement speeches that were booed, Eric Schmidt being 1 of them, and 2 others by maybe less notable folks. When you hear young people booing AI vociferously, why are they doing that, Friedberg? What’s your take on the overall PR problem and how to turn it around?

That’s a long answer to that question. It relates in some ways to your concerns about socialism and polarization.

Answer? Yeah. What’s the long answer?

I mean, why do people hate technology?

Why do they hate this technology? They love their phones. They love the internet. This technology they hate.

I think there’s an underlying view that technology creates leverage for a small group of people, which creates power imbalances, and nothing represents that more than AI. A small number of people who control, profit from, and benefit from AI are going to end up getting outsized returns relative to the broader population.

Ultimately, all technologies commoditize and diffuse, but the time to diffusion of the technology here is such that it’s going to be extremely asymmetric for society. I think there’s something fundamental about that.

Nuclear bombs, I think, really created this moment in people’s minds in the mid-20th century that, by the back half of the 20th century, gave everyone a high degree of skepticism about technology and science generally. Those who have the knowledge and those who engineer solutions with that knowledge can create outsized advantages for themselves, and it puts the rest of us—the rest of the world, the rest of the population—at risk.

Because those questions about when this benefits me and how it benefits me can’t be answered today, the economic benefit that’s accruing to the few today becomes the narrative. It becomes the story, and it becomes this power system that a few people take from the many.

There’s something deeply disturbing for the average person about that. They don’t understand how it works, why it works, what it’ll do for them, or when it will do it. All they’re being told is that some people are making trillions of dollars.

So, I think it’s pretty obvious why this has such a backlash. Secondly, I think there’s a deep amount of external energy fueling this antitechnology sentiment in the United States, and it has been for decades.

I think to Gavin’s point, I don’t think it’s just China with NGOs today. I think there is a long history of state actors intervening in media activities in foreign nations to try to create and fuel a sentiment that reduces progress in a competitive state. I think this goes all the way back to KGB disinformation during the Cold War, and it’s been refined, honed, and improved over time.

This is not just some conspiracy theory. There are plenty of great books about this. The techniques of what’s going on specifically today, I don’t know enough about. I don’t have any great details on that.

But I don’t think there’s no foreign interest in seeing technology advancement slow in competitive nations. The United States probably does similar things to other nations, and I think that’s probably a key part of this.

And then I think this third piece is, when the Copernican Revolution happened, it was a mind—heliocentricity was a totally new way of thinking for humans, and it was deeply disruptive to the church. It was deeply disruptive to the power centers, which were the centers that could tell people, “Earth is at the center of the universe. We’re in control. We’re the direct channel to God.”

The idea that the sun is at the center of the solar system, that we spin around it, and that we’re a tiny speck in the universe was very hard for people to grasp. There’s something about AI that’s very not human-centric, and it kind of shifts away from the ego of the human. It’s almost anti-humanist. I think that’s a deep psychological current for a lot of people and their disdain for this technology. It fuels it. It’s not the cause, but I think it fuels it.

So, I think there are a lot of complicated aspects to this, J Cal. I don’t think there’s a simple “put Cham on a podcast and he’ll solve the problems of AI right now.” I think there’s a real set of shifts happening and a real set of global competition underway where various state actors and interests are competing with each other.

Yeah. Cham, do you think that we should slow down?

I don’t think you can.

No, no, no. Do you think we should slow down?

No. I was just talking to some people on Zoom right before this, but I think after the Manhattan Project, research labs were stood up to retain our scientists who worked on the Manhattan Project from effectively leaking back to Russia, Germany, and other places that were adversaries to the United States.

They were all against the nuclear bomb. They worked on it because it was necessary for the United States’ security, but when Russia got ahold of the secrets, they were leaked because people were worried that if the U.S. had all the power, there would be no counterbalance to that power. The nuclear secrets were leaked to Russia for that purpose.

Then, when Russia had the nuclear secrets and began developing hydrogen fusion bombs, and it was clear that they were going to race ahead, the United States raced ahead with developing nuclear bombs as a counterbalance to Russia. When the proliferation began, there was no stopping it. It began, and you had to have this balance in the world. Otherwise, you have an asymmetric power that can do whatever it wants globally.

I think there’s that moment in the world right now where, if the United States does not advance its AI technology—the availability of it, TBD, industry, taxation, all these things that we’re talking about doing—someone else will. And if someone else does, we can go through what would happen.

There’s a complicated game theory on this, but what would happen if China had sufficiently advanced models and sufficiently advanced scaled deployment of those models relative to the United States? As you do that analysis, you realize, “Wait a second, that’s probably not a healthy place for the world to be.” It’s also probably not a healthy place for the United States to be the only one with AI.

So, I think what we end up seeing is that if we do try to slow down AI, we lose this moment of balance that’s necessary when you have a technology proliferation like we saw with the arms race after World War II. We’re going to see it again here.

Chamath, you bring up a good point. Should we slow it down, or could we slow it down? There have actually been some discussions about ways to do this.

One of them would be: with self-driving, people are scared that all these cab drivers are going to lose their jobs—Uber drivers, cab drivers, bus drivers, truck drivers. This is over 10 million people in the United States driving things for a living. Would you be in favor of announcements that it will be a paced rollout and won’t happen all at once? In other words, those people would be given some amount of job security to stay behind the wheel with it.

Another example that’s been given is, if you put Optimus into Amazon factories—or the Figure robot just did a week of sorting packages—if Amazon deploys those, there’ll be a tax on them per hour. We’ll tax humanoid robots in some ways and then use that for, say, retraining people.

Those are 2 very specific conditions and approaches that people have been promoting. Do either of those resonate with you in any way?

I think it’s interesting that in all of those discussions, I’ve yet to see an actual survey of only the truck drivers and only the package sorters.

The question that I would have is: do the people who do these jobs want these jobs? If they do, then there’s a reasonable claim to make to keep those jobs the way they are. If you’re saying, “This is the job that I do, I love it, and I’m able to provide for my family,” great. That’s a very different argument than, “Amazon has 35% or 40% churn inside of its warehouses, and we should probably ask the question: why is that?” Because if it were such a great job, I suspect the churn would be 3% or 4%.

So, what exactly is it that we want to protect? Have you asked them? I think this is, again, a bunch of people in the peanut gallery who want to take a moral high ground and try to make some other group of people feel guilty or feel bad. At no point are we actually having the conversation that we should be having.

It’s interesting to me that there was supposed to be an EO, a presidential executive order, announced today. Then it was pulled. It was scrubbed at the last minute.

Did you guys notice that? Yesterday, what was leaked was everyone who was attending. It was all the big AI lab CEOs, and it was all the big hyperscaler CEOs, including friend of the pod Nikesh Arora. Shout-out to Nikesh.

Then it was scrubbed an hour ago. Why was it scrubbed? The president said that there were aspects of the order that he didn’t agree with. As far as we can tell, those aspects would have required some amount of supervision, insight, and review from the federal government.

Of language models specifically. Of these frontier models is what I read.

Not language models, because I think it’s just AI. There are going to be many different kinds; they’re not always going to be language models.

Look, I think Friedberg is right. We are in a proliferation with China. I think it’s actually good that China is less than 9 months behind us. I think it allows us to find a détente where we have a certain magnitude of capability that they also have. That allows all of us to then seek peace and abundance.

The fact that we are orthogonal societies—we are organized differently—increases the probability of finding peace, using René Girard’s framework of mimetic theory, than if it were us and another country that was exactly similar to us.

So, I think what we need to do is probably adopt KYC. That should be something that the U.S. and China get together and say: “You don’t want it to get into the hands of people you can’t control.” You probably already KYC those models anyway inside of China. You already review those training runs before you allow these models to be released. We already know that’s happening.

