What really worries me is that I keep hearing "cooling is cheap and easy in space!" in a lot of these conversations, and it couldn't be farther from the truth. Cooling is _really_ hard and can't use efficient (i.e. advection-based air or water cooling) approaches and are limited to dramatically less efficient radiative cooling. It doesn't matter that space is cold because cooling is damned hard in a vacuum.
The article makes this point, but it's relatively far in and I felt it was worth making again.
With that said, my employer now appears to be in this business, so I guess if there's money there, we can build the satellites. (Note: opinions my own) I just don't see how it makes sense from a practical technical perspective.
Yeah, I don't see a way to get around the fact that space is a fabulous insulator. That's precisely how expensive insulated drink containers work so well.
If it was just about cooling and power availability, you'd think people would be running giant solar+compute barges in international waters, but nobody is doing that. Even the "seasteading" guys from last decade.
These proposals, if serious, are just to avoid planning permission and land ownership difficulties. If unserious, it's simply to get attention. And we're talking about it, aren't we?
You should read the linked article, they talk about it there. You radiate the heat into space which takes less surface area than the solar panels and you can just have them back to back.
In general I don't understand this line of thinking. This would be such a basic problem to miss, so my first instinct would be to just look up what solution other people propose. It is very easy to find this online.
But heat = energy, right? So maybe we don’t really want to radiate it, but redirect it back into the system in a usable way and reduce how much we need to take in? (From the sun etc)
I've always enjoyed thinking about this. Temperature is a characteristic of matter. There is vanishingly little matter in space. Due to that, one could perhaps say that space, in a way of looking at it, has no temperature. This helps give some insight into what you mention of the difficulties in dealing with heat in space - radiative cooling is all you get.
I once read that, while the image we have in our mind of being ejected out of an airlock from a space station in orbit around Earth results in instant ice-cube, the reality is that, due to our distance from the sun, that situation - ignoring the lack of oxygen etc that would kill you - is such that we would in fact die from heat exhaustion: our bodies would be unable to radiate enough heat vs what we would receive from the sun.
In contrast, were one to experience the same unceremonious orbital defenestration around Mars, the distance from the sun is sufficient that we would die from hypothermia (ceteris paribus, of course).
Assuming merely attitude control, sure only radiative cooling is available, but its very easy to design for arbitrary cooling rates assuming any given operating temperature:
Budget the solar panel area as a function of the maximum computational load.
The rest of the satellite must be within the shade of the solar panel, so it basically only sees cold space, so we need a convex body shape, to insure that every surface of the satellite (ignoring the solar panels) is radiatively cooling over its full hemisphere.
So pretend the sun is "below", the solar panels are facing down, then select an extra point above the solar panel base to form a pyramid. The area of the slanted top sides of the pyramid are the cooling surfaces, no matter how close or far above the solar panels we place this apex point, the sides will never see the sun because they are shielded by the solar panel base. Given a target operating temperature, each unit surface area (emissivity 1) will radiate at a specific rate, and we can choose the total cooling rate by making the pyramid arbitrarily long and sharp, thus increasing the cooling area. We can set the satellite temperature to be arbitrarily low.
Forget the armchair "autodidact" computer nerds for a minute
Temperature is a property of systems in thermal equilibrium. One such system is blackbody radiation, basically a gas of photons that is in thermal equilibrium.
The universe is filled with such a bath of radiation, so it makes sense to say the temperature of space is the temperature of this bath. Of course, in galaxies, or even more so near stars, there's additional radiation that is not in thermal equilibrium.
> Temperature is a characteristic of matter. There is vanishingly little matter in space. Due to that, one could perhaps say that space, in a way of looking at it, has no temperature.
A perfect vacuum might have no temperature, but space is not a perfect vacuum, and has a well-defined temperature. More insight would be found in thinking about what temperature precisely means, and the difference between it and heat capacity.
