Above the Cloud: Building Data Centers in Space - Richard Campbell - NDC Copenhagen 2026

[00:00:00] Speaker 1: so the talk you saw this morning is a very polished keynote i've done it a few times you're seeing the first draft of this talk because it's one i've wanted to do for a while ever since this bit of nuttery coming up about the idea we should put data centers in space like i got to do the numbers on this and the only way i'm going to get around to doing it is by making an hour on it that's just the way i work uh my name is richard campbell i make a lot of podcasts any uh any dotnet rocks listeners in the room oh that's awesome my friend carl franklin started the the show dotnet rocks back in 2002 which is about three years before the word podcast existed that's why we called it an internet audio talk show because we're not clever enough to come up with the word podcast and we just passed 2 000 episodes uh every week sometimes more than once a week talking to a luminary in the industry just trying to understand what dotnet developers need to know today and that's an ever evolving subject area need to say ai is a big part of it i also make a show called run as radio for sysadmins because i felt like the content available for sysadmins was terrible back in the day of course starting a sysadmins show in 2007 just as vista shipped seemed like a good idea but that one's every wednesday since april 11 2007. and for a brief interval while we weren't sure if dotnet rocked or not uh we made the tablet show where we're mostly talking about mobile and tablet content eventually rolled that back into dotnet rocks and if you are a dotnet rocks listener then you know i make these shows called geek outs and the geek outs are whatever geeky subject the audience wants there's over a hundred of them now and for a while there i was doing one almost every month and uh it's where this whole class of talk comes from because back in 2017 i did a talk about building bases on the moon it was derived from a european space agency paper that i'd read which i thought was pretty cool and so turned into an hour long talk about what's actually practical about building stuff on the moon admittedly this is almost 10 years ago now and then a few months later i get a call from nasa which is a pretty good day right and to be clear it was an email an email i ignored at first because come on but no it was really a nasa.gov email and i finally got charlie on the phone i'm like hey what can i do for you says hey we really like your podcast about building bases on the moon i'm like cool thanks that's awesome please well we were wondering if you'd ever considered doing a talk about building bases on the moon i'm like i guess now i didn't know that you know i spent a lot of time doing talks he thought it was a podcast you see do you know powerpoint we kind of like powerpoint at nasa like yeah charlie i know powerpoint i could probably make that work this is what we'd love if you come down and and talk to us i'm like to you yeah nasa yeah you know it's uh it's your stuff right yeah yeah we know it's like okay so you want me to talk to you about your stuff yeah okay so later that year i went down to the ames research center that's outside of san francisco this was the big board in the front of the of the property as you drive in my mother loves this picture she's very proud and i did this presentation about building bases on the moon to a bunch of moon guys which was an interesting experience and then then we got to tour the facility which was awesome like really a cool place uh and then the other conferences found out that i'd done this talk for nasa like why would you not talk for us and like because it's not about software development like yeah you should do that talk for us like okay so i've done a few space talks over the years and then i because the geek outs were big on alternative energy i've also done a few alternative energy ones as well and so this talk falls into that category it's very very much geek out talk although much more relevant to us today since all of us are using this technology and as a group of folks out there things we ought to put these things into space so i'm telling you all this to say like i've done a lot of work on space related data i knew where to research to say what's it going to take to do this what's really feasible like these don't look like space-based things right these are large concrete buildings with a lot of cooling and a lot of backup power like i you can't fly that into space like that doesn't make any sense at all not that people aren't trying so there's a group a y-combinator startup they're now calling themselves star cloud and late last year they actually launched an h100 gpu into space it was on uh one of the spacex rideshare missions called bandwagon 4 this is in november there was 18 microsatellites on of it and one of it was star cloud one there you see it leaving and it's got an h100 on it so it's actually the most compute ever put into space at once uh it's one gpu they can run they have run a nano gpt on it just to prove the point it's only going to last less than a year about maybe 11 months nanosats have limited functionality so it's cool but it's not particularly practical and like i said star clouds work it is part of y-combinator so they've made a filing to the fcc to fly 88 000 satellites which seems unlikely that's really a lot of satellites they also did these mock-up drawings these animations of giant five gigawatt data centers the data center is the little thing in the middle the rest is all the solar panels you'd need to generate five gigawatts four square kilometers of solar panels and they had the same little animations of this thing coming up little ship coming up and deploying modules of compute that would attach to that data center piece in the middle of it this is infinitely impractical that's a lot of of satellites a lot of solar panels you're going to have to lift them up in segments and assemble them you're probably going to want to put this out in geostationary orbit which is a bit tricky difficult to get to you're probably not going to be able to dodge a lot of micro meteoroids along the way so expect it to take a beating it's this is not how you're going to do this right it's it's cool makes for good visuals but it's not very practical the orbit matters and we talk about different categories of orbit when it comes to satellites geostationary special at 35 786 kilometers up at that altitude your spacecraft is orbiting at the same rate that the planet is rotating so effectively creates this illusion that directly over the equator the satellite is just hovering there and it's far enough away that if you set up three satellites at 120 degrees apart you can cover the whole planet