We should probably do some sort of KYC so some crazy person doesn’t create some biological weapon. I think those are reasonable ground rules, but otherwise Friedberg is right. You have to take a little bit of a deterministic view here, which is that we are in this existential race, and we need to get to the place where each of us—meaning us and China—can look each other in the eye and say, “All right, weapons down,” so to speak.

Kevin, I’m going to hold you to answering 2 questions. One: should we run frontier models? That’s specifically what was mentioned in the leak about the EO—frontier models, the powerful ones. Should they be run through some sort of testing before they’re released, and should there be some regulatory framework for that? That’s my first question. It’s a yes-or-no question, and then you can explain your answer.

Kevin Weil

Jeez, I just think it’s such a complicated topic. It feels like we’re a little early for that. I don’t love the idea of the United States doing it and no one else doing it. I think in a world where we hold hands with China, that’s much more palatable, and we are aligned, trust each other, and have verification capabilities.

Right. Yeah. Let me then rephrase it. Should China and the U.S. come up with a simple battery of things that have to be tested before these go out, including bioweapons, terrorism, and, in general, in that vertical, just really known dangerous things—just like the FDA might test for poisons or contaminants in a food or a drug?

Would you be in favor of that? I'm curious.

So, a few things. One thing that's great about America is that we have other forms of regulation.

Self-regulation, sure.

We have self-regulation. We also have the courts. If an AI model company behaves responsibly, they know that there are ways that people who have been harmed can seek recourse. We already have a system that encourages responsible behavior on the part of the model makers.

That's a great point, because OpenAI is being sued right now by a kid who killed themselves after talking to OpenAI's model. So, you're correct in that. Yes, after the fact.

And we will see more of that. To me, once you give something—give a power—to the government, it's almost never taken back, and it seems to grow. It's kind of a one-way ratchet.

One-way ratchet. Yeah. And then, the second question: Chamath is saying, “Hey, nobody listens to the cab drivers or maybe the people sorting the packages. Do they want their jobs or not?” Actually, in the UK, there was just a 60-minute special, and Boston and New York are pretty adamant that they want humans to stay. They want to ban self-driving in those locations, severely limit it, or maybe limit it in some way to let those people keep their jobs. How do you feel about that possibility? Is that something you think society should be open to—some gradual licensing?

They're going to get sued for wrongful death. When somebody runs over somebody else, and you could have implemented a solution that has a zero-death rate, that's very different from package sorting.

Okay.

Go talk to the package sorters, is what I say. Go talk to the people inside the Amazon warehouse. Ask them what they would rather do at Amazon. Ask them.

Yeah, sure. But Gavin, what are your thoughts here on either one of those examples?

I think going to a city where you can't get in a Waymo or a Cybercab is going to feel barbaric and unsafe until you—

Agree. I strongly agree.

I don't know if you remember, but in the early days of Uber, sometimes you would go to a city where there was no Uber.

Yeah, that'd be incredibly frustrating.

I'm not going to come back until they have Uber. It's so inconvenient. Whatever individual municipalities decide, I do think, one, Chamath's point is really powerful. There are 50,000 automotive deaths per year in the United States, if I recall correctly, and 1 million globally. That's not tolerable, and there will for sure be wrongful-death lawsuits. From a convenience and quality-of-life perspective, I just don't think it's going to persist.

That's another great thing about America: You have this patchwork of different states and municipalities, each one doing things in a different way. I'm not suggesting that's good for AI, but it does tend to—as historically, the Curley effect aside—lead to more positive outcomes where cities and states compete. The Curley effect being that—

Yeah, that is a really important point you're making, David. With Flock Safety as but one example, we have an AI tool called Flock Safety. It's cameras that use AI to monitor people who are committing crimes. There's a privacy issue around it. It is bottom-up. You just do it by town. It's not top-down, and states can regulate it. The same thing will probably happen with self-driving, and states will probably have some say in how AI is deployed, even if maybe some centralized governments don't want to do that. I really think this comes down to—

The Flock thing, I think it's so good, Jason. Crime is now a choice.

Yeah.

You know, I think the Cambridge City Council voted to turn off gunshot detectors 2 days ago.

Which city did that? Which municipality?

Cambridge. Cambridge, Mass.

That's the place where Harvard is.

So, the geniuses coming out of Harvard in that town decided gunshot detection shouldn't occur.

We don't want gunshot detection. Just so we're clear.

It's wild because there's a theory that it disadvantages—that it might lead to an illegal migrant who's shooting a gun being apprehended, and we don't want that.

Got it.

And a16z had a great essay on Flock. We can really solve crime, and it's just a choice. Different states and municipalities will make different choices to be pro-crime or anti-crime. I'm sure they don't cast it as pro-crime. There's some sort of moral or ethical reason they may make that choice, but people will vote with their feet over time, and then voters will vote with their votes, and we'll see what works.

Have you guys been to Vegas recently? My wife and I went to visit Vegas, and we spent the afternoon with Ben Horowitz and his wife, Felicia. She has done this incredible job with the Las Vegas Police Department. It is one of the most impressive things I've ever seen, and to your point, crime is an option, and they've said no.

What happens there is they have gunshot detectors and drones that get deployed off the roof of the police building. We were sitting inside mission control, where you see it happening, Jason. If something happens, they have eyes on-site within minutes. They can track offenders and bad guys all the way to wherever they're hiding. You walk out of it, and you feel incredibly safe, like they're really on top of it.

When you understand the level of investment, it doesn't take billions of dollars compared to the cost of the crimes. It's de minimis. It's de minimis, especially when compared to the cost of the crime occurring.

Exactly. If you gave the Las Vegas Police Department $30–40 million a year, it would be the safest city in America, and that's all it would take.

Yeah, exactly. And for the privacy concerns, Friedberg, there are very simple solutions to this. I am a privacy advocate myself, of course. We all want some level of privacy. I had the Flock CEO on This Week in Startups twice in the past 10 years. He's very considered.

The way they do it with Flock is they allow you to have a rolling database, and I think there's a maximum amount of time you can save the license plates for. They don't do facial recognition. I don't see why not, but let's put that aside. You can only keep it for 2 or 3 years, and then they insist on having an audit trail. There are all these little things you can do on the back end to protect privacy with audit trails.

We have a lot more to get to in the docket. I just want to give my final thoughts on what we're talking about here in terms of the AI problem and the PR problem. I think we have to recognize that the layoffs that are occurring in Big Tech and in a lot of these places are not just the bloating issue anymore. I'm just going to point to 2 factors that I think are scaring the bejesus out of people, and we just have to admit that this is occurring, as opposed to continuing to debate it: Is it occurring? Is this just cover? Are we AI-washing?

The first one I want to give you an example of is Matthew Prince, who's the CEO of Cloudflare, an incredible public company. Two weeks ago, he said, “I laid off more than 20% of my workforce. I didn't do it because Cloudflare is struggling. We posted record revenue growth, have strong free cash flow, and are adding an unprecedented number of customers,” yada yada yada. He says, basically, that he's getting rid of measurers. Measurers are the people who manage people and who measure data. He just says, “We're getting rid of all those people. They're unnecessary because of AI, and we'll be adding people in other positions.”

At the same time, Zuckerberg did another round of layoffs, and they were done in a way that people felt was not considerate and a bit dystopian. Here's Zuckerberg for 30 seconds:

“In general, the average intelligence of the people who are at this company is significantly higher than the average set of people that you can get to do tasks if you're working through these contractors. So, if we're trying to teach the models coding, for example, then having people internally build tools or solve tasks that help teach the model how to code, we think it's going to dramatically increase our models' coding ability faster than what others in the industry have the capability to do who don't have thousands and thousands of extremely strong engineers at their company.”

Okay, so what Zuckerberg did was, at the same time, tell everybody, “We're laying off these 8,000 people.” A lot of those people are incredibly talented. Some of them are on H-1B visas, which creates all kinds of chaos for them in their personal lives, and obviously they're having record profits there as well.

