Sonnet 5 发布、中国4900美元机器人、Fusion首座电站获批｜Philip Johnston 对谈 | #268

It’s pretty clear that Sonnet 5 now is a way to kind of fill this gap until Fable 5 is back out. A kind of mediocre capability at a high price point, but people will still need to buy it. Anthropic’s flagship model, Fable 5, has been offline for 15 days because the US government pulled it. Now, Axios reports it may be back within days.

Historians will look back and say this period marked the period toward the middle or the endgame of recursive self-improvement. Helion has cleared the required Washington state regulatory approvals for its Orion fusion power plant. It looks like fusion is finally here.

Alex

If you were watching the right metric or the right figure of merit over the long term, you could predict when this is going to happen, and it’s imminent.

All right, mates. Let’s jump into data centers and space. And for that, we’re pleased to bring a friend on.

Thanks so much for having me. It’s a huge honor. I’ve been a longtime fan.

Now that’s a moonshot, ladies and gentlemen.

Welcome back to Moonshots, everyone. Your front row seat to the coming singularity in the age of abundance. I’m here with Dave, our managing partner of Link Exponential Ventures and number-one funder of MIT and Harvard AI startups; Salim, our global trotter; our CEO of OpenExO; and, of course, Alex, our in-house ASI. I’m Peter Diamandis, your host and your abundance evangelist. Gentlemen, good afternoon, good morning, good evening, wherever you are.

So, where is Waldo today, Salim? Where on the planet are you, and where have you been?

We dropped our son off at a camp in San Sebastián, in Spain—or near there. I’m in Biarritz right now for a few days.

Okay. And you were in Germany before that. I mean—

I’ve been in 5 countries in the last 3 days. It’s really been nuts.

Of course. Of course. And, Alex, all well with you?

Alex

Yeah, I’m GDP-maxing, or doing my best. Always be GDP-maxing.

I’m happiness-maxing, gratitude-maxing. And Dave—

Good to see you, Peter. I am at Link Studio. Just a word for the wise: we have a ton of Northeastern and Princeton activity, in addition to MIT and Harvard, these days. Of course. We had TechCrunch killing it out in San Francisco. Lip-Bu Tan, Andrew Feldman, the number-two guy at NVIDIA—they were all there on Friday talking to the troops. It’s really rolling.

Amazing. This is the thing we did with Eric Schmidt, Eric Bolson[?], and Daniel Larus[?].

Today, we’re going to cover a bunch of new stories. We’re going to catch up on robotics, energy, and data centers. We’ve got 20 stories across 6 fronts. A lot’s happening, and a lot of capital is flowing. All right, let’s jump in. For those of you joining us for the first time, our mission here at Moonshots is to keep you informed and keep you optimistic about the future that we’re creating. I have 3 stories I want to hit on the abundance front.

Wait, wait. I’ve just got to make a quick point to everybody watching. If you’ve not seen the last episode with Emad, I’m about three-quarters of the way through it. It’s tough trying to keep up with our own episodes when I miss one, and it was ridiculously amazing. Just a comment there.

Yeah, and Emad’s made some great releases in the last 48 hours on his latest model combinations. All right. So, we’re going to open our first story on robotics.

Before we do, because I want to talk about predictions for how many robots we’re going to have on planet Earth, I was having a conversation with a dear friend, Ramez Naam, today. Ramez is one of the earliest Singularity University faculty members and futurists. He’s extraordinary, and he shared this chart with me that I’d love to discuss. It’s a look at how experts consistently underestimate exponential growth.

In this chart on the left, we see new solar growth, and that yellow exponential line is the actual growth in solar. It’s been growing at an extraordinary rate. What we see in these departures that go horizontal are the predictions that the experts make every year, showing linear or just small incremental amounts of growth. Over and over again, they underestimate it.

The chart in the center is the experts’ predictions on EV growth, and again, we see exponential growth in EVs. The forecasts consistently underestimate the actual growth. Finally, we see the same chart going on in battery sales.

It’s an interesting phenomenon that, while we’re living in this exponential growth, the experts—who are the experts in the way things used to be—are not projecting the growth. They’re staying very, shall we say, sublinear in their estimates. Salim, you and I have seen this before, and we’ve discussed this. Any thoughts?

Oh my God. Every presentation I ever give has a segment with several slides showing this. The poster child is a story that came out in 2013 saying Moore’s law would end by 2022. You can go back and look at the technology press, and every 2 years, that article appears. It’s been happening for 60 years, right?

Experts are really good at measuring the technology. They’re terrible at measuring the compounding ecosystem around it. There are some really dramatic examples from Ramez around the energy ones, where solar is vertical and every expert for 10 years goes horizontal in that thing. It’s an endemic problem.

We have that whole headline in the original book, Peter, which we put together, saying, “Beware the experts,” right? And this is the immune system, because when you’ve got somebody with 30 years of experience in something, they’ll tell you how not to do something.

Yeah, agreed. Alex, any thoughts on this particular note?

Alex

No, I think the moral of the story is you should always take the logarithm of the actual history before you hand it to experts for their linear extrapolation, so you can get the right answer out.

Cute.

Alex

I define an expert as someone who can tell you exactly how it can’t happen, right? It’s so true that experts today are so ingrained in the past, because if there’s a disruption, if there’s a revolution that comes, they’re no longer the expert, and so it’s against their best interests.

I want to tie this story to our first robotics story here, and it comes with 2 predictions. The first was Morgan Stanley. Morgan Stanley had originally predicted 14,000 Chinese robots coming out of China, and they upped it to 26,000. Now they’ve just upped it to 50,000, projecting 500,000 robots by 2030.

But the fact of the matter is, there are 140 humanoid robot companies developing hardware in China today. At the same time, you’ve got Elon projecting tens of millions—50 million robots by 2030—and billions going into the early 2030s. On the flip side, what we’re seeing here on the right-hand side is a chart from Andreessen Horowitz that shows we’re going to be seeing about $16 billion of hardware investments in Q1 of 2026.

The US is finally catching up. Alex, I know that you’re heavily committed to this, and Dave, your thesis is that we’re moving from an AI-centric entrepreneurial ecosystem to a hardware-centric ecosystem. I’d love your thoughts on this. Alex, you first.

Alex

As I’ve mentioned on the pod in the past, I think superintelligence is set to spill out of the data centers into the streets. I think the most obvious vehicle for that is autonomous vehicles on the one hand, and humanoid and near-humanoid robots on the other.

I think as we start to increase the number of humanoid—or just, say, general-purpose—robots per capita, there are going to be certain regimes. At the low number of humanoid robots per capita regime, it looks like robots performing industrial applications: robots in factories, robots doing logistics. As we start to get closer to approximately 1 humanoid robot per capita, it looks like domestic robots everywhere. It looks like an iRobot-style regime where everyone has domestic staff.

Interestingly—and this is a point that I don’t think I hear frequently enough—as we start to push well through the approximately 1 humanoid robot per capita regime to 10 or 100 humanoid robots per capita, at that point, I think, to a hobby horse of Salim’s, we start to end up in some pretty exotic futures where there’s no longer necessarily justification for the humanoid form.

We end up with microrobots and nanorobots. There’s a natural sense in which the humanoid form is no longer natural in a world where we have 1,000 general-purpose robots per capita, and we can start solving all of the grand physical-world challenges that would maybe be uneconomical if we only had 1 humanoid robot per capita.

I think we’re going to very rapidly scale through the low-per-capita regime, to the approximately 1-per-capita regime, to the many-per-capita regime. It’s going to be a very exciting scale.

So, Dave, take a second and walk me through your thesis right now. We’ve been investing together in AI companies mostly, and you’ve said that you expect that to sort of fall off, with more investment going into hardware in the next couple of years. Why?

Yeah, I think we have to think in terms of 10-year investment themes. Ten years in the age of AI is like 100 years in any normal world.

I do think the next 1–2 years are still dominated by white-collar automation, AI algorithms, chip design, AI that designs chips, and the beginnings of data centers in space. But if you think beyond 2 years in the future, what are the investments that are going to really be big 3, 4, 5, 8, or 10 years from now? The automation of construction of data centers is all going to be robotic.

Biotech, chemical mixing, experiments, reading gels—that's all going to be robotic. And so, the machines that do that—you take a guess: How big do you think the U.S. gutter-cleaning industry is? People who go on your roof and pick the leaves out of your gutter. How much do you spend per year? I do. I looked it up.

Alex

How big?

That use case alone is a billion dollars a year. If you built a robot that cleaned gutters and washed windows, those 2 tasks would represent a $20 billion global market. A $20 billion global market. So that's a theme that will support this. This slide says 140 humanoid robotics companies. There's room in China.

Alex

In China?

Yeah. There's room for thousands and tens of thousands of robotics companies specialized for various use cases, everything from wafer movement inside a chip fab to chemical mixing and biotech to gutter cleaning to just everything. Construction has thousands of individual tasks, so it's a very broad, long-term investment theme.

Kitchen work, too. We're already starting to invest in that one. Kitchen-work automation is compelling because fast-food restaurants can buy at scale, and so they'll co-develop with you. This is where Travis Kalanick is focused with his company, CloudKitchens, just fully roboticizing kitchens.

You know what's interesting is this particular note: $16.5 billion in the last quarter was in 500 deals. I used to think that—I'm a huge Star Trek fan—I used to think that Star Wars was kind of silly, with all those hundreds of variations of droids out there, but it looks like it's coming.

Peter, it's such an interesting point. I'd be curious to hear your thoughts on this. Historically, the robots have been missing from Star Trek, other than in the first few series, like some type of Data robot. There were no robots. Maybe it was—what are your thoughts on that?

I think they wanted to create a very humanistic series, and Gene Roddenberry was all about the societal implications of technology in this future. Even the computer—right, the computer, whose voice was Gene Roddenberry's wife, Majel, playing that role—was very roboticized. It didn't predict these incredible empathic voices we have on our models today. Yeah, they missed that.