so we generally put satellites up there only for stuff like weather certain categories of one-way broadcast because they're also 600 milliseconds away latency matters and the speed of light tough to beat and so you wouldn't want to do telecom out of there because you got a half a second delay for everything and that gets really annoying very very fast now medium altitude is above 2 000 kilometers below geostationary that's a nice orbit for again very much one-way communications because you're talking you know at least 100 milliseconds or so of latency the satellites you usually find up there are stuff like gps satellites gloss nas those types of satellites in arrays they do fly in a grid around the planet mostly just sending signals down they don't take a lot of signals up low earth orbit is where it's at when it comes to communication because you get the ping time down right down to 20 milliseconds or so that's the exciting space to be in and that's what's getting very crowded because we want those satellites to be in in tight now the downside to leo is atmospheric drag there's not a lot of atmosphere at four or five hundred kilometers there's also not none right it's very tenuous atmosphere and it changes as we get different levels of solar radiation as our sun acts up and emits coronal mass ejections and things like that it actually makes the the atmosphere expand it energizes it makes better aurora borealis that's all energy coming from the sun and that increased energy makes the the atmosphere expand increased drag on satellites alter their orbits so we can't compute low earth orbits a long way in advance we have to constantly monitor them and keep track of them one more aspect about orbital mechanics that's important for us if we think about using these devices is especially in low earth orbits inclination so it's basically the angle relative to the planet of how we're going to orbit this is a a flat visualization for the international space station and you can see it goes fairly far north and fairly far south that's a 51 degree inclination that's where the international space station is and that's to make sure that it flies over russia who wants to be able to command control on their modules if you put it in a zero degree inclination which you can do that's directly equatorial that's where all your geostationary satellites are they need to be directly over the equator if you go 90 degrees well not exactly 90 degrees because of weird things you can do a polar orbit so now you're flying pole to pole that's where spy satellites go they want it they want to be able to map the entire terrain and so they'll be up in high high inclination orbits takes more energy more complicated to get there now the biggest satellite network we've got today is starlink and starlink it started back in 2019 was commercial by 2021 by december of 22 they had a million customers today they have 10 million customers right and this is internet downlink for remote areas they've launched about 12 000 satellites total of which 9400 or operational 9500 depending on how you do the math they basically come in two different generations about 3 000 of the generation ones and about 7 500 of the generation twos and the generation two what they call a v2 mini to still fly on a falcon 9 they're about 800 kilos each and maybe you know roughly four meters by two meters so big thin heavy but not particularly huge it's important to note that 9600 satellites that elon's operating the total number of satellites for the rest of the world combines about 4600 so he operates more satellites than the rest of the world combined you know there's an argument that we've only flown 10 000 satellites before starlink and they've flown 12 000 no signs of stopping anytime soon but starlink's essentially done their goal is actually get to about 12 000 satellites by this year and that'll be considered operationally complete they weren't supposed to finish this before starship flew but starship's been taking its time if you've been following that so we want to build a data center in space we got to start with a computational unit that little single h100 is not going to cut it we can't fly a whole data center that's way too big so let's pick a unit of work that we can do the math on and this is what i spent a lot of time figuring out on so i picked the nvl 72 from nvidia it's about a three million dollar unit it's a 50u rack space uh it's a complete unit with a set of uh grace processors 36 grace cpus 72 black wells and it can do about 72 720 petaflops so they can handle a trillion parameter gpt right gpt4 would run inside this one rack now normally when you talk about an ai data center you've got dozens of these even hundreds of these connected together with infiniband to do your big model builds but again that's too big for us to lift one of these racks should be able to do inference for us so i don't think we're ever going to be able to build models in space but we could move the inference workload up there there's still some problems let's work our way through it a couple other important numbers if we're going to think of this in terms of space yes it's liquid cooled we're going to need to do that because cooling is really a problem in space it's about 135 kilowatts of electricity per rack that's going to make for some interesting numbers and it's 1.3 metric tons right i know that number from their specs because you got you've got to care about that stuff when you put it on a lifted floor all right so we've got our trillion parameter model let's talk about building satellites building satellites is easy here's how the gps2 satellite works i love lockheed martin but let's be clear satellite construction is a very specialized art there's lots of pieces involved and i'm going to walk you through all of this not with this graphic this graphic was designed to scare you and it worked let's start with the bus side okay we've defined the payload the payload is at mvl 72. so now what's all the ingredients that makes a satellite work well you need some kind of structure to hold it together typically made of aluminum as light as you can possibly get it you're going to need command control computers that are going to deal with your power management they're going to deal with the pointing systems how we keep it aimed the right direction do your health monitoring telemetry all of that sort of stuff not big those are little pieces attitude determination and control so i need to figure out which way i'm flying i don't want to be tumbling i need to be able to keep my my solar panels pointed in the right direction and collector like collect electricity my cooling system pointed in the shade to be able to disperse cooling i need to be able to communicate with some things and i got to do