Data Centers in Space: The Future of AI Compute | Philip Johnston, @starcloud-inc (YC S24)

Thanks so much for having me.

Yeah, for sure. So, I mean, maybe there's a few different stages before we get to these very large data centers. I think I can start with what the very long-term sort of multi-gigawatt data centers look like, if that's helpful, and then work back. So for a very large data center, you're, you know, for example, a 5 gigawatt data center, you're going to need around 100 of these 50 megawatt modules. Each module is going to be what you can launch on one full Starship payload bay. So the big difference between, there's two big differences between a terrestrial data center and a space-based data center. The first is in space, obviously, you don't have gravity. So you don't need to build in this form factor where everything is supported from the base. And the second is you only have launch capacity for about 100 tons with the Starship new launch vehicle that's coming online. So everything needs to be shipped up in these 100-ton sized modules. So yeah, the concept that we have for the larger data center is in the center we'll have a core which a bunch of modules plug into. So these will look like almost like large shipping containers which on the end have a docking port. So this docking port will plug into a central spine. And in fact, if people want to see a video of this, a sort of concept video of this, they can go to starcloud.com and on the front page of the website, we have this concept video of these modules docking into a central spine. And then from there out comes this absolutely enormous solar panel with a radiator built in on the back side of it. And so for a 5G data center, you're looking at around a 4 km by 4 km solar panel with around a 1 km by 4 km radiator on the back side to reject all of that waste heat. So yeah, that's the main difference in terms of what it looks like for the smaller versions. So we're launching a demonstrator satellite which is a much smaller, like a 60 kg smallsat the size of a fridge, that's going to launch in November. That will be the first to have high-powered GPUs, particularly H100s from Nvidia, on board. And the purpose of that is just to prove that our radiation shielding and thermal management techniques allow us to operate terrestrial data center grade GPUs in space. But yeah, essentially we scale up every year, you know, 5 to 10x the size each year until we get to in the early 2030s this huge 40 megawatt module that can be connected together in space to form an Omega structure. Does that answer the question?

Yeah, for sure. So in order, you know, as you mentioned, you can't launch this whole 5 gigawatt data center together in one go, and so it needs to be broken down into these manageable chunks. And we think it's around 40 or 50 megawatt, 100-ton Starship launch. So that each one of those modules initially will contain everything that it needs to be self-sustaining and self-sufficient. So each module will have enough solar panels and radiators to sustain that module on its own. And so even if you didn't hook them up in space, you would have, you know, you would have the modules being able to operate on their own and provide a useful service. But yeah, essentially inside the racks, you have coolant coming from the radiators from an outside loop that connects to an inside loop. That coolant then runs past the chips. And then, you know, and it's very similar to the way that liquid cooling on chips, direct-to-chip liquid cooling works today. That is pushed back out to the radiators. And then yeah, so essentially we're providing power, cooling and connectivity to these chips, and then everything else is done by, we hope to partner with somebody who's going to actually operate the data center themselves. But it shouldn't be too different an environment than a terrestrial environment in that regard.

Russell Long was the guy.

So usually they don't show us the investment memos. So actually I don't really know what is in those investment memos. I can tell you what the top five complaints that people have are. And you sort of, the first time you pitch, you hear like one of these things, like you roughly have an idea, but what I do is every time I hear a new question, I add a slide to the back of the deck. And so by the time I've done like five or six pitches, I basically have a slide for every possible, you know, like objection that people have on the idea. The top five are: how are you going to keep this thing cold because space is a vacuum and it's hard to dissipate heat in a vacuum? How are you going to make the chips work in a high radiation environment? How do you handle the problem of orbital debris and micrometeorites hitting such a large structure? How do you handle security, data security or physical security from adversaries? And how do you handle the regulatory environment for permitting such a large structure in space? So essentially, you have to have a well thought through and plausible, like an actually good answer to all of these points. Otherwise, you know, you're not going to get very far. I won't run through that whole list, but unless you want me to, I'm very happy to if you want me to, but maybe I'll wait for the questions and then when people, you know, have... Oh, the last one is how do you handle maintenance when you don't have access for five years? So that's going to suck.

Yeah, and then, oh, and then lastly, how have you proved that there's, what's your commercial validation? How have you proved that there's people that actually want to pay for this? And then people say, another one is people say, you know, what's the competitive landscape? What's to stop somebody like SpaceX or Blue Origin doing...

Yeah, actually there's probably like 20 different questions that people ask that you have to have a pretty coherent answer to. But I would say though, the first time we pitched to Russell, I didn't have a good answer to any of those. So I mean, I'm still pretty grateful that he, I think he might have actually been the very first investor call I took, and he wrote the best check. And I thought after he wrote that check, I was like, 'Oh my god, this is going to be the easiest fundraise ever.' And then it was not that much easier after that anyway.

Thanks very much. That's fantastic.

Yeah, so the...

