Elon Musk explains how SpaceX could build AI data centers in Space

People probably struggle to visualize a little bit when you say, like, data center in space. Like, we're not going to slap engines on a building and fly it up there. Like, these actually look, like, pretty different. And so, kind of walk through how you take something that's in a giant building on the ground and turn it into something that's functional in space. Yeah, I think it's pretty interesting. A lot of people don't actually know what the inside of a data center even looks like, right? Yeah, it's some, like, mythical place where the internet's in the cloud or something. Yeah, some people envision wires, some people envision boxes. But, like, effectively it comes down to a set number of chips and the things that we need to launch into space are actually quite small when we look at it. The more challenging part is figuring out how do you get the power for it? And that's where a lot of what we've worked on for existing, like, Starwing technology, the solar arrays, are what we want to utilize that expertise to be able to build a satellite that can actually launch the critical components of the data center into space itself. We like to look at this and say, like, what is the actual engineering problem here? And it's really a combination of delivering power and then taking the waste heat and energy away and sending it into the vacuum of space, as you mentioned. Yeah. Now, the AI satellite is actually much simpler than a Starlink satellite. A Starlink satellite has gigantic phase ray antennas. It's got parabolic antennas. It's got a lot of laser links. It's much more complicated than an AI satellite. An AI satellite is essentially a lot of solar cells, a radiator, and you still need some laser links, but you don't have all of the super complex antennas that you have on a Starlink satellite. So, I mean, given the two, the easier one to design for is the AI satellite. It's just a little bit bigger. It's bigger. Just makes stuff bigger, yeah. I was like, so we've got, this is our AI one, if you guys want to walk us through. Yeah, so the first thing that we're really looking at here is, like, first you've got to make something compelling, right? And we thought that the right place to start is around the 150 kilowatt, like, peak power level. But as we look at the workloads with our experience with XAI, we get to actually see that we can also support about 120 kilowatts of average compute. There's a difference. What we're showing here is kind of a draft version of the version one of the SpaceX AI satellite. The AI one, I guess you could call it. And it seems like a reasonable place to start is 150 kilowatts peak power, 120 kilowatts sustained power. And to give you a sense of what does that actually look like in terms of the size of the radiators, the size of the solar panels, the assumptions here are 250 watts per square meter for the solar array and about 1,400 watts per square meter for the radiators. So the radiators, these are double-sided. Radiators are radiating both sides. They're oriented knife-edge to the sun. And it's 1,400 watts per square meter is a very achievable goal. Over time, we think we can probably do above 250 watts per square meter and above 1,400 watts per square meter for the solar panels and radiators, respectively. But this gives you, this is pretty much what the satellite is going to look like. It's a lot of solar panels, radiator, and then everything else is pretty small, like the virus. And these are like evolutions of things that we have actually already launched in our Starlink constellation to date. Yeah. That's really, I think, the cool part to me is that we're looking at solar technology that we already are going to use on the V3 Starlink vehicle. So I'm really excited to then just take those and make it bigger. Yeah. Part of what we want to convey here is that there's not some magic that's necessary that doesn't exist for the AI satellites. As Ian said, a lot of this is technology we've already made for the Starlink V3 satellites. So basically, we don't think this is a super hard problem compared to things we already do. There would also be probably something on the order of a terabit of laser link connectivity from the satellite. The 150 kilowatt peak power level roughly matches what, say, an NVIDIA GV300 rack would do. So if you've got a GV300 with 72 GPUs, its peak power, I think, is around 140 kilowatts. But it's almost impossible to be at that peak power. A more reasonable operating envelope would be around 120 kilowatts average power. But it can peak up to 150. So it's basically, think of it as a rack of compute in space. And then you can connect these racks of compute to either each other by the laser links or directly to the Starlink constellation. So you can close the link with the Starlink constellation. And then Starlink can then send that data to the ground using the existing KA and KU antennas on the vehicle. It also has laser links to the ground as well. And this would not be at a particularly high latency. We're talking about maybe being around 600 to 800 kilometers above the Earth. And light travels 300 kilometers per millisecond. So that's about three milliseconds away, basically. It's not very far. Won't worry about that too much. Sometimes people think there's going to be some high latency. I'm like, no, light moves pretty fast. Light moves pretty fast. It's a tall one. Yeah. I think the cool thing also is the radiators themselves are about the same size as the existing solar arrays for the V3 vehicle. Kind of in that realm where we're flying today. Yeah. So, I mean, they got about a 70-meter wingspan. So these are fairly large. And we're talking about building a lot of them and putting them up there. But you like to say, like, space is in the name. Like, there's a lot of space up there. And so even when you're talking thousands or even up to a million satellites, you got plenty of room to move around up there. Yeah, space is really big. So it's not like space is going to get crowded. Space is enormous. Like, if you zoom in close, the satellite looks big. But if