Alex

LCARS, right?

Well, you know, we're going to have Rod Roddenberry with us at the Moonshot Summit. He's going to be on stage, one of the judges of our Future Vision XPRIZE. So we'll sit down and ask him. I think those are important questions to ask: Why did they miss that part of the future?

Alex

Was it just a low production budget? Because they added them in more recent Star Trek series.

Yeah, they retconned them. But still, nothing close to Star Wars. And I think a lot of it had to do with the theme, right? Infinite diversity, infinite creativity. They were focused very much on human interaction—all the aliens that were humanoid, not just because you could put actors in those suits, but because you were always dealing with the interpersonal elements in the plot lines.

So, yeah, we've got lots of companies that are building robots now. Famously, in the U.S., we've got Tesla with Optimus. We have Figure, we have 1X, and in China, probably the robot company that's gotten the most publicity has been Unitree.

We had one of the founders of Unitree on stage with us at the Abundance Summit. I want to play their latest video. It's got 11 million views, and this is a glance at the R1, which is incredibly cheap, selling for $4,900. I mean, this is the price of a cheap used car, which is saying a lot. So let's take a look at the R1.

So, I mean, it's crazy. My point is, it's an extremely capable robot, but the work here is going to be on the software layers, right? You buy this robot. I don't think it does that out of the box, but I think you could probably buy the algorithms that enable you to do that.

But I see an explosion in the number of people experimenting with these robots. At $5,000, that's affordable by almost anybody who has a reasonable income. Alex, where does this go for you?

Well, I think the elephant in this particular room, as with Elon demonstrating to the world that you could drive the cost of heavy lift to LEO down to effectively near zero through reusable, propulsively landing rockets, is humanoid robots that are able to assemble other humanoid robots.

If you take the cost of assembly down to near zero, what we're left with is the cost of raw materials and the cost of energy. And that's going to be effectively de minimis. So I think, as we drive—and by “we” in this case, I really mean Chinese organizations, because the West is woefully behind at the moment—as humanity starts to drive the unit cost of general-purpose robotic embodiment down to near zero, at some point we will need to cross the threshold of robots being able to assemble other robots in order to keep driving that cost down.

And then, at that point, we have physical labor too cheap to meter. As I pointed out in the past, approximately 2/3 of the service economy constitutes some sort of physical labor. We can do for the physical world what AI agents are right now in the process of doing to knowledge work, which is basically driving the cost of knowledge work, and soon physical work, down to near zero.

Can I make a point, actually, that we learned at the Gigafactory that I completely think was lost on me, and I think it's critically important? When Elon says humanoid robots building humanoid robots, the CNC milling machine or the auto lathe already exists. It's already making the parts; it's just a file loaded into the machine.

Those machines were designed for a human to go and take the part out of the machine and put it in the next machine. And so the humanoid robot does not need to literally create something with a file and a piece of metal; the automation's already there. It just needs to do the part that the human is doing today, which is moving the part from machine to machine and doing the final assembly.

So it's a much easier problem than “robots making robots” sounds like.

Yeah. I think the point I want to make here is we're heading toward commodity pricing on these things. $4,900, right? And so, as we move to commodity pricing, the question is: Where's the value layer?

Well, the value layer clearly will be in the software, and the exact point I want to make is—yeah, I think that's a really important point, and people are going to build apps. Look, if you go back to the original personal computers, you put them out there, and we didn't know what people would do with them. Then, over time, you had more and more applications built.

Now these things show up worked out of the box. There will be profound new skills emerging all the time from these things. I think it's going to take a lot longer than people think, because driving took us 20 years, and that's a very bounded domain space. Humanoid robots doing gardening, et cetera—there are a million edge cases.

What's different here is that—and this is sort of, what does it mean for the entrepreneur out there? You can buy this and begin to build on top of it. This is not something that requires permission from anybody. It's not something that requires a massive corporate budget.

An entrepreneur can buy this for $1,400. It's like the Raspberry Pi moment, and you can start hacking and publishing software to these robots that becomes a new revenue engine. I think that's what's most interesting for me: the explosion of applications that came on top of the iPhone now coming on top of these robots.

I agree, but I still think we're going to spend a lot more time figuring out the industrial use cases—the dull, dirty, dangerous jobs—and getting those automated. There's so much scope. It's going to take a decade to get through that before we can get to somebody coming over and doing gardening for you.

You know, we talked in a past pod about Royal We. I mentioned the idea that the export controls we're seeing at the moment imposed on frontier models are, in some sense, regulations on importing foreign—or exporting, depending on your perspective—superintelligence, if you look at them through the lens of immigration policy. But that's the software layer.

I do think—admittedly, maybe this is a spicier take—that as the cost of robotic embodiment, primarily from China at the moment, starts to come down, I would not be surprised to see similar import controls or other national-security-motivated restrictions start to kick in. It's not just about dumping. It's also about taking the embodiment for superintelligence and moving them across borders.

Well, you've heard the Trump administration saying they want to invest in the robotics industry. We've seen, obviously, Tesla and Brett Adcock from Figure getting massive investments to support the growth of these systems. It becomes strategic for the US, and we're going to find out in our next story. Let me just go there for a second, where we're going to start to see robots being used in a number of different areas, including law enforcement.

A drone, of course, is a robot. I'm going to share a particular video here that Alex shared with me. This is out of Orlando, and this is US law enforcement beginning to use drones as first responders. Let's take a listen to this video.

News Anchor

And new at noon, the Orlando Police Department is now using drones as first responders, sending them to some 911 calls to give officers a live look at the scene. A new eye in the sky is now responding to some of Orlando's most serious calls. Orlando Police Chief Eric Smith is addressing a big question from the community.

Community Member

What would you say to the citizens who either are seeing this as an invasion of privacy or an overstep?

Eric Smith

We're not looking in people's windows. We're not spying on people. We're not just flying around just to fly around.

News Anchor

When a qualifying call comes in, a nearby drone can be dispatched to specific GPS coordinates. Then, from the crime center, an FAA-certified pilot can control the drone, giving officers a live view of the scene, including whether a suspect runs, hides, or possibly has a weapon.

So, interestingly enough, this was demonstrated on June 17 as a first deployment. It had what they call 9 docks at different locations and 11 Skydio networked drones. Skydio is sort of the US manufacturer today. It used to be DJI out of China, but we put in import controls on DJI for a number of security reasons. And so, even though Skydio is more expensive—something like 3 to 10 times more expensive than DJI drones—they're getting it.

Interestingly enough, Rick Smith, who's one of my Abundance members, is the CEO of Axon, the company that makes the Taser and the body cams. Axon has the contract here, and they're coordinating all this. They did a trial with a single drone, and it beat patrol officers about a third of the time in getting to the live location. It provided useful information 97% of the time, they claim.

But you can imagine that we're going to have drones on buildings throughout the city, and a drone will get there almost immediately. I can imagine there's fear that people have on this subject, but I think being able to get to emergency locations, making sure you can assess a situation when you need to have medical personnel there, is going to save lives. I want to acknowledge the fear that people have, but I think, on the whole, as long as you have good policy and good governance about how the data is retained or used and who gets access, I think this is an important step for law enforcement.

Alex

It's huge, and we did our very second sprint back in 2015 with Interprotección, which is the largest insurance company in Mexico, with 30 million users. They actually deployed drones because they were so much faster than ambulances at getting to an accident scene, and they would scan and map the whole area so that, before people moved the cars and everything like that, you had full information before anything happened. It was kind of an amazing experience.

So we've seen this trend over a long period of time. I think we can expect to see this accelerate pretty radically just because of the practicality of it. Alex

If I might add, there are regimes where, as the cost of robotic embodiment trends toward near zero and becomes too cheap to meter, we in the West are unaccustomed to this, but China, for a variety of reasons including demographics, has seen it for a number of years.

For example, this is well publicized: the Chinese Communist Party maintains party members or officers on a per-block basis. Certainly, this was the case during the pandemic in China. The West doesn't really have any concept of this.

If you look at how first responders are geographically distributed in the US, it's on a per-municipality basis or by precinct. There's no notion of, say, 1 or more officers per block who are just permanently stationed on a single block. But with drones, this becomes possible. We could have literally drones as densely distributed geographically as fire hydrants are. You could literally just, if there's a problem in a block or a part of a block, remotely activate the drone, and then you have an instant point of presence.

Let me share the second story; it comes out of Sacramento. There was a suspect who had a knife, and they deployed a drone with a magnet to grab the knife and make the scene safe for the police to enter. Let's take a look at this.

Here we see the drone going. The guy is holding a knife. A magnet, or electromagnet, is attached to the knife, and it pulls it out of the guy's hand as he's apparently sleeping. This is interesting. This is a drone that's disarming somebody.

If you guys remember during the Abundance360 Summit, Rick Smith, again with Axon, showed us his Taser-equipped drone, and I went onstage wearing a suit to protect me, and he tased me from his drone. I think this is coming: this use of drones in law enforcement, both for observing and for trying to de-escalate a situation.

The drone market right now is about $100 billion around the world. That's going to be increasing. We're seeing drones being used, obviously, in Ukraine, very famously. Eric Schmidt's been funding a drone company to help the Ukrainians in their fight for independence. Dave, any thoughts from you?

Dave

Yeah, Scientific American did a great research study on why crime rates are down by half in the US, and they keep coming down. It was entirely connected to deployment of first responders in the right place at the right time, largely driven by GPS. But now, with the drone footage, you can get much more accurate.

You know, the first responders want to be there and they want to help, but it's about getting the right people to the right place at the right time. I'm a huge believer that the video footage is going to be massively impactful, and the resolution just keeps going up and up and up and up. The physical side of it, where you're disarming somebody, is a little ways out. You know, that was kind of a—

Niche case.