This came after tens of thousands of layoffs, which were obviously because of bloating. At the same time, he said, “We're putting recording software on every single person's computer in the company to study and train our model.” People were like, “Oh.” People who had been there previously said, “During the AI hackathons they had months ago, I built all these AI tools to make my job more efficient.” And then Zuckerberg laid me off.

So, the perception people have now—and it's quite correct, I think—is that the most you can hope for here is that you keep this job for some amount of time and train your way out of it. Hopefully there's some more work for you, but they're studying you, and Zuckerberg just said it plainly there: “Hey, we're going to study everybody here, and that's going to lead to more replacements.”

This is scaring the bejesus out of people, and we need to have an answer for it. Yeah.

I thought the Matthew Prince note was horrible.

Okay, explain.

This was like from the PR school of retards.

Okay, here we go.

You could not have written a worse memo. It's like you reduce humans to a label called “the measurers,” and then you're like, “I'm going to lay off all the measurers.” I just think that part of this, again, I'll go back to maybe the Shyam Sankar quote that I'm thinking about, should be extended beyond the model makers. Can you just play this for 1 second, and I'll tell you why I think this is just so—

And then we'll go on to the S-1 from SpaceX.

Shyam Sankar

We're listening too much to the inventors of AI. I know that's appealing. They're geniuses. They're smart. We need to be listening to the frontline factory workers who are using AI, saying, “Wow, I was able to add a third shift. I was able to hire more workers.” Or the ICU nurse who says, “I have more time to spend with my patients. I'm able to ensure they don't code during a shift change.”

So, look, my point is the first part of what he said applies here, which is, who cares what Matthew Prince thinks? Because the reality is that if this is the way that you're going to message something as critical as this, I think you did a horrible job. Now you label these people and put a scarlet letter on their backs. So now when they try to get a different job, they're like, “Oh, you're one of the Cloudflare measurers?” How does that help anybody? It didn't need to be done this way.

There's enough of these tech CEOs that are now public. You can hear them, and you can understand them. I think what we're learning is, man, they're really good at one thing, and they're not necessarily as good at all the other things.

Yeah. Okay.

And so I would say, shut up. Get behind the keyboard, just do your job. If you need to manage something, just manage it, but don't write these missives. You're terrible at it. All of you. You're all terrible. You suck at this.

All right.

End of my TED Talk. Thank you for coming to my TED Talk.

Thanks for coming to Chamath's 18-minute TED Talk. We'll keep moving on.

Sorry. And sorry. When everybody gets upset, this will be why.

Yes. I do think the Zuckerberg announcement and Jack saying, “Hey, we're going to have half as many people, and everybody reports directly to me”—this is all building this fear in society, and I think people are rightfully scared. If the people building it tell you to be scared, your job's going away.

Wait till the next regulatory filing comes out from these companies and they authorize a massive share buyback and an increase in their dividend.

Yeah, and their cash pile grows.

All I'm saying is there's a right way to do this, make these decisions, and then there's a wrong way to do it, which is to message it in the way that they're doing it. Whoever is running PR and comms and approves and reviews these things are really bad at their job.

Mhm.

They don't understand the moment.

Oh, some breaking news here. It looks like Anthropic has hired 3 more people. Here we go. Personal job news from Sam Altman: he'll be joining Anthropic. I say that's pretty good. Okay, who else is joining Anthropic? Let's see if we've checked everybody's socials. Oh, Tucker. Tucker Carlson is also joining Anthropic. He'll be doing their PR and podcast from Anthropic headquarters. And who else? Oh, Chamath. Looking good. Well, you look like you put 5 pounds on, and you're using the pants—

First of all, this is not what I look like. And—

Did you gain 5 pounds? Looks like the same guy. Are you not wearing underwear? He's not wearing underwear, and he's wearing his khakis, but I think it's just he's trying not to show off those scrawny splints—those little slats he calls legs. He's covering them up now.

Hold on. I'm going to send Nick an updated picture of my legs. I knew that this was going to come up. I'm going to send Nick an updated picture of my legs, and you can deal with this.

You can Photoshop your legs all you want.

I didn't Photoshop—

You look like an ostrich, brother.

He's better than you today. I'm working on it.

He may not be Photoshopping. He may have been leg-maxing.

He could be leg-maxing. Are you leg-maxing? Are you BPC-157-ing your legs? What are you doing here?

First of all, I'm 6'2", you goons. Okay, so—

Ostrich legs.

All you little people—you know, Jason, you're 5' tall, so your ability to have legs—

I'm 5' on a good day with my platforms.

—to have legs is different, because my muscle mass won't show on my legs the way it shows on your legs.

Oh, I don't think you're helping, but look at what he's doing there. Friedberg, you see it, right? Look at how skinny the calves are, and then look at how he's wearing a—

Oh my God, Jason.

—a bra for your quads.

Oh my God, stop.

The equivalent of a push-up bra. He has those BOSU balls. He put BOSU balls under his hammies to—

You did it. You're using filters to pump those up. My legs are flat. Okay, you know what? Move on.

You know, I was just saying I think we should give credit where credit is due.

Yeah. Chamath's legs—

Better.

Chamath has been doing a lot of work on his legs.

Thanks, man.

It's better.

Thank you.

It's better.

Thank you. I'll give him better, but I do think that he's pumping them up here. Okay, let's keep going here.

All right, topic 2 is SpaceX just filed their S-1 on Wednesday. They are aiming to raise $75 billion at a $1.75 trillion valuation. This would be the largest IPO ever by more than double Saudi Aramco's $29 billion IPO a couple of years ago. Listing is expected in mid-June, likely June 12th. The ticker will be SPCX.

We got a lot of interesting information in the S-1 teardown. Obviously, SpaceX has 3 main business units. Starlink is the money printer right now, but there's a second one that's emerging. Starlink did $11.4 billion in revenue last year on 50% growth, with $4.4 billion in operating income. Over 10 million people are now subscribing to Starlink. That business could easily have hundreds of millions of paying subscribers, so that's a lot of growth potential there.

The space business is about $4 billion in revenue. It's growing at 17%, which would still be strong growth, but it had $650 million in operating losses. xAI did $3.2 billion in revenue, more than double year over year, but it had $6.4 billion in operating losses. SpaceX had $20 billion in capital expenditure last year. Over 60% was for the AI compute buildout. Obviously, they were trailing Anthropic, OpenAI, and Gemini in terms of xAI, playing catch-up. Elon did a big reboot of that, as we saw on Twitter.

But here's the big one: EWS, Elon Web Services, as we call it here on the All-In Pod, has exploded. Anthropic is paying SpaceX—wait for it—$1.25 billion a month to rent out Colossus 1 and parts of Colossus 2. It's a $45 billion deal over 3 years, $15 billion a year. In other words, they added a Starlink in terms of revenue to the party. Plus, if they buy Cursor, that's going to add another $2 billion or $3 billion.

Not “if.” I already told you they already bought it.

Okay. I'm just trying to dot the i's and cross the t's here. But when they buy Cursor, that adds $2 billion or $3 billion. That's not in the S-1.

That's also growing and doubling.

Yeah, that's probably growing 2x year over year. Who knows how much faster it'll grow? Polymarket gives SpaceX a 71% chance of closing its first day of trading with a market cap above $2 trillion.

I'll stop here. Gavin, you've been involved in the company for a long time. Chamath, you, I think, were a big investor in the satellite company that became part of Starlink, which is the revenue driver there. So you both have a lot to say about this. Gavin, your take on the S-1, and specifically Elon Web Services.

Well, I think what's important about Elon Web Services—it does make me laugh—but $15 billion means the AI business right there is going to quadruple. It has already effectively quadrupled.

I think what's important about that is there's a stat in it that, for the first data center, I think it was 122 days. The second one took them 91 days. The third one was, I think, 66 days. They build data centers dramatically faster than anyone else, at a lower cost.