you actually look at it relative to the Earth, the satellites are so tiny you can't even see them. So they're very, very tiny compared to Earth. And, I mean, we have about 10,000 Starlinks in orbit right now. We've got a pretty good idea of how to operate just really large constellations and do it safely now, right? We are the only operator that has any experience of that scale. It's a great thing that, you know, we have this background so we know how tightly we can pack the satellites and fly them safely. That's our number one goal when we look at the constellation. We're going to be building a lot of satellites. And we're going to be building them here in Bastrop, right? So we've got this. So we're in that building kind of in the middle. Yeah, we're sitting in that building right now. This is my first time here. The building is massive. Like, you come around the corner, you see it through the trees, and you're like, oh, wow. But we're about to kind of put this building to shame, aren't we? Yes, we're going to. In fact, we already have the solar manufacturing facility. It's under construction already. And then we will be building out the AISAT production building soon. And, yeah, so we expect to have the AISAT production, the solar production, and all of that operating at some reasonable volume by the end of next year. So if anybody wants to work on AI satellites, this is kind of going to become the hub of that. So, I mean, like right behind us, the machines are humming. We're still making all of our user terminals for Starlink here. That's not going anywhere. In fact, we're turning on new production lines for new units, right? Yes. In fact, these are the new Starlink terminals, which we've made in much higher volume than the current terminals. And ultimately, we think there's probably going to be a few hundred million Starlink terminals out there. And then the Starlink direct-to-cell constellation will connect directly to people's cell phones and enable high-bandwidth communication directly from your phone to space. All right. We're two limiting factors down. We've got mass-to-orbit. We've got putting solar in a few third-ones chips. Yes. So at least in the beginning, we can obviously launch the chips that are already being made. So our current reference design is for NVIDIA Rubin chips, or it could be either GB300 or Rubin chips. And we'll also have a reference design for TPUs. And essentially, you can put any existing chips into orbit. But the current industry seems like it's going to get to maybe around 100 gigawatts a year of AI compute. But that doesn't answer the question of, well, how do you get to a terawatt? That's why you need the TeraFab. Oh, he's looking a step bigger. Yeah. In order to get to the next order of magnitude, you need a gigantic ship factory. And to give you a sense of scale here, we expect that the TeraFab is going to be around 100 million square feet, which is 10 times the size of the Tesla Gigafactory Texas. And what, aside from just, you know, I'm going to need Starship point to point to get from one end to the other. Aside from just the size, what's going to make this unique, different from any other chip building operation on the planet? Well, I think over time, there's going to be a lot of technology evolution with the TeraFab. But fundamentally, it's about scale. So even if there were no fundamental technology breakthroughs, you could simply scale the existing chipmaking technology with a lot of difficulty to a terawatt of chip output per year. If you look at it just from the Logitech standpoint, that's like having a billion chips per year with a kilowatt per radical. So it's a billion full radical equivalent chips, each doing a kilowatt. And then you're going to need a lot of memory to go with that. A lot of people today even think orbital data centers were like a decade away. Yeah, I think we want to try to give people a sense of the timeframe. At least the timeframe we're aiming for. I mean, you know, people should take this with a grain of salt to some degree because this is just our best guess. So this is not a promise of what we'll do. This is what we are going to try to do and think we probably can do, which is to get to roughly an annualized rate of a gigawatt per year by the end of next year in terms of space AI compute. And then aspirationally scale that by an order of magnitude per year. So in two and a half years hitting an annualized rate of 10 gigawatts a year to space and three and a half years, maybe 100 gigawatts. And then depending upon what progress there is in chip making in the rest of the world and with the TerraFab going beyond that to scale to a terawatt per year, which is a thousand gigawatts. That's twice the current electricity consumption in the United States. I think there will be an appetite for that, but we'll see. It's a lot of satellites. I don't know what it's going to think about, but maybe do a lot of simulations or something. Yeah. So after we've, you know, broken through all the limiting factors, we've kind of topped out what we can do on Earth. What is the next step to, again, try and actually notch maybe some percentage points towards becoming Kardashev level two? Why stop there? Yeah. Why think small? Because a terawatt actually is very small. I don't want you to think small. Let's not think small. So there is, in order to get to another three orders of magnitude to a thousand X from a terawatt per year, the only way that we can really see that you can achieve that is on the moon with a mass driver. Essentially, where you do local production of photovoltaics and radiators on the moon. Maybe you bring the chips from Earth, or you could conceivably make the chips on the moon, but you need most of the mass to be made on the moon, so you don't have to transport it to the moon from Earth. And then, because the moon has no atmosphere and only one-sixth Earth's gravity, you can accelerate the AI satellites into deep space without a rocket. So you can basically shoot them into space using an electromagnetic gun, like a rail gun type. I mean, it's basically a linear electric motor is the way to think about it.