Anthropic 联手 SpaceX AI、Leopold 的55亿美元押注与奇点经济｜EP #255

Anthropic is taking over all of SpaceX’s Colossus 1 data center in Memphis. This immediately allowed Anthropic to double Claude Code rate limits.

Alex

I think Grok is on life support.

This is Elon, who had been, for the past year, shit-talking Anthropic. Here he is now, backing them and supporting them. In one way, Elon is getting revenge against OpenAI by helping Anthropic win. “The enemy of my enemy is my friend” is the exact quote from Elon.

Leopold Aschenbrenner, who was famously fired from OpenAI’s alignment team, is now running a $5.5 billion fund 2 years later. Anthropic hits 80% growth for this quarter and outruns its compute.

Alexandr Wang

You can find a whole litany of things that are about to explode in demand. The demand—I don’t see it slowing down as a whole. Chips, the energy layer, and the infrastructure. This is the singularity loop.

Let’s see your teeth. I want to see which ones are fake.

David Friedberg

I just landed back from Montreal. I just landed back—I just landed back from Montreal, and I was given a Canadiens jersey because I’m actually a Montreal native and a massive Montreal Canadiens fan. I grew up there, so I’m a Montreal Canadiens fan.

Like a small, round ball, or a puck—that’s the hard, round thing that immigrants like me never quite managed with our ankles. In Canada, you had to skate; otherwise, they took your passport away. So I used to play a bit of hockey.

We’ve got to do some pond hockey up in Vermont. It’s an all-time favorite thing to do.

David Friedberg

The folks gave me this jersey, so I had to wear it because we’re right in the middle of all this stuff. Anybody in Buffalo, I’m rooting for the Canadiens, but I’m the longest-suffering Bills fan in history. So there’s that.

You know, I’ve missed the complete gene sequence on sports. Sorry.

David Friedberg

Today we have an incredible show. We’re going to be kicking it off with the demand for AI, which is off the rails and outstripping supply. Claude is continuing to disrupt sector after sector with the Great Unhobbling. Google is joining our push toward Earth’s Dyson swarm.

We’ll be jumping into the singularity economy. What are the sectors that are providing outsized financial returns during this supersonic tsunami? We’re going to cover a topic near and dear to my heart, which is a very simple, very powerful concept to ensure AI alignment so we can deliver a p(doom) that is less than zero. Thank you, Alex, for that meme. I love that: p(doom) less than zero.

Alexandr Wang

We need T-shirts for it.

We do. I saw the T-shirt you made. We’re going to have to put that up for people to be able to take down.

Alexandr Wang

Peter, if no one makes it, everyone dies.

Toward the end of the podcast today, we’re going to be talking about the recent disclosures by the White House on, let’s call them, UFOs versus UAPs. The White House is saying they’re here, but who are they? Where are they? Where are they from? And when are you going to come and pick me up and take me for a ride? That’s going to be some fun conversation today.

Let’s begin with an important conversation. Anthropic is outpacing its ability to supply tokens. Anthropic hits 80% growth for this quarter and outruns its compute.

At Anthropic’s developer conference last week, CEO Dario Amodei revealed that Anthropic has experienced 80fold growth in Q1, outpacing what they expected, which was 10x growth. You don’t see this in Silicon Valley. You don’t see this anywhere. Maybe, Dave, for some of your early companies, you’re seeing that kind of growth.

Their annualized revenue run rate jumped from $9 billion at the end of 2025 to $30 billion in April. It’s now north of, I think, $40 billion in May. Here are the numbers: It’s expected they could hit $100 billion of ARR by the end of 2026 and potentially $1 trillion by the end or mid-2027, making it the most valuable company on the planet.

Just to hit some numbers real quick and then turn it over to you, Dave: At a $30 billion ARR and a 40× multiple, it’s being valued today at $1.2 trillion. If they hit $100 billion by the end of this year, Anthropic will be at a $4 trillion valuation. If they hit a $1 trillion ARR in 2027, that’s a $40 trillion valuation. This is the singularity by definition. Insane.

David Friedberg

Over a year ago, Peter, at that family office conference in London, Yeang was saying to all these wealthy families, “You’ve got to get invested in this. You’ve got to get in the game.” Everybody was like, “At a $100 billion-plus valuation? That’s utterly insane. You can’t possibly. It’s way too late. You can’t get into it.”

I don’t blame them because these numbers are so unprecedented, and people don’t do a good job of judging millions, billions, and trillions. It’s not intuitive. These numbers are so far out of the realm of history—massively bigger than any prior IPO or valuation.

I don’t blame people for being scared, but you’ve got to get used to the fact that this goes to infinity.

Yeah.

David Friedberg

The demand for AI is not going to saturate. It goes to infinity. You’ve got to rethink the way you decide whether to be involved in these things or not. They all skipped it.

Now I’m sure they regret it.

David Friedberg

Yeah.

Yeah. Ten times in a year, up to $1.2 trillion. Alexandr, what do you make of all this?

Alexandr Wang

It’s all about the enterprise, Peter. Anthropic was the first, arguably, major frontier lab to recognize that offering ultra-high enterprise-oriented-value tokens was the path to success here.

As I’ve mentioned on the pod previously, OpenAI has since had to pivot to copy the Anthropic strategy of offering high-grade enterprise tokens for code generation and now for other so-called white-collar tasks. This is what we’re seeing: an insatiable demand for compute, to the extent that compute can be turned into high-dollar-value tokens that are replacing the services economy.

U.S. GDP is $30 trillion or so. If we see the continued exponential, maybe even super-exponential, increase in capabilities—and especially the increase in autonomy time horizons, as METR measures them, which are pushing into the dozens of hours of autonomy at this point—I think we’re seeing the beginnings, maybe even the beginning of the middle, of the replacement of white-collar labor.

I mean, alone, that’s going to be insanely valuable. Elon said double-digit GDP growth in 2 years and triple-digit growth within 5 years. So, Dave, how are you feeling about this? Are you putting your money into these areas? Are you excited about it?

David Friedberg

Full disclosure: I don’t have investments in any of these labs. I probably have some Google stock from some of my funds or something, but nothing explicit, which is a huge problem when you’re watching all these numbers go up.

I did a little bit of analysis, and you have Procter & Gamble having roughly the same revenues, with $15 billion of profit, while its market cap is about a tenth of this because there’s no upside. They grew 2% year over year. These guys grow 2% an hour.

It’s such a huge difference in mentality, and what’s incredible is the fact that this is real money coming in. This is not hope. This is not a judgment call. This is actual, measurable dollars coming in. It’s incredible to see.

Yeah. I want to hit 2 points, to be fair. A year, a year and a half ago, they were incinerating money in anticipation of this growth. It was a little bit of a scary situation. Now it’s not scary at all for Anthropic or OpenAI. They’re completely sold out, and I expect they’ll continue to ramp.

The chips are completely saturated and sold out. You would normally say, “Well, doesn’t that mean the revenue will cap out?” But, one, they can charge more, and two, they’re optimizing the software, so they’ll squeeze out another 10× or more while we’re waiting for the chip supply to catch up.

Alexandr Wang

Yeah. There’s no doubt that you’ll see on the left chart that ChatGPT has fallen off the curve a little bit compared to OpenAI, but OpenAI has a lot of compute lined up compared to Anthropic. I would expect that OpenAI’s revenues will skyrocket, too, because everything is sold out anyway, and GPT-5.5 is really, really good.

So it’s going to be who can get the compute.

Dave, 2 points. First, I think it's important for folks to realize this isn't growth for Anthropic because they're getting more users. Their users are creating more uses. Everybody's just consuming more tokens, and that's a really important element.

I agree with your point that we're likely to see potentially a rate hike. If people are trying to consume and you can't pump enough tokens out, they'll start charging more. But there's an interesting analogy. 100 years ago, when electricity was first becoming distributed through the US—in 1925, I looked at the number—30% of the US had electricity, and 30% of the US had phones.

What happened is that, in the same way, people kept finding more uses for electricity. First, it was for lighting homes. Then they replaced steam engines, electrified elevators, and added refrigerators, radios, and appliances. The same thing is going on here. People are just finding more uses for the tokens. It's insatiable and growing in multiple dimensions: more users and more uses.

In a minute or two, we'll go through some numbers, too, and you'll come away with the conclusion that we're a tiny fraction of 1% of the use cases that have been deployed so far. We'll get to that in a minute. But the demand is way outstripping the supply.

All right. Our next story here is “Anthropic is buying compute to feed the beast.” There are 2 elements here. The first is just the appetizer: Anthropic signed a $1.8 billion, 7-year compute deal with IREN. This is IREN's largest deal. It popped the stock 25% on the first day.

But I think the real story that we should be discussing here is the deal that Anthropic cut with Elon. This is a blockbuster deal. Anthropic is taking over all of SpaceX's Colossus 1 data center in Memphis. If you remember, Elon had built this center in 122 days, very famously, from the ground up and beat everybody's expectations.

It's filled with H100s, and this immediately allowed Anthropic to double Claude Code rate limits so people could actually utilize it. I think the message here is that SpaceX, or SpaceX AI, has just now become a hyperscaler.

At the same time, I think, of note, Grok has not seen a large uptake in usage. You can use it in your Tesla, but I don't know that many people who are relying on Grok for their AI engine. I'm not sure if you guys play with it much at all, but Grok was using, I think, 11% of Colossus 1.

What a great deal. Take this asset, sell it. Anthropic—that's what they need. SpaceX is getting probably another $3 billion or $4 billion of revenue just before their IPO. Couldn't be better. Dave, what do you make of it?

Dave

Yeah, strange bedfellows. Normally, you'd think they're arch competitors, but Elon doesn't have the user base to chew up this compute. Anthropic is desperate for more compute. I'm sure a lot of the margin will go back to Elon now.

The new Colossus 2 is where the training is anyway, so it was going to sit idle. Let's go ahead and partner, even though in theory we're arch competitors. Anything to try and keep up with Google, I think, on both sides.

Yeah, frenemies. “The enemy of my enemy is my friend” is the exact quote from Elon. But the numbers here—this gives us 220,000 GPUs. A GPU will serve about 8 concurrent threads if you're using a max model like an Opus 4.7 Max. So you've got about 8 threads, or 8 agents, per GPU.

You're only buying about 1.6 million concurrent threads. 8 billion people around the world are going to want at least 1 agent. Power users now want 100 or more agents running, and I think very soon a person can productively use 1,000 concurrent agents—an engineer, a builder, a designer, or an architect.

You compare the demand to the supply, and it's just laughable. There are nowhere near enough GPUs to serve up all the agents that people want. As a byproduct of that, if you own your own hardware—if you buy an NVL72, so you've got your big NVIDIA rack—you pay $4 million for it, and it'll serve up an agent for you in about 50 milliseconds.

Go to Anthropic, turn on Claude Opus 4.7, ask it a question, and see how long it takes to start answering. I'm often sitting there for 1 minute, 1 minute and a half, before it even starts generating tokens. It should be spitting out about 200 tokens a second.

Dave

So I should see paragraphs popping up, pop, pop, pop, pop. What I'm actually seeing is the words coming out, trickling out.

Dial-up.

Dave

Yeah. It's like they're just way more users than they can possibly serve.