that like knowing where i am and so on satellites there's a bunch of different systems the first is their inertial management system so this is a set of spinning gyroscopes and they calculate the orientation of the spacecraft very precisely the problem is that it well it's accurate it's relative so it drifts over time the drag from the from the environment all those sorts of things affect its location and so to correct the imus we also have star trackers and sun trackers so the star trackers are cameras telescopes really that know the particular pattern of stars and so when they see a particular pattern of stars they know roughly from the imu what orientation they're in they can compute that's that set of stars we must be here unless the sun's around then they can't see stars at all and then the sun tracker kicks in but it knows what time of day we are roughly what our orientation is and so we can say oh if i see the sun there that's here those two bits of information combined routinely updated every few minutes depending on where we are in our orbit gets loaded to the imu keeps us in an orientation we understand they also will put gps antennas on these satellites and communicate with the gps network which is just listen to the gps network which would be above us if we're in leo it's in meo so we should be able to precisely target on that as well those combination of things let us know where we are and how we're oriented from that now we want to modify that we want to change our pointing angle we want to keep tracking perhaps do some maneuvering to avoid collisions with things your principal device is a thing called a reaction wheel there's a set of three of them and as you they they have momentum in them so if you they're set at different orientations and if you accelerate one it will force the satellite to rotate in different directions and so you keep varying the spin on that to keep your orientation correct but over time they'll spin faster and faster and faster and they'll become saturated as fast as they're safe to spin and then you actually need to desaturate them allow them to spin down while not messing up the orientation typically this would be done with reaction controls little jets but because we're in low earth orbit we can use magnetorquers which is a great name the magnetorquers essentially an electromagnet that we can fire against the gravity well of the earth to force the vehicle to rotate and desaturate the gyros and so that combination of the gyros and the magnetorquers allow us to maintain orientation against our attitude control system we're also going to have some kind of propulsion system and that propulsion system is going to allow us to change or modify our orbit as we need to do collision avoidance that sort of thing typically they're done with hypergolic fuels there's two compounds that are very poisonous but when you combine them they make a little burst of energy and so you bend them on a nozzle very precise you can do maneuvering that way there's also electrical thrusters what they call hall effect thrusters that's what starlink uses very lightweight very low power it doesn't move the satellite very much but it's very efficient they last they go through fuel slowly with that so those are all the bus elements every satellite has these things especially a low earth satellite so we've got this chassis we're going to be able to fit that big computer into it as our payload and this chassis is going to be able to point the directions you need to point may do our maintenance orientation operations now we've got to feed the payload the first thing we need to feed the payload is power and we need 135 kilowatts to feed the payload the math is pretty straightforward this when you're in low earth orbit and you visit view of the sun you get 1356 watts per square meter of energy from the sun you use triple or quad junction solar panels the expensive kind you're probably going to be able to haul about 300 watts per square meter out of that you need 135 kilowatts so that adds up to about 450 square meters of solar panel or four rays roughly 30 meters long and seven and a half meters wide big space station big though the space station runs on about 140 kilowatts so if you think about the space stations arrays we can use a little bit more modern than the ones that are used on the space station it's not unprecedented they're going to be pretty large and heavy they're going to have to roll out but we've made satellite we've made power plants that big so we're right in the realm that's enough power to run that rig next problem we need to cool the rig how much cooling do we need 135 kilowatts you put 135 kilowatts of energy into the fancy rocks you get 135 kilowatts of heat back out now i know the movies tell us space is cold and nominally that is overall but the reality is you're in a vacuum you're in a vacuum bottle there is no convection here there is no conduction here you have to radiate your heat off and that's hard to do it's a very slow process to radiate heat off so the way you do it is with ammonia you pump ammonia through the blocks to pick up the heat it vaporizes into gas the phase change absorbs a lot of energy and then you pump it out through plates sticking out the back of the spacecraft to radiate that heat into space this is the same again as a space station i'll actually show you the space station system for this that's a pretty big array we're going to need two of those each one of these does about 70 kilowatts of heat dissipation so we need a pair of these to push the heat back out by the way we got what a ton and a half worth of compute maybe a ton or so worth a chassis with the fuel and so forth this is three tons of sa of solar panels six tons of cooling so if you're keeping up we're coming on about 12 tons that's big it's not insane but it's complicated next up we need to communicate so we provide power and cooling to the rack now we need to move data in and out of the rack now traditionally if we were on the ground these things would be connected together with infiniband right infiniband comes in like three flavors these days 400 gigabit 800 gigabit 1.6 terabit we really don't have communication tech that fast the fastest working communication tech we have right now in space actually comes from starlink and it is their uh optical mesh system these are 100 gigabit communication lasers now so it's very doctor no we're putting lasers in space but these actually work and they can move about 100 gigabits of data now they're not like infiniband because infiniband has stunningly low latency like normally you'd be binding these racks together with infiniband moving 400 gigs at sub millisecond speeds we still have to deal with the speed of light in orbit so we're going to talk about five to