Dave

Yeah, very niche case. But the video side is right here, right now. And as Alex was saying, you could easily, cheaply have as abundant a fleet as there are fire hydrants. That would cost next to nothing.

Alex

Yeah.

So that's imminent. I'll also point out, if I may, this was all foretold by Minority Report. You remember the scene in Minority Report with the police officers deploying spiders to search an apartment complex for Tom Cruise's character. We're starting to catch up with that now.

Right now, it starts with a couple of police precincts in the US. They're using drones for first response, and my understanding is that it's far more frequent in China right now than in the US or the West in general. But project forward a few years, when there are a variety of new form factors. Maybe we get spiders. Maybe we get drones.

I have to imagine the drone—the flying form factor—is a good deal more versatile for interacting with hostile scenarios. These are going to get smaller and cheaper and more plentiful. We haven't even seen what happens in the West from a first-response perspective when police can deploy swarms of drones rather than just individual drones. But as cost goes down, we will absolutely see swarms.

Dave

Yeah, I mean, again, I think the public is going to have a bit of a fearful reaction to this. It really depends what the drones are armed with. It depends on what the guidelines on the use of drones are, and we need to address that.

I mean, getting medical equipment to a site of an accident rapidly—you know, we're going to see eVTOLs, flying cars, delivering ambulance personnel there—but getting a defibrillator, for example, to a location that's jammed by traffic, drones are going to play an important part of this, and they're getting better and better. You were going to say—

Alex

I'm just remembering the counterpoint here. There was a fellow from the Dutch police force at one of our Singularity executive programs, and they were combating the fact that drug dealers were using drones. So they trained a bunch of hawks to drop mesh wires onto the drones to wrap them and tangle them. It's so retro to be training up birds to be attacking drones. It was totally surreal.

That's actually a real problem below a certain size. There are a lot of birds that go after these things. So the really small ones actually have a little bit of a problem in—

Dave

In Ukraine, they're having a huge problem because they've got strands of optical fiber from all of the drones that have attacked them lying everywhere.

Oh, yeah.

Alex

It's a massive issue.

Yeah. Well, I'll tell you, the number-one use of drones I'm excited about came from the Wildfire XPRIZE I talked about in the last pod, where drones are able to get to a fire at inception, put it out rapidly before it causes hundreds of millions or billions in damage and causes the loss of significant life. So—

Dave

And we didn't get our flying cars in the end. I mean, we have flying-car companies, but they're relatively sparse.

Alex

But I do think we're going to have skies over the next few years that are densely filled, I should say, with these drones.

Well, the drone ambulance is the coolest thing ever because no one's going to stand in the way of a drone ambulance, right? It's there to save somebody's life, but that'll unlock all the technology, all of the airspace, all of the regulatory barriers, and also prove the efficacy.

Alex

So that's going to be a great, really cool stepping stone.

One of the Substacks I put out was about the fact that we're heading toward a point where you can know anything you want, anytime you want, anywhere you want. We've got orbital satellites. We have the 200 satellites from Will Marshall at Planet. Then we're going to have an aviation layer from the flying cars and these drones, imaging everything at centimeter and subcentimeter resolution, and then all the autonomous cars gathering terabytes of data on the road. So everything is going to be imaged very soon.

Alex

You know, the spooky side of that, obviously, is loss of privacy, if you believe you have privacy. The positive side is that there's no crime. I put out a part of that blog saying, "When people are observed, they act better." I got a lot of negative feedback on that.

I can imagine.

Alex

It's true. I think when there's a CCTV camera pointing at them, right, they do less—

They do less.

Alex

They behave differently on the global stage. One of the foundations I used to support was the Lindbergh Foundation, which would fly drones over herds of elephants, and the poachers would stay away when the drones were flying over them.

Well, you want to hear a funny story from China? Sean, my son, who just got back from China, was talking to a guy who was mansplaining the entrepreneurial vibe in China and how to build a great company. Sean said, "Well, it's all about the team, right? This is what we preach at Link Ventures: You get great people, they succeed every time."

He said, "No, no, no. It has nothing to do with the team." You're like, "Well, then is it the business plan?" He said, "No, no, no. It's what the government needs next. That's the only thing that matters."

Alex

Wow.

You're like, "Wow, is that discouraging?"

Alex

So I think with the loss of freedom and privacy also comes the loss of innovation. I don't think the drones are going to be taking away all of our privacy and all of our freedom. I don't think that's a real issue. But in general, the slippery slope does kill innovation and entrepreneurship.

Alex, should we talk about the innermost loop?

Alex

Let's do it.

All right. Our first story is out of Switzerland, and it's an important one. Switzerland just voted to lift its ban on nuclear plants. After Fukushima, back in 2017, they phased out nuclear completely. Now they're reversing course.

Just to give you a sense, nuclear has been very slow. The country that succeeded so incredibly well is France, which has 57 operating reactors. The UK has 9, Spain has 7, and Switzerland has 4 aging reactors that supply 40% of its power. They were all due to be shut down, but they're going to be upgraded instead.

I was on Zoom earlier today with Ramez Naam, and Ramez is one of the most extraordinary thinkers in energy. We should have him on the pod.

Alex

We definitely should.

He would do an extraordinary job giving us an overview of all things energy, across solar, batteries, and so forth. He was saying the reason that France actually succeeded as well as it did is because it mass-produces a single reactor design instead of starting from zero, where the costs escalate and get out of hand.

I asked him whether he thought other European countries would follow suit and be able to implement nuclear, and he was like, "Nope, not going to happen." But it's interesting that the buzz on nuclear is beginning to soften, and the need is significant. Alex, your take on this?

Alex

Europe's in a bit of a bind. So maybe here's a really relatable story. Whenever I'm in the Swiss Alps and it's not the winter, it's very difficult to find air conditioning. I think Switzerland and a good portion of continental Europe have a real energy crisis.

They've lost access to cheap Russian oil thanks to recent events. Except for France, they underinvested in nuclear energy. They aren't this amazing native producer of their own solar PV. And they have a culture that one can sort of theorize where the culture comes from, but a culture arguably of energy scarcity.

Now, as global temperatures are rising and power consumption is increasing, Europe is having to do an about-face and discover—sort of learn to love nuclear energy, learn to love energy in general. And the risk, as we've talked about previously—cite the EU 2031 scenario and other scenarios—is that Europe is going to need to start to radically increase its power consumption and power production per capita. Nuclear fission in particular is a very attractive way to do that.

Otherwise, Europe will smolder under heat domes, including the one that Europeans have been suffering under over the past week or two. I forget the exact statistic, but thousands of Europeans are dying due to heat overexposure every year. It's a startling statistic, and it's unnecessary with better air conditioning and higher energy per capita. So I think this is the obvious trend of the future for Europe.

Yeah. Yeah, I mean, the good news is AI demand is going to turn energy from an environmental issue into a capacity issue, a commercial, profit-driven issue.

Alex

Well, but at least the need is a national capacity issue. Every country has to deliver enough energy.

And it's going to get it from fission in the short to medium term, until fusion or whatever comes along to cover the baseload.

So let's talk about fusion in our next story. Interestingly enough, the joke about fusion has always been that it's 50 years away and holding. Well, it's now here. There are some 50 privately funded fusion companies that have raised about $6 billion.

2 US companies lead the pack: Helion, which we're going to talk about, and Commonwealth Fusion Systems. I had Bob Mumgaard, the CEO of Commonwealth Fusion Systems, on stage with me at the Abundance Summit. They're building a tokamak-like design, and they're expected to build their first 400-megawatt plant circa 2032.

The second story here is Helion. It's a Sam Altman-backed company. He was the largest early investor back in March 2012, and until just 2 months ago, he was the executive chairman of the company. Apparently, he's stepped down now so Helion can actually do some large-scale partnerships with OpenAI.

On June 16, the news here is that Helion cleared the required Washington State regulatory approvals for its Orion fusion power plant, which is intended to supply Microsoft with 50 megawatts of power starting in 2028. If successful, this is the first fusion plant coming online. It's relatively small. 50 megawatts is—you know, we talk about gigawatt-level plants. This is 50 megawatts.

They've raised about $1 billion at a $5.4 billion valuation, but it looks like fusion is finally here. I've got a video showing how Helion works because it's a unique design. I think it's worth discussing, but Alex, do you want to comment before I show the video?

Alex

Yeah, maybe just a comment. Fusion—or the lack thereof—has long, in futurist circles, been the whipping boy for why long-promised technologies never happen.

But if you look at one of the figures of merit for fusion, the so-called triple product, which is the product of the density of the plasma, the confinement time of the plasma, and the temperature of the plasma, there has been steady progress for the past half-century toward self-sustaining and net-positive, in terms of power production, fusion reactions over the past 50 years.

This has been sort of—not just—it's not the case that there was suddenly some recent unlock, although arguably, economically, there has been in the form of high-Tc, high-temperature superconducting tape that's very helpful for certain architectures of fusion reactors. There has been continuous progress this entire time.

So I think there's an interesting parallel that one can draw between fusion, which is arguably achieved by compressing enough matter into 1 volume that you achieve net power output, and AI/ASI, which is arguably achieved by taking enough human knowledge and compressing it into a small enough information-theoretic footprint until you achieve really a phase transition that produces prompt engineering and large language model behavior.

All of that has strong parallels. I think they're both inevitable, but they're both inevitable as—I'll make a stronger analogy—which is, if you were watching the right metric or the right figure of merit over the long term, you could see both of these from 50 years away. You could see—or maybe 30 years, at a minimum—

Slow linear growth over time.

Alex

You just watch the compression over time. Arguably, with large language models and AI, if you were, say, watching the Hutter Prize—if you're Marcus Hutter and it's the late '90s—and you're watching the ability to compress the English Wikipedia over time, you could see LLM and AGI happening from decades away.

Similarly, with Helion and all of its competitors, if you're watching the triple product, you could predict when this is going to happen, and it's imminent.