Now that you have a clear off-take partner, and I would expect partners to become partners, there is no reason they can't start stamping these data centers out really fast. Having watched Jensen for a long time, it is important to Jensen that his GPUs be used. GPUs will be allocated to whoever can plug them in, turn them on, and start converting electrons into tokens.

I think this business can grow dramatically faster than maybe what anyone could have contemplated 3 months ago. But $15 billion from Anthropic is extraordinary.

Important note: it can be canceled by either party with 90 days' notice. I just want to make sure we also have that in there. That means Elon might want his compute back, or Anthropic may find another solution. So they both have an out.

And I think that's probably an important provision for everyone. But I think the other thing that came out this week, which was not in the S-1—Nick, can you throw up the Pareto frontier? Maybe don't include the email and the names and everything, but the Composer 2.5 stat. I think this is really extraordinary.

Cursor’s Composer 2.5 model came out this week, and this is Pareto dominant. This is just 3 or 4 weeks of doing reinforcement learning on Colossus 2 with Cursor’s data. We will never know, but Cursor allegedly has more tokens of coding data than exist on the public internet. That was a statistic from more than a year ago, so I’d imagine it’s grown significantly.

I think Cursor and Anthropic probably have the most proprietary tokens of coding data. What this jump from Composer 2 to Composer 2.5 showed us is that when you do an appropriate amount of reinforcement learning using that data, let alone injecting it into the pretraining of a new base model, you can get extraordinary results. Composer 2.5 is the same base model as Composer 2, which is Kimi K2.5. This is amazing. This is 3 or 4 weeks, and it is Pareto dominant.

The Pareto frontier, if you draw a curve of the blue dots, shows that Composer 2.5 is literally well outside the Pareto frontier. And that’s after 3 weeks. What’s going to happen next is that you’re going to have a new base model with a Cursor model in it.

Yep.

Then you have the Cursor model RL’d using the biggest coherent compute cluster in the world. And I think this is—I think this may—

It’s significant, yeah. And Cursor was dead in the water in terms of access to compute. They were falling very far behind Codex, Google, and Anthropic, and then Elon let them on Colossus, and boom, instantly their models are growing faster. This could be— we could be sitting here a year from now and they’re the dominant player.

Could we be sitting here, David, in a year with Elon selling compute to Google and OpenAI? Is that a possibility or not?

Well, I think it’s much easier to see him selling compute to Google, and I think that’s going to happen. There have already been posts about that. For sure, Google is going to want to be part of orbital compute.

It’s very funny: the only people who are skeptical of orbital compute are those people who are not involved in data centers or space. Google, Anthropic, Amazon, and NVIDIA are all very convinced that orbital compute is going to be a reality, and obviously SpaceX is extraordinarily well positioned for that. But I do think that Composer 2.5 data point is really powerful.

Keep an eye on it, yeah.

The other thing that’s come out is Grok Build. What Grok lacked that a lot of other models had—and I do think it’s important to remember that the newest version of Grok 4.3 is on the Pareto frontier for all frontier models. You’re either on the frontier or you’re not.

The companies on the frontier are xAI, with one build of Grok 4.3, which is a 500-billion-parameter model; Google, with Gemini 3.1 Pro; and then OpenAI and Anthropic. And that’s it. Those are the companies on the frontier.

And then the 4 horsemen, yeah.

They each have one dot on the Pareto frontier. Obviously, you want as many dots as possible, but Grok lacked a harness. Claude had Claude Code, and OpenAI had Codex. Now, with Grok Build, there is a harness available to Grok.

It’s a downloadable app that translates into English and has integrations with all your favorite stuff, whether it’s Notion, Gmail, Slack, and so on. If you don’t have that, it’s just like using a basic chatbot from a year ago. Now they have their downloadable app in market, and they’re cooking with gas on it. They’re playing catch-up, but they’re moving fast.

It’s more than just an app. It’s a runtime; it’s an environment. It manages state and memory. It makes these models dramatically more useful, to the extent that I think the people at the frontier all agree that the harness is essentially as important as the model, especially in an agentic world. The harness and the model need to be developed together.

The release of Grok Build and the pace at which they’re iterating is, I think, also really encouraging. Now you have Cursor, you have the Cursor data, and you have a clear existence proof that the Cursor data is really important. Composer 2.5 is now Pareto dominant and the most selected model on Cursor, and that’s also important because these evals don’t capture everything. This is why people on X talk about the vibes, and the vibes on Cursor 2.5 are also really good.

They’re immaculate, yeah.

That together—

Yeah.

— with Grok Build, I think these are really important developments.

Yeah, I think Elon was incredibly frustrated by the state of affairs at xAI. He was very public about that, and he’s less frustrated now. He’s shipping a lot faster, and so I think that says something. He has been very focused on it.

Friedberg, your thoughts on the SpaceX IPO and what this collection of companies might look like a year or 2 from now, especially if, like many people believe, Tesla and SpaceX merge? What do you think of $ELON as an entity, and what impact might it have? If those 2 were put together, the market cap would make it the 4th-largest company in the world.

We can revisit our earlier conversation about an anti-tech, anti-AI, anti-progress world and society ahead. If there is a concerted, organized effort by governments to stop or block access to information, restrict freedom of speech, restrict freedom of purchasing or buying things, and control more things, I think there’s a trend line in this direction right now globally.

The internet has always been lauded as this kind of system that provides an open alternative to physical commerce. You could create digital commerce, digital information, and digital media that you could share. It’s almost this digital representation of society, but the internet has to sit physically somewhere.

The assault on data center build-outs in the United States right now may indicate the importance of having an alternative internet from the ground layer up. If you have a communication network that isn’t restricted and controlled by a government on Earth, it’s almost like a backup for civilization, but it’s a backup for progress.

I don’t own any SpaceX shares, and I’m not trying to sell the bull case for SpaceX, but I think there’s an important aspect to whether you can create a system that’s not under the control of governments as a way to ensure humanity’s progress and civilizational continuity if things go south—if things aren’t good, if things are restricted, and if fundamental forms of tyranny start to restrict speech, commerce, information flow, and whatnot.

I think having a space-based communication network, space-based data centers, and space-based communication back down to Earth wirelessly is generally a good thing. It’s good to have a backup.

Most people don’t remember this, but when Elon was starting SpaceX, the original idea, when he was running around with the DAO and they were looking at some rockets and getting carriage from Russian rockets, was to back up the biosphere. He came back from that trip, and I remember talking to him about it. He said, “I think I just have to make my own rockets because that’s actually where the problem is, and it would be easier just to make my own rocket to back up the biosphere.”

He wanted to put geodomes—geodesic domes—in space with all the plants, wildlife, and creatures. What an incredible vision. Then there was the necessity of actually getting that up into space, and that’s the unknown origin story.

I will say this, Chamath: the idea of putting data centers in space seems completely doable, even though there are a bunch of people who are saying it’s not, when you compare it to what happened with SpaceX and Starlink, which people said also wouldn’t work. Now he’s got 10,000 Starlinks up there. The difference between a Starlink satellite and a data center satellite is really not that different.

No, they’re pretty different.

Well, conceptually, yeah. Of course they’re physically different, but conceptually, Elon put 10,000 Starlinks up. Is he capable of putting 10,000—

No, look, the size is much—the size is much bigger, Jason. The foils are much bigger, and the wings are much bigger.

But my point is, it’s not different if he has the new Starship, because that’s 10 times bigger.

You can’t just scale like this. That being said, it’s technically possible. I think he will be the first one to figure it out. But I’ll just take a much more pedestrian take, which is: if I’m asking myself, Chamath, how do I underwrite SpaceX at $2 trillion? Here’s the basic math that I would do.