ten milliseconds at least just between the satellites so we're not going to be able to bind these racks together as if they were on the ground there's no way to do that we're going to have to stick with the inference model and we don't actually have to move that much data up and down the initial load of the model that's going to take some time but 100 gigabits should be enough to communicate back to the ground we're probably not going to do it directly with these satellites we'll most likely be communicating through starlink and starlink will provide the downlink for us which is why spacex isn't a very good place for making all of this happen so set put a satellite zone like you could put more than one you probably will if they don't consume a lot of power they're not terribly heavy right in theory you could gang four of them together and get 400 gigabits but the latency is going to kill you here in terms of actual performance for building models but for the from the inference perspective if we're going to use this with more than one person at a time we're going to need one or two laser uplinks for each of those connecting into starlink down linking to people so you're doing your queries they're going through starlink up through the network relaying to one of these satellites doing the processing for a certain amount of time relaying back down last component for a normal satellite in low earth orbit your orbit periods about 90 minutes 45 minutes of that's going to be in sunlight 45 minutes that's going to be in dark so if you want that to work you need big ass batteries 135 kilowatts worth of it right and i'm budgeting yeah it's only 45 minutes so we could do slightly less i don't think we should we're probably going to need that all we probably won't actually do this because even high density batteries here we're talking about a ton and a half worth of battery but then we also have to put on the charging equipment and fast charging those batteries in 45 minutes needs a lot more cooling so it might only be a ton and a half of batteries but it's probably another couple of tons of cooling equipment there's an alternative use the right orbit we can use an orbit that's always in the sunlight i'll talk about that in a sec so i put the components together there's a couple of mock-ups of what the satellite might look like this is one from spacex the four array approach with a big block in the middle i don't know why they got lightning bolts on it i think ai went nuts on this one that's not how that works uh another one that that uh came from elon from their site directly was what they called the ai sat mini which i love because it's sitting beside a starship v3 which for the measure that's 140 meters so actually i think my math's about right instead of doing four 30 meter panels he's doing two 60 meter panels plus a bit see the white parts in the middle that's the radiator system so i think i'd map this out pretty well it's the right amount of cooling roughly the right amount of power to one a rack in elon specs on these he was talking about 100 kilowatt rig so i'm going a little bit bigger because the mbl 72 is 135 but we're in the ballpark and we skip the batteries by flying in a sun synchronous orbit so sun synchronous orbit is a special kind of polar orbit you actually launch at 98 degrees not 90. at 98 you're a little past the maximum and it's because the gravitational effects of the earth distort your orbit you're going across the bulge in the middle of the planet and so that offset helps you stay in place but you also launch right at the night day terminator so that around three 400 kilometers up you'll be in sunlight all of the time that's cool we don't have to deal with batteries we're always in sunlight makes our cooling problems a little bit more challenging you're going to put that big array out the back in the shade to try and dissipate enough heat you'll always have power but you're limited to essentially one inclination of orbit so we only got so much room here right and if we go higher slows down our communication so we only have a certain range like we're limiting the amount of space also when you're in polar orbits like this you also have other inclination orbits that are crossing so we're going to have to talk about collision challenges here because it's not a simple problem to deal with all this we're also again i think we came up to about a 70 to 150 meter wide satellite coming somewhere between 12 and 15 metric tons now that's not an unprecedented size of satellite there are military spy satellites that are bigger than that there are no commercial satellites that big the biggest commercial satellite is actually a brand new one the jupiter 3 it's about nine and a half metric tons but we're in the ballpark like the math more or less makes sense for this right that we can go if we go into the sun synchronous orbit we can put a certain number of satellites up here with sufficient spacing interfacing with the starlink network and push it down but a 12 ton satellite that's heavy you know those star links are only about 800 kilos each right as falcon 9 can launch 25 to 28 of the star links it's only going to lift one of these data center satellites at 12 12 and a half metric tons if we switch up to the vulcan heavy we can get to five of them uh if we just go over and talk to our friends at blue origin who aren't very happy at the moment blue origin uh the new glenn could only lift three starship v3 could lift eight so to get to that big data center on the ground with several hundred of those racks there's a lot of flights sort of densify that and it does speak to like how many satellites do you want to fly on this how many of these data centers do we need elon talking about millions the star block guy said 88 1000 like that's just a lot of satellites you know we've we've already dealing with an awful lot of satellites these days and so i asked the logical question like how many satellites are too many satellites what is the reasonable limit like i'm we're looking at a map of all of the debris orbiting the earth right now more or less it's a few years out of date so start with starling right they they want to get to 12 000 satellites and they probably will this year which is great it's just understand that that's more satellites than everybody else combined but not for much longer you know spacex told the fcc they could go up to 34 000 satellites but these days they're talking about we're just going to finish this network up to 12 000 and we think we're good like that that's going to be sufficient for now but amazon is working on leo formerly project kuiper and they're all they've already flown 300 satellites they're supposed to fly on new glen which that's going to be difficult like if you've been following