Jusssst had this conversation two nights ago with a smart drunk friend. To his credit when I asked "what's heat?" and he said "molecules moving fast" and I said "how many molecules are there in space to bump against?" He immediately got it. I'm always curious what ideas someone that isn't familiar with a problem space comes up with for solutions, so I canvased him for thoughts -- nothing novel, unfortunately, but if we get another 100 million people thinking about it, who knows what we'll come up with?
I got really annoyed when I first realized that heat and sound (and kinetic energy) are both "molecules moving," because they behave so dramatically differently on a human scale.
And yes, obviously they aren't moving in the same way, but it's still kind of weird to think about.
Author also forgot batteries for the solar shade transition period and then additional solar panels to charge these batteries during the solar "day" period. then insulation for batteries. Then power converters and pumps for radiators and additional radiators to cool the cooling infrastructure.
Overall not a great model. But on the other hand, even an amateur can use this model and imagine that additional parts and costs are missing, so if it's showing a bad outlook even in the favorable/cheating conditions for space DCs, then they are even dumber idea if all costs would be factored in fully. Unfortunately many serious journalists can't even do that mental assumption. :(
Cooling isn't anymore difficult than power generation. For example, on the ISS solar panels generate up to 75 W/m², while the EATCS radiators can dissipate about 150 W/m².
Solar panels have improved more than cooling technology since ISS was deployed, but the two are still on the same order of magnitude.
So just 13.3 million sq. meters of solar panels, and 6.67 million sq. meters of cooling panels for 1 GW.
Or a 3.651 km squared and 2.581 km squared butterfly sattelite.
I don't think your cooling area measures account for the complications introduced by scale.
Heat dissipation isn't going to efficiently work its way across surfaces at that scale passively. Dissipation will scale very sub-linearly, so we need much more area, and there will need to be active fluid exchangers operating at speed spanning kilometers of real estate, to get dissipation/area anywhere back near linear/area again.
Liquid cooling and pumps, unlike solar, are meaningfully talked about in terms of volume. The cascade of volume, mass, complexity and increased power up-scaling flows back to infernal launch volume logistics. Many more ships and launches.
Cooling is going to be orders of magnitude more trouble than power.
How are these ideas getting any respect?
I could see this at lunar poles. Solar panels in permanent sunlight, with compute in direct surface contact or cover, in permanent deep cold shadow. Cooling becomes an afterthought. Passive liquid filled cooling mats, with surface magnifying fins, embedded in icy regolith, angled for passive heat-gradient fluid cycling. Or drill two adjacent holes, for a simple deep cooling loop. Very little support structure. No orbital mechanics or right-of-way maneuvers to negotiate. Scales up with local proximity. A single expansion/upgrade/repair trip can service an entire growing operation at one time, in a comfortable stable g-field.
Let’s say you need 50m^2 solar panels to run it, then just a ton of surface area to dissipate. I’d love to be proven wrong but space data centers just seem like large 2d impact targets.
Space hardware needs to be fundamentally different from surface hardware. I don't mean it in the usual radiation hardenrining etc, but in using computing substrates that run over 1000c and never shut down. T^4 cooling means that you have a hell of a time keeping things cool, but keeping hot things from melting completely is much easier.
I think the point is, yes, cooling is a significant engineering challenge in space; but having easy access to abundant energy (solar) and not needing to navigate difficult politically charged permitting processes makes it worthwhile. It's a big set of trade offs, and to only focus on "cooling being very hard in space" is kind of missing the point of why these companies want to do this.
Compute is severely power-constrained everywhere except China, and space based datacenters is a way to get around that.
Of course you can build these things if you really want to.
But there is no universe in which it's possible to build them economically.
Not even close. The numbers are simply ridiculous.
And that's not even accounting for the fact that getting even one of these things into orbit is an absolutely huge R&D project that will take years - by which time technology and requirements will have moved on.
I've done some reading on how they cool JWST. It's fascinating and was a massive engineering challenge. Some of thos einstruments need to be cooled to near absolute zero, so much so that it uses liquid helium as a coolant in parts.