Last year it did $18–19 billion. It’ll probably do $25–30 billion this year. Okay, so I’m buying this thing at a fairly costly premium, right? What am I buying? I’m buying probably the most important internet infrastructure project that’s happened since the internet itself. That’s going to scale to hundreds of millions of users, and the reason that’s going to scale to hundreds of millions of users is that it’s just very useful and it’s just going to become cheaper and cheaper and cheaper.

So that’s number 1. I’m buying a delivery infrastructure.

But I think over time, GDP plus 10, GDP plus 15—kind of a grower. So, good business, valuable business, but it’s the underlying platform that allows everything else to happen. And then I’m buying an AI business, which will be, at the top level, the apps, but at the bottom layer, all the compute capability.

And I think when you scale that out, why is Colossus so valuable to Anthropic? Maybe that’s a good question to ask. It’s because, if you look at who’s actually capable of delivering a gigawatt data center, these guys are the closest. Like, an actual gigawatt. And the reason is that this stuff is very complicated and very, very hard.

I think you’ve probably heard this famous story where Jensen was like, “Yeah, he was the one who figured out this one thing that nobody else could figure out, so that you could string a bunch of racks and drive a bunch of east-west traffic and make the whole thing work together.” So, I suspect what happens is next year it’s probably $40 to $45 billion. And then the year after that, it probably doubles again.

So now I’m buying it at 20 times revenue. And you would say, “Well, why can you buy a company like this on revenue versus earnings and cash flow?” I think the reason is because what the revenue does is give him the operating leverage to go and invest in all of these other businesses that ultimately consolidate his differentiation and his competitive moat.

Because what he creates is a capital moat that then accelerates a technology moat, that then accelerates an execution and a learning moat. And that flywheel, when it starts to spin, spins very quickly. And you would say, “Hey, hold on a second. It’s probably spinning quickly now.” I would say we’re at the beginning of the beginning.

Because he still has all these disparate assets. I still don’t like the fact that Tesla’s over here. And as I’ve told you, that will get merged in. And now you have this incredible corpus of physical capability, movement of all kinds, X, Y, and Z. You have learning capability. You have infrastructure. You have all the connectivity.

That thing will look very cheap, I think, in a few years. And he has this one thing that nobody else has. If you look at the big CEOs, who steps on stage where you’re always curious, “Okay, what does he have up his sleeve?” You know, the Steve Jobs, “Oh, and one more thing.” This is the only guy at the scale of civilization, out of left field. He’s the guy. Whether you like him or hate him, he’s the guy. And there’s a premium that is well-deserved that comes with that.

So, if you had to pick an underwriting case, Jason, I would flex the revenue and realize that terrestrial data centers alone are $100 or $200 billion of revenue by 2030 or 2032. And that means just building it. Already, you’re buying it at 20 times revenue just for that business. Everything else is gravy.

On the ground, earth-based.

It’ll be Colossus 3. No, no, forget space for a second. It’s Colossus 3, Colossus 4.

It pencils out with that, yes.

Look, getting a 1-gigawatt nameplate working is freaking hard, man. And then, by the way, there’s all the stuff that he can do on land that he’s the best positioned to do.

I’ll give you one example. There’s a great push that Jensen’s making, which he needs a partner for, and I think Elon becomes a natural partner, to do DC to DC. Forget all this DC to AC to DC nonsense that goes inside of a data center. All the losses.

That is in English, yeah, for everybody.

Just look, you go through a bunch of power transformations to actually deliver the electrons into the rack so that, as Gavin said, you can generate the token on the other end. Today, it’s very inefficient, it’s very costly, it requires a lot more power, it requires a lot of cooling, and it requires complexity.

What people have said is, “Wow, if we could just do DC to DC—if it comes in as DC, direct current, it goes right to the rack as DC.” But it requires a fundamental rearchitecture. Jensen needs a design partner and a thought partner to get that done. He’s probably the only one.

So, I just think there are a lot of reasons why you can underwrite this to a multiple of revenue, plus the X factor, which is just the creativity and the one-more-thing factor.

Love it. And then here are 2 charts, and I’ll have you comment on these, Gavin, afterward. Here are the rocket sizes. Just in terms of scale, most people have not actually seen a Starship in person. When you see this thing in person—and I’ve been inside that rocket—I think you were with us, Gavin, when we were in the first build. Inside of that, you can fit 300 people.

If you thought of a commercial aircraft, that’s what it feels like when you’re inside, right? Like a 747 in terms of the amount of space in it. Especially when you compare it to the Falcon Heavy, which is their workhorse, correct, Gavin?

Yeah, and Starship’s going to get bigger. Based on their roadmap, it’s going to get a lot bigger.

A lot bigger. And then this one is the most interesting, and it started trending this week. This is cumulative payloads launched from 1957 to today. This is really what exponential growth is about, and this is what disruptive technologies are about. Just from 2012 to today, SpaceX is about to dwarf the rest of the world’s cumulative payloads into space.

So, Gavin, maybe take the other side of it. When do these data centers in space happen? What has to happen for those to be a reality? When does that hit SpaceX’s bottom line? We’ve heard from Chamath, “Hey, here are all the things that hit the bottom line in the short term and midterm.” But I think data centers in space would be a midterm-to-long-term play. 3 years is what I’m hearing. So, tell us about that business in relation to the 2 charts I just shared.

Well, the one thing I would just say—first, all those charts about launch are before Starship was operational, and most of that mass to orbit was done by Falcon. The Falcon is reusable. Starship is designed to be rapidly reusable. And this is a critical difference.

Let’s say Blue Origin successfully solves reusability. They’re where SpaceX was 10 years ago. Let’s say China solves it 10 years from now. Reusability means that you extensively refurbish the rocket—the engines, everything, the fairing. It takes a lot of time. Maybe you can fly that rocket again in 30 days or 60 days.

Rapid reusability means that you can land the same rocket multiple times per day. I think it would have been not trivial, but much easier to have Starships working if it were just designed to be reusable.

Mhm.

That’s not enough for what Elon wants to achieve: a moon base, a colony on the Moon, a colony on Mars, and mass drivers on the Moon. You need rapid reusability, and that is why Starship is such an engineering challenge and will be such an impressive achievement when they have rapid reusability.

When do you predict they’ll have rapid reusability to space, do you think?

We’re going to find out. I’m going to be at Starbase today for the whole launch.

Yeah.

Gavin McCready

So, we turn over cards, and it’s important for everyone to remember: Let’s just say it’s a fireball. SpaceX will still learn from this. They learn from failure. If you don’t fail, you’re not learning. In the same way, if you’re not wrong, you didn’t learn anything that day.

This is a brand-new rocket, a brand-new booster, and a lot of new technology. There’s a lot of instrumentation on it. Whatever happens today, SpaceX is going to learn and rapidly iterate. I don’t know when, and I don’t want to make a prediction. I would guess a year or 2. Maybe sooner.

I think that’s the consensus. A year or 2 is, I think, perfect consensus, yeah.

Gavin McCready

We’ll see. Even if everybody else solves reusability, mass to orbit from everyone else will quickly asymptote to a very small number.

As far as when orbital compute will be a reality, I would say it’s important to realize there is a working H100, an NVIDIA H100 GPU, in space today. Andrej Karpathy both trained a model on it and used it for inference. There’s a working GPU in space today.

And NVIDIA’s making a space-designed version of this, which will be different because the heat sink has to be different. There’s a bunch of weight that you put on it when it’s in a data center that you don’t need in space. And you also have to reinforce it for the journey to space because these things are going to shake and break apart.

The data center ones are not made to have that many Gs put on them. So, you’re going to need an industrial-strength one that gets to space and has a different profile. Yeah, Gavin?

Well, one of the things that’s been so magical about SpaceX is they’re very good at engineering the rocket and the payload so that you can use semiconductors that are not designed for space or satellites in space, and those semiconductors are a lot cheaper.

My firm, Seraphim Space, is an investor in a company called Excite Labs that, as a matter of public record, is going to be in essentially every Starlink. The chips were not designed to go to space. They’re not radiation-hardened.