the news there was one of the biggest explosions that's ever happened on a pad ever in the new glen test fire and that that rocket does not exist anymore they have another rocket but they don't have another pad so like stage zero is a real thing you need a pad they're going to be a while before they have a pad again but the fcc filing for amazon leo is 3 300 satellites to build in a competing network essentially the starline their satellites are bigger so they don't need quite as many and they're in a slightly higher orbit so they'll be slightly more latent and again that's not the only one then we get to the chinese the chinese are actually working on three different mega constellation sets the main one is called thousand sales it's out of a company in shanghai they've already flown 750 satellites this is one of their launches their goal is to get to global coverage by 2027 so another year and a half 15 000 satellites by 2030. then there's a military network and a and a different form of commercial network that they're working on that both are in the thousands of satellites so if you're doing the math here with me we're at about 35 000 to 40 000 satellites total and collisions are a problem we really don't want them now the thing we care most about when we talk about satellite collisions is the international space stations the largest object in space by a long way it's the size of a football pitch weighs 450 tons and it has 6 to 11 people on it which we'd like to keep alive now if we back up to 20 before starling started back in 2019 space station was already up there there was about 2 000 operational satellites in the world back then 20 000 objects being tracked in total pieces of old spacecraft dead satellites explosion debris from anti-satellite missions that kind of thing and at that time we have ground-based radars that are tracking all of those objects and maintaining a map and they would communicate we would let a satellite operator know typically up to 60 days before a potential what they called a conjunction so your satellite and another object are going to become within less than a kilometer of each other for the space station it's three kilometers and so in those times again this is pre-starling so 2019 they would expect to move the space station which has fairly powerful thrusters once or twice a year for the decrease of chance of debris they would actually get notifications to potential collisions every day maybe a dozen but as soon as they get that notification they would aim radar at the area where they thought the collision might be get more detailed maps of the objects in motion and some of these objects are less than 10 centimeters like they're small and once they knew more precisely how they're moving most of the time it's like okay they're not going to collide it's fine once in a while like many satellite operators would go with no requests to move at all back then once you add starlink to the mix like a spacex lowered the price of flight so the number of satellites has just gone up even without starlink right there were less than 2 000 satellites in 2019 today there's 4 500 plus starlink's 9 000. but starlink satellites all fly in related shell orbits and so they're in constant communication with each other and literally do tens of thousands of maneuvers a year most starlink satellites maneuver almost every day now just to keep that network of 9 500 satellites in their various shells from colliding with each other it's put a lot more pressure on the iss the iss now gets radar checked for debris around it three times a day and while largely nasa stopped talking about how often they maneuver the current estimate is they maneuvering every other month they're avoiding debris so our collision risks have grown dramatically there's more debris and there's more objects these days they figure we're tracking about 50 000 objects so of those 50 000 we've got about the 15 000 active satellites about 1500 defunct satellites around 2 000 rocket bodies which are the real problem they're one to eight metric tons and then everything else is debris 10 centimeters and bigger anything smaller than that we can't track so your individual nuts and bolts and flakes of paint and so forth which are all still hazardous they're basically not trackable but stuff over 10 centimeters a it's trackable and b it impacts hard enough it will destroy things so we track it all so we've gone from 20 000 objects 25 26 years ago to 50 000 objects today there was a research group called mitre that did a study what they called the satellite carrying capacity this is back in 2024 what they really the question was how many satellites in low earth orbit is too many and one of the things they talked about was these coordination systems so the same way that starlink wrote some fairly advanced software to have the satellites avoid each other less direct human intervention if we had all of the satellites doing that all subscribing to the same network and i'm including the chinese satellites which is going to be tricky then we could carry as many as a hundred thousand satellites and lower carbon that was mitre's estimation now that's a lot of satellites but remember we're almost going to get to half of that just with the communication constellations that are on the books right now so where's the room for the ai data centers now elon again was talking about millions well elon is prone to exaggeration and realistically we're talking maybe ten thousand twenty thousand at most before we get close to carrying capacity and that's not mitigating the debris the satellite sensors are not capable of detecting the debris that has to be done from the ground and relayed to the system we can put automation around it to try and make it better but we have to unify all the networks to make that work and the issue here is something famously called kessler syndrome and kessler syndrome is this idea that any given collision generates enough debris to make more collisions inevitable over and over and over again until literally whole orbits are littered with debris remember that every collision imparts enough energy that you're going to put particles in other orbits they're going to spread up and down and so we get yet to a point where it's very hard to deal with all of that debris now for years we've been trying working on debris issues the first and most important are what we call rocket bodies this is actually a one of the very few photographs of a rocket body second stages typically and the point with these second stages is that they should be deorbited after they're used so they're supposed to hold back enough fuel and this is normal today at least for western rockets so a second stage off of a falcon 9 even the new glen rocket they will very deliberately after