Space Data Centers Are Dumb.

[00:00:00] Speaker 1: If threatening economic collapse and the annihilation of human civilization weren't enough, AI data centers are now presenting a new problem, resources. A large data center will now put out enough heat for 100,000 homes into an increasingly hot atmosphere and slurp up enough water for a thousand families. Obviously, something is going to have to give if we want to continue to build out this infrastructure that no one actually wants. That's why recently some AI tech bros have suggested moving data centers here to space, but that's a stupid idea. For almost every reason. It's kind of embarrassing, actually. Here's why. [00:00:45] Aria: Now entering the facility. [00:00:48] Speaker 1: The idea of space-side AI data centers seems almost exclusively popularized by tech billionaires who believe that they would be doing physics if they weren't so busy not paying taxes. And if you don't understand simple physics, the idea does sound pretty smart. AI data centers need a lot of cooling. Space is really cold. Boom! AI data centers need a lot of power. There's infinite power from the sun in space. Boom! AI data centers need a lot of space. Space has all of that. Boom! You're welcome, shareholders. But like other very futuristic, but ultimately totally impractical megastructures, AI data centers fall down as an idea if you, you know, know the physics. [00:01:31] Aria: Don't you also fake being a physicist? [00:01:34] Speaker 1: Yes, Aria, but hoarding the equivalent of an entire country's wealth isn't the main reason why people take me seriously. It's the smolder. Ooh, try to resist that. You can't. I won the staring contest. I did. Let's start with the most reasonable reasons for space data centers and work towards the least as they are all intertwined. Some SDC proponents make the case that moving artificial intelligence into space would be a boon for satellites that right now collect a lot of Earth observation data and then beam it down to the planet for processing. If that processing could be done by GPU clusters in space instead, it would make everything a lot faster and more efficient. This seems like a fine, but niche, use case. However, what makes the headlines are not the niche cases, but mega to gigawatt sized AI installations in orbit doing what they do on Earth. [00:02:35] Speaker 3: A little push right here. [00:02:40] Speaker 1: Like make stupid chiropractic videos and destabilize the world economy. And almost everything about this idea is intro to engineering level wrong. There are three ways to transfer heat in this universe. The first is by conduction. Through innumerable energy conserving collisions, two objects in contact will reach an equilibrium temperature with each other as their atoms jiggle against one another. The second way is convection. This is like conduction, except that heat is moving from one place to another via the movement of fluids. Like air swirling around a sweet turk in the oven, or water flowing past you in a pool. That's why a summer swim is so refreshing. Water is a very good conductor, and new water is constantly conducting heat away from your body as you move. You looked especially refreshed after that one time. It was cold, Aria, okay? It's not always, moving on. There's basically nothing in space. [00:03:33] Aria: Citation needed. [00:03:34] Speaker 1: So the dominant mode of heat transfer in the void is the third way: radiation. Everything with a temperature emits radiation as charged particles in matter speed up, slow down, excite, de-excite, and release energy in the form of photons of light as they do so. However, this is the least efficient way to transfer heat energy around. And so the main reason that everyone gives for moving data centers to space, cooling, is either ignorant of thermodynamics or ignoring it. We can put numbers to this, because we already have stuff in space with significant cooling challenges. According to the AI ruining Google search, the very best GPUs put out about a thousand watts of waste heat. That's about 10 times more waste heat than your body produces, which is kind of crazy to think about. Today's data centers can have more than 100,000 of these GPUs, which means 100 million watts of heat that has to go somewhere. Why not space? Well, the biggest and best heat radiating technology that we have in space is on the International Space Station. The radiators are right here in between the solar panels and have a surface area of 42 square meters. According to NASA, this area can reject around 14,000 watts of heat into deep space. Do you see the problem? It's with the number of zeros. To cool just one space data center like we do the ISS would require over 7,000 times the surface area, 300,000 square meters. This is the size of the International Space Station. And this would be the size of a single space data center with the requisite cooling panels. The ISS was the most expensive construction project in human history. And you're telling me that to solve the AI resource problem, we're going to put hundreds or even thousands of these monstrosities into orbit in five to 10 years with current technology? Someone slept through Thermodynamics 101. [00:05:40] Speaker 3: But Kyle, shouldn't we at least move AI data centers to space so that they can have infinite sun energy and not use any of our energy on Earth? Actually, fun voice, no. [00:05:52] Speaker 1: And for the same simple reason. Surface area. Most of you have never been to space, but that doesn't mean that you don't know how heat transfer in space actually works. In fact, I'd reckon that many of you have had first-hand experience with it. Case in point, this is a thermos. This is my Yeti thermos. It's fantastic. You can also get these cool demon core and nuclear waste warning sign stickers at shop.kylehill.net. Anyway, how a thermos like this can keep ice cubes ice for like 24 hours is the fact that they use the vacuum, just like space has. So inside of this Yeti thermos, you have two stainless steel walls. In between those walls is nothing. It has been evacuated of air. Now because of this, there is no physical medium through which you can have conduction or convection. So on the inside, on the inner wall, you have ice water and then you have nothingness. And then you have this outer wall