And then it’s like, well, we’ll see how they do with radiation testing, and they just so happen to pass. That is one of something that’s very underappreciated, I think, about space.

One of their specialties.

Gavin McCready

Yeah. But I think the second half of 2028 to the first half of 2030 would be my point prediction.

All right, let’s do Nvidia and then the market recap, since we have you here, Gavin, and since Friedberg, you wanted to get in on that. Nvidia blew out its earnings again. Q1 performance is just mind-boggling: $81.6 billion in revenue, up 85% year-over-year and 20% quarter-over-quarter. High growth in the stock market, for those people who don’t participate, would be 20% for a high-growth company year-over-year. They did that quarter-over-quarter.

They had $58 billion of net income and $48 billion in free cash flow. They’re doing all this at 75% gross margins. They’re growing massively, and they’re obviously the most valuable company in the world, at a $5.3 trillion market cap. The stock’s up about 16% this year with all that growth. That’s a magnitude of that 16%.

They’ve announced another $80 billion in additional buybacks on top of the $100 billion in buybacks they did at the start of 2023. So they’re buying back about 4% of the company. They raised the quarterly dividend 25 times, from 1 cent a share to 25 cents per share. Their CFO said they’re going to return 50% of the free cash flow to shareholders. Never been a company like this, huh, Friedberg?

Mhm.

Yep. Don’t seem to—

They have—first, there’s your market report from Friedberg. You don’t seem so enthused.

Yeah. He’s got potatoes in the oven.

I have a question for Gavin. He did a really interesting talk with Patrick O’Shaughnessy, and there was this one thing that I wanted to ask you about because I thought it was so interesting. You said when you look at the revenue multiples of the chip companies and you look at the revenue multiples of the DRAM companies, both cannot be true.

In the context of Nvidia’s earnings, can you just explain that in plain language for folks? I thought it was so fascinating because it explains, I think—just to set it up—where is value over the next 5 years? I think if you looked at Leopold Aschenbrenner, his fund has gone from $0 to $5 billion overnight, and it looks like he’s just got massive puts on the chip sector and has kind of rotated. So just give us context, Gavin. Where’s the puck going?

Well, so maybe I’ll take the question in reverse order. For Leopold, who’s clearly a brilliant man—I think he was a Rhodes Scholar at 19—and my understanding is that he’s putting up pretty extraordinary numbers, I’ve yet to meet him. He actually shares an office in San Francisco with a friend of mine, so I think I’ll probably meet him sometime soon.

That 13F that he filed was at the end of the first quarter, when I would say we were in the thick of geopolitical fears. I think you saw a lot of puts on a lot of 13Fs, and I don’t know that those puts are still there.

Okay.

Gavin McCready

I think a lot of people wanted to be hedged for it, and now I think it’s a little more clear. So I wouldn’t read Leopold’s 13F as being super negative on semis.

Chips, okay.

On chips. Second thing, I think cross-sectionally, if you look at the valuations for all these AI names, they can’t all be accurate. You have memory makers at 3 to 5 times P/E, you have Nvidia at a really low P/E, and you actually have some other accelerator companies at reasonable multiples.

Then you have everything else: everything in power, everything in cooling. When I say power, I don’t mean utilities. IPPs are actually quite reasonably valued. But power, cooling, and even probably some of the optical names are discounting very different things.

If the multiples on the power, cooling, and optical names are correct, Nvidia and memory are going up a lot. If the multiples on Nvidia and memory are correct, everything else is probably going to underperform. The AI market is cross-sectionally inefficient right now, which is what I was trying to say.

As far as the Nvidia quarter, they went to a new reporting structure, data center and AI, and then with no data center and edge, and then within AI they have hyperscalers, and then I think they call it AI clouds, industrial, and enterprise.

What I believe is that, if we were to make a true apples-to-apples comparison—and Broadcom, you know, there’s a narrative that Nvidia is losing share to the TPU—Broadcom guided for 143% year-over-year growth in its AI semiconductor revenue in the quarter that it will report. That’s comparable to the one Nvidia just reported.

I so wish they had reported slightly differently. I wish they’d done hyperscalers, AI clouds, and then industrial and enterprise.

I think the segment that is comparable is the sum of hyperscalers plus AI clouds, stripping out China, because Broadcom just did not have the China business that Nvidia did. And I think on that basis—in other words, in the Western AI world, within data centers that are being built, whether they’re being built by CoreWeave, xAI, Amazon, or Google—Nvidia’s AI business is growing faster than Broadcom’s, and faster than a lot of other companies that are seen as part of this ASIC share-gain story.

He’s become very vocal about that. What is going on here? You’re putting up record numbers, and we’re getting no credit.

Yeah, I get it. How can there be a share-loss narrative if I’m gaining share? It is indisputably true that he’s growing faster than hyperscaler CapEx, even without these adjustments.

Yeah.

And I think the other thing that’s so frustrating to him is that these other ASICs—

Mhm.

—are not being submitted for benchmarks. They’re not in SemiAnalysis’s InferenceMAX. They’re not in MLPerf. And I think the reason they’re not being submitted is that they will lose.

Mhm.

And you can’t fight shadows. Until we see a clean benchmark of whether it’s GB300s versus TPU v7s or, you know, versus—

Versus Inferentia, yeah. Yeah, yeah.

—versus Trainium, we’re not going to know. And that’s why a lot of these other chips—

I think Trainium’s in a great spot.

—aren’t being submitted. But nonetheless, if Nvidia’s doing well, once you become the largest company in the world, you tend to trade by observation, which would be at stair-step patterns. The multiple compresses, compresses, compresses because people are skeptical of the size.

Then you have a re-rating.

And then you re-rate.

A new floor is established at a higher rate.

I think there was one other really important thing in the Nvidia quarter: they said that they thought their CPU business was going to be $20 billion this year.

Yeah, yeah, yeah.

Put it there.

It means overnight you’re one of the world’s largest CPU manufacturers. I think that is a testament to Nvidia’s unique position. They’re the only company that works with every lab, and so that puts them in the best position to architect their chips—they call it co-design—for where the models are going. I think that $20 billion CPU figure is pretty extraordinary.

This is the thing: at the end of this Groq transaction last year, my prevailing thought was that we were going to move to these domain-specific architectures. I thought that was a fait accompli; we were just waiting to see which models. But the reality is that this DSA market evolution is actually happening inside of Nvidia. That’s what’s so insane to me.

That was my takeaway from the quarter as well, which is, holy—these guys actually have domain-specific architectures because they’re doing these design programs with everyone. This is why, going back to when he does DC-to-DC with Elon and Colossus 3 or whatever, it’s another game-changer for everybody.

He doesn’t get a 1-year flip. He makes 9. And then I don’t think the cost at which you can finance these chips and these useful lives is really important.

You had an incredible insight, which is the amortization schedule for CoreWeave and all these guys. You may want to just explain what that is and why you said that. I thought that was a great insight.

No, thank you, Chamath. I appreciate it. When CoreWeave and all these neoclouds came public—and by the way, this goes for the hyperscalers too—there was a big bear case that these guys were amortizing their GPUs and CPUs over 4, 5, or 6 years, and that’s way too long of a lifespan. The true lifespan of a GPU is more like 2 years, and therefore the profits of all these businesses are overstated. The reality is—

This was Michael Burry who put this out there.

Yes, to be clear. And thank you, Michael Burry. We need bears. Thank you.

Yeah, that’s like asking Gerardo about modern music.

Well, I don’t want to cast aspersions on Michael Burry. He’s a brilliant man, but we need bears.

Hey, somebody call Vanilla Ice and ask him what he thinks.

Oh my God. Milli Vanilli, check your mail.

What a waste of time. What a waste of time.

Oh, poor Michael. Come on the program anytime, Michael Burry. Go ahead, Gavin.