Now JWST is at near L2 but it is still in sunlight. It's solar-powered. There are a series of radiating layer to keep heat away from sensitive instruments. Then there's the solar panels themselves.
Obviously an orbital data center wouldn't need some extreme cooling but the key takeaway from me is that the solar panels themselves would shield much of the satellite from direct sunlight, by design.
Absent any external heating, there's only heating from computer chips. Any body in space will radiate away heat. You can make some more effective than others by increasing surface area per unit mass (I assume). Someone else mentioned thermoses as evidence of insulation. There's some truth to that but interestingly most of the heat lost from a thermos is from the same IR radiation that would be emitted by a satellite.
The computer chips used for AI generate significantly more heat than the chips on the JWST. The JWST in total weighs 6.5 tons and uses a mere 2kw of power, which is the same as 3 H100 GPUs under load, each of which will weight what, 1kg?
So in terms of power density you're looking at about 3 orders of magnitude difference. Heating and cooling is going to be a significant part of the total weight.
For some decades now I’ve heard the debunk many times more than the bunk. The real urban myth appears to be any appreciable fraction of people believe the myth.
But space isn't actually cold, or at least not space near Earth. It's about 10 C. And that's only about a 10 C less than room temperature, so a human habitable structure in near earth space won't radiate very much heat. But heat radiated is O(Tobject^4 - Tbackground^4), and a computer can operate up to around 90C (I think) so that is actually a very big difference here. Back of the envelope, a data center at 90C will radiate about 10x the heat that a space station at 20C will. With the massive caveat that I don't know what the constant is here, it could actually be easy to keep a datacenter cool even though it is hard to keep a space station cool.
As you intimated, the radiated heat Energy output of an object is described by the Stefan-Boltzmann Law, which is E = [Object Temp ]^4 * [Stefan-Boltzmann Constant]
However, Temp must be in units of an absolute temperature scale, typically Kelvin.
So the relative heat output of a 90C vs 20C objects will be (translating to K):
383^4 / 293^4 = 2.919x
Plugging in the constant (5.67 * 10^-8 W/(m^2*K^4)) The actual values for heat radiation energy output for objects at 90C and 20C objects is 1220 W/m^2 and 417 W/m^2
The incidence of solar flux must also be taken into account, and satellites at LEO and not in the shade will have one side bathing in 1361 W/m^2 of sunlight, which will be absorbed by the satellite with some fractional efficiency -- the article estimates 0.92 -- and that will also need to be dissipated.
The computer's waste heat needs to be shed, for reference[0] a G200 generates up to 700W, but the computer is presumably powered by the incident solar radiation hitting the satellite, so we don't need to add its energy separately, we can just model the satellite as needing to shed 1361 W/m^2 * 0.92 = 1252 W/m^2 for each square meter of its surface facing the sun.
We've already established that objects at 20C and 90C only radiate 1220 W/m^2 and 417 W/m^2, respectively, so to radiate 1252 W per square meter coming in from the sun facing side we'll need 1252/1220 = 1.026 times that area of shaded radiator maintained at a uniform 90C. If we wanted the radiator to run cooler, at 20C, we'd need 2.919x as much as at 90C, or 3.078 square meters of shaded radiator for every square meter of sun facing material.
The temperature that you raise to the fourth power is not Celsius, it's Kelvin. Otherwise things at -200 C would radiate more heat than things at 100 C. Also the temperature of space is ~3 K (cosmic microwave background), not 10 C.
Will these space-based data centers run on rad-hard silicon (which is dog slow compared to anything on Earth) or just silently accept wrong results, hardware lockups and permanent failure due to the harsh space environment? Will they cool that hardware with special über-expensive high-temperature Peltiers that heat the radiators up to visible incandescence so that the heat can be shed with any efficiency? There's zillions of those issues. The whole idea is just bonkers.
> For ML accelerators to be effective in space, they must withstand the environment of low-Earth orbit. We tested Trillium, Google’s v6e Cloud TPU, in a 67MeV proton beam to test for impact from total ionizing dose (TID) and single event effects (SEEs).