Yeah, no, it's a very good question because the capex, actually the biggest expense in everything that we're doing is the same as the expense in a terrestrial data center, which is actually the cost of the chips. And so one of the biggest levers in making the business case close is the depreciation timeline, as in how long are these chips going to generate cash for you and like at what rate does the cash they generate decline over time. So no, it's a super important question. But so, you know, terrestrial data center chips last between four or five years. They have a useful life for about four or five years. We expect it to be similar in space. And in fact, there's a case to be made that it might be slightly longer because we don't have an oxygen atmosphere which is quite corrosive. We have a zero cost, a zero marginal cost of energy. So running them longer doesn't cost us as much as it does terrestrially. And we also don't have some environmental factors like earthquake tremors which can make things fail over time. We do have radiation, but over time the radiation doesn't, the larger the satellite, the less of a problem radiation becomes. So essentially five years for the chips. The infrastructure can last longer than that though, the radiators, solar panels, and or networking infrastructure can last a bit longer on the order of between 10 and 15 years. And so that's why we have this dock-in, dock-out architecture for the later versions. For the initial versions, we are expecting not to reuse the solar panels and radiators for more than the five-year lifetime. So the first few satellites, after they finish working, will be the same as Starlink satellites. So they'll demise in the upper atmosphere on re-entry essentially. Or later versions, if we have these extremely large 100-ton versions, it may become a problem to make things demise in the upper atmosphere. And at that point, either we put them in what's called a graveyard orbit, you know, an orbit where they're not going to cause any problem, or we re-enter them with the Starship launch vehicle, which will have the capacity to re-enter various things. Won't be too cheap to start with, but over time, I think that cost will come down.

Yeah. Yeah. Yeah, the orbital maintenance part is a, or maintenance on orbit part is an interesting question as well. So for the first few iterations of the satellite, we're not going to have access to it. So it will be very similar to the way that a Starlink satellite works, for example, which is that you have redundancy on the critical systems, you know, triple redundancy in certain instances, and then you have overprovisioning on aspects that you expect to degrade gracefully over time. For example, the solar panels and the radiators, you make them slightly larger than they need to be so that if you, you know, you're losing about 2% per year on the efficiency on the solar panel. So by the end of life, you still have enough power to power the whole module. So for the first three or four satellites, that will be the approach. There will be no access to the, for the data center on space suit. You just have to operate like any other satellite. Over time, the space industry is moving towards robotic maintenance. You see this on the International Space Station, but also on other missions. And actually, terrestrial data centers are also moving towards robotic maintenance. I had a conversation with somebody recently who said, you know, it wouldn't be surprising if within five years we see humanoid robots maintaining terrestrial data centers. And I personally think that makes sense. The interesting thing about humanoid robots is they actually operate with basically zero modification in space. So if you put a humanoid robot in a spacesuit that takes care of thermal management or radiation shielding. So, you know, it wouldn't surprise me, this does sound a little bit sci-fi, I admit, but it wouldn't surprise me if on the order of five years we start seeing humanoids building structures in space. Yeah, we're not reliant on that, but that's my hunch of where this thing's going.

When you say on Earth, what do you mean by...?

That's a good question too, but we're not going to be doing our own data backhaul. So we were very... I have to be a little bit careful because it's, we're under NDA. But we will very likely be using one of the mega constellations for our own data backhaul. And that means, yeah, we don't, we won't need to rely on ground stations or use or rely on our own inter-satellite connectivity to get data back to and from the satellite. So I don't think we'll have a ground footprint at all really, except for, for example, this facility behind me.

Yeah, that's a good, yeah, that's a good question. Or it's something we think a lot about is how, you know, it's quite a capex-heavy business, especially at the beginning. And so yeah, there's a few different approaches. One is you can spend an enormous amount of capex building something very elaborate and then four years later launch it and hope it works, and if it doesn't, then basically shut the company down. The other is you can launch every year and launch something that's a more incremental step, but proves out a certain part of the technology and is not such a big risk at the end. We've taken the approach, the first approach, which is, sorry, the second approach, which is we're launching incrementally. So first launch is going up in November this year. That's going to have the first H100s on board. It will be about 100 times more powerful GPU compute than has been in space before. The second one is going up in October next year. That's going to be about an order of magnitude more power generation, sorry, power generation. And we'll have the first Blackwell of the Blackwell architecture from Nvidia. That one will be the first one that provides a commercial service. But to be very frank, we are taking an approach where we want to generate cash from the second satellite onwards. We're aiming to generate more cash than it cost us to design, build, and launch that satellite. I mean, also to be fair, it's, we're doing it pretty cheaply. Like the first satellite, we raised a $2.5 million pre-seed, and the whole of the first satellite we designed, built, and will launch for that $2.5 million, which is like dirt cheap compared to any other space startup I've ever heard of. The second one, yeah, we can do with the seed money that we raised. So yeah, I mean, the plan is essentially always to have enough cash to be able to survive the next two launches at any given time. So right now we have enough cash to get through till mid-2027, that gets us through the second launch at the end of next year. But yeah, we, and so we'll basically try and raise after the first launch. And if for some reason there were to be a failure on the first launch, we are not then totally screwed. We have enough cash to be able to get through to the one after. But yeah, I'd broadly say that our investors are supportive, and I think they're supportive because it derisks it a lot to them if we can iterate quickly rather than if we have to wait until there's one big one and have all of this risk on just one satellite.