Yeah, I think there’s something incredible here because this is such a foundational technology for abundance, right? This is the abundance thesis, for sure. That’s the foundational technology for it because once you have clean, cheap, and dense energy, the cost of computation, desalination, transportation, manufacturing, agriculture—I mean, everything becomes the cost of just the materials at that point. So this is such a big deal, and it’s one of the hardest conversations I have with CEOs and companies, especially with the public sector: energy is becoming abundant over the next few years. And when energy becomes abundant, all sorts of other dominoes fall.

Alex

Energy is the number-one correlate to GDP, to health, to education. The more energy a nation has, the better it is across the board. This is something that Europe needs to learn.

Interestingly enough, the challenge here is the fission plants, right? The small modular reactors and the Gen III plants are still not going to come online really until the early to mid-2030s. And the fusion plants—getting up to 400-megawatt plants, like Commonwealth Fusion Systems, or getting Helion up to that level—again, those are not going to be coming online until the early to mid-2030s.

And so the question is, where do we get the energy from? Now, I had that conversation with Ramez, and he says it’s from the grid. It’s going to be from the grid, and we just need to make better use of the grid. He’s got a company, Alex and Dave called Aentic [?], that basically is sucking down energy to batteries in the middle of the night, between 1:00 a.m. and 6:00 a.m., and then pumping that energy out during peak hours. So there’s plenty of energy on the grid if you can time-shift it. I think that was fascinating.

I think one of the questions that I don’t hear enough people discussing is: What is the killer app of fusion going to be? It seems obvious we’re going to get it, barring some surprise that we haven’t anticipated. But will fusion arrive in time to be transformative for terrestrial data centers? Maybe, maybe not. Will it be helpful for orbital data centers? Maybe. But there’s also a lot of solar.

Alex

There’s a 93-million-mile-away fusion plant that works really well in space.

That’s right. Sorry. Go ahead, Alex.

Alex

I think space propulsion is actually one of the killer apps. If we get compact fusion reactors, that’s a wonderful application.

My 9-year-old science-fiction self loves that.

Alex

Yeah. Hand in hand with that, if any materials scientists or chemical engineers want to work on this, if we have fusion, then any storage mechanism—even if it’s inefficient, as long as it’s clean—is suddenly viable. So if you have cheap, cheap, cheap, cheap, virtually free fusion energy, and you can put it into a car in a cheaper way than a lithium battery, right now you have to have some degree of efficiency. You don’t want to throw away electricity. But post-fusion, you won’t care about the efficiency of the reversible reaction. So anything is good.

And that’s true for launching rockets, too. Once you’ve got fusion energy, you can create any reversible reaction very efficiently. Then you can port it out to your space station or to your moon base, have it do whatever it’s going to do, come back, and recharge it. You don’t care if it’s only 10% efficient.

If you remember Bob Mumgaard, the CEO of Commonwealth Fusion Systems, when he was on stage, his goal, once he gets his unit working, is to pump them out—to create the machine that builds the machines. The same thing here for Helion. I’ve got a short video that explains how Helion works, and given the fact that it may be the first fusion plant coming online, and it’s unique in how it works—using magnetically propelled plasma and then magnets to pull electricity out of the plasma—let’s take a listen to this. I think it’s valuable for our listeners to hear about Helion.

Narrator

Helion’s pulsed fusion device directly recovers energy, which is used to generate zero-carbon electricity from fusion. It starts with Helion’s fusion fuel: deuterium and helium-3. These fuels are injected as a gas into Helion’s formation chamber, where they are superheated into an ionized gas called a plasma. The machine’s capacitors are charged and send electricity to magnets that wrap around Helion’s device. The magnets invert the plasma’s magnetic field on itself into a toroid, or donut.

The device’s magnets fire sequentially, accelerating the plasmas toward each other at a velocity greater than 1 million mph. They collide in the fusion chamber and merge to become 1 hot, dense plasma. In the center of the device, the machine’s magnetic field is rapidly increased, compressing the plasma with a powerful force over 10 tesla.

These fusion reactions within the plasma convert matter into new energy, which strengthens the plasma’s magnetic field. As the plasma’s magnetic field gets stronger, it pushes back on the magnetic field of the machine, causing a change in the machine’s magnetic flux. In accordance with Faraday’s law, this change in flux induces current in the machine’s coils, which is directly recaptured as electricity and returned to the capacitors that originally charged the magnets around the machine.

You know, we really are living in the future. When I see that, it’s extraordinary. And their goal was to mass-manufacture those Helion plants.

Well, if I had to call out 2, it’s that Helion raised $1 billion, and so did Commonwealth Fusion. Remember, we had dinner with him in Riyadh in March, and he had just raised $1 billion. When we were at MIT, the budgets for this were in the tens of millions.

Alex Wissner-Gross

Research budget. Yeah.

Research budgets. Now, suddenly, we’ve said this on the pod many times, but we’re actually truly investing in the commercial sector in hard science for the first time in my lifetime. Something great will come out of those 2 $1 billion investments, for sure.

I mean, Alex, look, to bookend this, once you make energy abundant, every other scarcity becomes negotiable. It’s probably also worth pointing out what the so-what of that explainer video is. Unlike many other fusion architectures, the whole point of Helion’s architecture is direct recovery of energy from the fusion plasma.

In a more conventional, say, tokamak-style or other fusion reactor, there’s a bucket brigade of energy production. You create the plasma through inertial confinement or through magnetic confinement. Then the plasma will be used to heat something, maybe water, and that produces vapor. Then the vapor goes into a turbine, and you turn the turbine, and you recover electricity from the turbine, inductively inducing, via magnets, currents in wires. It’s like a 10-step process.

The whole point—and what’s potentially quite seductively attractive—about the Helion architecture is that you’re just directly recovering from magnetic fields that are being induced by the plasma, and then currents induced by those magnetic fields. You’re almost directly recovering free energy from the plasma. So you’re skipping a whole bunch of steps, and it’s potentially a lot more efficient.

I was going to say a moment ago, if I had to pattern-match to a Mr. Fusion from the Back to the Future Part II architecture and identify the archetype of any one of the now many fusion startups that are out there, I think Helion is the closest to being a Mr. Fusion startup. All of those extra steps that are being skipped could lead to potentially radical compactification of the ultimate fusion implementation. So it’s very exciting.

Alex Wissner-Gross

Yeah, it is. It’s a beautiful design, and again, something Helion can mass-manufacture. Where does it go? It goes into every township, depending on the size; every city; every place that you need baseload energy production. Compact fusion is going to be a thing.

Well, if you told me I’d be listening to a little chipmunk voice explain Faraday’s law…

Alex Wissner-Gross

I know. It was like, “Am I running this? Am I running this at 1.5?” No. That’s the voice they chose.

All right, let’s jump into AI and a really fun story to kick us off. There’s a $1.8 million incentive prize founded by Nat Friedman, former CEO of GitHub, and Daniel Gross. Matt, I know Nat is your friend and roommate.

Matt Welsh

He was my first roommate at MIT. Fun stories.

He was on stage in March 2023 at the Abundance360 Summit, and he announced the Vesuvius Challenge. Here it is being won some 3 years later.

The challenge was this: There are these scrolls that were basically buried and burned under Mount Vesuvius back in 79 AD. The scrolls are fully carbonized. You can’t open them and read them without destroying them. So they asked, can we use technology? Can we use CT scans to gather the data and then use AI to read them?

For almost 2,000 years, they’ve been unreadable, and the challenge has just been won. You can see here in the image these scrolls of ancient Greek that have been linearized and laid out by the AI: 22 columns of ancient Greek text. And there are still hundreds of scrolls that can be read. I mean, this is using AI to basically look back in time. Alex, this must be a favorite one for you—

Alex Wissner-Gross

And not just because of the Nat connection. Objectively, I think computational archaeology powered by AI is going to be utterly transformative in the future.

I've argued from time to time before.

Sure.

Alex Wissner-Gross

Yeah. The killer app of the singularity is superpowering computational archaeology. I'll inevitably cite Nikolai Fyodorov, one of the fathers of the strain of philosophy called cosmism—the idea that humankind's common task is to essentially resurrect every human who's ever lived using technology.

I see in the Vesuvius Challenge the very beginning of a larger arc of technological progress that may require the singularity we're in to fully run its course. But imagine, just as a thought experiment, if we could do what the Vesuvius Challenge did—not just perform high-resolution scans of the positions of ink, or X-ray analysis of small blotches of ink, in order to recover scrolls that were otherwise preserved by the eruption of Vesuvius—but do this at a planetary scale.

Imagine that somehow—we're not going to say specifically what the mechanism of action would be—we're able to scan the Earth and gain fine spatial and temporal precision, or position, momentum, canonical coordinates, if you like, for every atom on Earth. Imagine if we fed the entire state of the Earth into an AI and what we might be able to recover about Earth's history. I think the answers would be quite transformative.

One of my favorite anecdotes is environmental DNA. If you go for a walk outside, you may or may not realize this, but you're just drowning in DNA that's been aerosolized from animals, alive and dead. If you dig into the soil—say, soil near a cemetery where human bodies have been buried—DNA has a surprisingly long half-life, even under environmental conditions.

There's a lot of state left over from Earth's past, not just in these scrolls that were preserved by volcanic eruptions, but in general. I think, again, an idiosyncratic position here, that with strong enough AI in combination with strong enough scanning technology, at some point in the future we will be able to recreate large fractions of our past light cone.

Well, you know, this is what Colossal is doing in a way, right? Going and extracting DNA from fossils, bringing back the dire wolf, bringing back the woolly mammoth in a very limited slice—and this is taking that to extremes.

Alex Wissner-Gross

Yeah.

Yeah. There's another angle to this story, too. If you're watching this podcast live and you look at the image and really zoom in on it, AI is very, very good at interpolating these fragments, these little pieces. If you look at the characters, no human being could ever reverse-engineer what that original character was, but the AI is really good at filling in those blanks.