>
> The results were promising. While the High Bandwidth Memory (HBM) subsystems were the most sensitive component, they only began showing irregularities after a cumulative dose of 2 krad(Si) — nearly three times the expected (shielded) five year mission dose of 750 rad(Si). No hard failures were attributable to TID up to the maximum tested dose of 15 krad(Si) on a single chip, indicating that Trillium TPUs are surprisingly radiation-hard for space applications.
At Satellogic, we famously flew mostly just regular cellphone parts on orbit. We did have higher rates of various kinds of failures than is usual on Earth, but hardware failure can generally be masked by software redundancy.
It's very important in this case to specify which orbit the satellite is going to be in. If you're in LEO like the international space station you spend all day inside the Van Allen Belt protected from all those charged particles that the sun is pumping out. You're still lacking the atmosphere's protection from cosmic rays but that's not a huge dosage.
If you go out to MEO then suddenly you're outside that protective magnetic shield and you have to deal with charged particles smashing into you and you want a large mass of water or wax shielding if you don't have radiation tolerant electronics.
SSO, a low earth orbit whose plane is perpendicular to the direction of the sun so it gets constant sunlight, is harsher than normal LEO orbits because it passes over the poles where the protection from the Earth's magnetic field is weakest, but it's still a lot better than higher orbits. This is probably where you want a datacenter to get constant sunlight and as much protection as possible.
Starlink is however operating at ~500km where radiation is less of a concern, but where the lifetime of a satellite is only 2-3 years.
The unit economics of orbital GPUs suggest that we'll need to run them for much longer than that. This is actually one of the few good points of orbital data centers, normally older hardware is cycled out because it's not economic to run anymore due to power efficiency improvements, but if your power is "free" and you've already got sufficient solar power onboard for the compute, you can just keep running old compute as long as you can keep the satellite up there.
My understanding is non rad hardened method get around this by basically doubling or some multiple of repeating calculations and chexking data often.
Random errors will occur you just need to be checking fast enough to fix and update that bad bit flip.
I am sure there's all sorts of fun algorithms in this space but I am under the impression there is SOME tax to doing this. What is the tax? Is it 10% ir 60% I have no idea would love to know!
Your other options of fault tolerance typically achieved by doing everything at least twice and being willing to reboot (and accepting attrition from total ionizing radiation) or lots of shielding are fine for building functioning space hardware but suboptimal for building datacentre business models...
Say what you will about the data centers in space idea (I think it's transparently stupid), but ML is generally resistant to random undirected noise. It's almost a requirement by definition that a machine which takes pictures and accurately outputs the probability that they are pelicans has to be pretty robust to nigh-infinite amounts of minor variation. That's part of the reason all the super low precision stuff works. It's only in the control logic or maybe the absolute precise chokepoints of computation where flips are dangerous, so most of them are harmless.
Orbital data centers are impractical for a lot of reasons (to put it mildly) but radiation shielding isn’t one of them. Proportionally less shielding is needed as one scales up, due to lower surface/volume ratios.
There are ways in which shielding in space can do harm: really energetic particles get trapped and produce a shower of daughter particles and rays over a greater area. So you'd need even more shielding. Or you accept that such things will happen and use rad-hard parts, redundancy etc. When you have the whole atmosphere above, it's much less of a concern.
Besides, that's even more mass to be lofted. Pushing the economics further into the ludicrous end.
I think “won’t”. I could be wrong of course, but I imagine efforts to put servers into orbit will die before anything is launched. It’s just a bad idea. Maybe a few grifters will make bank taking suckers’ money before it becomes common knowledge that this is stupid, but I will be genuinely surprised if real servers with GPUs are launched.
I don’t mean to be facetious here. But saying “will” is treating it as inevitable that this will happen, which is how the grifters win.
> This is all to say that the current discourse is increasingly bothering me due to the lack of rigor; people are using back-of-the-envelope math, doing a terrible job of it, and only confirming whatever conclusion they already want. Calculating radiation and the cost of goods is not difficult. Run the numbers.
