Does It Make Sense To Put Data Centers In Space? Can They Really Cost Less To Operate?

[00:00:00] Speaker 1: Hello, it's Scott Manley here. This week, a small Silicon Valley startup called Lumen Orbit published a white paper entitled, Why We Should Train AI in Orbit. Their pitch is that they want to put data centers in space. In their paper, they do some back-of-the-envelope math, which explains why they think that data centers in space are the way forward for giant compute jobs, such as training AI data sets. Now, they are by no means the first people to suggest putting data centers in space, but they are the ones that have been given a couple of million dollars by investors saying that maybe this could actually turn into something. And as a Silicon Valley tech person myself, I was surprised to see that the people backing this are Y Combinator, a tech accelerator startup that's been responsible for companies such as Reddit, Twitch, DoorDash, Airbnb. And while that is an impressive pedigree, you have to remember this is the world of venture capital. And for every big success that comes out of one of these, you're going to have nine companies that didn't quite make it. And I will be completely unsurprised if Lumen Orbit is one of those nine that fail. But they are all about spaceflight, and I'm really interested in talking about this. So what is the science of putting data centers in space? Why do they think that this makes sense? Well, right now, we've had this big AI boom, where all these different big companies are spinning up clusters to train AI, LLMs, all sorts of products. And one problem that has emerged is that these data centers have so many CPUs and GPUs that are just requiring power, that there are data centers that may have space, but they don't have the power and cooling capacity to handle new hardware. Data centers are being built where the power and cooling requirements can be satisfied. So with power being the major constraint, Lumen Orbit are arguing that if launch costs get low enough, then solar panels in orbit, which are exposed to sun 24 hours a day and are above the atmosphere and above any weather, will produce power more cheaply than solar panels on the ground. And so their baseline in this is saying that they can launch 100 tons to orbit for $5 million, say on something like SpaceX's Starship. And look, that's hard to tell because we don't know what Starship is ultimately going to cost, but look at their numbers here. They're comparing $8.2 million over 10 years versus $167 million. Even if Starship costs $50 million to launch, they're still going to come out of it ahead, right? Right? Well, no, because I don't think all these numbers are correct. I think there's a number of problems that they haven't really addressed, but hey, they're a company with money, they're going to spend time addressing these. I just don't think the final numbers will be anywhere near as good as this. So look, here's their big pitch. This is a five gigawatt data center with a four kilometer wide solar panel, right? That's 16 square kilometers of solar panels. Now, this is not their first step, but their first step is going to be launching a smaller spacecraft next year to demonstrate compute in space. Some of their studies are talking about 40 megawatt systems, which can be launched on a single rocket. That's more likely to be a first step. So now imagine you want to build a 40 megawatt data center on earth. Well, the solar panels are going to be cheaper, but you're probably going to need about five times as much, assuming you pick the place which has pretty good weather. Most of the year, you're still going to have night and you can't get away from clouds completely. And the atmosphere absorbs some of that solar radiation, reducing the power at the surface. So anyway, their value for the price of the solar array for 40 megawatts was like $2 million. So yeah, terrestrial version would be 10, but hey, you got to make that work overnight. So you need battery storage. You'll need 40 megawatts for 12 hours. That's 480 megawatt hours. If you're charging maybe $100 per kilowatt hour, which is way below what, you know, it would normally cost. That's, you know, $50 million. Now, of course on earth, you have other power sources available, but their math says, well, that's about 4 cents per kilowatt hour. That works out to $140 million over 10 years. Now, the thing is, even in low earth orbit, you're still going to have day and night cycles. You'll just get them every 90 minutes. And either you're going to have to have a bunch of batteries onboard your spacecraft to keep it running for the 40 or so minutes on each orbit, which is behind the earth, or you can pick a special orbit, which goes over the poles. And this polar orbit will keep it in sunlight all the time. So you don't really need to have any storage, except in the rare occasion that something like the moon gets in the way of the sun. Now this doesn't fully solve the problem because as the earth moves around the sun, the plane of the orbit is actually going to stay the same with respect to the stars. And eventually you'll get the orbit passing into night. So instead you need to use a sun synchronous orbit where a 97 degree inclination orbit experiences a small