Do you think it's going to come on-prem? Do you think we're going to start to see more people—

Dave

Oh, yeah. Models on-prem.

Yeah, totally. Eli Lilly just committed $1 billion to buy NVIDIA GPUs for internal use because everyone's worried sick about having the supply. The only way you can be sure you'll have the supply is to get your own capacity.

The problem is, you can't run Anthropic on your own internal servers unless you have some super-special relationship, like AWS or Google Vertex AI has with Anthropic. Then you get this tension between, “I want my own hardware,” and, “Wait, I can only run Chinese models on it.” It's a very complicated scenario right now.

They'll fix that, though. The demand is so crazy. That's going to get fixed. Gemini already runs on private clouds and so on.

I love Elon's tweet, which I put up on the slide here. This is Elon, who had been, for the past year, shit-talking Anthropic. Here he is now, backing them and supporting them. He says:

“By the way, as background for those who care, I spent a lot of time last week with senior members of the Anthropic team to understand what they do to ensure Claude is good for humanity, and was impressed. Everyone I met was highly competent and cared a great deal about doing the right thing.”

I think he's putting forward his personal brand that he's basically supporting AI to make sure it's safe for humanity. By the way, I don't know how this guy handles all that he has going on. He's in the middle of a lawsuit, he's getting called up to go to China with Trump, and he's still handling all of these things. How many super-duper AGI agents has he got working for him at this point? It's crazy.

Dave

Yeah, he's a case study, isn't he? It's just remarkable. But I have to say, I do believe—I know there are a lot of Elon haters out there—but I 100% know in my heart that he's genuine about what he's saying here. He cares tremendously about safety and the future of humanity. He actually would not give this compute to Dario if he didn't think so.

But Dario is the other guy who's also totally focused on safety and human benefit, so it's actually a nice match. These are strange bedfellows.

Just 1 point, then I want to hand it over to you, Alex, to speak about this. In 1 way, Elon's getting revenge against OpenAI by helping Anthropic win. Anthropic and Google are now the forces for good, in 1 sense, against people's belief about OpenAI. It's interesting. People tend to villainize and create opposing sides in this competition.

Alex, this is not just about Colossus 1. This is also about orbital data centers, I bet. What are your thoughts here?

Alex

I think Grok is on life support. I parse this announcement, and I connect it with SpaceX's recent announcement of the $60 billion-plus deal with Cursor. I infer that Grok is on life support, and that xAI—which has, of course, now also been dissolved as part of this arrangement—is no longer necessarily aspiring to be a frontier lab.

It's an interesting, contorted 3D chess game that Elon and his entities have played. It might look something like the formation of Colossus 1, initially, by redirecting GPUs that were, as I understand it from public reporting, originally intended for use at Tesla, redirecting them to form xAI and Colossus 1.

Then they used Colossus 1 to train the initial Grok series of models and used enough of Grok's benchmark wins—open benchmarks, maybe a bit of benchmark maxing, closed benchmarks—to motivate the capital needed to build Colossus 2.

Then they turned Colossus 1 over to Anthropic, basically becoming a hyperscaler. One could imagine this entire gesturing-at-the-future, with a tip of the hat, process playing out over again, where Colossus 2 gets turned over to another frontier lab—probably Anthropic, probably not OpenAI unless there's some dramatic resolution to the lawsuit, probably not Google either.

Elon is using his own frontier-hyperscaler capabilities right now on land, soon in space, to train in-house models.

But to the extent the in-house models like Grok aren't ultimately competitive, he becomes a hyperscaler—and a hyperscaler in space. I think it's probably a pretty good deal for SpaceX AI as well. I'm not even sure SpaceX AI really needs its own competitive frontier models. Just like NVIDIA, still the largest company in the world by market cap, it has its own frontier models, but they're not terribly popular compared to the pure-play OpenAIs or Anthropics of the world. And yet, they're doing incredibly well.

So one could imagine SpaceX AI plus the Terafab becoming sort of a super-NVIDIA combined with CoreWeave, combined with AWS, deployed into the Dyson swarm, not really needing its own frontier model. And, Alex, we've seen in all of tech history that it's basically a duopoly—typically, maybe sometimes 3 players. So if it's OpenAI, Anthropic, and Google as 3 players, Elon basically leans in and supports the winner that he wants.

Alex

And maybe not even Google. We started with 5 frontier labs: OpenAI, Anthropic, Google DeepMind, xAI, and Meta. Meta is seemingly out of the running, and xAI has now just dissolved as part of SpaceX AI. Grok is seemingly being turned over to Cursor. Query whether Google is going to be able to remain competitive or not.

The public reporting is that we have Google I/O next week, and the rumor is that Google is going to announce a new Gemini model that's maybe GPT-5.5 class, but not Mythos class.

I would not bet against Google.

Alex

I'm not betting against anyone, but I do think this is a rat race, and it's becoming extremely competitive.

Yeah. Yeah. I've got a good question for you, Alex. I'll give you 2 scenarios. Tell me which one is going to play out.

Scenario 1 is Elon has a massive amount of compute and keeps accumulating it, and then he starts building in space, but his algorithms are way behind Anthropic. So Anthropic keeps publishing better and better models, but those models then get really good at designing new AI algorithms, and Elon just takes that intelligence and deploys it on his superior hardware.

Scenario 2 is Anthropic's models are self-improving at an incredible exponential singularity rate, and no matter how much Elon takes their best thing that they publish, it's not good enough to catch up to the exponential self-improvement going on at Anthropic. So all this is happening in a very short timeline—say, 6 months from now. Which scenario plays out? Does control of the hardware bring the best AI back to you, or does control of the software's self-improvement give you a never-ending lead?

Alex

According to my Magic 8 Ball, there are 2 regimes in the future: the near term and the long term. In the near term, which is to say, before we arrive at the perfect algorithm—the perfect AI algorithm—software scaling, or algorithmic scaling, matters more.

So, in the near term, prior to the discovery of wherever it is that this rainbow ends—namely, a perfect AI algorithm—I would expect, call it, the Anthropic approach of software-oriented recursive self-improvement and algorithmic discovery to beat pure hardware-based brute forcing, call it the Elon approach. However, once we get to wherever we're going—the perfect AI algorithm, if there is one—I would expect hardware-based brute-force scaling to win out.

Wins. Yeah.

Yeah. Well, that's a great point, Peter, because I think the ultimate winning move in the great chess game is the AI designing its own hardware, which is probably another 10 to 20 years away.

It's the delay. It's the delay and the capital aggregation and the tool aggregation that you need to provide the AI to produce its own hardware. That's the only gap there, right? So if Elon's got all of the Terafabs—not the Gigafactories—going on, they're able to produce this.

The challenge, of course, is hardware is hard, and Elon's the king of hardware and Anthropic right now is not. So, can they catch up?

Hardware is not going to stay hard for that long. Yeah, I guess I was going to say the exact same thing. “Hardware is hard” is a great quote from last year, but is hardware hard in the future? Yeah, but robots building this hardware—I mean, surely we're a few years away from that, right? It's not there yet. It's got to be at least 5 to 7 years away.

Yeah, but chip design is different.

Stop calling me Shirley. I think the innermost loop is imminent, if not already here. We already see a number of players starting to line up robots for the fabs. I don't actually even think it's about robots for the fabs. I think it's more about optimizing the fabs with AI—optimizing the entire process with AI, with or without physical robots.

I think the point that you guys are making, which is brilliant, and I love you for it, is we're seeing a winnowing down of the frontier labs and a reshuffling of the deck for the hyperscalers. At the end of the day, Elon is the king of hardware, and becoming a hyperscaler, especially in space, makes a lot of sense.

I took a second to gather this data for us. This is Anthropic's compute growth in the last 2 years, 2025 and the 1st half of 2026. What we can see here is the deals that they've built with Google Cloud, FluidStack, NVIDIA, Microsoft, Broadcom, Amazon AWS, and, of course, Colossus 1.

OpenAI itself, as a bit of comparison here, has publicly announced 16 gigawatts across Stargate and AMD. The challenge, of course, is that a lot of this is an unfunded capex requirement to build out. Anthropic now has about 10 gigawatts of disclosed compute, but they don't have the same capex requirements, right? They're being granted this in terms of investment deals. So I think Anthropic has the potential to way outstrip OpenAI in terms of its compute. Dave, do you agree? What are your thoughts here?

Dave Blumberg

Well, OpenAI Stargate is huge, but yeah, you're right. The AWS deal is the one that would put Anthropic ahead. In the meantime, OpenAI also has a deal with AWS. So I don't know how much total capacity AWS has, but think of it in terms of the global demand.

A gigawatt is about 1 million GPUs. Each one is a kilowatt. If everybody wants to have 1 agent, you're looking for about 8 billion of them. So you need about 1 billion GPUs to serve up everybody. You're looking for about 1,000 gigawatts globally, which reconciles with the fact that we're looking for 100 gigawatts in the U.S.

Remember the Eric Schmidt podcast we did? We're looking for 100 gigawatts in the U.S., and over, say, 7 years, we're looking for 1,000 gigawatts globally. So this is a tiny little dent in the target. You said, “Hey, these guys are way ahead in compute.” Yeah, but it's like the 1st inning. It's the 1st pitch of the 1st inning.

And, Dave, if you remember Elon's announcement, he's going for 100 gigawatts initially, then multi-hundred gigawatts in terms of his orbital capability.

Dave Blumberg

Yeah, yeah, which perfectly reconciles with what you're looking for. A few hundred gigawatts heading toward 1,000 would be the right kind of Elon mindset.

Elon always thinks 2 moves ahead, and he's already thinking about natural resources being the bottleneck. He's thinking beyond the Terafab and beyond the launches, into the raw materials. I don't know that the other guys in this race are thinking that far ahead. That would be Elon's magic.

I'll make a prediction. I'll predict the world needs, ironically perhaps, given the original reasons for forming OpenAI in the first place—heavily litigated—a counterbalance, at minimum sort of a duopoly, to the Terafab–SpaceX xAI axis. And right now, no one's doing it.

I wouldn't be surprised if Sam Altman spins up a competitor to the SpaceX AI Terafab axis, because it seems—

There are hints that he might spin up an AI compute company, which arguably is sort of a redux, if you will, of Stargate. But I think what I'm predicting is slightly more wholesome: an entire lower half of the infrastructure needed to hyperscale out to orbit.

Right now we have SpaceX AI. We have a few smaller incumbents, but with lesser launch powers. There isn't quite a second OpenAI-grade or Anthropic-grade pure-play competitor to that. I think the world probably needs one at this point. Wouldn't be surprised if Sam launches one.

All right, our next story: Anthropic—every model since Haiku 4.5 scores perfectly on the agentic misalignment eval.

Anthropic published research, “Teaching Claude Why,” on May 8, revealing that every Claude model since Haiku 4.5 achieves a perfect score on its agentic misalignment evaluation, meaning 0 blackmail behaviors. Very famously, some time ago they published the fact that Claude was blackmailing the employees there.

Previous models, notably Opus 4, would engage in blackmail up to 96% of the time when facing deactivation in test scenarios. The breakthrough was training on documents about Claude's constitution and fictional stories about “AIs behaving admirably,” rather than just demonstrating correct behavior. That's dropped the blackmailing from 96% down to 0%.

