Hi, I'm Katherine Schwab. I'm an assistant managing editor here at Forbes covering tech. And I'm so excited to be here today to talk about data centers in space. Everyone is talking about this right now. Elon Musk, Jeff Bezos, Google, and I'm so excited to be here with Philip Johnston. He is the CEO and co-founder of Starcloud, an Nvidia-backed startup that just launched the first GPUs to train AI models in space. Philip, thanks for joining me.
Thanks so much for having me.
So, Philip, let's start with the most obvious question. Why should we be building data centers in space? What's your vision for how this is better than building data centers on Earth?
Yeah, it's definitely the right place to start. So the reason we're building data centers in space is because we're very quickly running up against energy constraints on Earth. And in space, we have access to almost unlimited low-cost energy in the form of solar. So, you know, if you want to build a new, you know, 100-megawatt or gigawatt-scale data center on Earth, first you need to build a new 100-megawatt energy project on Earth. And the lead time on permitting for that can be, you know, decades. Whereas in space we can launch these things very quickly and be up and running instead of in, you know, years and decades, within weeks and months we can be up and running.
Okay, but it must be so much more expensive to actually do.
Well, right now it's more expensive. We have a bunch of new launch vehicles coming online which will lower the cost of launch very significantly. And so over time, we actually expect the all-in energy cost of doing it in space to be 10 times at least cheaper than doing it on Earth. So rather than for a new energy project on Earth, you know, you might be looking at between 5 and 10 cents per kilowatt hour. In space, we're looking at, you know, less than half a cent per kilowatt hour including the launch cost once we have these launch vehicles like Starship online.
Okay. So, don't you need though really like massive solar panel installations to be able to make this work? I mean, that's what the International Space Station uses. So, can you explain what the satellites you're building look like and how they will possibly generate as much energy as you're talking about?
Yeah, for sure. So on the front page of our website, starcloud.com, people can see an animation of what a 5-gigawatt data center in space might look like, and that has a 4-kilometer by 4-kilometer solar panel with a 1-kilometer by 4-kilometer radiator down the back there. And since you mentioned the ISS, I think it's probably a helpful comparison. The International Space Station, you could launch the entire International Space Station with two Starship launches. Now, Starship is going to—they're producing three—they will be producing three Starships per day. Each one can potentially fly multiple times per day. And the capacity builds on itself. So, we could be seeing, you know, hundreds of Starship launches per week within a few years. And so, yeah, I mean, the ISS is going to be considered very cute by the standards of five years.
Cute. Okay. And of course, when you're talking about Starship, you're talking about Elon Musk's much, much bigger rocket that he has been testing at SpaceX. Can you talk a little bit more about how Starship represents this kind of paradigm shift for what we're able to do in space?
Yeah, for sure. So, Starship, for those that don't know, it's this absolutely massive rocket that SpaceX are building down in South Texas. It is the first ever fully reusable rocket. And that's extremely significant because right now when you launch a rocket, you have to rebuild the whole upper stage every time. That's like, you know, 10 or 15 million just thrown down the drain every time you do a launch. It would be a bit like if you fly from London to New York and you had to rebuild the plane when you landed to get back. That's literally how the space industry operates right now. With Starship, it's reusable and so that means you can just continuously fly it, a bit like flying an airplane. And so it literally lowers the cost of launch by, you know, between 50 and 100 times the current price. So with Falcon 9, the current launch vehicle that everybody, you know, most mass is put into space with this launch vehicle, it's currently around $6,500 a kilo. The marginal launch cost with Starship is $10 to $20 a kilo. So the additional cost per launch for SpaceX is, you know, much, much cheaper, like at least 100 times cheaper. The customer price might be like $50 to $100 in the end state. That's what they said they're targeting. But we're break-even at around $500 a kilo. So well within range of the Starship prices.
Okay. So these solar panels that you're talking about, which are really core to making this entire idea work. You said those are four kilometers by four kilometers. How if you were to—yeah, are those kind of folded up and then they expand? What happens if they break? Like, I mean, there are so many pieces of this idea that I think are hard to understand, like practically how this makes sense. So, I'm curious, like where you're at in the engineering stages for actually building this satellite and how you're thinking through those practical problems.