It's not regular LLM AI. It's not your Anthropic or your OpenAI transformer. And this is where, Peter, you know David Siegel, right?

Sure.

He's on the board of MIT. He has a project called Project Open Athena, which he's funding heavily. It's designed to give AI compute resources to people who have alternative versions of AI that are not necessarily transformers, but are very, very good at these types of problems.

There are many, many, many of these. As Alex was describing, if you wanted to look at fragments of DNA that are lying around and reverse-engineer what happened in that room, that's a really good use case. All of these world events and historical events leave a little trace that's scattered around.

There's usually only one interpretation of history that could have created that trace. It's impossible for humans to glue those fragments together, but it's not LLM AI. It's core neural-network AI that's built from the ground up to solve that problem.

Let me give a shout-out to the non-technical founders out there. If you have an idea for a technology or a company, but you're not a technologist and it's your dream to make it happen, imagine being able to use an incentive prize like this to aggregate the best experts in the world to come help you solve your problem.

In this case, a $1.8 million prize probably brought, on the order of tens of millions of dollars, if not more, of genius to apply and solve the problem. Think about this: This is the basic principle of XPRIZE—to get people from around the world to focus on solving a problem. We get 30 times the prize money spent cumulatively to solve a challenge like this. Salim—

I wanted to stress that the prize model is incredibly powerful. What really strikes me in this is the before and after. The image of the before, and getting real data and information out of it, is so mind-boggling.

But there's an audacity in thinking that you can do that, and I think this is the abundance thesis again: Challenges we never thought were solvable fall as a result of the technologies that we're building.

All right. Let's move to the SpaceX universe. Grok 4.5 is coming out based on a 1.5-trillion-parameter V9 foundation model. Interestingly enough, Elon has made the claim that he's going to iterate a new model and release it every month for the rest of the year.

Alex, let's go to you on this one. He said something even spicier, if I understood his announcement correctly: that he was going to start pretraining every month—not just distillation cycles, or not just post-training or fine-tuning cycles per month, but start a new pretraining run every single month. That's—

Alex

Completely a new model from scratch.

From scratch, which means pretraining, which is, I mean, it's audacious. It's brute force. It's exactly, I think, what the world expects of Elon: a brute-force attack.

Alex

I've had a number of folks, since we first discussed in an earlier episode my comments about Grok being put on life support in favor of SpaceX's hyperscaler resources being handed over.

You got a lot of hate mail on that one.

Alex

I got some spicy comments. I had people saying, "Okay, well, Elon's announcing this, Elon's announcing that. Are you retracting your comments about Grok being put on life support?"

I think, again, my comments may have been misconstrued. I want the frontier to be competitive. Right now, we are arguably in a duopoly between OpenAI and Anthropic. They're just running away with the race, with Chinese open-weight models a few months on their heels.

I want there to be a competitive frontier, and I think Grok is one of the possible competitors, along with Gemini. Maybe Meta will come up with something eventually. I want it to be competitive, and Elon's strategy historically vis-à-vis xAI has been brute force.

Hopefully, as SpaceX brings more and more compute online, this sort of brute-force approach—where he eventually has more compute than everyone else, in combination with off-the-shelf algorithms—maybe this will work. I see the beginnings of that with Grok 4.5 in combination with Cursor.

Right now, the Cursor acquisition is taking the form of post-training models that he already had. In the near future, Elon has promised that Cursor is going to be part of the pretraining recipe. If he can make that work, I think Grok has a fighting chance to join the frontier through brute-force compute efforts.

There are a few things he's doing. First of all, he's got tens of billions of dollars of dry powder now to focus on this. The second thing is, he's announced that he's bringing in his smartest players from SpaceX and Tesla to work on xAI.

He is by no means giving up the ghost here. He wants to be number one. He wants to beat OpenAI. He wants to create Grok as the ultimate, greatest seeker of truth. And again, I would never bet against him.

Alex

Yeah. I think the best bet here is that he will brute-force his way back to the frontier. That's what I'm hoping will happen.

Yeah. Well, he also doesn't see it as a two-horse race.

Yeah. What you're describing as a two-horse race, he doesn't perceive that at all. He thinks it's still a race against Google, specifically because training directly toward a customized chip is a 10-to-100-times unlock, and he doesn't perceive Anthropic to be near there.

OpenAI has some activity there, but it's nowhere near it. Google, however, has already got its whole vertical monopoly stuck together. He doesn't think he's going to lose because he thinks he's going to be the first to have a reasonably good model with custom silicon that supports the model instantaneously, with the design of the chip being AI.

You know, we're seeing verticalization win every race here. We'll see it in the space industry. We're seeing it with Google, and we're seeing it with xAI. I mean, it's fascinating. Salim, please.

Yeah, there's an idea here that could allow him to leapfrog even faster. We're moving from models that were built on human artifacts toward models being trained on the actual process of human-machine work, which is where the Cursor data becomes really useful.

That creates a flywheel where you have a better model, more usage, a richer workflow, and then a better model. I think that's going to serve him very well in the future.

Alex

I’ll comment on that point narrowly. I think it’s an interesting debate one could have. The Cursor acquisition—I think I viewed this at the time and commented on this on the pod previously—I viewed it as a brain transplant for the future trajectory of Grok, given that ChatGPT is also going through the same brain transplant, making Codex essentially a model—or class of models together with scaffolding—that were optimized for code generation and recursive self-improvement. That’s becoming the new mainline ChatGPT; by analogy, xAI and SpaceX acquiring Cursor post-IPO to make Cursor essentially the new mainline Grok, I view as an analogous move.

However, the dataset from Cursor, which, in my understanding, consists in part of lots of reasoning traces driven by developers who wanted more code generation, I think buys Elon sort of a leap to the near frontier in terms of code generation. But it will still be incumbent on Elon, xAI, and SpaceX to achieve their own recursive self-improvement loop. You can only get so far by post-training on developer or user traces. At some point, the models need to start developing better models. That’s been, historically, I think, a strength of Anthropic—probably a weakness of xAI—but maybe he can brute-force himself to the front of the recursive self-improvement loop.

Let’s jump into a few Anthropic stories. Anthropic’s flagship model, Fable 5, has been offline for 15 days because the US government pulled it for national security fears, and hopefully we’ve talked about this ad nauseam. Now Axios reports it may be back within days. Secretary Lutnick credited Anthropic for working on the risks through the Pentagon and NSA. They still haven’t signed off on that.

Interestingly enough, Stripe recently reported that they ran a test using Fable 5 to overhaul 50 million lines of code in a single day, work that would have taken engineers many months. The government is treating Opus 5 as a commercial AI model, like a controlled munition, taking it offline and re-permitting it for users. We’re going to have to see how this evolves in the coming days. So let’s talk about Opus 5 coming back online, and Anthropic is another story. Alex, if you want to cover that one too.

Alex

Sure. I think there are a couple of interesting notes coming out of this if Fable 5, as I expect, eventually becomes available again. One of the more interesting takes, I think, is that this will have been a period of a few weeks when allegedly Chinese organizations that were leveraging access to Anthropic’s frontier models or near-frontier models for reasoning-trace distillation will have been denied that access. So we talk on the pod all the time about how the US maybe has a 3-month lead or a 6-month lead or an 8-month lead, depending on how you count. There’s a certain sense in which this 1-month-ish shutdown—future perfect tense—may have, or will have, denied China at least a month of catch-up time. That’s a generous interpretation.

A less generous interpretation is that historians will look back and say this period marked the period toward the middle or the endgame of recursive self-improvement, when months counted and the permitting of frontier intelligence became almost a Zone of Thought, to borrow from Vernor Vinge, or a bloc system, to borrow from the Cold War. Before, anyone could access frontier intelligence; after, you have to be a US person, there are strict export controls for capabilities, and there’s a nonproliferation regime where, if you’re a non-US person, you gain access to models that are maybe a few months behind the frontier.

But I also think a few years from now, when we look back on this time, I think this will have been—yes, there was a phase change in terms of the diffusion of frontier models—but I do think sometime in the next few years we’re going to get to the end of the recursive self-improvement rainbow, and there’s going to be a perfect model. We’ll look back and say this was just a period of months—a delay—but ultimately everyone is going to figure out what the perfect model looks like.

Was this 1-month period also a chance for a lot of critical systems to safeguard themselves against Fable 5—the most essential ones, maybe?

Alex

By and large, yes. There was a stand-up, both within the US government, of vulnerability scanning, and outside the US government, we saw 3 or 4 independent nonprofit or for-profit organizations stand themselves up to do bulk vulnerability scanning using Fable 5. But I think, to the extent pause-ism has had its day in the sun, I don’t think this actually decelerated AI at all. I think this is a net accelerant.

I’m not thinking about decelerating AI. I’m thinking about decelerating black hats from being able to get in there and penetrate.

Alex

I don’t think so. I think black-hat capabilities are proportional to capabilities overall. And I think what we saw during this time is that Chinese models like GLM-5.2 began to give anyone who wants near-frontier capabilities the ability to do essentially whatever they want with them. They may not be as capable as Opus 5, but this creates enormous pressure on Chinese organizations and the Chinese frontier labs to catch up.

And so I think, as with the original PauseAI movement, it had the net effect of accelerating capabilities globally. Same idea here.

Two quick implications from this slowdown. I just want to point out one for investors: regulatory risk is now one of the first-order variables when you’re looking at companies, because it’s real. And the second, for technical founders, is don’t build your product or your company on a single model. You have to make sure you’re able to swap out models, because you have no guarantees as we’re going forward. And Dave, you saw that. Dave and Alex, do you want to talk about the Sonnet announcement that came an hour ago? Dave, you first.