they deploy their payload turn around and burn to slow down enough to know exactly where they're going to drop that rocket body the thing with these rocket bodies is they're big enough and they're hard enough that stuff makes it to the ground and so it's important to drop them in the ocean and for the most part we do that the notable exception are the chinese the chinese are flying a lot of rockets not anywhere near as much as spacex but enough and they tend not to de-orbit their second stages now you can't always de-orbit a second stage there are certain orbits especially geostationary orbits where it you put the rocket up so high you don't have enough energy to bring it back down and there are still things you can do to make it far less dangerous the main thing is to dump its fuel and depressurize it because an old rocket body with fuel in it still is prone to exploding and one big five ton piece of metal is hazardous but five tons of shrapnel is far more hazardous and so well-behaved rockets depressurize if they can't de-orbit i mean everything is going to come down eventually but once you get much above low earth orbit the the the de-orbit times are in hundreds of years these big options and there's a couple of thousand of them they are big but they're very trackable because they're large and they're somewhat reasonable to manage for satellites same thing as satellites age out before they run out of fuel which is typically the age limit of a satellite satellite burns a certain amount of fuel every year to do station keeping collision avoidance and so forth when that fuel load gets too low they should de-orbit themselves so they use the last of their fuel to this is actually the picture is wrong they're facing the direction of flight this would make your orbit rise not lower they should turn around burn the other way to put themselves down into the atmosphere quickly de-orbit safely there are also passive de-orbit systems that we've been experimenting with like solar sails solar sails are real the so the pressure of the solar of the solar wind can actually affect the satellite and so one of the proposals and these have been experimented with but not put into production is a solar sail that you unfold and then orient the satellite so that the solar pressure slows the satellite down to de-orbit it rapidly there's also tether methods they're all fairly passive techniques for de-orbiting and that presumes that you're doing the responsible thing and mitigating your satellite before it becomes debris so it's not going to do much to the debris that's already up there but it's a way to limit the amount of debris starlink runs in such low orbits that just normal atmospheric drag will de-orbit their satellites within five years which means they now have a constant process now to replace the satellites they're losing we only have five-year lifespan on 12 000 satellites you know divide by five you get an idea of just how many satellites they've got they're going to place a couple of thousand satellites a year just to maintain the network but it decreases the issue of needing to de-orbit them for the inerts or the real debris we have a bunch of experimental proposals nets and harpoons different systems to capture that debris and de-orbit it so for the biggest pieces this like the rocket bodies this somewhat makes sense for the smaller pieces it's more laser based i know it sounds cool but it's actually a fairly practical thing that if you heat an object up enough it will actually decelerate it you heat it in the right direction to cause it to de-orbit more quickly probably going to be a combination of these things when we talk about the 2000 rocket bodies the most likely thing to do here is to actually do mechanical de-orbiting so robot capture and deliberate re-entry it's expensive but it gets those rocket bodies out of play and then for your mid-sized debris the stuff that's you know 100 centimeters to 10 centimeters these netting systems so so forth would be the way to get that and everything small lasers and ion and ion units to slow them down to try and de-orbit them so try to reduce the overall debris decrease the risk to try and maintain these networks and again even before these problems have existed before the ai data centers in space concept we're going to do these multiple communication networks we're going to have to to manage our debris better than we're doing it but we're still trying to figure out if any of this makes any sense i've worked through all the numbers for you i'm not saying it's impossible we can debate the particular costs we know that the things they're proposing mostly are going to be limited by communication they're only going to be able to do inference as long as they're big enough you know what were they actually trying to accomplish here you know there's one argument that the threat to go into space was to push more municipalities to allow more data centers to be built it's not like we're actually going to run out of land but we are having problems with power and to be clear no power plants have been built for ai data centers so far it takes about five years to build a power plant and this craziness is only going on for three the reality is no data centers have been built either other than data centers that were already being built so a lot of this is pure hype right this is taking the slope of the graph pointing directly upward and going we need more of everything doesn't mean that it's real just that it's a good way to convince investors to put more money in and for the most part the only people are really talking about this are the y-combinator types the star clouds and so forth and i've got nothing to lose except for spacex spacex keeps talking about this and funny that we're right at a time when they're getting ready to ipo the company hmm now let me remind you all of a few things when we talk about this ipo and i'm sorry this is so bright first when elon created spacex he said very specifically i am not going to ipo this company because my goal is to put men on to build a city on mars that's not a business and ipoing it would make that impossible so he's clearly thrown that story away because if you're really going to ipo this thing no shareholder is going to put up with the amount of money you need to spend to put people on mars with no possible revenue stream in sight so why is he doing it let's talk about what it means to be a billionaire so where there's only one way to become a billionaire and that is to own a huge interest in a company that gets valued for lots and lots of billions and when you do that you don't sell your stock because that reduce the value of the stock and you have to pay capital gains and all those other fun things so what you do instead is you borrow against