exposed to the environment. Because this interstitial space is effectively the vacuum of space, heat can only transfer through radiation. Meaning that heat transfers very slowly. Meaning that your water stays cold. In fact, I've done some calculations about heating and cooling in space before. And I think you would take the human body, which is mostly water, like over a day, if not many more days to freeze all the way through. If you ejected a body into outer space, which isn't happened yet. Anyway, it's fun engineering examples like this that brings the void a little bit closer to the vacuum that is your mouth. That's how you drink stuff. You create a vacuum in your – whatever. Ah, I forgot to put ice in it. You get it. Yes, there's a giant fusion reactor in the sky. But getting power in space isn't exactly a walk on cake. Or whatever humans say. Solar power is good, but it requires a lot of surface area. And the panels themselves are heavy. And that makes them expensive to get into orbit. Chemical fuel is very energy dense, but it's also very heavy and therefore expensive. You're going to say something about nuclear power now, aren't you? No, Aria. Even the nuclear technology that we currently have in space, radioisotope thermoelectric generators, are relatively weak in the amount of electricity that they can produce. So how are we going to actually make all those short films that get ratioed on Twitter by moving those hot, giant racks to space? Excuse me? I – GPUs, Aria. What did you think I meant? The best comparison is, again, the International Space Station. It has the largest solar panels ever deployed in space. Two 1,000-kilogram retractable blankets of 33,000 solar arrays each. These engineering marvels can produce 200 kilowatts of power in direct sunlight. An NVIDIA H200 GPU draws 700 watts at full slop. You don't have to act like you understand physics to do the simple math here. Just line up the zeros. It would take at least 300 ISS-sized solar arrays to power just one large space data center. And probably many more than that, due to inefficiencies, cooling, redundancies, etc. You're telling me that in the next few years we're going to successfully launch nearly half a million kilograms into space, per data center, safely, efficiently, and economically? [00:09:29] Speaker 3: Really? Really? Really? [00:09:33] Speaker 1: This is not a serious idea for serious people. And honestly, it's exhausting. We are constantly stuck in a loop of some guy with too much money proposing something ridiculous, and we all have to waste our time hearing about it, the news has to waste their time covering it, and scientists have to waste their time backtracking to first principles and interviews about it. It's just—it's just— Deep breaths, big man. [00:09:57] Aria: Hey, why don't you talk about radiation? You love that. [00:10:00] Speaker 1: Oh yeah, good thinking. Radiation in space may not immediately incapacitate astronauts, but if space data centers are to constantly and carefully process terabytes of data in the void, radiation in space and in orbit is going to be a real risk to sensitive electronics and data integrity. An unstoppable solar storm wiping out the majority of humanity's compute in just a day seems like a bad thing. And the shielding against that thing is going to be heavy, and therefore expensive, and therefore hard to get into space. Honestly, space is one of the last places where I'd want the majority of humanity's digital throughput to be. It's hard to get stuff there. It's hard to power stuff there. It's hard to maintain stuff there. It's hard to move stuff there out of the way if debris is heading at it at mock . It's going to be hard for astronomers to see stuff in the night sky if they're filled with all these shareholder-driven monstrosities with these giant panels so many times larger than the International Space Station. Why are we still talking about this? Sorry. I was slop-mopping again. Look, data centers in space is a fun idea, but it's a sci-fi idea, like space elevators or Dyson spheres. If all of the oxygen wasn't continuously being sucked out of the room by money vampires, we could entertain better ideas. For example, some companies are looking to put AI data centers at the bottom of the ocean, where it's really cold, there's excellent conduction and convection, and there's near-infinite water for cooling. But they're being drowned out by all this other stuff that we're forced to hear about. Or, and this is just a thought, we could not cover the entire world in GPUs that we don't want. But what do I know? I don't have big crazy ideas that'll never work. [00:11:52] Aria: You'll see in five to ten years, we'll have them. [00:12:02] Speaker 1: I'll talk about it on a podcast and get a billion dollars. You'll see. You're just not kind of playing the chess that I'm playing. I'm playing 4D chess. You're such a... I'm Sigma. I'm a Sigma. You're not Sigmas. [00:12:16] Aria: Now exiting the facility. [00:12:19] Speaker 1: Thank you so much to the very nerdy staff here at the facility for their direct and substantial support in the creation of this here video. If you want to join the facility, if you want videos early, if you want to enter our private discord through the blast doors, if you want private live streams with me, if you want each and every episode of the Office Hours podcast and to drape on a not evil hoodie, go to patreon.com/KyleHill and join the facility today. And hey, if you support us just enough, you're going to have your name in every single video. There's so many of you, I don't know how I'm going to pass this little remainder of the time. Look, for niche applications, AI is going to be amazing. Like we have niche applications for nuclear technology in space and Antarctica and on the moon and things like that. You could have AI for very niche applications like solving protein folding problems and stuff like that. But when you suggest the most ludicrous extrapolation of the technology and say we should put them all in space, it very clearly indicates to me that you aren't a serious person thinking through these things seriously or at the very least surrounding yourself with people who are serious people thinking about these things seriously. And that should seriously scare you and me and disappoint us. Until next time.