pool from the oblateness of the earth. The fact that the earth is fatter around the middle and that twists the orbit to keep it lined up with the sun. I like to call sun synchronous orbits the result of fat earth theory. And sun synchronous orbits are used by many spacecraft these days. In particular, earth observation satellites like that this keeps their observing conditions constant over a long period of time. But this of course leads to another problem in that sun synchronous orbits are very busy. And if you've got a busy part of space, the last thing you want is a giant four kilometer wide solar panel that is going to hit every piece of space debris. And by the way, yes, I did this rendering using Kerbal space program. I used tweak scale to scale up those solar panels as big as I could, but they still looked very small. So I also used scaling to scale down the starship to be very small. Even then that didn't get me all the way to the true grandeur of the lumen orbit dream. But when I did actually try to get something that size, the Kraken kicked in and we got a pretty good simulation of what happens if, say, a catastrophic debris event happens. So look, I think their big problem is that they rely on sun synchronous orbit, they absolutely require it, but they are too big to maneuver. Everyone else would have to maneuver around them. Realistically, with the amount of space hardware there is, that's not going to happen. Now, they could solve this problem by perhaps going to a higher orbit. And also a higher orbit is probably needed because nowhere in their white paper do they talk about the costs and requirements of orbital maintenance. And if you have a very large solar panel, you're going to be getting a lot of drag, even in the tenuous upper atmosphere, you're going to need ion propulsion, some sort of thrusters to keep your spacecraft from falling down. This benefits if you can go to higher and higher altitudes. And so if you go high enough, you're going to get above a lot of that low earth orbit debris that you'll get above all that atmospheric drag, but you're going to have to pay a lot more per pound for the launch because you're having to carry the payloads to higher and higher orbits. So I think they need to recalibrate their costs based on a higher orbit. And of course, the reason why they have almost no competition in these higher orbits is because there's much higher rates of radiation from the Van Allen belts. Now they have actually talked about radiation shielding in their data center modules, and I'm sure they can make these things proof, you know, against the kind of cosmic rays or the radiation from Van Allen belts. But they can't do that for the solar panels. The solar panels are going to be out there and they're going to be getting hit and they're going to suffer higher rates of degradation because they are exposed to this higher radiation flux. They just, they need to be higher. They need to handle that higher radiation. But the most important kind of radiation when you're putting data centers in space is thermal radiation, black body radiation, because you are going to have to keep everything cool. And the only way to get rid of heat in space is through radiators that radiate thermal radiation, right? And if you mention the idea of data centers in space to the average, you know, space nerd, they're going to say right away, how are you going to cool this thing? We have big problems cooling like astronauts in spacesuits. While the average person on the street knows that space can be very, very cold. It's not very good at cooling things because it's a vacuum. That's why vacuum flasks are used to keep things cold or warm. But the truth is, it's not an intractable problem. I mean, this International Space Station keeps itself cool using radiators. And these are actually all over the station in various locations. But the way they work is you have a cooling loop, you have like a cooling liquid, which is piped, pumped through little pipes in the surface, heating these panels up and they will passively reject that heat into space and then colder fluid comes back. And that's how your cooling loop works. Now the rate of cooling via radiation goes as the fourth power of the temperature of those radiators is very temperature dependent. In the lumen white paper, they specifically say that a one square meter panel at 20 Celsius will emit 850 watts of power. And their quick estimate says that they would only need one third the area of radiators compared to the solar panels. But I will point out the video that they show does not include any radiators on display. Ideally, you want the radiators to sit in the shadow behind the solar panels. You want them oriented at 90 degrees to the sun or any other illumination. And guess what? The Earth is actually a source of illumination. It is emitting infrared. It is reflecting sunlight. And so your thermal management doesn't just have to account for the sun but also the thermal power from the Earth. But look, forgetting that the Earth exists is really nothing compared to the fact that they have more or less just said, oh, you know, cooling isn't going to be a big deal. And I'm going to say, no, it is absolutely going to be the biggest of big deals. I'm going to say, no, it's