> References: Gemini, Gemini, ChatGPT, ChatGPT, Gemini, ChatGPT, Gemini, ChatGPT, Grok, Gemini (There are sub-references from these services in the GitHub.)
I think, if you're going to make statements like this - especially from a position of expertise, you should be personally verifying the numbers and citing their sources directly. What good is asking the reader to trust an AI on your behalf? They should trust you.
(To be clear, I suspect the conclusions drawn are still correct.)
>This isn't about talent. It's about integration... Vertical integration isn't a nice-to-have. It's the whole ballgame.
I'm going to assume there's tons of logical errors and oversights in the math, considering the author couldn't even be bothered to write the text of the post himself.
The short of it is that cooling is likely the biggest problem, given you will need to pump the heat to the backside and radiate it away, and the amount of mass you will need to dedicate to cooling works against deployments and increases the cost per unit significantly. Not to mention, the idea of these huge deployments runs into potential space debris issues.
Whenever one of these ventures actually manages to launch a proof of concept, I think we'll be able to quickly discern if there is actually a near-future here.
On the ground you don’t have to fling tons of sensitive electronics to orbital speeds, you can have someone pop by to check whether everything is ok if you get an alert, you have a thick atmosphere cutting down solar radiation, you can cool everything relatively cheaply, and you can run fiber and power cables directly into the building. In space ou have to design super robust and self maintaining machinery. It’s a cool marketing stunt, but I don’t get the economics behind doing it for real.
This is an interesting analysis, and I like the sliders that let you instantly show the impacts of system trades.
The one glaring hole that I see is the challenge of moving the data to/from the datacenter while it's on orbit. Bandwidth to/from space isn't free. FCC/ITU licenses are required, transmitters/receiviers/modems/DSP/antennas all add to SWAP (size, weight, and power). Ground-stations are needed to move the data up/down, but those have recently become a commodity too. Still, they're not free. (see: https://aws.amazon.com/ground-station)
There is also the added latency between earth-based users and space-based datacenters, which may be a deal breaker for some applications.
Another issue I don't see covered are the significant differences between space-based hardware and terrestrial hardware. The space stuff needs to be radiation tolerant, and that usually makes it a lot slower and a lot more expensive than the terrestrial stuff, all other things being equal.
In the end, space-based datacenters are highly impractical even if you assume that Starship can put anything into orbit very cheaply.
The most chased workload will inevitably be cryptographic research, proofs of mathematical statements are hard to find the proof for, but tend to be short and easy to verify once a putative proof is presented. Just send the proofs back to earth.
Here's some math on how affordable that abundant LEO solar energy is:
First you have to pay energy to get to LEO
A Starship Launch costs[0] 51.75 TJ of energy in terms of its methane fuel.
It will be able to take a payload of 150 tonnes or 331,000 pounds[1].
How many computers is that?
One online estimate says a computer weights 80 lbs or 35 kg.
So 150000 kg / 35 kg/computer = approximately 4285 computers that we can launch into orbit per Starship.
51.75TJ / 4285 computers = approximately 12.08 GJ per computer to place it in orbit.
Let's say each computer is a H200 and consumes 700 watts continuously. How long would it need to run in orbit before it used as much energy for computation as it took to launch it?
12.08 GJ / 700 W = 12,080,000,000 J / 700 J/s = approximately 17,257,143 seconds.
Or about 6.5 months to break even on energy.
That sounds pretty good, except my estimate for the weight of each compute unit and associated power system & cooling etc. are probably underestimates by one or two orders of magnitude. In which case you'd be looking at 5 to 50 years to break even on energy, by which time the chips are obsolete and need to be replaced anyway.
You are just launching computers, with no propulsion, no attitude control, no solar panels, no radio/laser systems, no radiators. So all of that will take mass away from the computing power. A starlink satellite already weighs about 1000kg, and that really is just the supporting infrastructure you need before you start adding computers...
