AI data satellites need cheaper launch and much greater scale

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 9:00

   satellite very simple take a chip take a memory basically take a compute module not a memory take memory sorry take a chip and memory and put it inside of a satellite that's quite a bare-bones satellite that's extremely large and generates a huge amount of power but is very simple in that it doesn't have you know great pointing accuracy it doesn't have the ability it doesn't have very many unique capabilities. It's basically just a powered shell. So just a box that has a huge solar panel, a tiny battery, some interconnection, you know, networking devices, some attitude control devices, a small thruster, but basically a very bare bone satellite. That's what an AI data satellite is. Roughly speaking, an AI data satellite needs to be about $5 per watt for it to be competitive with a data center on the ground, because the data center on the ground

2. 9:50

   costs between $10 and $15 a watt and lasts a little bit longer. Okay, so if you want to make it better, it has to be about $5 a watt. And that's all in. So that's building the satellite. That's every single component that's out the door and into space. So roughly speaking, to make a $5 a watt satellite, something like $1 per watt needs to be launched. It needs to be the launch cost. So if we break this down and we look at the numbers and look at, let's say, a 200 kilowatt class data satellite, This satellite should cost about a million dollars in total, of which something like 800,000 will be the cost to build the satellite, and 200,000 should be launch.

3. 10:29

   If 200,000 is a launch, and the satellite weighs something like 1,000 kilograms, which is quite optimistic, actually, for the weight of the satellite, it's actually very optimistic for the weight of the satellite, then launch needs to cost $200 a kilogram. Right now, SpaceX, well, when they were in the market, would sell you launch on Falcon 9 for single-digit thousands of dollars a kilogram, and it would cost them something like $1,000.

4. 10:54

   So you can see how Falcon 9 is not a sufficient vehicle to make- It's like 5X off. Yeah, it's 5X off, exactly. Additionally, if you look at any of the companies that are currently deploying data centers at large. Okay, look at OpenAI, Anthropic, any NeoCloud, SpaceX themselves actually with XAI. They're deploying data centers on the order of sort of one to 10 gigawatts a year. So let's do some more math. If a 200 kilowatt satellite is one ton, and a rocket like Falcon 9 can carry about 10 tons to sun-synchronous orbit, which is where AI data sets need to be in order to be useful, each Falcon 9 launch can carry two megawatts. What that means is that to get to two gigawatts, okay, which is a medium number in comparison to one to 10 gigawatts, it's a thousand launches. So Falcon 9 is 10x short on scale and 5x short on cost. And so what this means is that

5. 11:56

   you need a roughly one order of magnitude improvement in both scale and cost. cost just to get to the baseline where the AI data sat is both better at cost-wise and even modest in scale, right? People talk about how AI data sats are the solution because, you know, solution to needing lots of AI compute because they have infinite scale. And this is true. More or less, they have infinite scale once you have infinite, you know, pseudo-infinite launch capacity. But to get there from Falcon 9, you need 5x cheaper cost and 10x higher scale just to get to the baseline. And then you need 100x scale to get to appreciable capacity and 1000x scale to get to really serious capacity. So a rocket like Falcon 9 is limiting. And it's limiting because it's too expensive and it's too slow to launch. The thing that makes its cost high

6. 12:44

   and the rate of launch low is the fact that the second stage of the rocket is not reusable, which is to say it gets thrown away on every launch. So to launch it twice, you have to build two second stages, which is time consuming. And to launch it twice, you have to build two second stages, which costs money. So the solution to this problem that is known and has been known for decades is to make both the first and second stage of the rocket reusable. So in simple English, rockets that are useful for deploying AI data sets are defined as rockets whose first and second stage are reusable. Right now, there are exactly two companies on earth that are seriously developing a two-stage orbit fully reusable rocket one of them is spacex who we already described has roughly no incentive to sell services and the second is a company called stokespace which i

7. 13:35

   think is an amazing company that is also developing a rocket of similar categorically similar performance although it's a bit smaller which doesn't really matter and so it is for this second stage of the future where we are immensely scared because um if you look at the optimal sort the game theory optimal logic for each participant in the market it's as follows first spacex has a rocket that works let's say they have stars let's say starship works in about a year so let's say spacex has a rocket that works they have no incentive to sell they deploy their own data sats because super intelligence is valuable great if another company makes a rocket and that rocket works and the other people who are also building rockets and by rockets here i'm talking about Useful rockets for data sets, so reusable ones.

8. 14:21

   The other reusable rocket makers are not selling services. Then the game theory optimal choice for them is the same. My rocket works. I have no incentive to sell it. Scale AI data sets because superintelligence is valuable. Then if another company builds rockets that works, their logic is the same. So it's for this future that we're worried. So basically it's never optimal for any of these companies to sell launch once.

9. 14:50

   On an instantaneous basis, this is true. And as soon as multiple companies start selling launch, this dynamic changes.