Starcloud's Philip Johnston: Why the Cheapest Compute Will Be in Space

0:01 Thanks so much for having me. my name is Philip Johnston and I'm the co-founder and CEO of Star Cloud and just like the previous company we have also been abusing GPUs in ways they were not designed for. >> >> so yeah, we're building data centers in space mainly for the energy that we can draw and I will spend the next 5 minutes explaining why it will soon make much more sense to build data centers in space than it does to build them on Earth and then I'll take 5 minutes for questions. So please start thinking of some questions. Before I do that though, I want to show a quick video which is actually the deployment of Star Cloud 1 and this was it had five in video GPUs on it but the most significant was the in video H100 chip and I'll just quickly play the video first.

0:50 Star Cloud 1 separation confirmed. So we don't you don't normally get as great a deployment video by the way. Half the time they'd like deploy it into the shadow. so the reason this was so significant is until this point many people thought you actually couldn't run state of the art terrestrial data center grade GPUs in space for two main reasons. One is the thermal dissipation. So they they're very power dense. They produce a lot of heat and the second is the radiation tolerance. So people thought that you would have bit flips at too high rate and so by with this chip we were the first to train a model in space. We actually trained nano GPT from Andrej Karpathy and then we also were the first to run a version of Gemini, the first to do high powered inference on SAR data. So other satellite data.

1:37 and so it was a very significant step in proving that we can actually run the state of the art terrestrially. but the yeah, I think maybe to make the case for why it will soon make more sense in terms of energy cost, I'd like to quickly draw a comparison with with a solar project on Earth since solar is the cheapest form of energy that we have on Earth. So if you want to build a solar project to power a new data center, you have three main costs. So the first is the cost of permitted land and in fact in North America that's actually the largest cost or can be for most new solar projects.

2:12 The second is the cost of battery storage and backup power because we're only you know, we only have peak power for about 4 hours of the day. So we need to charge those batteries to use at night. And then the last is the cost of the solar cells themselves. So how does that compare to building a similarly sized solar project in space? Well, in space number one we don't need to pay for permitted land. So your biggest cost is gone. You don't need to pay for battery storage and backup power because we're 24/7 in the sun. So your second biggest cost is gone. And then you need eight times less solar cells cuz 1 square meter of solar panel in space produces eight times the energy of 1 square meter of solar panel on Earth.

2:49 So the only additional cost or the main additional cost we have in space is the launch cost. >> >> And so you can clearly see there's a break even point where the launch cost comes below the cost of permitted land, batteries and solar and we see that break even cost to be around $500 a kilo. So about a 10x reduction from where we are today. But that's well within range of the launch vehicles that are coming online. So for comparison, Starship is designed to produce launch cost of around 10 to $20 a kilo.

3:18 And so I think I'll just finish by playing you a one final concept video and this shows a constellation that we're building now. So we've just filed with the SEC for a constellation of 88,000 satellites. Each one's about 200 kilowatts. It will enable us to deploy on the order of 20 gigawatts of new compute capacity. Really just scratching the surface with this new constellation. and it will enable it's basically for all inference workloads. And so this could be and yeah, maybe I'll I'll start the video and you can get a sense of it.

3:51 So in this case it's to generate a 3D video but it could also be for back office business processing agents, code generation agents. they will come up via optical link to this constellation in this dawn dusk sun synchronous orbit. Means it's always in the sun. 24/7 power. Sub 50 millisecond latency to anywhere on Earth. all optically linked. >> >> And this really is the start of the largest infrastructure project ever. I mean we're talking about just for this constellation of 88,000 we're talking about a hundred billion dollars of capex spend which is actually much lower than it would cost to do to do terrestrially.

4:26 and not only is it the start of the largest infrastructure project, it's also in my opinion the start of a Kardashev type two Dyson sphere type civilization and potentially Kardashev type three. I will finish there and we have about 4 minutes for questions. So any questions? Yeah, we'll start at the front. Yeah. the intuition on the availability solar is obvious. can you just give us the napkin math on the the radiator equation again? So like dissipating heat for anyone who's thought about it, it always feels hard. And then and also please say something about the availability of dawn dusk. That orbit is finite, right? Yes. Yeah. Yeah.

5:01 it's a great question. So because space is a vacuum, it's actually much harder you know, space is only 3° Kelvin. So very low ambient temperature. But because it's a vacuum as you rightly point out, it's actually quite difficult to dissipate that heat and what it requires is a large surface area so that you can emit that in infrared. So everything that's warm is glowing in in infrared all the time. If you had an infrared camera on my face, you'd see that I'm glowing. and so the rough math on these surface area is you would your solar panels generate around 200 watts per square meter and the radiator if you keep it around 50° C will dissipate around 800 watts per square meter. So what that means is if you've got a you need about a quarter again the surface area in radiator than you have on on solar panel. So if you had a 400 square meter solar panel, you'd need an additional 100 square meters of radiator to dissipate that heat.