Does anybody actually work with H100s and the like? Their failure rate is so high, I dont understand why anybody will even consider it feasible to put the machines in orbit or even the sea. By my ballpark estimate, if you have 800 H100s, after 6 months, about 100 would be overheating or throttling, and a few will disappear and one or two will crash the machine with load.
> Does anybody actually work with H100s and the like?
They don't. The expectation the cloud develops in people is that magic computers just appear. They're living at a virtualized layer where all the nitty gritty of real machines going down and needing to be serviced all the time is handled by unseen minions (sorry SREs and DC staff) and cluster management and provisioning software.
The reality is that datacenters in space is mind-boggling stupid, just from the infeasibility of maintenance alone.
The analyses here miss the economic realities of building datacenters. "Just use land", "just use nuclear", "just use water". All of this is contested. A system of lawsuits and regulations turns negative externalities (even ones you aren't convinced of!) into costs you can weigh against. So, like hydrogen vs. RP-1, it's not enough to pick a handful of physical performance metrics. It has to win holistically.
If you can produce any kind of economically productive compute node and add it to (for example) the Starlink network, and launch on a reusable vehicle, you carry on installing them as fast as you can build them.
So, the move is to turn the problem of contested land use into a manufacturing problem.
This is not so easy to pin down on a spreadsheet, and will be decided at the level of the business unit. If SpaceX can put a GPU/TPU on the grid more economically than the other guy, then it doesn't matter if they have ammonia in the pipes instead of water.
> That "why" is almost missing from the public conversation. People jump straight to hardware and hand-wave the business case, as if the economics are self-evident. They aren't.
But then he never answers that fundamental question, and jumps straight to the hardware and power and cost? What problems are orbital data centers trying to solve? What optimizations are they intended to deliver? Are these optimizations beneficial to everyone who uses a data centers, or just operators or users of orbiting satellite constellations?
> But the knock-on effects are why this keeps pulling at people. If you can industrialize power and operations in orbit at meaningful scale, you're not just running GPUs. You're building a new kind of infrastructure that makes it easier for humans to keep spreading out. Compute is just one of the first excuses to pay for the scaffolding.
This seems to be the closest we get to a “Why”, but it doesn’t make much sense. A constellation of 40,000 satellites with GPUs “infrastructure that makes it easier for humans to keep spreading out”? How?
> The target I care about is simple: can you make space-based, commodity compute cost-competitive with the cheapest terrestrial alternative? That's the whole claim. … Can you deliver useful watts and reject the waste heat at a price that beats a boring Crusoe-style tilt-wall datacenter tied into a 200–500 MW substation?
Isn’t the answer clearly “No”? The default settings of his model — which I assume he considers optimal — tell us that power for orbital data enters will cost 3.5X terrestrial ones. And that only SpaceX has the vertical integration to do even attempt to do this. So again, where is the competitive advantage?
Also, I don’t understand why he’s including satellite construction and launch costs for a 40,000 satellite constellations in his analysis, if he’s assuming SpaceX as he claims. Wouldn’t SpaceX simply implement these compute capabilities in the next gen of Starlink, so which would reduce costs significantly.
> It might not be rational. But it might be physically possible.
But isn’t that precisely what everyone has been saying? I don’t think the question has been whether orbital data centers are possible, it’s been whether they are rational. And that centers foremost h the unanswered question, Why is this a good idea?
> But then he never answers that fundamental question
The fundamental question is “is it economically viable”, and the answer from his model is “not really”
> A constellation of 40,000 satellites with GPUs “infrastructure that makes it easier for humans to keep spreading out”?
I think he’s claiming industrializing larger and more economical power generation in space, as well as the means to put it up there, would make it easier to transition to a theoretical space economy
> But isn’t that precisely what everyone has been saying?
From the article, he claims that people handwave the economics, so at least the people he has interacted with haven’t been saying that.
I love the sliders, but note that the numbers on this site literally came from ChatGPT, so there is plenty of room for disagreement.