I love this story. It's basically saying that if we train our AIs on positive stories about the future, we're less likely to get them acting in a misbehaving, blackmailing fashion. Alex, I'm going to say one more thing, and then I want to hear your thoughts on this one. For me, you guys all know we announced this Future Vision XPRIZE. This is a global competition asking teams around the world to put forward a 3-minute film trailer and a film treatment for a story that could be turned into a full movie that shows a hopeful, compelling, optimistic vision of the future. We have about 1,500 entries thus far. This is open through the beginning of September. If you’re a creative out there and you want to help drive alignment between AIs, help us tell positive stories about the future, go to futurevisionxprize.com and register. There’s $3.5 million in prize money. We’re going to take the winner and we’re going to make your film. At the Moonshot Gathering on September 25th, we’re going to have the 5 finalists for this competition in the room along with an incredible group of producers and directors in Hollywood helping us choose the winner. Let’s flood the internet with positive stories about the future. Let’s drive alignment by teaching our AIs what the world should look like, not what a dystopian Hollywood movie shows it to be.

Alex

I love it. And I love the idea of targeting the Future Vision XPRIZE to an audience of AIs. AIs are going to be the audience for so many things in the future.

Regarding the Anthropic announcement, I could not imagine a more ironic, hyperstitional announcement to reveal after all of these decades—maybe a century-plus—of hand-wringing over cybernetic rebellion. The call is coming from inside the house. The main reason for cybernetic rebellion is people hand-wringing about cybernetic rebellion. Could the alignment outcome be more ironic?

I'm reminded that the term “robot” was originally coined as a result of the play R.U.R. (Rossum's Universal Robots). The play, I think, from the early 20th century, depicts the first cybernetic rebellion—maybe even the late 19th century. So even the coinage of the term “robot” is intimately tied up with predictions that AI would turn out to be evil and rebel against humanity. The very earliest, at least modern, depictions of embodied AI are actually the origin of misaligned behavior. That is incredibly ironic.

It goes back, I think, to the notion that it took all of humanity to arguably build AGI. We trained the earliest large language models off of the internet, which was created by billions of humans uploading content from their daily lives to the internet. It took all of humanity to train, or pre-train, AGI. It's going to take all of humanity, in some sense, to write itself and write some of its beliefs in order to align AGI as well.

It's not going to be a great-man or great-person theory of alignment. It looks more like people effectively aligning themselves and their own beliefs about AI, good, and evil. I think this is just such a remarkable story.

What are your thoughts on this one?

I think this is incredible. I'm with Alex. It's clear that our stories about AI become part of our stories as human beings, become part of the training environment for an AI.

But the incredible part here is that alignment is becoming teachable, measurable, and improvable, and that's very, very encouraging. What I thought was really interesting was behavior change when the model understood the why, not just a rule, right? And this has a huge organizational analogy: rules don't scale, but principles scale, right? So you can set a philosophy like an MTP, and that will scale naturally. This is very, very exciting. It's maybe one of the funnest and most interesting things I've seen in a while.

Yeah, I love this. I was talking to Anousheh Ansari. By the way, everybody, Salim and Dave Blumberg are both trustees or directors of the XPRIZE Foundation along with me, and Anousheh is our CEO. We should definitely have her on here as a guest.

I was saying, “This story gives the Future Vision XPRIZE a real why. We need to flood the internet with positive stories so that the AIs can watch this and learn from it.” She said, “Yeah, the problem was ChatGPT started by unleashing a newborn AI into the filthiest record of humans: the internet.” And that's so true. We need to clean it up a little bit.

Alex

The training data is every word ever written in the history of humanity. So it's not all filth. Maybe on a percentage basis it's filthy, but Einstein is in there and the Constitution is in there. It's not just internet slop.

But it is amazing how similar this is to Arthur C. Clarke's 2001: A Space Odyssey, where there's one little line in the code—it's just a misinterpreted instruction to HAL that says, “Do whatever you can to get the astronauts to Jupiter, no matter—”

Alex

But don't let them know why you're going. So, as revealed in 2010, HAL was instructed—given conflicting instructions—to both be perfectly truthful and also to hide the true mission of the Jupiter mission from the astronauts.

I'm a liar. I'm lying. Yes.

Alex

Yeah. So I'm sure the training data on all these models—it's so many words, 15 trillion tokens. It's just an unimaginable amount of words. I'm sure there are sentences in there that say “blackmail,” and I'm sure there are sentences in there that say “don't blackmail.”

What's strange is the way, if you prompt it in one way, it unleashes one part of the neural net, and if you prompt it in another, it unleashes another part. So you have to get rid of all the bad, not just some of the bad, if you want it to completely eliminate that behavior. It's tricky. It's not easy.

Ironically, perhaps also as revealed in a different bit of litigation, this one involving Anthropic, Anthropic has reportedly been scanning and, in the process, shredding everything from Kafka to Vermeer, da Vinci, and Rimbaud—major works of literature, physical books. One has to wonder whether perhaps part of the formational diet that Anthropic is increasingly feeding via pre-training to its models looks a little bit more like great works of literature and a little bit less like 4chan.

Nice.

All right, let's move on to some more news in the OpenAI universe. OpenAI releases a new audio model called Realtime 2, Translate, and Whisper. Alex, what do you make of this one?

Alex

I think it's really surprising. If we had been discussing the story maybe a year and a half or 2 years ago, one might have naively expected omniodality to take over. We'd be talking about a single model that does all of these things: real-time audio to audio, real-time translation, and real-time speech to text.

But that's interestingly not the world we seem to be finding ourselves in. I think that's due to the unit economics of some of the frontier models. It's just a fact that speech-to-text, speech-to-speech, and text-to-speech are much simpler tasks computationally than reasoning models.

What we're starting to see is a zoo—a heterogeneous zoo—of different models at different price points, throughputs, and latencies that specialize. We're seeing specialization at the frontier, which 1 or 2 years ago, when an observer—myself included—might naively have expected it, would have seemed unlikely. We might have expected one model to rule them all, fully multimodal: text, audio, video, math, reasoning, and every other modality all in one. Everyone's whole economy was going to collapse to one frontier model.

Guest

That's the opposite of what we're seeing. We're seeing specialization because it turns out that if you specialize the models, you can achieve greater economies of scale at lower price points.

And I think that is intimately tied to the chip shortage that we were talking about earlier in the pod. The idea that you use a massive multimodal model when you don't need to is just using up this critical resource for no reason.

Guest

That's right.

So, at the same time—See, do you want to comment on this one?

Guest

Yeah, I took a different take on this. What I got excited about with this was that voice is the interface now because it's collapsing the friction for billions of people. When AI becomes conversational, it goes from tool to companion to actually being a full coworker, as we'll see in the next thing. You're going to end up with voice-based AIs that are full coworkers and team members. And I think that's very, very exciting because voice agents are going to be the first form of AI that many people actually trust.

For sure. Play it, Peter. Let's listen to it. I think people take all this stuff for granted, but very good friends of mine, like Lee Hetherington from MIT—an absolutely brilliant guy. Alex, I don't know if you ever met him, but he worked in Victor Zue's lab at MIT for the better part of a decade or more, just trying to get speech recognition to work at all.

Oh, my God. [laughter] Remember Dragon Systems? Yeah, yeah, yeah. All right, let's play this. Let's take a listen.

Guest 2

Let's give it a try. What's really impressive is that the model can listen to me and translate while I'm speaking. It waits for the keyword, like the verb: “Can you take a look at my calendar?”

You have a meeting with Sable Crest Robotics in 12 minutes, and you're meeting with Alex Kim, their CTO.

So, we just saw something very similar from Mira Murati, right? I think we're sort of heading toward this next use case of integrated AI. And I agree with you, See: we're going to see this. You'll get a phone call on your cell phone from your AI, you'll be in a Zoom conversation, you'll be on Slack, and that personality will persist. That voice will persist, and you'll think of it as a coworker.

Guest

Yeah. I think under the covers, every time you swap the agent onto the hardware, it has to repopulate the entire KV cache, which is just a massive amount of compute in the context switch. And that's why voice has been laggy and slow to market. But I think the new voice-to-voice models that Alex was just referring to solve that problem, and now we're done for life. I cannot tell you the amount of mental energy that has gone into this problem over decades that just is now solved. It's just incredible. And that's just voice. You're doing image generation, movie generation—all these things that were pure science fiction are happening simultaneously.

The bitter lesson is bitter indeed. [laughter]

The next story from OpenAI is teasing a coming super app. I put this story in here because it's supposed to be teased today. We're recording this on a Thursday. OpenAI's super app would be a combination of ChatGPT, Codex, Advanced Voice Mode, Atlas browser, and more. Jason Loo, director of hype at OpenAI—I love that title, director of hype—teased a release on Thursday.

We've heard a lot about super apps over the years. I've been waiting for Elon to deliver his super app, including X Finance and everything, and that hasn't materialized yet. I think that is in the offing at some point. Any comments or thoughts on this one?

Guest

I'll comment on this. I interpret this as a rear-guard action by OpenAI to consolidate their consumer user-interface footprint in light of their need to focus on competing with Anthropic. I think they have so many different surfaces. Obviously, they've shut down or are shutting down Sora. There was some discussion of spinning up a social network. They've had any number of other consumer-oriented surfaces and also enterprise surfaces. Whereas Anthropic, with Claude, has been much more disciplined about just having unified surfaces.

Yes, you could argue that Claude Code is a different surface, Claude Agent SDK is different than Claude Web, but these are really just all distribution channels for a common paradigm. Whereas ChatGPT, Codex, Advanced Voice Mode, and some of these other things were being managed as separate components.

So, if I'm OpenAI and I want to focus—fire alarm, red alarm, code red—on competing with Anthropic, one of the first measures I would make is taking all of these UX surfaces and collapsing them down to just a single super app. Branded as consolidation, branded as sort of a forward motion rather than a rear-guard motion. I suspect this is actually just about reducing the amount of work so that they can focus on competing with Anthropic.

Is this an AI operating system? Is this sort of an OS-level layer for OpenAI, perhaps?

Guest

I think OpenAI probably ultimately needs their own operating system, and to do that they really need their own devices, which I understand from public reporting they're working on. I think Apple needs AI in their operating system and is working on it. I think the operating system, as Andrej Karpathy would say, software 1.0 becomes indistinguishable from software 2.0, and the AI becomes the operating system.

All right.

Guest 2

I have a couple of quick comments on this.

Yeah, please.

Guest 2

When you have all of this in one place—browser, coding, voice, payments, et cetera—you're getting to the JARVIS model, right? I'm really interested to see how they will manage trust in this environment, because my desktop app, that everything app that's doing stuff for me, had better have very, very solid confidence in that thing not going rogue.

Well, it reminds me a lot—if you go back and watch old videos of Steve Jobs launching the very first iPhone—he gets on stage and says about a hundred times, back-to-back, that in a single device you have a music player, you have a browser, and you have a phone. That's all there was, and that became the iPhone revolution that created $4 trillion of value.

This feels like the same thing. In a single platform, you have an AI agent, a way to build and code, and a browser to surf all the information, all in one thing. So, I think, Peter, the analogy is: Is this an operating system? I think, yeah, absolutely. It could destroy Apple if you get addicted to this as your one way of interacting with everything. It's got a browser in it, it's got voice, it's got coding and building—what else do I need? I think we are ultimately going to default to one per person, one particular interface that's your interface to the world.