Dave

Well, it’s really obvious that AI is sold out, and that when Fable 5 came out, they doubled the price on us, but you had to use it because it’s just so good. So it’s pretty clear that Sonnet 5 is now a way to kind of fill this gap until Fable 5 is back out. But the price point is very high, given the amount of compute that they have to use to deliver it. People will still buy it because, again, AI is sold out.

So you see the revenues at Anthropic going through the roof, and it’s because the demand for AI way outstrips the underlying chip supply. The byproduct of that is a lot of things. Only the very top-of-the-mountain use cases are going to get access, and then after Fable 5 comes back out, I agree with what Alex was saying a minute ago. This moment in time will be remembered in history. This is the intersection of AI and the government that’s never going to go away now.

But not every person on the planet and not every company on the planet is going to be able to access the models, and it’s supply-constrained at the same time. So then there’s going to be preferential routing. Sonnet 5 is a kind of mediocre capability at a high price point, but people will still need to buy it. And then Fable 5 will come out at its extremely high price point. That’s my read on Sonnet.

Alex, any addition?

Alex

I think this is a bizarre announcement. Admittedly, this is a hasty analysis, since Sonnet 5 was released right before we went to air here. But I’ve been trained—as I think the majority of sophisticated users have been trained—to expect that the Sonnet series from Anthropic would represent some distillation of the Opus series. Similarly, the Haiku series represents a distillation of the Sonnet series.

As you go down toward smaller, lower-parameter-count, more-distilled models, you see some optimal frontier emerge in price-performance space where performance—at least throughput—goes up, price per token goes down, and performance goes down. Maybe I’m missing something, and maybe the answer will reveal itself in the next few hours. But just looking at the cost-versus-performance-at-agentic-tasks curves that Anthropic released with Sonnet 5, it’s a little bizarre.

On the one hand, Sonnet 5 is an optimal frontier, sort of a Pareto improvement over the last version of Sonnet, Sonnet 4.6. But Opus 4.8, which has been out for what, in these singularity times, passes for an eternity, is better. It’s superior on a cost-performance basis. So I’m not 100% certain I understand what Anthropic is hoping to achieve with Sonnet 5.

I can tell you, Alex, everybody’s working on these frameworks where you can bounce from model to model while keeping the context intact, and all of the work that’s piling up—the prompt history and the intellectual property—is piling up like crazy now. You have a choice between working in an open framework, but Claude Code and Claude Cowork are super compelling, with all the MCP wrappers and connectors already built in. So the easy choice—the Apple-like, “I’m going to pay more, but it all works” choice—

Alex

—is to go with an all-Claude stack. And then, when you’re working in Opus 4.8, if you have a simpler question, you go to Sonnet—or it automatically goes to Sonnet—and if it’s an even simpler question, it just goes down to Haiku. So that’s the easy way to go. But the more cost-effective way to go would be to bounce over to a different model, but then you have to use a third-party context-management platform. So that’s the tension. But this is kind of like Anthropic becoming the Apple of AI, where you’re overpaying by some insane amount.

I feel like there’s some branding going on behind the scenes that we’re just missing. My intuition is—there’s something there we’re not seeing.

Alex

Maybe, in the sense that Fable and Mythos are the new high-end models. Maybe there's some sense in which Sonnet is the new low end. Sonnet is the new Haiku, and viewing it through the branding of Sonnet is maybe incorrect. Maybe we should be thinking of it as the Haiku level, and it's just that Fable 5 isn't accessible. There's some weird Pareto-optimal frontier, I think, that's missing in order to explain why we've seen a reversion to this optimum.

We will find out, and we will get to the bottom of it.

All right, mates. Let's jump into data centers and space. For that, we're pleased to bring a friend on. Philip Johnston is the co-founder and CEO of Starcloud, a startup building space-based data centers. Starcloud has raised about $200 million, with its last round over $1 billion. Famously, Philip's company launched Starcloud-1, the first NVIDIA H100 GPU in orbit. Apparently, Philip, you've trained your first LLM in space. Starcloud launched in November 2025 on Falcon 9. Welcome, Philip. It's a pleasure to have you.

Thanks so much for having me. It's a huge honor. I've been a longtime fan, so I'm very privileged to be here.

Awesome. It's mutual.

I have a lot of stories we want to talk about in data centers and space to wrap up today's episode, but let's start with a little bit about Starcloud. Tell us about the company and your vision. Where are you guys, what have you done with the H100, and—yeah, let's go there.

Yeah, for sure. We started about 2.5 years ago, in January 2024. We're a team of about 20 engineers based in Redmond, Washington—actually, right down the road from the Starlink manufacturing facility. About half our team came from SpaceX, and the rest are from the data center companies up here, so AWS, Azure, and the others.

Then, yeah, we launched our first spacecraft, Starcloud-1, in November last year. It had 5 GPUs—2 from AMD and 3 from NVIDIA—but the most important one, as you mentioned, was the NVIDIA H100. With that, we were the first to train a model in space. We trained nanoGPT from Karpathy, which is a very tiny model, but it still counts.

Totally counts.

We were actually the first to run a version of Gemini in space. We ran Gemma, which is DeepMind's cut-down version of Gemini. We've done a few other things. Now we're doing much more useful workloads. We've just been doing high-power inference on SAR data—synthetic aperture radar data—in collaboration with various DoD entities. We also actually played Doom. We ran Doom on Starcloud-1.

As one does. A little bit of lag there, probably.

Before we get into it, I just want to say you guys have been ahead of this trend more than most. Thank you for your support. I know Alex is often vocal about his support, and Dave gave me a shout-out early on. Peter, I know, in the interview with Elon, you were also very supportive, so I really appreciate it.

Yeah. Well, you're building the Dyson swarm, and this is important for our great-great-grandchildren—for all of us. Philip, speaking for myself, I want there to be multiple Dyson swarms. We can't have a solar-system-scale monopoly.

Agreed. Maybe a Matrioshka brain, so we can have swarms inside swarms inside swarms.

So there's Jupiter and Saturn—plenty of atoms left to disassemble.

Philip, you gave us a sense of where you are right now in terms of Starcloud-1, but where do you go next?

Yeah. We've got 3 launches booked next year. We're launching Starcloud-2, booked for January. It's about 100 times the power generation of Starcloud-1. We'll have, by far, the largest commercial deployable radiator in space. That's going to have a whole bunch of H100s, also the Blackwell chip from NVIDIA, and some other interesting things, like some Bitcoin-mining ASICs. It will also have an AWS Outpost, which is an on-premises server blade, so we can run an instance of EC2 in orbit, which is useful for DoD customers.

As soon as possible, we'll be launching a much larger spacecraft, what we're calling Starcloud-3. It's a 200-kilowatt, 3-ton spacecraft, which will fit on the Starship rideshare dispenser form factor. Actually, if you guys can see this, I can even show you. We've got a version of it welded up on the ceiling—the chassis. With this, we'll have huge 100-meter deployables hanging off the side of that.

How much power are you generating to power that?

That's 200 kW.

3 tons, you said?

3 tons, yeah.

3 tons. Okay.

Yeah, and we can fit about 50 of them per Starship. So we're talking about about 10 megawatts of new compute capacity per Starship launch. We're hoping to be launching very frequently on Starship. They're building absolutely enormous capacity, as I'm sure you're all aware: 2 Starship gigafactories are designed to produce something like 3 Starships per day. Hopefully, there'll be capacity for us to launch. Obviously, we're looking at some other launch providers as well, but that's the primary one.

You have to ask the elephant in the room: When you're pitching investors and they say, “But there's Elon and his mega plans, and Google's far behind. Are you picking a niche in this area that you're going to be competitive in?”

Yeah, it's a great question. The main niche we're going after is to be more like an energy and infrastructure play than like our own cloud. For example, we've got an agreement with Crusoe where we essentially say to them, “Hey, we have a box, and that box has power, cooling, and connectivity. We'll work with you on whatever chip architecture you like, and you can sell to whichever customers you like, for whatever price you like. You just pay us a fee.” In the same way that you would pay a rental fee to somebody like Equinix, you pay that kind of rental fee to us, and you finance the chips. That's the approach we're thinking about.

I think, in the early days at least, it looks like SpaceX is primarily going to be serving their own Grok and xAI workloads. In the medium term, we'll be providing a cloud service to folks like Anthropic. I think it's probably a bit further out that they're looking at being just a pure infrastructure provider, though I think Elon did mention something about that relatively recently. But that's the idea.

In general, it's a good point because, of course, we're going to have a higher cost base than SpaceX because they own the launch. As long as we have a lower cost base than all of the other hyperscalers, I think we're in a reasonable position. If we have a lower cost base than OpenAI, for example, they're going to need to figure out a space solution. Either they pay xAI to run workloads on xAI or on SpaceX's satellites—in OpenAI's case, I think that sounds unlikely—or they start building their own satellites, and it's possible, but I mean, they're going to be way, way far behind.

Or the last option is they'll look around in 2 or 3 years, as Starship cadence ramps up, and they'll be like, “Okay, we're going to get left behind if we don't get on top of this.” They'll be like, “Okay, who's the most advanced in the market? Maybe besides SpaceX.” At that point, I think we'll have a very significant lead over anybody, perhaps besides SpaceX.

Interesting. Dave, Alex, you want to jump in?

Alex

I've got 2 questions.

Okay.

Dave

Okay. One was, 2 years ago, we didn't have data centers in space on our bingo cards at all. What had you do that? Some of us did, and definitely the mainstream did not. I know, Alex, you've been talking about Dyson swarms since you were probably 5 years old. But what made you jump to that and say, “We're going to do that?”

Yeah. We actually started off by looking at space-based solar. So, in mid-2023, I randomly took a trip down to Starbase, Texas, on a weekend, even before the first launch. Not as many people were looking at it back then, and I was just blown away by the scale of what they were building.

In my head, I was like, “Okay, all of the concepts from sci-fi that I remember reading about—Asimov was talking about space-based solar in the ’40s, even—are going to come true. It's just a matter of timeline now.”