Seems like according to this analysis it all hinges on launch cost and satellite cost. This site's default for Starship launch cost is $500/kg, but SpaceX is targeting much lower than that, more like $100/kg and eventually optimistically $10/kg (the slider doesn't even go that low). At $100/kg (and assuming all the other assumptions made on the site hold) then you break even on cost vs. terrestrial if you can make the satellites for $7/watt (excluding GPUs, as the whole analysis does).
OTOH SpaceX has a pretty good history of undercutting the industry on cost. If Starship full reusability works I would be very surprised if it only lowered launch costs by a factor of three. Of course it's not guaranteed to work, but clearly SpaceX's orbital datacenter plans are predicated on Starship working.
I am struggling with a why for this (other than “huh cool, that will get investors”). All the jurisdiction and regulation arguments and the “we could get the costs down” seem to meet the objection of “for the same investment we could do just as well or better on the ground”.
The one that does not is the physics of the whole thing. I struggle to work out how exactly but being slightly time dilated compared to the ground does not seem like a win, but being able to gather data from opposite sides of the planet slightly faster than cables does seem like a potential win. Most stock exchanges make a significant chunk of their revenues renting out data space, so it seems a possibility.
I’m not one for conspiracy theories, but since SpaceX is the only launch services provider that could actually put one of these in orbit, this smells a lot like hyperloop to me — an unserious proposal that serves as a distraction and furthers Musk’s aims, and benefits anyone who can get close enough to the piles of cash that VCs will drop on this.
You know what’s easier and cheaper than putting a data center in space? Putting one literally anywhere else other than space.
"Datacenters in space" make for a catchy narrative and an interesting demo, but the math simply doesn't work.
When considering factors like launch cost, maintenance complexity, and the cost of high-bandwidth communications (latency included), there is no realistic set of economic and engineering assumptions under which orbiting datacenters become cost-competitive with simply building conventional nuclear-powered (or renewable energy-powered) datacenters on the ground.
In fact we're off by 50-100x. Dramatic launch cost reductions still won't make it work. And of course if you invest a lot in specific lines of tech to make it work you then have to consider that the same can also be invested in better ground-based nuclear, bringing the cost of power down for everyone.
Did a similar back-of-the-napkin and got 5x $ / MW of orbital vs. terrestrial. This article's analysis is ~3.4x.
I do wonder, at what factor of orbital to terrestrial cost factor it becomes worthwhile.
The greater the terrestrial lead time, red tape, permitting, regulations on Earth, the higher the orbital-to-terrestrial factor that's acceptable.
A lights-out automated production line pumping out GPU satellites into a daily Starship launch feels "cleaner" from an end-to-end automation perspective vs years long land acquisition, planning and environment approvals, construction.
More expensive, for sure, but feels way more copy-paste the factory, "linearly scalable" than physical construction.
It becomes worthwhile if its actually cheaper (probably significantly cheaper given R&D and risk), or if you're processing data which originates in space and the data transfer or latency is an issue
You can set up plant manufacturing chips in shipping containers and sending them to wherever energy/land is cheapest and regulation most suitable, without having to seek the FCCs approval to get launch approved and your data back...
Aside from the economics, the question is why do it in orbit vs on land (or sea)?
What are the regulatory/legal gains? Lack of jurisdiction means open slather?
What are the national security gains? Redundancy and resiliency by each satellite being a "micro-compute" connected by high speed laser links? So more resilient to attack?
I think the main draw is its elegance. You have very efficient power from the sun, put that directly into your compute, radiate it out. Energy is ~free, no heavy infrastructure required, just a closed circuit for computing.
Elegance compared to a PV/Storage facility built next door to a data centre?
It doesn't make sense right now, and won't for at least 5-10 years.
By which time, this current round of hype will have burned up ~$1T if it doesn't fall apart from the current internal contradictions and lack of market/customers/uses.
We're still on the uphill ride of the Gartner hype cycle, not even at the "Peak of Inflated Expectations" yet.