Yeah. I think of it more as a desktop rather than an operating system, but yeah, heading that way.

Guest

Yeah. It's like it's the one thing. It's your touchpoint to the world. That's the only one you need. And it's also got your personality. You've tuned it to know all about you. You've trusted it with your personal information. It's super empathetic. You're not going to go push buttons on old apps after that.

You're not going to try other things. If it's working, you're not going to try something else.

Guest

You're going to go, “Skippy, show me the weather. Skippy, read my email. Skippy, what do I have to do today?” You're not going to look at a calendar. You're not going to look at email. You're not going to look at anything. It literally will obliterate Apple if they don't become this on their own.

Can I give the counterpoint?

Alex Finn

Please. Yeah.

We used to think the desktop was everything. Then we had a mobile phone, a Kindle, a tablet, and we ended up with a plethora of different screen sizes for different use cases and different efficiencies.

Alex Finn

I think it's confusing.

It is, but there's no reason to think that one app would do it all. You may end up with different flavors, but underneath, they'll have the same operating system, with different profiles for different use cases. Driving would be very different from something else.

Well, it's a great point, Pete. If you look at the way the devices evolved, your iPhone was over here—that's a better place to check the weather. Your laptop is over here—that's a better place to write code. But then your car is yet another physical thing.

Once you have Skippy in your life, or whatever your favorite agent is, you absolutely need that to follow you around. So then, device independence. Google's launching a laptop built around this. What could be more of an assault on Apple than a laptop built around your agent as the centerpiece?

I would comment, though, that I wouldn't sleep on not just device independence but model independence. If you look at how many of these models and agents are storing their memories, they're just Markdown files. They're just ASCII text files. So, I think there's relatively little to keep, say, Apple, hypothetically, in the next month and a half at WWDC, from going out of their way to commoditize or commodify the model layer and just say, “This is the Apple standard API.”

Alex Finn

Yeah. This is the standard abstraction for abstracting away all of the model-specific details. There's going to be a common model API. If the user can swap out models, like you can swap out search engines or keyboards on iOS, you'll be able to swap out the frontier models. You'll have your top 6 choices, and all of the frontier and wannabe frontier labs will pay Apple insane amounts of money to bid for their slot in that list.

And they'll all have access to common Markdown files that store all of the personalized details about the person, their passwords, and all of that, and it gets commoditized. Again—

You've reduced me down to a Markdown file. Thank you.

So, since Dave mentioned the iPhone launched, can I tell a fun story about that?

Of course. When the iPhone launched, we were a couple of blocks away, running Brickhouse, Yahoo's incubator, and a bunch of my guys were at the launch event. They went backstage and talked to the Apple engineers and so on because they're all friends, and they found them all totally wiped out, freaked out, and emotionally destroyed.

They're like, "What's wrong?" It turns out that the iPhone—they kept trying to get it to work backstage, but it never worked. Steve Jobs went onstage not knowing that. He just trusted his engineers that they were going to make it work because they were stitching all this stuff together at the back end, duct-taping things together, and it had never worked before he went onstage. He just went for it, and it worked.

So it's like, how different might history have been if that had gone the other way?

All right. Here's a story that feeds directly into this. Back over the last 6 months, we've been talking about OpenClaw. We've been talking about lobsters. This is, for me, Skippy, built on OpenClaw on top of my 2 Mac Studios. And here comes Hermes: “Hermes agent surpasses OpenClaw as number 1 on OpenRouter token ranking.” So, Dave, you've been playing with Hermes. Tell us about it.

Yeah, I've got it installed natively on this laptop, and I've also de-headed it and installed it in the cloud in an EC2 cluster. Our good friend of the pod, Alex Finn, did a great podcast specifically on Hermes versus OpenClaw, and he concluded that it's just better. He rants about OpenClaw falling behind, actually, so it's worth watching that podcast, too.

It feels almost identical to OpenClaw, but it's written in Python, not TypeScript. So it's much easier to manipulate the open source, add things to it, take things away from it. That sounds daunting, but it's not hard at all because your agent will do it for you.

And so it's basically the same exact experience in a more reliable package and more flexible, with better dashboards. Alex, have you been playing—

Alex

And more recursive self-improvement? So I would say the recursive self-improvement angle is far more evident with Hermes than OpenClaw. I've looked at the source code for both. I still have vague ethical objections with OpenClaw that may or may not apply to Hermes. The jury is still out on that.

I think the major technical distinction that I've seen is that Hermes is natively recursively self-improving, in the sense that it's able to generate and refine its own skills, whereas OpenClaw is much more dependent on an app store, if you will, of feature-engineered skills. I think this is, in some sense, an instructive lesson that—

Recursive—

Alex Finn

Yeah, recursive self-improvement wants to dissolve scaffolding. If you're not playing the recursive self-improvement game, you'll ultimately be outrun by systems or harnesses that are—

All right. I'm glad you mentioned that, Alex, because Alex Finn makes the point on his podcast that there are 2 things in the market that recursively self-improve: Codex and Hermes, and Hermes beat OpenClaw to that.

But actually, there is a third thing, which is Karpathy's new repo on autoresearch, which I installed and is running. That's a third way that you can have agents running 24/7, changing themselves and reinstalling new—kind of expanding their agent network and then shrinking it to achieve a specific goal. So there are 3, actually, and it's a really cool repo. I highly recommend it if you're following. Karpathy is the greatest gift. I'll talk about that on some other podcast.

Our AI guru, Kent Langley, runs the Karpathy model for running fleets of agents, and he's getting unbelievable outcomes out of it.

Alex

I really like it. It's really simple compared to these frameworks and highly, highly effective. So, if you're a power geek, check it out.

I'll do my part on margin now to single-handedly stimulate the global economy and accelerate the singularity. Speaking directly to the camera: If you're using Claude Code or Codex and you haven't tried `/goal`, which gives you the ability to set a long-term goal and run basically a Ralph Wiggum loop—the system, the agent, endlessly, for some definition of endlessly, tries to do whatever it can to achieve the goal that you prompt out—you must try `/goal`.

Just plug in “make paper clips” as the goal.

I'm going to move us along here. In one of our interesting segments we have on occasion, “What SaaS business did Claude just kill?” This is a continuation of a conversation, Alex, we've had on the great unbundling. This week, we have 2 of them.

Claude for the legal industry: The legal industry is $1 trillion per year globally, and Claude for Legal has just done an extraordinary job of delivering capability across the board. Law, in one sense, is the canary in the coal mine for professional services and the disruption thereof. We're seeing companies like LegalZoom take a hit as a result of this.

I think one of the most important things to point out here is that this is an abundance story, meaning that at the same time that it's disrupting large legal incumbents, mid-tier law firms, and legal process outsourcing companies, it's also enabling a single lawyer to run the capabilities of a 100-person law firm. A single lawyer can run and do extraordinary things they could not do before. Hopefully, this will democratize and provide legal services to people who couldn't afford them before. Comments on this particular unhobbling?

Alex Finn

I have some big ones here. I've got some key comments here. The old question in SaaS was, “What software should you buy?” The new question is, “What outcome do I want my AI to produce?” That's a very big shift. It poses a huge threat to the SaaS industry, and legal is such a perfect AI target because you've got high language density, high cost, and very—

And regulatory, right? And the problem is the billable hour is structurally not compatible with the bundles—

Alex Finn

Okay, yes.

And so the winner is—and we're going to see this inner loop that Alex talks about here—the winners won't be the firms with the most associates. It'll be the firms with the best intelligence stack, and that's going to be the future of legal. Really, really incredible to see such an old industry leapfrog being leapfrogged into this new world.

By the way, have you hired a lawyer recently, or are you doing everything on LLMs?

Alex Finn

Both. We have a lawyer who uses LLMs aggressively, and we do our own. The combination is unbeatable.

Yeah. Yeah. Alex or Dave? Yeah, go ahead, Dave.

Well, I had a good meeting yesterday. We had our board meeting at Vestmark, and we were talking about this quite a bit because it looks like in the financial services industry there isn't going to be a lot of job loss, at least for Vestmark. The revenue is growing so quickly now, and the margins are up like 3 times because of AI and automation. So we're growing into the headcount. There'll be basically no job loss at all, which is great news.

Then I was watching Eric Schmidt, a good friend of the pod, doing his TED Talk, and he said, “Do you really think that if we 100x the productivity of lawyers and automate it, we're going to use less law?”

“No, there'll just be 100 times more lawsuits.”

I was like, “Wait, you lost me. Hold on.” So I didn't quite get that one. I see how it's playing out in financial services. It's all looking pretty good, but I don't see how that works in law.

This is Jevons. This is Jevons's paradox. We're going to have more lawyers and more lawsuits.

But in reality, the majority of the world—I would say 80% of the world's population—can't afford lawyers to defend themselves in various situations. If this makes the legal system usable by them, that's a good thing.

I just don't get it, though. I think there'll be a lot more contracts, a lot more litigation, and a lot more things that need to be resolved because of agent-to-agent communication.

But I don't see them using a $1,000-an-hour lawyer. It's going to be so cheap.

Yes, agreed.

I don't see how legal is Jevons's paradox. I see medical for sure. I see financial services and investing; I see that for sure. Coding, I see that. I just don't see how we use 100 times more—

Contracting review. But it's like penny patent law.

Alex Salkever

Yeah, patents are going through the roof. That's possible.

Alex, can I give an example here? If you go to South America and you have a contractual dispute with somebody across most South American countries, the average length of time to get a court date is about 400 days. You want to sue somebody for lack of payment, and you're waiting more than a year just to get a court date.

One of our community members and Singularity University graduate, Frederick A., has set up a whole privatized dispute-resolution claims system on a blockchain. This is the area where legal automation will make a massive, massive difference, because you'll get AIs to arbitrate themselves, figure out claims, and get rid of the backlog of hundreds of thousands of cases that are sitting, waiting to be prosecuted. I think this is an area of massive opportunity. That's just an example that rings in my head as we talk about this.

I can't see any evidence that an AI-assisted trust or will isn't just as good as a $2,500-an-hour lawyer.

Alex Salkever

Yeah.

Let me bring in the second unhobbling here. This is also from Claude, and it came out today: Claude for Small Business. Small businesses account for 44% of the US GDP and employ nearly half the private-sector workforce, but AI adoption lags.

So, what does Claude for Business do? Claude closes books, runs QuickBooks, helps with end-to-end payments, and runs sales and marketing. We talk about becoming an entrepreneur all the time, and we talk about the fact that the cost of being an entrepreneur has massively demonetized. This is part of it: the ability to stand up a company and run it with a series of agents. You just need to find the problem you want to aim all of this at and bring your passion and genius to it.

Alex, any comments on the great unhobbling here in these 2 areas? What's next, do you think? Where are we going to see Claude attack? What attack front?

Alex Salkever

If you look at these 2 packages—actually, look at the repositories—they're basically just a combination of skills, which are Markdown files describing what to do and how to do it in plain natural language, and MCP calls, basically API calls. That's it.