Yeah, I know. Keep it on track. I'm happy to chat. But now that we've disaggregated inference, we have these different domain-specific accelerators. You can mix and match them, and I think the GPU stays, in a lot of ways, at the center of this constellation for a while. You can put a Groq accelerator or a Cerebras accelerator in front of old GPUs, and use Groq or Cerebras for decode.

Those older GPUs have a useful life of 10 or 15 years. This means that you can finance GPUs. I think CoreWeave's lowest financing—I can't remember if it's 6% or 7%.

6%. 6%.

Yeah. And if you can get an asset-backed loan, asset-backed financing for GPUs, at a lower rate than other chips—

No, that's a profound advantage.

That quarter single-handedly saved the neo-clouds. This quarter, Nvidia saved—I mean, they single-handedly saved them all. I think they should all say an incredible thanks to Jensen, because I interviewed the CEO of CoreWeave, Michael Intrator, and he was saying, “Hey, listen, people have no problem buying and financing these over a 6-year period.”

People are asking for things that are coming off, and he thinks they're going to have years 7, 8, and 9—they'll have some useful life in addition to that. So he's like, “I don't know what anybody's talking about here. This is just not informed analysis,” was his point. The game is on the field, and people are betting with their dollars with him.

He has them pay in advance and sign 6-year contracts. If they didn't think it had a 6-year lifespan, they wouldn't be signing a 6-year contract. Pretty straightforward. And they can't get enough of them.

Okay, let's end on this market update. The macro picture is not great. Oil remains elevated, although there might be a settlement. Every week, there is a settlement coming. Maybe this time, 16th time's the charm, and the Iran war is going to wrap up, but we're in week 12 of it. This was supposed to be 4 to 6 weeks.

Wars never get resolved quickly. That's one thing we've learned in our lifetimes. Oil is driving inflation massively higher. Polymarket says there's a 99% chance May inflation comes in at 4.2% or higher. The Survey of Professional Forecasters is projecting CPI hits 6%. You heard that right, folks. We weren't just talking about a 3% handle, which we just said.

Now people are saying 4%, 5%, and 6% in Q2. Obviously, that's a huge revision. The narrative was, “Hey, more Fed rate cuts coming.” Now we're talking about Fed rate increases. Inflation is causing, obviously, bond yields to rise. The 10-year hit 4.6%.

You remember we've had Scott Bessent on the pod multiple times, and his goal was to get that under 4%. Now it's significantly above that number. Also, if we go around internationally, Japan's 30-year is at a high of 3.1%, the highest ever recorded. UK yields are the highest since the Great Financial Crisis. Germany's are the highest since 2011. In Korea, retail investors are borrowing record amounts of money to trade in AI chip stocks. They also had an incredible run in Korea betting on crypto at the peak. So that's some sort of interesting signal.

Friedberg, is your alternate personality going to come out right now? Are you concerned? Is Dr. Doom making an appearance here, or do you think this is manageable? How concerned are you?

What is the point of being concerned when you have ridden the roller coaster to the top and it is beginning its descent? I don't know what there is to be concerned about. The force of gravity is inevitable. The roller coaster will roll down. We will throw our hands in the air and scream, “Whee!” as we go for the ride.

Global debt to GDP is 310%. Reserve-currency status aside, the spending problem at the federal, state, and local level, and the spending problem in every country to basically keep economies growing and support existing leverage, ultimately create a cascading effect. It ultimately breaks. As it starts to break, you have massive inflation because the value of your underlying currency collapses, and then you have money printing and all this other sort of stuff, which inflates the value of assets and allows you to keep servicing your debt. The spiral takes off. So we will just enjoy the ride.

The 30-year at 5.2%—this Japanese yield, some argue, might be one of those things that could be a catalyst for a credit crisis. You should talk to more active market participants than me, and probably some economists who trade the market, but I would think that this is one of those things that could be a catalyst for a credit crisis. There's a lot of people that are in this carry trade. And we'll see.

Okay.

This is water leaking out of the bucket.

There it is. Dr. Doom is here. Chamath, your thoughts on Dr. Doom's panic attack of the month. Is this time real? Is this the 17th prediction of the next 6 recessions? What do we have here, Chamath? Are you concerned? How concerned are you about these signals that are flashing?

I think that's exactly what that is, Jason. There are signals that are flashing. I think there are pockets of the market that still make sense, that you can underwrite if you want to buy businesses that represent the future. If you can find a few of those, get comfortable with them, and own them for 10 years, I think you buy those companies, and generally everything else, I think you should not speculate and should generally avoid. Not just because it's an up market, but in every market.

I've learned this the hard way. We've all kind of gone through this. As I get older, it's just not worth it. The vicissitudes of the market don't give me anywhere near the sugar high they used to on the way up, and it makes me feel horrible on the way down. So how I manage myself is, I have a few companies that I really believe in. I have extremely concentrated, large holdings in those.

Large for me doesn't mean large for everybody else. And otherwise, I just kind of stick to my knitting and keep my head down. It's a much more rational way to behave.

How many public stocks can you keep in your brain and still sleep at night while holding for the long term, Chamath? What's the number for you? Is it 5? Is it 10?

7. It's 5—it's 5 or less.

5 or less?

5 or less.

And so, what percentage of your net worth is your largest holding right now, which is a—

I don't know.

Or, like, top 2, maybe.

Yeah. Oh, top 2?

Yeah, like 1 and 2. 1 is 20%, 2 is 15%, or 1 is 40%, 2 is 20%.

I don't know. Again, it depends on the day, but I don't know.

Just curious. But I think that's really important.

My point is, there aren't 30 things that I'm tracking. I don't have the time. I'm not smart enough. There's too much information. There are about 4 things that I stay on top of.

Gavin, you do this for a living. How many positions do you manage, and what's your take on some of the flashing signs that are saying, “Hey, slow down,” or maybe there might be a wreck around the corner here? You know, when they do the checkered flag in F1, or whatever metaphor you want to use.

Well, I manage more than 100 positions at my firm. I do that with a team. We're over 30 people now, so it's not just me, and I work with some great people. I think 3 things can be true. Rates going up is very concerning. What is happening with AI right now, with Anthropic growing faster than any company in history at massive scale, is absolutely unprecedented.

And certainly any country.

Absolutely unprecedented. Yes.

They're bigger than 100 different countries.

For sure.

Exactly.

And compounding really fast.

And now profitable.

Yeah.

Which I think really changes the moment.

Which is bizarre.

Yeah, how did that happen?

So those things can both be true, and I think we should all remember the tech bubble happened with the 10-year and the 30-year much higher than they are today. The Nvidia of the tech bubble was Cisco, and it traded at 100 times forward earnings. I think Nvidia is probably at a low-teens multiple, kind of a low- to mid-teens multiple, of real earnings, if you take that as a buy-side consensus, not the trade number.

The third thing that I think is true is the Strait of Hormuz is being closed. While it's terrible for everyone, it is relatively the best for America because we are self-sufficient in energy, we are self-sufficient in food, we've become a massive exporter of oil, and we're now the world's largest not only oil and gas producer but also oil and gas exporter. Those 3 things can be true. What do they mean? I think it's probably hard for me to see, with America being the most advantaged by what is going on.

We're still the best currency.

Yeah.

The best economy. We still have the best public companies, and we have the best private-market companies.

Yeah.

And we are one of the greatest producers of oil in the world. So we're in good shape despite this international chaos.

I don't think a dollar crisis is around the corner. Now listen, if the Bundesbank was still in charge and you had the Deutsche Mark, and there was a currency with better fundamentals, we would be at very high risk, probably. But just because we're the best house in what is globally a bad neighborhood of high debt levels, and we have AI in our corner, and we have energy self-sufficiency, and every day the Strait of Hormuz is closed, I think it's relatively good for the reindustrialization of America.

You're saying it's a forcing function. It makes us, just like COVID did, more resilient and more self-reliant.