Philip, if I could clarify: space-based solar power satellites for beaming energy down to the ground?

Yes, exactly. Huge solar panels in space, and then either using infrared or microwave to beam power down.

The main problem—and we spent several months on the math, essentially, on the break-even—was that we wanted to know what the break-even launch cost was at which space-based solar makes sense. We came to a number around $50 a kilo, where that would make sense. Initially, we thought, “Okay, that's good enough. Starship will get there at some point,” and we started working on that.

But then we were like, “Okay, well, the problem with space-based solar is you lose 90% or 95% of the energy in transmission from space to ground.” We were looking around and saying, “Okay, once we get the power down, what are we going to be using it for?” Even 2 years ago, most new energy projects being built, particularly in the US, were being built primarily to power data centers.

So the thinking was, “Okay, well, either directly or indirectly, that power is going to be going into data centers. If we instead can find a cheap way to get the data center to space, let's rerun all those calculations to know what the break-even launch cost would need to be for that business to break even—to make sense versus terrestrial.” So we reran those numbers. We came to a launch-cost break-even of around $500 a kilo if we had a cheap way to get the data center to space, and that became the basis of a white paper that we put out in summer 2024. Then that essentially became the basis of the company.

Wow, that's great. That's a great story. It's an interesting entrepreneurial story for all the entrepreneurs listening, right? You're going down one road, and then all of a sudden you see a massive opportunity, especially when you can go deep enough to look at it, because the idea of space solar power satellites has been around since the ’70s.

Gerard K. O'Neill at Space Studies Institute looked at it. His solution was to build them on the Moon and then fly them to Earth orbit, where the launch costs obviously are minuscule.

I'm sure that would happen. I'm sure that will happen.

Dave

Yeah. Second question is, what are the couple of biggest bottlenecks you're facing right now?

We are actually very constrained on launch right now, as everybody is. We're so constrained that we're trying to book now on Relativity Space's first launch in—

January. All right. I'm really excited about that. I think that would be super cool to have a launch on there.

For those of you who don't know, we covered Relativity Space in a podcast a while ago. This is Eric Schmidt's company, where he's CEO. He bought it from Tim and Jordan when it missed the financing, and Relativity Space is about the size of New Glenn, half the size of Starship.

Well, if you're building toward the Pez dispenser, actually, that's a problem. That's really interesting. How's that going to work out?

Yeah, it is a problem. I think we're going to have to have 2 form factors. The primary one we're working on right now is for Falcon 9, but I think we're going to need a form factor that will also fit on Stoke Space. Stoke Space is the only one that has a reasonable upper stage, or the only one that's seriously working on a reusable upper stage right now. So we're also looking at a Stoke launch-vehicle form factor.

To be frank, the Relativity one is—maybe just to take a step back—the business has 2 phases. The first phase is while launch cost is relatively high. We're launching on Falcon 9 and others, and we're primarily serving edge and cloud, providing edge and cloud services for other spacecraft, particularly DoD and Earth-observation constellations.

Then, on a sort of 3- to 4-year time frame, as Starship ramps up cadence and production, that's when we switch over to competing with all terrestrial data centers on energy cost. So, when I say we're launch-constrained right now, I'm even talking about the first business. We've got 3 launches booked for next year, but if you want to book anything for 2028 on Falcon 9, there's just nothing available.

Wow.

The government's just plunked down 20 launches, which has bumped everything back. We're going through various channels to try to get some priority on that.

So this is GPUs for processing in space?

Yes, correct. We will receive, for instance, raw imagery—hyperspectral or SAR, or other types of satellite-sensing data. Instead of having to wait for a ground station and downlink enormous amounts of data, we can process all of that on the edge.

Actually, one of the demonstrations we've just done is to process a whole bunch of SAR data, identify the coordinates of a tank, and then just downlink the coordinates of that tank. Rather than having to wait—it could take 3 days to get enough ground-station passes to get 100 gigabytes of SAR data off a satellite—we can receive all that data optically in space, process it on orbit, and then just downlink the insight. That's the use case we're building toward right now.

I'm curious, Philip. The premise there is presumably bandwidth is scarce—scarcer, maybe, than launch. So you have to do a lot of edge inference in LEO or wherever you're doing this.

I'd love to ask you a similar question that I asked Will Marshall of Planet in a previous podcast episode. Let's project out 10 to 20 years, well past the current bottleneck in heavy-lift or heavy-launch capability. What do you think, as the founder and CEO of one of the incumbent Dyson swarms—plural—what do you think the Dyson swarm, or Dyson swarms, of, call it, 20 years from now look like?

Does it look like LEO? Does it look like sun-synchronous orbit? Does it look like the Moon? Does it look like a Dyson swarm around the Sun? Paint a picture for us of what, 20 years from now, the Dyson swarm or swarms look like.

That's a good question. I hope within 20 years we've started putting significant amounts of compute in a Sun orbit, probably starting with the Lagrange points, although you probably don't want to clog them up too much, but even just a distinct Sun orbit that trails Earth or is in front of Earth.

Certainly, you could fit about 10 terawatts of compute in the dawn-dusk sun-synchronous orbit, and then you're back to flying in orbits that have an eclipse—sort of 45 minutes of every 90-minute orbit. That then drives the cost up significantly because you need batteries and all these other things.

Scarcity of real estate. SSO is going to get crowded. You're one of the few people, Philip, I hear talking—this is like overpopulation on Mars—about SSO getting crowded due to the SSO Dyson swarm and then overflowing back to other orbits. That's fascinating.

Yeah. Well, because there's only one SSO orbit that is—you know, it's a very rare and beautiful orbit—the dawn-dusk sun-synchronous orbit, which flies over the terminator line.

I mean, 10 terawatts is a lot of compute. That's like 20 times the entire US power grid. My expectation, roughly, is that in 10 years we might hit a point where most new compute capacity is being deployed in space.

If you ask me what the total percentage of compute in space is at that point, it's still probably less than 5%. In the same way, right now, in certain parts of the world at least—Norway or whatever—most new cars coming off the production line are electric. But if you want to know what percentage of the fleet is electric, it's still like 4%.

It just takes a long time to replace all of the capacity that we're building on Earth terrestrially right now. So, in 10 years, I would say most new compute capacity will be in space. In 20 years, it could be—I would be surprised if more than half of all compute were in space even in 20 years.

But beyond that, certainly there'll come a point, and it's probably more like a 50-year time frame, where 99% of all compute is in space.

I also just have to ask: second elephant in an increasingly crowded room. If, over the next few years, heavy launch is the main bottleneck, is there some plan in a back room somewhere for you either to build or buy your own vertically integrated heavy-launch provider, if that's the main constraint?

We are in quite serious discussions to partner with various launch providers. Probably shouldn't go too much down that path. It's early days. If SpaceX can provide the capacity, we'll be very happy customers of SpaceX.

If not, we’ll need to figure out something. If you ask me, do I think SpaceX has a monopoly on launch in 5 years? The answer is yes. In 10 years, the answer is no.

I realize—and we’ll talk about Rocket Lab’s purchase of uranium in a moment. What I’d love to do is jump into a few stories, Philip, and have you comment alongside the other guests. The first story is a recent conversation by Elon about space-to-Earth telephony, so let’s jump into that.

I’m going to play a short video. This is a video clip of Elon speaking at the All-In Summit, and then we’ll talk about it afterward.

Elon Musk

The phones that are able to use the spectrum that was acquired probably start shipping in around 2 years. We also need to build the satellites that are going to communicate on those frequencies. In parallel, we’re building the satellites and working with the handset makers to add these frequencies to the phones.

Then the satellites and the phones will handshake very well to achieve high-bandwidth connectivity. The net effect is that you should be able to watch videos anywhere on your phone.

Fascinating story—vertical integration again. SpaceX owns the whole stack: launch, satellites, spectrum, and increasingly compute. Direct-to-phone is an interesting product. It’s going to be space-based internet for everybody on the planet. Thoughts on this? For investors, we’re going to see the telecom industry getting disrupted.

Yeah. He didn’t mention it here.

To a point on vertical integration, he didn’t mention it here, and I think he’s talking about working with phone providers. It would not surprise me at all if they either buy or start manufacturing their own phones as well. I don’t know—what was your read of that? My read was that he was talking about working with phone providers and not talking about building their own phone. At some point, I would be surprised if they don’t think about building their own phone.

It’s been rumored that Tesla phones have been coming for a long time. Elon famously does not work well with others. He tends to buy them or blow past them. Did you see the recent rumors about the possible acquisition of T-Mobile?

Yeah, the rumors are out there, and also with Charter Communications.

Starlink is the most unbelievable business. Starlink is going to produce hundreds of billions of dollars of free cash flow in the next 5 to 10 years. They’re going to have direct sales and unbelievable bandwidth. Unless you’re really in the middle of Manhattan, I think Starlink will be the best option for most people.

I can’t see a world where, on Starlink revenues alone, SpaceX isn’t the most valuable company in the world. We’ve been projecting $10 trillion by 2030 in terms of valuation and scaling toward $100 trillion. I’d absolutely back that.

Dave

My question, actually, is: if we have direct satellite-phone connectivity, downloading videos, and that implies tens or even hundreds of thousands of satellites in low Earth orbit, and then data centers are going to want that same space, is that actually going to survive the escalating needs of AI?

If you figure, as you said, 5 to 10 years from now, almost all computing will be in space, that’s also a lot of very high-value use cases for that famous LEO.

The orbits are slightly different because the Starlink satellites fly around 460 kilometers, I think it is now, but they are in a 50-degree inclination. They aren’t going over the poles; they’re mainly going up to just midway through Canada. They’re flying as low as they can.