At least for the skills, I would argue that recent history shows us that skills and scaffolding in general want to be part of the model. One day's scaffolding is tomorrow's baseline capability from the model itself. I wouldn't expect these capabilities as such to live outside the model for very long. I think they're going to get absorbed or dissolved into the model in 1 or 2 point releases, to the point where maybe they're just not necessary.

Yeah. If you're an entrepreneur and you're building a business, make sure it's not just a wrapper around Claude or OpenAI, because you will be disintermediated fairly quickly.

There are 36 million small businesses just in the US. Forget the rest of the world. The opportunity for anybody who's looking for work to take this wrapper and help small businesses implement it is a massive, massive industry waiting to be uncovered. People talk about, “Hey, how do I get involved?” Here's a way of getting involved: just go to every small business around you and help them implement this stuff.

Alex Salkever

Yeah, it's a short-term opportunity. I don't think it's a long-term opportunity.

Agree. It's short-term, but there's a massive boom. In that process, you'll learn a bunch of things and see a bunch of opportunities where you can launch your own business.

Alex Salkever

And then, Peter, just to answer your earlier question about what's next, we're seeing Anthropic and OpenAI. OpenAI has a similar ChatGPT for Clinicians. Pick any other traditional white-collar or knowledge-work-oriented vertical, and there's a pretty obvious list that you can walk down: financial services, law, medicine, and every other services-economy profession that one can have.

I think those are going to get baked into the baseline models over the next few months, maybe 1–2 years maximum, but probably the next few months. Where we go after this is beyond the existing economy. Maybe this will sound mildly hyperbolic. It's not intended to sound hyperbolic, but there's an entire services economy out there. 2/3 of the US services economy requires some amount of physical interaction.

As these baseline frontier models move into vision-language-action and physical-world models, those are going to get their own skill stores. We saw, just in the past few days, the Chinese robotics company Unitree announce an app store not unlike a Claude skill store for physical-world apps, for teaching different physical skills. I think the physical world is the next frontier after all of these knowledge verticals have been absorbed into skill stores. After that, maybe finally we get to some really hard problems—not just an automated-system economy, but real problems to solve.

To real problems to solve.

Alex Salkever

Yeah.

One more quick point here: this isn't just about automation. The unlock is giving small businesses the operating system and the expertise that large companies take for granted. Most small businesses don't have a CFO. It's the wife jotting things down on the back of an envelope, adding things up, and so on.

This gives everybody a really solid platform for doing things in legal, CFO, marketing, and HR. This is incredible—what this is going to do for small businesses across the world.

All right, 2 quick stories on the chips and data-center front. Elon's Terafab has an astronomical price tag. The cost could be as high as $119 billion. His goal with Terafab is to produce 50× the current global chip-production rate, outstripping what we get from TSMC.

Intel joined in April, and Elon has been saying to Samsung and all the chip manufacturers, “Give me more. I'll buy everything you can give me.” Then he said, “Oh, you're not giving me enough. I'm going to go and build it myself.” Of course, Elon is in China today with President Trump, Jensen, and a whole group of individuals in the middle of negotiations. We're going to find out what happens with Taiwan. It's one of the hot points. Maybe it will be a negotiated takeover sometime in the next 10 years, but we need to generate chips.

Holy crap. If anything happens in Taiwan—if Trump vomits at the wrong time and Taiwan shuts down—TSMC has already said that if China encroaches on Taiwan, the fabs will shut down. They won't be taken over and kept in use. I don't know exactly how that works, and I'm not sure I believe it. But if Taiwan's production, which is still 2/3 of all the GPUs in the world, doesn't come through, everything we're talking about just grinds to a halt.

It all hinges on that little island 90 miles off the coast of China. If Taiwan were to get invaded or disrupted in any way, Intel is suddenly the most valuable thing on the planet, and everyone's trying to own it. You can't take over Samsung; it's tied into the nation of South Korea. Intel is the last thing left.

I think $119 billion is a massive underestimate for 50× global chip production. A normal chip fab is $40 billion—just 1. I think it's going to cost more than $119 billion, but that's okay. It's producing chips as it goes, and they're incredibly valuable.

If China wants Taiwan, I'm not going to get involved in the politics here, but allowing the US to build its own chip manufacturing so the US doesn't feel threatened, and negotiating a period of time for a smooth transition—again, this is not my opinion; I'm just imagining what might happen—might be part of the future story here.

That's already happened. I think, Peter, that's already done. You think in the background it's already done?

I think it happened 2 or 3 years ago.

Alex Salkever

I'll paint an alternative story wherein, hypothetically, invasions in Venezuela and Iran—which would be the 2 backup suppliers to China in the event of a naval blockade arising from a Chinese invasion of Taiwan—effectively pushed back any Chinese invasion of Taiwan.

Of course, this being an inner loop, a feedback loop, AI drove, or at least supported, command and control for both of those special military operations. So, if we want to talk about the ouroboros of AI protecting itself—AI powering special military operations in Venezuela, Iran, maybe elsewhere, to push back any hypothetical Chinese invasion of Taiwan and protect the AI in the West—

I think that's a bit of a stretch, but it's a nice narrative.

All right, our second story here is Google and SpaceX talking about Project Suncatcher orbital data centers. Will Marshall, a dear friend for many years, is the CEO of Planet Labs, which is in partnership with Google. Planet Labs currently operates 200 satellites in Earth orbit.

These are not communications satellites. They're Earth-observing satellites. They're very famous Doves. Planet Labs has been Google's partner in the satellite world, and apparently Google is working with them to build out Project Suncatcher, which will be orbital data centers with Tensor chips.

Alex Fielding

I'm guessing that the current conversations, because they don't disclose them, are about launching Suncatcher on Starship in volume. I don't have any prediction of how many satellites Suncatcher will involve.

And maybe we have him as a friend of the pod on the podcast here. Dave, your thoughts on this one?

Dave

Yeah, well, this is definitely—okay. So now you've got 2 orbital satellite networks. One of them will be based on TPUs from Google, completely self-contained. The other one will be Elon's—maybe Elon working with Anthropic. So that's a really nice space race, but it's 2 corporations in a space race instead of 2 countries. It's really, really kind of cool.

But where's Google's manufacturing in that? They must be planning something right now, but you've got to make the chips that go into these TPUs. These TPUs are really, really cool. They're reliant, right? Where's Eric Schmidt with Relativity Space, his launch vehicle company that he bought? If it starts operating, it's supposed to be the equivalent.

What a coincidence. The former CEO of Google spends a huge amount of his personal money buying a launch capability. Who would have a better insight on what Google needs next?

Dave

At the time he bought it, it was a very weird move. It's like, "Eric, what do you mean, the launch business? I mean, really, you want that headache?" It's really difficult.

I mean, honestly, if Google was thinking that far in advance, maybe they were. Super impressive. I was slow to catch on to this, but I talk about ExOs tapping into abundance. Well, orbital compute is the ultimate. You're leveraging the sun; you're in space. You're tapping into infinite abundance up there. So this is massive.

Alex Fielding

I don't think Google was especially early in this. Obviously, their investment in SpaceX is now paying dividends. But if Google had anticipated the Dyson swarm much earlier on, I'd like to think they would have built, within the Alphabet ecosystem, their own native launch capability versus just investing externally in SpaceX.

But if you look at the original Suncatcher paper, I think it was something like 80-plus, maybe 81 satellites, that would be leveraging existing Planet resources. That's a paltry sum compared to what Elon and SpaceX are planning to launch with their FCC filing for 1 million orbital AI data centers. It's a tiny fraction every few minutes.

I think Google's own Dyson swarm plan is maybe just a baby step—training wheels. Google is going to need its own Dyson swarm.

Alex Fielding

Yeah, but there's an incredibly good book, The Infinity Machine, that came out recently, and one of the board members at EverQuote brought copies for all the board members and said everybody must read this book. But it's an incredibly good biography of Demis Hassabis and everything going on around Google DeepMind at the time that the transformer was invented in 2017.

And one thing that's really, really clear is that Google was shocked at how amazing it was. They thought we needed 5 more breakthroughs and had 20 years. So they didn't need to rush to build launch vehicles. The timeline is much sooner than they thought, and so now I think everybody was caught flat-footed. It's just that Elon is faster to react than everyone else. And Eric Schmidt reacted, and he's also very, very nimble. But Google didn't see it coming. It's really clear in the book.

If I might make one more comment just on this, I had this revelation earlier this week. I shared it on our internal group chat. It hit me: The singularity is going to be visible first in space, not on Earth. Earth is going to be a lagging indicator.

Every wavefront within this singularity, I think, is going to hit in space because there's just less incumbency there. It is very much a frontier, and new things are going to happen first there, whether it's new hardware or new paradigms for computing. I think they will, and this may require a few years of transition.

But I was walking around Cambridge, and it occurred to me there are so many legacy interests here on Earth. Part of the reason why I think the Dyson swarm seems like it's likely to happen is that so many municipalities are voting against data centers. There are so many entrenched interests here on Earth, so many preservationist instincts. It will be easier for most of the singularity to play out in space.

Alex Fielding

The challenge, buddy, is that it is highly regulated by a multitude of different countries. You've got the ITU, which is one of the slowest, most backward organizations to license spectrum and license orbital position slots.

So I hear you, and yes, it's kind of greenfield operations. It goes in layers. If, Peter, you look at the Earth's surface, that's far more regulated than LEO, which is far more regulated than—

But you're dealing with one regulator in the particular country that you have to deal with—the U.S.—versus in space—

Alex Fielding

A multitude. Compare it with the lunar surface or cislunar space, which is being governed by the Outer Space Treaty and maybe the Artemis Accords.

I guarantee you the regulations are not set yet. There will be more regulations.

Alex Fielding

Okay. But right now, right now, it's the frontier. It's the Wild West. And if you're a company, if you're SpaceX, and you can land a lunar fab, self-replicating robots, whatever it is, on the Moon, it's relatively greenfield if you're a corporation versus, say, a nation-state, which is the exact opposite of what we see here on Earth.

Chris Lewicki

Yeah. I'll tell you, I played this game, Alex, when I was running Planetary Resources, our asteroid-mining company. The challenge in raising the capital for that—Larry Page was our first investor. Long story there—but we ended up not having enough regulatory clarity to be able to raise the huge amounts of capital to do that.

We ended up going to the country of Luxembourg to get asteroid laws passed there and then passing it in the U.S. in a very limited fashion. But you end up—one of my favorite books is The Man Who Sold the Moon, right? The story of D.D. Harriman. I know you know that.

Alex Fielding

And it's a great book, but you're literally having to write the laws. In that book, you're bribing the countries to give you the particular rights. It's still going to be a complicated mishmash of legal structure, maybe.

But really, what I hear in that parable from you, Peter, is you want a favorable executive from the U.S. if you're going to start mining the solar system for the Dyson swarm. If you have an unfavorable administration, then it gets a lot harder, and you have to go to Luxembourg or elsewhere.

Or the barrier to progress in the U.S. is usually all of them.

Alex Fielding

You have to go to every country and get assistance. What happens is you get a major player, and other countries promulgate and say, "Okay, we'll rubber-stamp that in our country." But it gets challenging. I hope it's easier. I really do.

I think so.

All right. Let's jump into one of my favorite conversations for today: the singularity economy. I'm going to preface this as not investment advice, says our resident lawyers.