Yeah. So the initial customer set are people that will pay for the R&D up front, and the only people that do that are not really commercial customers because commercial, they want to see a product on orbit before they pay. So it's mainly DoD, so Department of Defense in the US, and government satellites where there's all of these different programs for funding R&D as you go through that journey. For example, the SIBR program. That's the main people that we're working with for the first few iterations of the satellite is Space Force, Space Systems Command, AFL, Air Force Research Lab, Space Commercial Office, TAC SRT, a bunch of these guys, DIU, Defense Innovation Unit. Yeah, it's mainly defense to be frank. Because so yeah, the first iterations of the satellite, we are providing an edge workload for other satellites. So we're providing cloud and edge services for other military and commercial satellites. That business on its own can close without needing very low launch costs from Starship. And in terms of competitors, you know, there's lots of people putting various levels of compute on orbit. As I understand, we're the only ones putting terrestrial data center grade GPUs on orbit. So things like H100s. Everybody else is putting, you know, Jetson Orin or other types of sort of space-rated or space-grade or ruggedized chips, which are much more reliable but about 100 times less performant in terms of compute capacity.

Yeah, true. No, I wouldn't say we're, for the first three or four years, we're not really competing at all for terrestrial workloads. So when people talk about latency, they're talking about sub-20 millisecond latency through Starlink, and that allows you to do gaming, for example. This is honestly not really, not really related to the first few iterations of the satellite. Later it will become relevant for sure. The first few iterations of the satellite, we're mainly talking with Earth observation satellites who produce enormous amounts of data on orbit. So what they currently do, so for example, BlackSky or Planet Labs or Spire, or maybe not Spire, but, you know, any of these Earth observation constellations, is they will take an image and then they will have to wait, or they'll take, you know, lots of images. They will then have to wait for a ground station, and that could take half an hour before they get to a ground station. Then they'll downlink as much imagery as they can, which usually is about 1% of the total imagery that they can take, because there's very constrained links between space and ground using these RF antennas. And yeah, so that's our main first customer set. We will connect with those customers using a very high bandwidth optical terminal in...

Space to space optical has a much higher bandwidth than space to ground RF, and also space to space optical you don't have to wait for ground station. So yeah, we'll essentially take enormous amounts of raw imagery from customer satellites, process it on orbit. So for example, run an inference workload. So let's say for example there's a satellite that wants to detect wildfires in California, so it has a constellation which is continuously passing over California. Right now they're just taking petabytes of imagery of just green forest essentially, and they have to wait till they get all that imagery to the ground before they can run any inference on that to see if there's a forest fire in a certain location. In future they'll be able to ship all of that imagery to us, we run the inference workload, and then we just in almost in near real time we can downlink the insight, and the insight might be there is a fire in this location moving in this speed and direction.

But yeah, that's the so when we talk about latency, we're talking about instead of waiting an hour and a half to get your data back, you might wait, you know, one second to get your information back.

One thing which people are not really aware of is there will come a time in the not too distant future, within our lifetimes probably in the next 50 years, where building data centers on Earth becomes impractical. Even if we solve nuclear fusion and we have almost free unlimited energy terrestrially, there will come a time when the sheer waste heat that we're pumping out from these data centers is enough to warm the earth to the point where you're either boiling the oceans or making the planet extremely uninhabitable.

And it's on the order of between 10 and 100 times the current energy capacity we're using for data centers. And so whether we like it or not, we have to, we will, if we want to scale compute beyond a certain point, and as I say it's around 10 or between 10 and 100 times the current capacity, we will have to figure out building data centers in space because this is the only place you can dissipate that heat in infrared into the vacuum of space. I think not many people are aware of this or see this as a significant problem right now, but I think over time on a sort of 20-year time frame at least, this will become very apparent.

Great question. I will answer it aspirationally. Ideally we do not do anything inside the box. So we would like to be considered more like an energy provider than a data center provider. So we will, and we are in the process of signing a very large LOI with the biggest data center builder in the US, where essentially we will provide an environment, an earth-like environment with power, cooling, and connectivity, and whatever anybody wants to do with that. We'll provide power at let's say 3 cents per kilowatt hour, and whatever anybody wants to do with that is up to them. The biggest cost, as I say, by a mile, is the cost of the chips, and we would much rather not be in the game of financing and operating these for people who have very strong opinions about how they want them to be operated. So yeah, we would rather not be doing that. In the initial days, we have to do that because nobody else knows how to do this stuff in space at the moment. But yeah, over time I hope we don't do that.