Actually, with the AI satellites, you want to fly them as high as you can, almost, because even in this dawn-dusk, sun-synchronous orbit, if you’re flying at 600 kilometers altitude, there’s about a month of the year when you have a 10% window that is blacked out. The ideal altitude to fly is actually 1,200 kilometers. That’s the lowest you can fly where you don’t have any blackout, even throughout the year.

Between 400 and 500 kilometers is going to get extremely crowded. I think where the AI satellites fly will get very crowded, but it’ll be for a different use case. I don’t think they’re going to be competing for the same real estate, essentially, because—

Dave

Are you going to run into Kessler effect problems if you go that high?

I mean, at scale, we’re talking—yeah. The main reason people don’t want to fly that high right now is actually radiation from the Van Allen radiation belt. Do you mean because there’s less drag, so if you have any collision there, the debris is going to stay up for longer in the—

Dave

Yeah, for millions of years.

Yeah. Well, even at 1,000 kilometers, most stuff will probably deorbit within about 50 years, which is not great. But it’s not as drastic as I think a lot of people sometimes think.

For example, I don’t know if you know about this, but in 1970, the U.S. government dispersed 400 million needles at about 3,000 kilometers altitude because they wanted to bounce radio frequency off them, which is insane to think about today.

Dave

It was like an artificial ionosphere.

Yeah, it was the most insane thing ever. You can’t imagine it being done today. But actually, every single one of those needles has now deorbited, which people don’t realize. People think that if you have quite a large satellite, it will stay up at 3,000 kilometers for quite a long time.

Things that are much smaller and don’t have their own propulsion, because of the way the Sun and Moon and all of the solar wind interact together, every so often they get closer to Earth and get dragged in and dragged in and dragged in. The Kessler effect is something we really need to pay attention to, but it’s not as drastic as I think a lot of people sometimes think.

That’s good news. Sim, you got a question.

Dave

Wait, but what about cooling? I mean, cooling and radiation—oh, sorry. Go ahead, Sim.

Sim

No, go ahead. I was going to ask that exact question.

Well, we were talking about this—I forget whether it was in Riyadh or at A360, Philip, but you had this all-aluminum cooling system, and I was like, “Wow, that’ll be incredible if that works.” But now it sounds like you’re going to move to liquid cooling or some kind of liquid-cycling process.

It’s a government project. What else do you expect?

Yeah, it’s liquid through aluminum. That design is the same as what we discussed in Riyadh, I think. Exactly. It’s a very large, low-cost, low-mass deployable radiator.

Radiators work because the International Space Station has been using them for 20 years. The problem with the ISS radiator is that it is both expensive and heavy.

Expensive, heavy, and late. The core IP of our company is making this radiator cheap and light. It’s not a new physics problem; it’s a manufacturing and engineering problem.

Our radiator design—and we’ve got it working; it’s been fabricated and gone through TVAC and everything—is 10 times less mass per watt of dissipation than the ISS radiator and about 100 times less cost.

Yeah, about 100 times less cost per watt of dissipation than the ISS radiator. I’m excited that that will fly in January, and it will be a big milestone. It’ll be, by far, the largest commercial deployment.

I wanted to double-click on something you said. Did you say that, over time, 95% of the compute we use will be done in space?

Ninety-nine point nine percent, I think, over time.

Wow.

But that’s on Alex’s Dyson-swarm-type time frame.

That’s right. Masayoshi Son just came out with a statement saying he disagrees with the thesis of orbital compute because energy is only 7% of the cost compared to everything else. Where do you come out on that? What’s your answer to him?

I did a post about this this morning because people kept tweeting it at me, so I was like, “Okay, well, he also sold all of his Nvidia stock in 2019, so he’s not always right.”

He is right that energy is a very small proportion. Energy and infrastructure, though, are actually quite significant. It depends: right now, chip cost is very high, so that is by far the dominant cost. But if you include energy and infrastructure, you’re talking about at least 30% of the cost.

The main problem is that, even to build a new energy project terrestrially, you’re looking at a 5- to 10-year lead time just for the permitting.

Yeah.

The main advantage is that we can deploy this stuff extremely rapidly. So even if we were breaking even on energy and infrastructure, it would still make sense to do this. But we’re looking at doing this about 10 times cheaper on both energy and infrastructure.

When I say infrastructure, what I mean is that we don’t need batteries, cooling towers, big chillers, or backup power. All we need is a dirt-cheap radiator. Our radiator is really dirt cheap, and a lot of the other infrastructure costs are gone.

It's only then that the launch cost is the additional piece we have. But as I say, that is very rapidly trending toward a much lower launch cost.

Philip, if I might, please just pull the thread a little bit on launch. Projecting conservatively 20 to 30 years out, where do you think launch is going to come from? Will we be using railguns to launch from the lunar surface? Will we have, optimistically, self-replicating von Neumann probes that are disassembling our solar system to build more compute? Where is all the matter, energy, and launch coming from 20 to 30-plus years out, in your mind?

I mean, Elon has this great quote, something like, “Optimus is the von Neumann probe,” and I kind of agree with him. If you can get 100,000 Optimus robots to the lunar surface and get them to build an Optimus factory on the lunar surface, then we're off to the races. Then you have this insane exponential curve in terms of development and pace of development.

Hyperexponential.

Yeah. Absolutely hyperexponential.

So, I do think we'll have mass drivers on the Moon, and I think they'll probably come sooner than most people. If it was less than 20 years, that would be maybe surprising to me. If it was more than 20 years, I'd also be a bit surprised. I think, yeah, around that time frame.

Dave

All right, I apologize. I have one of the senior executives from State Street Bank waiting for me outside the door here. I'm going to love watching this podcast because all of my questions are in Alex and Sem's heads, but I can't wait to hear your answers. I'm super excited about what you're doing, though. Congratulations.

All right, Dave.

Thanks so much, Dave. Appreciate it.

See you very soon again on our next pod recording. But let's move to our last story on the docket here.

Rocket Lab is acquiring Iridium, creating yet another fully integrated space powerhouse. Rocket Lab, for those of you who don't know, is a company started by Peter Beck—actually, Sir Peter Beck. It's a $6.4 billion company now. Kudos to them. Peter had no background in launch, and he built arguably the second-tier provider for launch after SpaceX.

They have their Electron launcher. It's a smaller-sized launcher, but it's launched 91 times and has very high reliability at this point. They're building their next vehicle, called Neutron, and it's planned for a first launch by the end of this year. Like Falcon 9, it's got a first stage that's reusable.

I know Iridium well. I was playing in the big LEOs in the early '90s when it got its license and started commercial service in November 1997. It's a 66-satellite constellation orbiting at about 780 kilometers.

A fun story—I don't know if you know this, Alex—it was originally called Iridium because it originally had 77 satellites, which is the atomic number for iridium. When they changed it to 66 satellites, they did not change the name to dysprosium, which is atomic number 66. Good marketing move there.

Alex

Also, iridium, I think, is a little bit more stable as a nuclide.

Yeah, and it sounds a lot better than dysprosium.

So, what makes Iridium interesting is that it's got 10.5 megahertz of bandwidth at L-band that's globally coordinated. They go through the ITU and get it. Importantly for our viewers here, you can get 10.5 megahertz in the U.S., but can you get it in every country around the world? That's what makes it a prize.

Spectrum is the prize, and vertical integration here is becoming the winning structure for the new space economy. Owning launch and manufacturing—they build their own satellites at Rocket Lab as well—and getting spectrum and operations is a winning combination. They're playing on the SpaceX handbook. Pretty extraordinary.

Philip, thoughts for you?

Yeah, I mean, it's a very smart move, I think. It's particularly smart because I think there's been quite a bit of commentary—people saying, “He's not trying to compete directly with SpaceX with this.” If he was, it probably wouldn't be as smart a move. He's really carving out a niche.

The other thing is his share price has gone to an insane multiple of his revenue, and he's capitalizing on that because I'm pretty sure all of this is going to be in Rocket Lab stock. It makes sense to start paying for cash-generative, profitable companies in Rocket Lab stock when you're trading at a however-many-hundred-times-revenue multiple, which is what they're trading at.

I have to ask Philip a question just about the spectrum side. There are a bunch of obvious questions I could be asking about vertical integration, and whether spectrum and LEO compute want to inevitably own, or be owned by, heavy-launch capability, but I just want to focus on spectrum.

To the extent that part of the Iridium–Rocket Lab story is the acquisition of RF spectrum, do you think radio has a future, or will we find ourselves 5 to 10 years from now with it all being optical-frequency direct laser links and radio having approximately no future?

It's a great question. I would lean more toward the second of those 2 options. I wouldn't say radio really has no future. It's useful because it's cheaper, because you don't need gimbals and things, but laser is where everything's going.

We've got 3 laser terminals on our second satellite, launching in January, with gimbal lasers. We've actually just signed a contract with SpaceX for the next 25 of our satellites. We'll have 2 Starlink—they call it PPL, plug-and-play laser terminals—on our satellites, and then 1 SDA-compliant laser that can connect with the government satellites.

So, yeah, lasers are the future for space comms, and it's also unregulated, which makes it amazing.

All right, guys. I apologize. I've got a hard out here as well. Philip, a pleasure, and I'm excited to watch StarCloud 2 and StarCloud 3 make it to orbit. Thank you for joining us, Sem. Are you coming home eventually?

Sem

I am. I'll be here for a few more days, and then I'll be back.

All right. And Alex, how about you? What's your travel schedule looking like?

Alex

Well, I don't know. I'd love to visit LEO or sun-synchronous orbit sometime soon. Philip, we should chat.

We should. We should. And Peter, I think you're going to be in Paris in a week, so I might see you there. I think I may be there virtually. Oh, no, I see.

Wait, where am I? I know. I'm in Calgary.

Alex

Calling the kettle black.

Yeah. Calgary, Germany, and Greece. Yes. Okay, okay. All right. Love you guys. Again, thank you, buddy. Thank you, Alex. Thanks. Take care, guys.