going to be a big deal. And if you have a 40 megawatt electrical system that has to dissipate that heat and say you're pumping ammonia around and you have maybe a 50 degrees Celsius difference between your hot and your cold loop, you're needing to pump 160 kilograms of ammonia through your system every second. Pumping that amount of stuff around is going to have a significant electrical cost that you need to factor into, you know, how much rack space can actually be involved in compute versus supporting the cooling system. There will likely be a need for refrigeration loops and that will take extra power too. Radiators will tend to be heavier than solar panels because you have to actually have a cooling fluid pumped through them. The radiators have to be situated closer to the data center, again, because you have to move the heat around through fluid, whereas it's much easier to move electricity along around through wires. So I think all those things together are going to work to make that 40 megawatt cluster a whole lot lower power. First of all, they won't be able to launch as much mass because they're going to a higher orbit. Secondly, more of their mass is going to be involved in cooling than they expect. Finally, a bunch of the power they generate is going to have to go into running that cooling system. And finally, since they're in a higher orbit getting higher radiation and their panels are degrading faster, they may not get 10 years worth of life. They might have to reduce that down. Do they have enough margins to still make this viable? That's a good question. And there's another thing that kind of works against them and that is the scope of this giant data center, this four kilometer, you know, size thing. Why are they making something this big when, you know, they can launch 40 megawatts in a single launch? Well, this is big data and big data really works best when you can put all the data in one place, right? You want the nodes to be able to communicate between each other very rapidly and very efficiently. Otherwise, they spend a lot of time idling, waiting for data to come across the network. So they really need to pull all this together into one giant satellite death star, data death star, what a cool thing, to actually make it viable. And that's not to say this can't happen. I'm just saying that they need to think about scale if they're going to beat the data centers on the Earth. One thing I do like that they mentioned is the fact that they might be able to move your petabytes of data up on a dedicated physical shuttle rather than sending it, you know, via transmissions. Load all your data onto a spacecraft and send it up there. It's sort of a space version of a bus, a data bus. And I should also say that, you know, all this idea of training AI in space, this is largely predicated on current technology, current AI technology. And maybe something changes. Maybe, you know, we get some more insight. Maybe we develop new lower power ways of training things. I mean, if you look at the last few generations of iPhones, they all now come with a small machine learning accelerator built into the processor. There's millions of people carrying highly efficient chips that are designed to do machine learning. And it's entirely possible that breakthroughs occur, rearchitectures, redesigns that mean that you can do this kind of training without nearly as much CPU processing power. But hey, even if we remove the need for massive AI machine learning data centers, data centers in and of themselves are still kind of versatile, they can be reprogrammed to do other things. And sure, I could totally see somebody doing the math and deciding they want to mine Bitcoin in space. And look, I don't want you to take away from this. I'm super negative on the whole thing because I would really like to see a future with space industry happening. And all these technologies are for solar panels, giant stations, radiators, these are the kind of technologies that any space industry would require the kind of space infrastructure, but data centers in space, uniquely, they have the advantage of not needing to send your materials, physical materials up and down. And by the way, for those of you wondering about the environmental impact of sending rockets to space, carrying solar panels, you know, is it a good thing to do this? Well, if you take the fuel that goes into Starship, right, that's about 1200 tons of methane, and you look at the energy density of it and figure out how much methane you'd have to burn to generate 40 megawatts of power, then for 1200 tons, you would get a couple of weeks worth even if you had perfect efficiency. So you're basically talking a couple of weeks of carbon emissions to get 10 years worth of the same amount of power. So there is absolutely like an objective gain to be had by doing this, if you can solve all those other problems. So would I invest in a company like Lumen Orbit? Probably not, because I think that the timescale for their work is just like too long, they won't be able to use patents. But you know, I have a long history of consistently choosing the wrong tech startups to work for and missing the big buck payout, so you probably shouldn't trust everything I say. I'm Scott Manley, fly safe.