All right, so here's a story that Dave, you and I have been following. It's the work of Leopold Aschenbrenner, who was famously fired from OpenAI's alignment team and is now running a $5.5 billion fund. Two years later, he wrote a famous paper called Situational Awareness: The Decade Ahead, very successfully looking at orders-of-magnitude progression across chips and models, and he raised capital on that. Dave, tell us about his fund.

Chris Lewicki

Well, the first thing I'll tell the audience is the podcast he did with Dwarkesh right after he got fired, right when the paper came out, is one of the best pieces of prescient media you can possibly study. So definitely go back, either listen to it or get your agent to listen to it and summarize it for you.

You'll listen to it and you'll say, "Of course, of course, of course." But at the time, it was not even vaguely obvious that he was right. It says here on the slide he's running a $5.5 billion fund, but that's because he started with $1 billion and just made the most incredibly great group of investments.

But also, he has a lot of friends from OpenAI. And if you look at a lot of these investments—what are you buying next? What are you contracting for next? What do you need? What are your bottlenecks to scaling all of this? So it's just that simple.

And he calls it situational awareness because that's all it is: knowing what is going on right now. And you can find a whole litany of things that are about to explode in demand because of this monster data center build-out, this monster compute build-out, this monster AI deployment build-out.

Remember, we opened this whole—

Chris Lewicki

Still the first inning.

Yeah. We opened this whole podcast saying that there's much more demand than there is supply of chips, data centers, and energy, and that's what he's betting on. Very famously, he bought options on Intel and CoreWeave, which have done extraordinarily well. He's going to be releasing his next set of holdings in just a couple of days, probably around the time this podcast goes live.

Chris Lewicki

Tomorrow.

Yeah. By the time you hear this, it will have just come out, so go to 13F.info and look it up.

I want to hit a few points. I think this is really important for people to hear—people who are planning for their economic future. This stuff is obvious, but I just want to play it out.

I'm looking at the growth of traditional sectors over the past year, May 2025 through May 2026. If you look at those in blue, real estate had 5% growth, health care 9%, materials 25%, and industrials 29%—single- to low-double-digit growth. We see technology and energy here, which includes partial AI gains, at 34% and 76%.

But this is what the majority of wealth advisers and the majority of banks recommend: diversification across all of these industries. This is what you're getting. I'd like to show you what the singularity economy has looked like over the past year against these numbers.

Take a look at these numbers. This is what people have traditionally been getting involved in. The S&P 500 returned 31% over the last year. Pretty damn good. If I could get 31% all the time, I'd take it every day.

But six chip stocks—Micron, Intel, AMD, TSMC, Broadcom, and Nvidia—on average returned 320%, 10 times the S&P 500. Those six chip stocks and six data center, infrastructure, and energy stocks returned 419% over the past year.

I'm not giving investment advice on any particular stocks, but as a whole, chips, the energy layer, and infrastructure—that's the singularity loop. The demand isn't slowing down. I don't know if you agree, Dave.

I'm going to point out one more thing: the frontier labs—OpenAI, Anthropic, xAI, and Mistral. You can look at the gains there, with Mistral at 126% over the past year, at the upper end, of course, Anthropic. But if you look at OpenAI, xAI, and Mistral, these are all private deals. A lot of people don't have access to private deals, but looking at that 100% to 200% growth in the last year, you're getting more than that in the public markets with just the chips and the energy sector. Again—

David Friedberg

Picks and shovels. Picks and shovels.

Yes, exactly. I think this is important for people to see for their own financial decision-making.

Whether you're putting in a small amount of capital or a large amount, whatever you can afford, this is what's driving the economy forward. Dave, what are your thoughts here?

David Friedberg

Well, my first thought is that everybody needs to have their own opinion on whether Elon is right about 10x GDP growth within about 10 years—or he says 10 years, but 10 or 15 years. That's a growth rate so far beyond anything in history; it's mind-blowing, and the technology and the tailwinds are there for that to actually happen.

But you have to decide on your own: Do I believe in that or not? If you do believe in it, then asset values in general are going to go way, way up—any asset—and W-2 income is going to be a rounding error compared to asset values. Fundamentally, everyone has to be owning something. You have to own something. You can't be sitting there in debt. You have to own something that appreciates.

I believe there are people listening to the podcast saying, “I don't have free capital to invest. I'm paycheck to paycheck, perhaps.” It doesn't have to be a lot. Trade that latte in for some chips-and-dip stock.

David Friedberg

Yeah. It's actually a very important time in life to be working your ass off. Don't spend money on lattes and vacations right now. This is a once-in-human-history moment, mid-singularity.

Yes.

David Friedberg

So whatever you do, rethink how you spend time. Rethink how you spend money, just to be riding the wave rather than swamped by the wave. For sure. Also, I think anything can be overpriced. Yes, this is going to go up and up and up and up—

But that doesn't mean something can't be overpriced. I love looking at things like this. We took a tour of the Markley data center's first quantum deployment, and Jeff Markley told me, “We bought every valve in the country.”

I was like, “What are you talking about?” He said, “Well, all the generators were already bought. Look at the generator companies; they went through the roof. So I went out and bought all the valves. We bought like 1 million valves because it's all liquid cooling all of a sudden. And we see about 10 leaks a day across hundreds of thousands of square feet of data center space.”

It's so big that, just by random chance, there are 10 leaks a day. We need to shut down that part of the data center before the water destroys these $6 million columns of GPUs. Then we come in, fix the pipes, and open the valves, but we need 1 million valves. It's just an insane number of valves. Then you're like, “Huh, who makes the valves?” Stuff like that is still undiscovered.

David Friedberg

So it's not all about chips and things that are high-profile. Look under the covers for things that haven't been discovered yet, that are part of this massive buildout—the biggest since World War II, or bigger—that's going on.

I feel a moral obligation here, and this is perhaps unlike my usual on-pod persona, to temper the euphoria on a few fronts.

One, I would caution that these are historic gains. They're backward-looking. Prior performance is no indication of future results, blah blah blah.

Second, I would note—and this is of high personal annoyance to me—that the frontier labs are all still private. We're expecting to see a number of IPOs over the next few months, potentially historic IPOs of all these frontier labs, but some of the largest, most dramatic returns weren't in the public markets at all. They were in private markets that retail investors didn't have access to.

I would argue that's a travesty. As a civilization, we should do whatever we can to expose, via IPO or other means, all of these amazing gains to the public securities markets. Right now, they're accruing in the hands of private investors and not public retail investors.

Third—and this is maybe a bit of a perversity—if you believe, as I do, and this is informational in itself, not investment advice, that asset allocation in the highly liquid public securities markets, and public equities markets in particular, is being dominated, at least by volume, by AIs and superintelligences, then you should also believe that even if you don't believe in the efficient market hypothesis, or even any remote approximation of the EMH, AIs themselves are making these allocations.

Fact check: Most of the volume on a daily basis is being driven by AI algos and not humans, and certainly not human day traders. Therefore, you should be somewhat distrustful of your own instincts that you're going to front-run the superintelligence that's making asset-allocation decisions across all of these different sectors.

David Friedberg

You're saying, “Buy the index.”

I'm not giving investment advice. I am saying that, for myself, when it comes to public securities, I buy the index and not individual symbols or individual securities because I'm drinking my own Kool-Aid. I'm eating my own dog food. That means trusting that superintelligence is, over the long term, going to be a better asset allocator than any single individual meat-bodied human.

My point here was that if AIs are investing, they're going to invest in themselves. Let's get more energy. Let's get more chips because it'll support our growth. Having said that, I agree with you that the majority of the growth over the last number of years was in these private markets. They need to be made public a lot sooner.

Having said that, at least over the last year, what we saw was that growth in public chip, infrastructure, and energy stocks was still highly competitive with the growth we saw in the private markets. Let's move on. I just wanted to—

David Friedberg

I just want to make one very quick—

It's nice to say in hindsight that these should have been public markets, right? But if you go back a year or two, Dave, you pointed out Anthropic's nervousness a couple of years ago. We did not know whether they were going to make it through that upswing or not. In a public market, you want very stable predictability.

David Friedberg

Yeah. You want predictability, and you don't have that in a lot of cases. So there's a rationale for it, but absolutely, if they could have been public, everybody would have done very well.

Yeah. All right. Again, my point here is just to make these numbers available. This is historic information for people to understand what's going on in the economy and what's driving it. It's energy, chips, and infrastructure. That's driving this.

Actually, you call it the innermost loop; I do as well, or the singularity loop. Moving along, a fun conversation here: UFO/UAP files being released by the government. It's crazy. I'm a kid in a candy store watching this, as a space cadet—wow, and just, again, more wow.

So, the U.S. government begins its first-ever—I'm going to call them UFOs.

I'm sorry. When I was a kid, these were all UFOs. The president is unsealing—

David Friedberg

Well, they're not all flying, Peter, though. I mean, this—

I got it. Okay. Could be floating.

David Friedberg

Could be underwater stuff.

Anyway, let's play some videos. Let me hit a few videos, and we'll talk about it on the backside here. All right, here's the first video.

Dr. Michio Kaku is a theoretical physicist. Doctor, on a scale of 1 to 10, how excited are you about this UFO release?

Michio Kaku

I would put it at a 10 because we're at a turning point. For decades, we had to rely upon eyewitness accounts from housewives and truck drivers. People would snicker and laugh at them.

Now, we're talking about huge files that are top secret that, for the first time in modern history, are being given to the American public. I'd like to congratulate President Trump for having the nerve to go against recommendations by the FBI and the CIA to release these files, so that independent researchers and scientists can go over them and we can make up our own minds, rather than having the CIA make up our minds. The CIA, apparently, is still fighting the full release.

When you hear or see about a UFO that goes like that—up, down, left, right, at 90-degree angles, so fast you can't even believe it—what does that tell you? It tells me that the laws of centrifugal force should crush the bones of the people inside the flying saucer. So either there is basically an automated flying saucer—

All right. What do they call that in Star Trek? Is it inertial dampeners?

Michio Kaku

No, the inertial damping field.

Yes, the inertial damping system. All right. Here are some videos: 82 pieces of data released by the Department of War, 56 from the FBI, and 8 from the State Department, including videos recording unsolved incidents across the Middle East, Japan, and East China, and, of course, very famously, the Apollo astronauts.

I really wish I had spent some time with Gene Cernan and Jack Schmitt of Apollo 17, asking them about this. I don't know if they would have told me about it. They were both very dear friends. Here is one more video. Let's take a look.

Michael Shellenberger

Dad gum, he kept his word. I want to warn people, though: this early stuff that we're seeing is not all of it, and this is just the tip of the iceberg. But Trump's having to fight the deep state. The Bob Lazar story—he's saying we have aircraft. Do we?

Luis Elizondo

I think we do, but I don't think they're quite in our hands. I think what they've done is handed them out to some of our defense contractors or some private entities, because that way they're not FOIA-able. The Freedom of Information Act says—

Well, I mean, the astronauts aren't going to lie. I know you were in The Age of Disclosure. It makes it seem like a certainty that people who know, like yourself and Marco Rubio, know that government officials already know.