5:56 there's a very nice equation called the Stefan-Boltzmann equation which basically says that the rate that the the thermal dissipation is proportional to the to the fourth power of the temperature. So if you can jack up that temperature instead of being 50° to 80°, which is like a 10% increase in Kelvin, then you you can actually half the surface area of your radiator or close to half the surface area of radiator. And so that's what we're working on with Nvidia now. If anybody was at GTC, you'll have seen Jensen walk out to this to the deployment video of Star Cloud 1 and then he spent 5 minutes talking about the the new space Ruben 1 chip that we're working on and it's designed to run at a hotter temperature without having a a higher failure rate. And the reason you want it to run at a hotter temperature is so that you can lower the mass on the radiator.

6:40 Great question then. Yeah. Yeah. All right. Kessler syndrome, the favorite thing, right? Now we're going to get more satellites there yet more in Kessler orbits cuz everyone will want What happens then? Yeah, it's a great question and it's also related to the question asked about the space in space. I think so it's something we take incredibly seriously. You know, everybody needs to be a responsible user of space. We do and and everybody else. We're you know, keen keen to make sure that space is usable forever.

7:10 For the first few satellites, so you can solve it in a few ways. If you fly at a relatively low altitude, the chance of a Kessler type effect is is extremely low. So our first satellite we're flying around 400 km altitude. That means that it will naturally deorbit within a few months. And so if you were to have a collision at that altitude, by the time it gets around to the next orbit, you're already a a few hundred meters below where you had the collision and the chance yeah, chance of Kessler is very very low.

7:37 Yeah, as you fly higher, it's actually extremely unpopulated those high orbits because then you start to edge into the the Van Allen radiation belt. but I mean we actually have a pretty good case study for this and that is SpaceX is now operating around 10,000 satellites without ever having a single collision in in low Earth orbit. And the the way that you do that is by having ex-pretty sophisticated collision avoidance. the the the the other reason I think people think this is more of an issue than it than it actually is and and the reason that space is so much larger than it looks is when you see a map of all of those satellites, each dot on those maps is about the width of California and you're representing something that might be this wide by something the width of California. And so people can often think that space is very congested. It's actually it you you we can easily fit on the order of terawatts of compute in just this dawn dusk sun synchronous orbit without having you know, huge problems with collision avoidance.

8:36 Any other questions? Yeah. is radiation like bit flipping, is that something you actually have to think about or consider? How does that impact stuff? Yes, it is something I have to think about. so the way that we're solving it is just an enormous amount of ground testing. So we've done four rounds of testing at the cyclotron down in Knoxville. it's a high velocity proton particle accelerator. and we take all of that telemetry and then that informs our choice on shielding. And then for heavy ions we have to go to the Brookhaven National Lab and we basically run all the chips through through the the space environment. So if over 24 hour period, you can put it through 5 years worth of radiation dose and then we take all that data and we then use that to inform shielding but also software development choices for it.

9:21 Yeah. Yeah. Do you do contiguous compute or are you doing something else? this is almost exclusively I mean actually for the foreseeable future will just be for inference. and the reason the reason it's for inference is number one, inference is going to be like 99% of the compute market very soon anyway. So even if we you know, we wouldn't want to write be running a large training set well, running large training sets will be a very small percentage of the total in 5 to 10 year time of AI workloads.

9:50 But secondly, it's very hard. We would need a yeah, we We need to dock together a large 5 gigawatt kind of structure. The I actually have a video of that here, but I won't waste everyone's time with it. >> >> Unless somebody wants to see a video of a 5 gigawatt data center in space. Do you guys want to see that? All right. >> >> We we made this video cuz we didn't want people to be like, "Oh, you could never train a model in space." So, this is what a 5 gigawatt 4 km by 4 km structure in space would look like.

10:20 So, this would be a Starship launch vehicle with a 40 megawatt is what you can fit per Starship launch vehicle. Which will connect to a central spine. Which is connected to this enormous solar panel. On the back there, we have a 1 km by 4 km radiator. Yeah, that that that's how you would train a large model, but as I say, it'll probably be at least 15 years before we get to anything like that. One more I think we're Okay, 40 seconds.

10:50 Or do we Okay. Oh, sorry. by when do you think the majority of the data center will be in the space? Oh, that's a great question. And I actually wanted to ask all of you. Let's run a poll. So, the the question I want to ask is when do you think it will be cheaper to run compute in space for anybody? It could be for Space X or for us. And the four answers will be the next 5 years it will be cheaper, in 5 to 10 years it will be cheaper, sometime after 10 years, or never. Okay, so who thinks within 5 years it will be cheaper to run compute in space than terrestrially?

11:24 Interesting. Who thinks 5 to 10 years? Who thinks beyond 10 years? And who thinks never? Brave. Okay, that's that's an interesting for me to see. I don't Yeah, I think that's we are out of time, so I will leave it there. Thank you very much for your time. >>