Yes, and electricity is a base input to every manufacturing or industrial process—essentially all of them.

Mm-hmm.

And what we make electricity with in America, overwhelmingly, is natural gas. You can look it up: NG1 is down this year.

Mm-hmm.

The input cost for electricity in the rest of the world is a lot of different things, but LNG is a very important one, and it's up 100–200%. The Strait of Hormuz is absolutely bad for everyone, but relatively good for America and relatively good for Trump's policy goals.

Mm-hmm.

And that's why I think he's in no hurry. For whatever reason, he does seem like a relative thinker. Every day the Strait of Hormuz is closed is relatively good for America. It's terrible for Europe. It is terrible for Asia.

Japan and China need that oil. The Philippines needs it. India needs it.

Yeah, so all these things can be true, but the one thing I do want to say is that rates going up and inflation going up is never good. But we have to hold what's happening with AI, where the fundamentals are getting a lot stronger, in our minds. The one thing I would just add is that AI has been seasonal, and the market's seasonal. It often follows “sell in May, go away.”

AI fundamentals also appear a little seasonal. In the past, that's been because college students use a lot less ChatGPT and Claude in the summer. Generally, people maybe work a little less hard when the weather's nice. Now, with generative AI, will the fundamentals still be seasonal? We will see.

Oh, that's a really interesting point. We see that in e-commerce, apps, and subscriptions. As investors in a lot of these companies, we would always have these board meetings: “Chamath, Q3—yeah, people are out gallivanting, and they're not playing Candy Crush or buying Calm or whatever. But, hey, Uber and DoorDash went up. People are traveling,” et cetera.

Okay, final story of the week. We had a 48-hour jaunt by a bunch of tech CEOs and the president to hang out with Xi. A lot of high fives, a lot of handshakes, and a lot of great vibes, but coming out of it, we haven't seen anything definitive, Friedberg, in terms of policy. This was supposed to be some big breakthrough. It would have downstream effects on tariffs and on us selling chips to China.

We did see a little movement there, and on the Strait of Hormuz, we would become Wonder Twins with Xi and Trump reopening it, but nothing really definitive other than some soybeans being sold and some H100s, A200s getting sold to Baidu and some of the top folks. Maybe some planes got sold, too. So, other than a little BD, a little business development, Friedberg, what's the outcome here? Or was it just a bit performative in your mind?

There was a question of whether the administration would leave China with a grand deal that made everyone feel like there was a long-range view on a partnership, but I don't think that's what happened. There were a few announcements, obviously, around an intention to continue to work together in a cooperative way and find a path to partnership, an intention to establish additional trade deals, and there were some purchases of aircraft and some agricultural product commitments.

Fundamentally, the grand deal—the big deal that would reduce or de-escalate tension—probably didn't manifest as some had hoped it would. I don't think it's any surprise that Putin is with Xi today. This is also performative: following the U.S. visit, there's now a relationship-bonding moment happening between China and Russia.

The story continues. There is no happy ending, and there is no rainbow-colored chapter 3 in this book. It's going to continue to be a dramatic arc as this rising power continues to challenge the United States. I think the story continues.

Were you expecting a happy ending?

Can't answer that question.

I mean, to the story of the China visit—not anything else going on in your life—but it seemed like some planes and soybeans got sold, and some H100s, perhaps. It wasn't like there was some grand deal that occurred, but it's nice to see them together, right? I mean, that is nice.

I think it was successful. I think there's what you see on the surface, and then there's what happens behind closed doors. Without speculating too much, I think that it was a useful and productive trip. I think the biggest thing that they probably got alignment on is, geopolitically, the tic-tac-toe of what has to happen next. I think there's some amount of agreement there. I'm just guessing.

Yeah, and so that guess, if I was going to unpack it: We have Venezuela, we have Iran, and Taiwan. You have lots of things to—

Here's the geopolitical chessboard, and here's what I would say, just very generally: I think that there's a way to divide up the game board in a way that helps them and helps us.

Mm-hmm.

Gavin, any thoughts on specifically Nvidia being able to sell more chips into China? Is it material for the company? Good for America? It's obviously a pretty debatable issue.

I think I disagree with Chamath on this. I think selling depreciated Nvidia GPUs to China lowers the odds of them developing their own alternative ecosystem, which would be a lot more power-hungry. Because you use optical, you bring in optical a lot earlier for scale-up fabrics. I think there are sound arguments that this is stabilizing for the world and is the best, highest-probability path for keeping America ahead in AI and kind of keeping control of AI.

That's almost a shame we've had to have this debate, because people like me have said this many times. If they didn't understand it, they probably do really understand it now. By the way, that's not to say we shouldn't have had the debate. But that is what I believe. Reasonable minds can disagree.

No, wait. Where do you think we disagree? I'm not sure I agree with you.

Oh, good. I'm glad we—

I was going to jump in. No, no, you start. You start with, “I'm in agreement on—”

No, no, no. I'm like, “Sell—”

I'm like, “Sell everything to them.” No, what I was just saying is that there's what you see on the surface of what they can speak to the press, but I think the most important thing was the negotiation of, “Hey, listen, we're going to do these things; you do these other things.” That's never going to get put out in a multi— you know, memoed press release. That's my point. That's all I'm saying.

Yeah, and I would just say, listen, America and China talking is only good.

Only good.

The Thucydides Trap that is being discussed—China talks about it a lot. They're very aware of it, and they brought it up.

Xi brought it up by name. Talking is an integral step toward avoiding this situation and having a nice resolution. From my perspective, the greatest superpower Trump has is his ability to bond with dictators, monarchs, royal families, and Gulf monarchies. He's just great at it. They see eye to eye. They vibe. He has no problem going to see them. He has no problem inviting them to UFC fights.

If Xi comes to the United States and is sitting courtside with Dana White, that's when we know things are going to be okay. I do think he's probably giving them the green light on, “Hey, Taiwan's yours. Just let's not have it during my administration. Maybe we do a 30-year deal or a 20-year handoff deal.” I wouldn't be surprised if something like that happens. Or a 100-year deal or a 200-year deal.

But the one thing I would just say that I'm sure was communicated is, “Hey, wars consume a vast amount of oil. You buy your oil from Iran, Venezuela, and Russia. Russia alone can supply a fraction of what you need. And now it should be clear to the world: Two of the three are off the chessboard.”

If you do something we don't like, Venezuelan oil is gone for you. American oil is gone. Brazilian oil is gone. And we'll say to all of our good friends in the GCC, “We're so sorry, but we have to close the Strait of Hormuz again.” So Iran—all of Middle Eastern oil—is gone for China. Now you just have Russia, and good luck fighting a war with just Russian oil against us, Japan, South Korea, Australia, the U.K., France, Germany—the world. I mean, who knows about Europe? But for sure Japan would be there. For sure Japan would be there. Oh, and Australia for sure, Korea for sure.

And so I think it's going to be a more stable world on the other side of Iran, however it resolves, and I think that's nothing but good.

Yeah, I think that's a good insight. I think it's a great insight. All right, listen. We missed you, Sacks. Come back soon. And Gavin.

Big shout-out to bestie Gavin. Thank you.

Yeah, you're so great. Thanks for coming. We appreciate it. Hey, and your father-in-law—what's his name again?

Jeff Painter.

We love you. Thank you so much for all the kind words. We'd love to invite you to come to Liquidity or the Summit. I know you're a big fan of the show. We want to give you a shout-out here on the show.

I used your fandom of the show to leverage Gavin, who was like, “I can't make it today.” And I was like, “Tell your father-in-law we're going to get him backstage VIP tickets to the next 2 events if you show up today. And if you don't, I'm going to back-channel it to him.” And all of a sudden, Gavin made it to the show.

Love you boys. A little bit of pressure. That's my way of keeping my hormones straight.

Yeah, so I'm always there for my father-in-law.

Absolutely.