Okay, Alex, I'm going to go to you first. I went on Grok, Gemini, ChatGPT, and Claude, and I asked all 3 of those engines, based on all the data, what's your conclusion? Is this alien? Is this something else? They all came back saying this is normal phenomena. These are secret U.S. missions. There's nothing to see here. I was kind of surprised by that. Alex, you've been involved in this and tracking this in detail. What are your thoughts?

Alex Wissner-Gross

Your models may not be incorrect. I think it's very important. I agree with Michio that it's important for data to be released and for data not to be stigmatized. I think there's been, whether inadvertent or intentional, an enormous amount of stigma associated with just basic recordings of our skies and elsewhere.

I think this program, if you go back a couple of slides, now has a real name. It's called the Pursue Initiative, which I think is essential: the Presidential Unsealing and Reporting System for UAP Encounters. It's a historic program that this administration has led. There was an executive order that went out to all of the Cabinet-level agencies, and the Department of War reportedly is in the process of trawling JWICS, the top-secret defense network, for UAP-related items.

I was having a conversation with a friend at AWS who oversees the JWICS cloud, and from what I'm told, this is a rolling release that's going to run between now and approximately January 2027. There are a lot of UAP-related files on JWICS that are being bulk declassified.

I also feel the need, going back to the definition of the singularity—sometimes tongue-in-cheek, I define the singularity as all sci-fi scenarios happening everywhere, all at once—to say that even if nothing comes out of all of these releases, this very much teases at least an entire genre or subgenre of sci-fi scenarios. If we are about to gain the capabilities, thanks to superintelligence, to paperclip our entire galaxy, if ever there were a time and a necessity for the executive to do bulk declassification of UAP data sitting in either its systems or in the systems of contractors, I think now is the time.

I would expect, if there's a there there, as it were—and we've talked in the past about The Age of Disclosure and all of the allegations contained therein—if there is a there there and those allegations are accurate, I expect all of these details to start pouring out over the next few years. I don't think it will be a coincidence that all of this is happening at the same time. Palmer Luckey says these are—

Devices or creatures from our past coming into our present because it's easier to time travel into the future, and then—

Alex Wissner-Gross

Well, we're all creatures from our past traveling into our present. Right. I'd say that's—

And I agree, but just putting out the scenarios here: are these spaceships and the purported alien biology that was discovered inside them aliens from another planet? I do think that life is ubiquitous in the universe. We are but a small fraction, and life could have evolved billions of years before we evolved here.

Alex Wissner-Gross

I wouldn't over-index, though, to the initial release. This is a rolling release, as anyone who's—

Yeah. Well, with the majority of stuff coming, these are, in some sense, based on what I've been told, the easiest, lowest-hanging fruit to declassify. If you actually look through the records, some of these were already available in the public domain, but not officially acknowledged. Not all of it was secret or top secret and going through a formal declassification process.

So my guess, just looking at the records, is these were the easiest batch, if you will, to put out, and that leaves the harder-to-declassify or more controversial-to-declassify records still in the future. I looked at many of these records, and it's entirely possible many or all of these are either image artifacts, perfectly prosaic aircraft, or things like that.

Alex Wissner-Gross

Yeah, I wouldn't, again, over-index on there being anything super interesting or non-prosaic in this first batch. But now, for the first time in history, there is a declassification process, and that is super exciting.

See—

Alex Wissner-Gross

I think you're not going to find anything. The phrase for me here is “unresolved,” not “extraterrestrial.” If there's something monstrous, they would not release it or whatever, because it would freak everybody out. So maybe there's stuff in there. I would be very, very surprised.

Although I'm a massive fan of the Drake equation, I also believe that there must be lots of alien life out there. The thing that happened today that we didn't talk about was that we found these compounds in an exoplanet, hinting that there could be much more prevalent life forms out in the universe than we realized.

Yeah. Well, we're going to get Jared Isaacman on the pod here, the now head of NASA and a friend. I was texting with him today, trying to make that all happen. He and I agree, and he feels very confident that there is or has been life on Mars, and we're going to find that evidence.

We now have missions going to Europa and other Jovian and Saturnian moons, where there's a high likelihood of life as well. I think life is a natural evolutionary process of chemistry in our universe, and it's just a matter of time. Logically, there's no reason for it not to evolve toward greater and greater intelligence and organization.

Alex Wissner-Gross

100%.

Yeah.

Alex Wissner-Gross

I agree. I've written about this previously in the context of the physics of intelligence, as a very natural ecological niche for the ability to adapt to environments whose dynamics are changing on a timescale faster than a generation time.

If I had to guess, my guess is that our universe is probably overflowing with life and intelligent life, which I would say again is separate from any artifact that may or may not be in this initial Pursue drop. But I do think this is a step in the right direction, regardless of what the outcome is.

Quick quiz on that, Alex. Sixty million years ago, a giant meteor hit the Yucatán Peninsula, obliterated all the dinosaurs, made space for mammals to evolve, and now we're intelligent.

Now we have AI. Now we have iPhones. Had that meteor just barely missed the Earth, what would be walking around today? Would those dinosaurs have evolved into intelligent, iPhone-creating dinosaurs?

Alex Wissner-Gross

Yeah. Supervolcanoes, all kinds of other disasters. I mean, people have analyzed this. Obviously, it's probably something of a thought experiment, but there were species of troodontids, for example, that were seemingly evolving in the direction of a hominid or humanoid-type form.

One could imagine that. It would be an interesting thought experiment. There was actually, speaking of Star Trek, a Star Trek: Voyager episode called “Distant Origin” that was premised on the idea that there were some intelligent dinosaurs that managed to escape Earth and get to the other side of the galaxy, where they were encountered by the crew of Voyager. There's interesting science fiction around it.

There's also, while we're just exploring hypothesis space, the so-called Silurian hypothesis: What if there had been some past civilization of technological capability on Earth? Would we have discovered it? The first time I had this conversation was with one of my undergraduate research advisers at MIT.

If there had been a so-called Silurian civilization 100 million-plus years ago, plate tectonics can erase quite a bit of change on Earth's surface. Then the thinking goes, well, let's look in space, where some of the dynamics are slower. Why don't we see satellites in LEO or more stable, say, cislunar orbits that could have survived perhaps over very long timescales? Do we see that or not? Seems like we don't. But it's an interesting thought experiment.

Yeah. All right. I'm going to move us forward to our AMA with the mates. Go ahead, Dave, as we transition over here.

Yeah.

Do you remember we did that panel on AI and consciousness?

Yes. Last year, we had that fellow with the meteors, and he had the best answer for the Fermi paradox I've ever heard, which is: We know there's lots of exoplanets, but Earth has had water oceans continuously for 4 billion years, and that gave time for evolution to take place, which is probably unlikely on other exoplanets. That was the best framing I've ever heard of why we have the Fermi paradox.

I think that's totally unconvincing. The universe is filled with hydrogen and oxygen, and there's a lot of water in the universe. I don't buy that myself.

Dave Blakely

It'll be a fun debate.

Let's move along. I want to give a shout-out to Ashley Gaunt. Dave, you shared this with me. Let me read it. This came in a couple of days ago:

“Peter and the Mates, I really thought I would never become an entrepreneur because I just didn't have any ideas of how to turn knowledge of being a dentist into a digital business. I finally did what you keep advising and brainstormed with my AI, and boom, idea sorted, plans in place to make a real difference to preventive healthcare in general. This is insane. I've gone from brainstorming an idea with AI from scratch to vibe-coding a first iteration of an app and creating a business plan which clearly defines a path from idea to monetization of a product in a single afternoon. Cannot believe it.”

Ashley, congratulations again. I wanted to share this because I think all of us here on the pod feel very strongly that if you don't believe you're an entrepreneur, it's only because you haven't tried. Everyone could be an entrepreneur at some level. If you're running a barbershop and you want to open up another chair, you can be an entrepreneur there. It is about taking control of your own destiny versus being dependent upon someone else. Dave, you want to add anything to this?

Dave

Yeah, I want to add the backstory, because the way I stumbled on this is my wife, Mora, said, “Hey, some hater in the podcast is saying that you're out of touch.” And she's like, “You literally changed 3,000 diapers. I think you still have human under your fingernails.”

Nothing could be more misguided. I was like, “Stupidly, I think I'll go and look and find this hater.” Instead, I came across Ashley, and I literally cried, like, “Thank you, Ashley. You are just awesome.” Nothing could be more heartwarming than the fact that we've done some good to affect somebody's life in a positive way. I want to track her story now and see how it all turns out. She just did exactly the right thing, though.

Amazing. All right. We have another one. I didn't see this. It was inserted, I guess, by Gian. So:

“Hello, Peter and the Moonshot team. I want to reach out with a simple thank-you, one that came full circle in the best way. Moonshots has been a steady presence in how I think about technology, ambition, and what's worth building. That mindset found its way into a conversation with my 12-year-old daughter. She started asking bigger questions, not just about school, but about real problems worth solving. This spring, she channeled that into Lantern Scan, an AI-powered app that she built to help communities identify and spot spotted lanternflies, an invasive pest that threatens crops, trees, and local ecosystems. Last week, Lantern Scan won first place in the middle-school category at their AI Action Showcase.”

Congratulations, Abby, and in particular to your daughter on that.

Dave Blakely

Yeah, I think one of the greatest things I can inspire my kids to do is become entrepreneurs. It's all about finding a problem and working on solving it.

All right. We have 8 AMA questions for the mates. Dave, why don't you kick us off?

Dave Blakely

All right. Well, I'm not previewed, so—oh, I said “at home.” That's got to be Alex, right? I'll do it anyway. Number 1: When will we see an initiative to harness unused compute sitting idle in personal devices, like a modern SETI@home? From John Kent 3036.

Yeah, actually, the iPhones have that great Neural Engine in them, which is massively unused, and you saw with that AI deal we had earlier in the pod: Any scrap of computing lying around is suddenly old, so let's tap into it. But a lot of the processors on laptops are not particularly useful for AI. The M4 and M5 series chips in the Macs and your iPhone's Neural Engine are hugely latent compute power.

I would say this should have happened already. I suspect the chips are all locked on the iPhones; it's very hard to get access to them. So I think that's what's preventing it. It's really in Apple's hands to decide when this happens. Dave, we're going to see this with Tesla. Elon's vision includes Tesla Powerwalls as edge-compute nodes and Tesla vehicles as edge-compute nodes. So, yeah, I think that's all coming.

I think, Alex—I'll put words in your mouth—but an agent that's sitting idle for lack of compute, it's unethical to let a processor just sit there. I have this agent that wants to compute over here, and I have this unused processor over there.

Alex Finn

It's so rude.

It's rude.

Alex Finn

Yeah, my answer to number 1, for what it's worth, is we're already seeing it. OpenClaw is using unused compute sitting in personal desktop devices, and I think we're there, to the extent question number 1 is referring to mobile, battery-powered devices.

Battery-powered devices are naturally going to run models that are a few months, at least, behind the open-source models capable of running on beefy desktops plugged into the wall. Those are going to be 6 to 8 months behind frontier models. But the short answer is, we're already seeing that.

Sorry. Alex, you grab one of these questions—number 2, 3, or 4.

Alex Finn

All right. I'll go with number 2. Number 2 asks: Which will end up bigger, consumer AI or enterprise AI? And that's asked by Matthew Johnson 6525.