your stock you basically commit chunks of your stock at a certain valuation to be sold when it hits that value but you don't want to sell it you just maintain the loan and elon has borrowed billions against tesla it's only publicly traded company and i don't know if you've noticed but lately tesla's not been doing that well now you know his goal with tesla and he said this out front was to make the electric car real and he's done that the problem now is a whole bunch of other companies making electric cars that are pretty darn good and he's not been doing a lot of innovation in that space you notice he's not talking about electric cars to tesla at all he talks about robo taxis and robots because we don't know how to value those so it keeps the stock price up there's been lots of folks talking about how close he is to his strike prices because that would be a catastrophe for him to be forced to sell all that stock to pay off those loans they would tank tesla and it would tank his net worth and we wouldn't want that but if you ipo another company you could move the loans over and unload tesla from that burden huh the trick here is to make sure you make that ipo as big as it can possibly be and here's the problem when you're spacex you already owned the launch market you're the cheapest flight into space you fly everybody that can fly on you already you dominate utterly and your main customer is you it's starlink and you finish that network well how the heck can you get a trillion dollar valuation out of a business you already completely owned you need another business and so in the ipo offering they listed out all of the different businesses their space enabled solutions the starlink broadband i mean they're worth billions they're big companies starlink mobile and then the ai stuff is trillions current proposed values 26 trillion dollars now also in the ipo is a little section where they go there are some concerns about the number of satellites that can be offered in space that might impact the value of this offering but that's several pages down you probably won't read that part and we are weeks away from this ipo so this is the hype cycle run amok essentially to create a valuation large enough to do a multi-trillion dollar ipo all so this guy can protect his money impede the mission propose things that are questionable at best it makes sense from that perspective does it make sense to build these things look we're going to need computers in space but not to use on earth we'll put them around the moon when we need compute around the moon because being a couple of seconds away from the earth doing compute back and forth that's going to suck too right you're going to want compute up there sticking in a low earth orbit is crazy sticking it purely in a sun-circuitous orbit because you have a limited amount of battery space that's even crazier now you're increasing your collision risk you're narrowing the space where those satellites can fly it's even gonna be hard to de-orbit them with that much traffic going up there and they're limited they can only do certain workloads i'm not convinced that they won't fly a few of these because it's interesting business i'm not convinced that they won't try and optimize there's more to be done here we haven't talked about radiation problems or any of that stuff we don't need to we already got enough complexity as it is but does it actually make a substantial difference to the overall network here's a question for you how much of the internet travels over starlink it's 9 000 satellites it's huge each one of those satellites can handle optimal 860 gigabits of traffic so combined if we round them all up together it's about 600 terabits less than one percent one undersea cable one modern undersea cable is 480 gigabits two of them is more than all of starlink combined we have 1700 cables worldwide they just can't move that much data it's it's hard to beat the wire and the wires are short they're fast low earth orbit's pretty quick but it's only so dense we can only haul so much data so it's not practical at large scale the amount you know you know the math i don't have because nobody wants to share the numbers just how much compute have we got in these data centers today what fraction of it could possibly put into orbit i'm betting it's going to come in at less than one percent for that amount expense complexity and risk it's very suspect but i can see a motivation for why we're talking about it right now i just wonder if we're going to be talking about it next year i hope this was useful you thank you so much for coming out to this and i hope you have a great day tomorrow thanks i'll happily take any questions if you guys want to talk further we're just at the party next door so whenever that noise comes we can go from there this is my first attempt at trying to pull all these numbers together i hope we i put some bits together if there's something missing please please let me know anyone have questions sir what oh the actual cost i mean the satellites have never been the expense the launch platform has always been the big expense right and you figure it falcon 9 is about 1800 a kilo so for flying you know 15 ton satellites they're gonna you know we're now we're talking 50 60 million per flight per satellite the actual cost of the vehicle of the of the satellite itself three mil for the for the uh main plant the solar panels are not are barely going to be a million dollars the cooler is going to be more expensive because that cooling system is pretty elaborate but i think all up these satellites are maybe going to be 20 30 million dollars like they're not wildly expensive and still the lift is going to be twice the price of the satellite and that's that tracks right until you get to national space assets like those spy satellites and so forth billion dollar satellites is the only time you see those kinds of things these ones are going to be tens of millions of dollars i think but we'll see we you know i'm trying to anticipate all the problems but i don't know that i have sir a space tether i've done a whole hour on that in the geek outs if you look for it it's pretty science fictiony so we have material that's sufficiently strong for the tensile strength that would be necessary to do a cable from the surface of the earth to geostationary above but we've only built those in sub-millimeter lengths right the carbon nanotubes and you need 60 000 kilometers of that and a whole bunch because we're gonna need some of them to hang out together here's the biggest problem with a space tether just totally off track but if you guys are okay with this i can do geek out all day now i put you in a cart on a 60 000 kilometer long line up to a station up above for you to take off from how fast can i make