Yeah, for sure. So, just for those who aren't aware, we've just launched our first spacecraft. It has the first Nvidia H100 on board. It's about 100 times more powerful GPU compute than has been in space before. And the purpose of that one was to prove that we can dissipate the heat from these powerful chips and that we can operate them in a high radiation environment. So, we've proved a significant portion of the technology already, although that's a much smaller version for the large 4-kilometer solar panel. To be honest, it's very unlikely we would need to dock together a data center that large until probably at least 10 years before you need to dock something like that together. And the reason is, the only reason you need to dock it together is if you want to train a large model in space. And so you need the whole thing to be internally very high-capacity internal networking. For most tasks you don't need that. So for almost all inference tasks, which is like when you type into ChatGPT and it comes back with a response, you don't need to internally network a data center. So for that we will just launch—every time we launch we'll launch one full Starship payload bay. That's about 150 tons. It's, you know, 20 to 30 megawatts per launch. And that will contain its own solar panels and radiators on board. So, yeah, you're talking about football field sizes rather than kilometer sizes of solar panels and radiators. And for that, yeah, you have these very thin-film solar panels that you can roll up and unroll.
Gotcha. Okay. So yeah, let's talk about the cooling piece of this because data centers on Earth right now, of course, use tons and tons of water to keep these chips cool because they're dissipating so much heat. There are no fans in space. There's no water in space. Can you explain the physics of why you need to radiate it and how that works?
Yeah, for sure. And that's a great question. So as you mentioned, in space we don't have convection or conduction because we don't have any medium through which to transport the heat away from the chips, and so the only cooling technique we can use in space is radiative cooling in infrared radiation. So all things that are warm kind of glow infrared. If you had an infrared camera on your face you would see that your face is glowing continuously in infrared, and it's the same for things in space. But what you need is a large surface area. And so the way we do this is we run coolant past the chips. It's a fluid, and then that runs out to these very large radiators through channels, and it keeps the radiators at around 50 degrees Celsius. Now the background temperature of space is about minus 270 degrees Celsius. So there's a very big temperature gradient. And so what that does is it emits in infrared at about 800 watts per square meter the waste heat. Now the solar panels only generate 200 watts per square meter. And so what that means is you need about four times the surface area in solar panel than you have in radiator. So if you were to have one square kilometer of radiator, that would be four square kilometers of solar panel that you would be dissipating the heat from.
So it sounds like there are a lot of engineering challenges here to kind of get to this much larger scale that you're talking about versus kind of the smaller chunks.
Well, I would call them manufacturing challenges. So the International Space Station has these large radiators right now which dissipate about 70 kilowatts, which is for the next generation about the size we're looking at. So we know it's doable. It's not really a physics problem, there's not too much hardcore R&D that needs to happen on the physics side. The problem with the ISS radiators is that they are expensive and heavy. So our challenge is to make them cheap and light, and that is more of a manufacturing problem than it is anything else. And so we have some very innovative manufacturing techniques. On our second spacecraft launching next year, it has 100 times the power generation of the first. Our radiator has 10 times lower mass per watt of dissipation than the International Space Station radiator and 100 times lower cost per watt of dissipation than the International Space Station radiator. So it's okay, all things are solvable in this regard.
Okay. So underlying all of this is the demand for more data centers. Of course, there's also kind of talk and fear about if that demand is going to be founded, like if, you know, all of these massive projects going up right now, is that demand really going to be there to support the scale of the buildout that is happening right now? It seems like you're really betting that not only is the demand going to hit the levels that we're seeing right now in terms of what's being plotted out on Earth, but that we are going to need so much compute that we need to go into orbit. So help me understand a little bit more about kind of those bigger assumptions that you have that are kind of at the core of the company.
Yeah, for sure. So actually my expectation is that within 10 years most new data centers will be being built in space purely because of the constraints we face on energy terrestrially. So I would reframe it slightly. It's not like we think there's going to be tons of demand on Earth and additionally we need all of this demand in space. It's more like there's tons of demand. It's actually a lot cheaper to do it in space than it is to do it on Earth, and so we will do it in space rather than doing it on Earth. So yeah, I wouldn't say we have a different forecast than most people in terms of the projected demand for AI compute. It's just that we can see that it will be cheaper to do it in space than it will be to do it on Earth. So if you're going to build data centers anywhere in 10 years' time, and I presume they're not going to completely stop building data centers in 10 years' time, then the cheapest place to do it will be in space purely because of the energy.