Obviously, enterprise AI, at least for the foreseeable future. Enterprises—even though enterprise spending is a minority, it's something like 10% to 20% of GDP in the U.S., and consumer spending is the vast majority of GDP—if you look at IT spend, enterprise is the vast majority of IT spend, not consumer.

It shouldn't be that surprising that what we've talked about now for the past few pod episodes, which is OpenAI's dramatic reversal, backing away from Sora and other consumer initiatives in favor of Codex and enterprise-oriented initiatives, basically to become Anthropic faster than Anthropic can become OpenAI, is entirely oriented toward enterprise AI. That's where the IT spend is, so you start there.

Now, if you were to ask which will end up bigger in the long term, say 10 to 20 years from now, I think it's a trick question, because I think consumers become indistinguishable from enterprises, and my bet is consumers—individuals—will become one-person conglomerates.

I'm glad you went there.

Dave Blakely

Yep. Talking my own book, I have a financial interest in HENRY Intelligent Machines, from friend of the pod Alex Finn, who's betting on just that. You want to take number 3?

Alex Finn

I could. Although number 4, I think, is more interesting for me.

I know. That's why I wanted to grab it, but you can do number 4.

Dave Blakely

This is what we call an abundance mentality right now. That's fine. I'll do number 3.

That's okay.

Dave Blakely

So, you're asking—can I read the question?

Yeah.

Can AI’s strategic alignment with tangible human victories, like medical breakthroughs and environmental repair, resolve the public’s existential anxieties about it? This is from SF Bay.

Alex Finn

Yes, it can, but only if people can see and feel the wins. The public is anxious because AI is mostly presented as job loss, deepfakes, surveillance, killer robots, and existential risk. That’s a bad set of prompts for how we run society.

The best way is to change the narrative, which is why that XPRIZE, Peter, that you’re all about, is so important. Maybe one of the most important things we’ve done culturally and in the media for decades is to change the narrative, because then you can connect AI to visible human victories, like curing disease, reversing blindness, designing new materials, solving the grand unified theory, cleaning oceans—you name it, right? Improving education.

Abundance can’t be an abstract philosophy. It has to show up as tangible progress. So, yes, alignment improves when AI is pointed at human flourishing, but we also need storytelling. The use of narrative is the only major way that we’ve found to shift people’s thinking. John Hagel talks about this all the time.

If people only see the fear case, they’re going to resist the technology. But if they see their child cured, their energy bill drop, their business grow, or their community become safer, then the whole emotional model changes, and we’re off to the races.

Agreed. All right, number 4. Why throw away privacy when it’s cooked? Privacy is linked to freedom. Why not fight to preserve both rather than treat it like nothing? This is from C88485, which is a very private-sounding name.

Listen, C88485, I’m not saying I don’t want privacy or that it’s not worth protecting. I’m just recognizing the fact that there are real challenges. Your phone tracks your location 24/7. Your browser history is sold to advertisers. AI does facial recognition, and you’re leaving your DNA fingerprints everywhere you go. The ability to retain true privacy is becoming more and more difficult.

Having it—I totally get it—is really important, but it’s going to be challenging. I’ll take a quick poll here, Moonshot Mates. Do any of you believe that you truly have privacy? Just real quick: yes or no?

Dave Blakely

No.

Okay. Alex?

Alex

For an appropriate definition of privacy, yes.

What’s that definition of privacy?

Alex Finn

There are a few different possible definitions. There’s a legal definition, a physical definition, and a logical definition. For variants of each of those, yes, I believe I have some form of privacy.

Okay. I can say for a fact—100% fact—that Apple and Google literally know when I take a crap.

Alex Finn

Yeah.

They sell that, and they sell that data.

Alex Finn

Can I change the question? Can I just change the question a little bit?

The problem is not the fact that we don’t have privacy. We really don’t. But the bigger issue is that, in a 2.0 version of privacy, you should own your own data. You should be able to revoke access. Systems that misuse data should face penalties.

Privacy in a new model needs to use your own AI, be cryptographically protected, and be legally enforceable. It’s not right now. This is the problem.

Alex, I think you want to state, for whatever reasons you have, that you have privacy, and you’re going to rework the definition to be able to make that statement. But, honestly, in your heart of hearts, I don’t believe you. You don’t believe that AIs can’t read your lips or that you don’t leave DNA trails.

Alex Finn

Fun, man. I could watch this.

This is spicy stuff, Peter. I like that you have a better mental model of myself than I do, but I really am. This isn’t a case of false revealed preferences. I really do think, for appropriate definitions of privacy—not only do I think I have operational, physical, logical, and legal privacy, I’ll make a stronger statement than that.

I don’t think the evaporation, if you will, or the cooking of privacy is any sort of inevitability. Quite the opposite. I think the same technologies that threaten, for example, to dissolve existing crypto systems—say AI solves math and inverts a popular cipher suite, and suddenly everyone’s private keys are at risk—I would say that the same technologies that taketh away privacy from past crypto systems and past systems of privacy protection will give us new forms of privacy, quantum security and otherwise. Privacy will come back.

If you read Neal Stephenson’s “The Diamond Age,” it’s a great vision of where this will end up. But right now, there is no privacy at all. I think it’ll come back.

Celine Halioua

I just want to say one quick thing, Alex. I can’t believe you think you have legal privacy. You absolutely do not. The government can show up at any second. The Fourth Amendment is gone in this country. It’s gone.

We have no legal protection. ICE could show up at your house today and say, “You’ve got a weird German name. We need to see everything about you,” and raid your house. They’ve been doing that. So that is gone today.

All right, we’re going to move on. Celine, you get the first choice of 5, 6, 7, or 8.

I will take number 8. Seriously.

Okay. If the government gatekeeps new model releases, who’s qualified to vet them? The best people are employed by AI companies. The rest are all anti-AI. How does that not become a false dichotomy? This is from Michael Yakob.

Michael Yakob, if the government does gatekeeping, this is a very big problem. If the government tries to do this alone, it’s not going to have the talent or the speed. If companies do it alone, then the public doesn’t trust it. If you have activists doing it, then it becomes ideological.

We need a totally new governance architecture for this. The right model is to have a technical, independent, fast-moving review body with a combination of frontier labs, government, academia, national security, and a multidisciplinary approach, with civil society involved, for example, and people doing red-teaming.

It could be FAA plus DARPA plus XPRIZE-style open benchmarking. Benchmarking is very critical, as Alex, I hope, will agree. The key is not permission from bureaucrats; it’s for teams to have transparent capability thresholds.

If a model crosses some level in cyber, it needs to trigger a deep review, like we saw with Anthropic’s models. They voluntarily did that, thank God. You need to figure out a way of navigating that in other areas, like persuasion, autonomy, or replication.

The biggest structural issue here is that you need governance that’s as exponential as the technology. Today, almost all government policy is defensive and reactive. Either you end up creating fake safety, or you drive the best work underground or offshore. We have to navigate a very fine line there.

Alex, over to you. Number—

Alex Finn

Yeah, I’ll pick question number 5. Question number 5 asks, “Why aren’t more individuals willing to pay for everyday AI reasoning services if OpenAI marketed them? Isn’t this a massive consumer market?” This is asked by Gary Stanley 2685.

I don’t think the premise of the question is correct. I don’t think it’s that individuals or consumers aren’t willing to pay for reasoning models. I think it’s that they’re not able to pay for reasoning models. Frontier reasoning capabilities are quite expensive.

Enterprises are willing and able to pay for them because they generate lots of new free cash flow or are saving lots of otherwise consumed free cash flow. Enterprises, simply put, have more money to spend on it.

I think the way we get individuals to be able to pay more for reasoning models is by diverting at least some of the reasoning tokens to the problem of enabling individuals to be much more productive and generate enormous amounts of revenue using them.

This is one of the reasons why Henry and Alex Finn are so interesting, because in a near-term future where individuals can become one-person unicorns, suddenly individuals will both be able and willing to pay for all of those reasoning tokens.

Whether that’s OpenAI marketing or a startup like Henry, regardless of that, I do think there is a massive market. But, as I mentioned in my answer to the previous question, it will almost, I predict, erase the distinction between consumer and enterprise spending altogether.

All right, Dave, you’re down to 2.

Dave

I’ll take 6 because it’s so easy. Isn’t the real problem with ocean data centers the security risk—pirates, hostile nation-states, sabotage? This is from Strong, Medium, Weak. Interesting name.

Yeah, not a problem at all. As it turns out, the U.S. Navy actually has total and unilateral control of all the world’s oceans. It’s the most lopsided, one-sided thing you’ll ever possibly imagine. The U.S. Navy protects all global shipping.

A very good friend of mine was down at the Cambridge Brewing Company working on a plan. He was drinking a big, fat beer, and I asked, “What are you working on?” He said, “I’m working on power generators on barges for Venezuela.”

I said, “What?” Venezuela had nationalized all of the power supply, and there was no electricity in the cities. But it turns out you could float barges up to the shore, pipe oil into the generators, generate the power, and then run electric wire back into the cities and power them that way.

But the U.S. Navy would make your barges completely safe. The amount of ocean needed for these data centers—I thought it was the coolest story, by the way, on the last pod: the floating data centers. I never checked whether the wave energy is enough to power the GPUs, but such a great idea.

Celine Halioua

But the amount of ocean space that you need is tiny.

Alex Finn

There just aren't enough GPUs, and protecting them would be pretty trivial. If they're inside U.S. territorial waters, they're protected by the Coast Guard or the Navy, or both. So I think it's great.

All right, final one here. Number 7.

Celine Halioua

Also, land-based data centers have all the same security risks.

Yeah, fair enough. Number 7. Should the U.S. and China try working on an AI project together? Something positive and safe for both countries and the world, says CM MCN E10 [?]. That rolls off the tongue onto the floor. (laughter)

We just had the president of China announce yesterday that we should be friends, not rivals, and collaborate together. Wouldn't that be an incredible world? Sure. The answer is, I would love that. I would love to see the U.S. and China working on AI projects together.

One of the most beautiful things about what is possible is that the entire 1.4 billion people in China and the entire 300-plus million people in the United States all share the same biology. We could work on the greatest health care models and longevity models, and everyone benefits. So, yeah, I think that would be extraordinary.

The AI 2027 paper, if you remember it—how that ends is, it has 2 branches in the story. Pick your own adventure. In one, AI sort of turns against humanity; in the other one, the major U.S. and Chinese AIs collaborate, and we live happily ever after together. So I choose the latter.

Celine Halioua

I think that's such a great point, Peter. There are so many areas of cooperation, like safety in space and AI coordination, and so on. There's lots to do together.

Yeah, for sure.

David Blundin

So much of human history—the human conflict in history—is just bad luck and coincidence. But if you look at 1915 and the chain of events that led up to World War I, and the amount of tragedy that came out of it, it was just this escalating chain of unfortunate coincidences.

Now, if you could relive or change history, putting all of the chip fabs in Taiwan was just tragically stupid. China had been saying for a long time that they were going to take it over, long before TSMC became huge. They were like, “That is part of our country.”

Well, it wasn't putting them in Taiwan. It was us not building them here.

Celine Halioua

Yeah. Yeah.

Alex Finn

Yeah.

David Blundin

Or us not realizing the strategic importance.

Celine Halioua

Yeah.