that cart move and i'll tell you why that's important because going up that cable we're going to go through both van ellen belts at their thickest points with high levels of radiation so let's make us book it like hauling ass 2 000 3 000 kilometers an hour you're still going to spend days going up and the radiation is going to kill you before you get there now how do we do how do we do this with artemis 2 because we just did it with artemis 2 right in april we flew two people through the island belts well we didn't actually what we do is we put them up in a high inclination orbit so they went when they took off for the moon they went through close to the poles where the van ellen belts the smallest we were also moving at 50 000 kilometers an hour and so we cleared those belts in a couple of minutes so the big problem with space tether on the earth is going through the atmosphere right and dealing with all of that but now that we have reusable rockets or almost reusable rockets it almost doesn't matter right if starship works the way it's supposed to we're going to be down in the hundred dollar per kilo space and that's not almost price competitive with the tether the place for a tether would be the far side of the moon so this is another geek out but there's a geek out about building civilization into space like actually become a space fearing civilization and the far side of the moon it's this magical place it's got a lot of resources it's relatively quiet the gravity is low enough that a tether will work a charm you can use kevlar for that tether it's easy but the main thing is once you're out at that caval at the end of that tether you only need a couple of thousand delta v to virtually any point in the solar system it's super efficient to come and go from and if you're going to start adding asteroids which i think you should do the logical place for you to bring the resources to is that same spot a chance of collision with the earth is extremely low clients colliding with the moon is higher we don't care that much about the moon it's going to be okay and it's a great place to build spacecraft from so as we build up the infrastructure actually operate on the moon we start taking advantage of that far side those are the kinds of things we'll put up there and become a space-faring civilization live and work in space not just do science but explore do more more sir five thousand years yes well i mean for the ai we don't have to do that in five thousand years right we're we've got missions to to asteroids now we could be mining in a hundred years you know realistically there's no technological barriers to doing this it's a bit of engineering left to go so it's fairly feasible the math for de-orbiting mass back to earth is bad so in general it makes very little sense to take resource and bring it back to earth manufacturing products in space with those resources makes a heck of a lot more sense right like there's several products we could be building in uh geostationary space or space in in the leo space system between there between the earth and the moon that would actually be better than stuff we can make on the earth starting with solar panels anytime you need to grow a crystal gravity is your enemy and so we need to do more experience and we've already done some of this in the space station with doing more crystal growing up in in low gravity environments same for fiber optic cable we had a bit more room to work in fact all the ingredients for making solar panels are on the moon it'd be a great place to make them they would naturally be superior quality just because it's easier to form crystals up there and that's the real thing that's going to happen over time is we're going to gradually learn about a set of products we can make in these different environments that are superior to the earth environment and that's the stuff we'll be willing to build back down don't send on the raw materials make the superior product that'll take some time but it's possible and we already have some scope on a few of them data centers on titan well we got the cooling part nailed because it's like negative 200 degrees there look we're going to want data centers wherever we have compute requirements right the only place that this makes no sense is leo pretty much everywhere else that we're going to need compute requirements we're probably going to move to move a certain amount of compute i don't see how now now that we're watching generative ai models be such a key part of science you don't think we're going to call a couple of racks up to the base on the moon that makes pretty good sense right just do those workloads locally run them on solar power on the on the small nuclear are we're going to get humans to titan one of these days you're going to need a warm jacket it's negative 200 something it's cold there but the gravity is low and the atmosphere is thick thick enough that with the right pair of wings you should be able to flap and fly uh but it's going to rain ethane on you and i would say don't worry about lighting a match because there's no oxygen to be found anywhere there it's uh it's in the ice you're going to have to dig the ice out to operate all that but we got a couple of missions going to titan now i'm looking forward to those so i may live long enough to see the dragonfly that that'll be exciting yes ownership of space oh you don't want to open a can of worms now do you so we got the outer space treaty of 1967 right back when it was only the russia the soviets and the americans and we kind of made this deal said hey we're going to treat space like antarctica antarctica artica gives us a pretty good example of what is most likely to happen yeah nobody owns antarctica but you don't really go into the american sector without letting the americans know and they get pretty upset if you do and same for the russian sector and the norwegian sector you know that's the normal pattern and what we're seeing in the conversations about stuff happening on the moon is exactly that that you get there put some stuff down you can request and demand an exclusion zone hey don't come around here and the chinese seem to be playing this game fairly well they're working on landers specifically to put payloads down most likely radioactive payloads to create exclusion zones so yeah we're going to have problems about that uh and admittedly that treaty is 80 years old it probably needs some updating uh people are going to find a way to work around it to some degree because treaties and agreements are only as good as their enforcement just ask the gdpr any more uh should i just start doing geek out talks at these shows where i just like geek out whatever you guys want bring your questions and we'll go maybe next time okay thank you thank you so much