Very doable with Photonic compute in orbit:
Way less waste heat (huge when you can only radiate it away), lower power draw per FLOP (smaller solar arrays), and photons don't give a damn about cosmic radiation causing bit flips.
Economics for who? The builder of the data center and plethora of all contractors and sub-contractors would see great economics though. Even the sponsor/owner of the data center might see economics work out, if you consider the reputational gain (why did we land on Moon? what's the economics?), experience gained and considering the burn of someone else's money. The money of mega companies that go into this kind of "monuments" is not exactly theirs.
When Starcloud put together that whitepaper the first thing I looked at was the launch costs[1]. It references a $5M cost to launch, which right away made absolutely no sense to me. Just a cursory search shows launch costs are around $50M per launch, if not more.
It's great that this site drills down even further to demonstrate that there is absolutely no point at which the launch costs ever make this economical or viable, so I really don't understand what people are doing.
Especially because this site was harping for years about the cost of launches and putting things in to orbit, the whole reason why SpaceX got started and has grown as it has. As soon as that became an inconvenient number, we now just make things up (Just pretend that launch costs are 10% of what they actually are to get people to invest?).
Spinlaunch is also promising drastically reduced cost per launch. The payload size for their first launcher is pretty small and they appear to be struggling to get the kinetic launcher online.
We know the upper bound for most of those numbers. SpaceX already achieves internal marginal launch costs of ~$1000/kg, for instance. We know their rough costs per satellite. In contrast, we know little to nothing about the inputs to the Drake equation.
The numbers don't quite work out in favor of orbital datacenters at the current values. But we can tell from analyses like this what has to change to get there.
The cooling and density challenges of datacenters here on earth are not trivial , but in space this is multiple magnitudes more difficult.
These numbers are just random bullsh*t numbers.
And what problems do orbital datacenters solve?
They still need uplink, so not libertarian we can do what we want, you have no jurisdiction here thing.
This is just a sci-fi idea that is theoretically possible and is riding the ai bubble for users and investors that don’t know better.
I’d love to see these variables fitted to learning curves. That would give you a forecast for when, if learning continues as predicted, the economics could be competitive. (If it doesn’t, you need a new paradigm first.)
https://en.wikipedia.org/wiki/Project_Natick shut down in 2024 (though apparently it hadn't been in the water since 2020.) It seems like it basically worked, but it wasn't clear that the cooling advantage was all that big relative to the hassle of having them in a difficult-to-maintain environment.
The FCC regulates satellites launched from or communicating with the US, including stuff which extends beyond spectrum licensing like mandatory 5 year deorbiting capability for newly launched LEO satellites. Europe, China and India are not regulation-free utopias either.
You've actually got more option to jurisdiction-shop with underwater data, but I'm not convinced that's the major issue with building datacentres anyway.
Ultimately there are latency and minimise data-transfer arguments for doing certain types of data processing on local machines in space, but the generalised compute and model-training argument only works if the unit economics stack up as sufficiently good to cover the risk and R&D, and they're not obviously favourable compared with cold place on earth with clear skies and access to cold water even assuming launch costs become minimal. (It's slightly amusing to see how much some advocates of that other controversial futurist vision of spaced-based solar power - whose chances of success equally depend on low launch costs - viscerally hate the latest wave of datacentres-in-space hype...)
Yeah try tell average eco joe you are planning to warm up oceans by 0.00000001% of what sun does already.
(I agree right now it probably makes sense, but decades and centuries away we probably don't want to warm up earth anymore. If anything space datacenters could provide shade for earth lol.)
AFAIK compute heavy datacenter in space don't work. But if you already have a vast fleet of laser connected LEO satellites throwing some efficient SSDs into them can make a lot of sense. A large portion of the traffic is fairly static, e.g. video based content or even model weights. Caching that will save you ground to space side of the transmission. This will let you put more user beams on satellites and use less ground stations.
This is very well done! I love including all the sliders so that we can play with the (reasonable looking) assumptions the author has made. Like the author, I share their surprise that the result did not come out even more in favor of terrestrial GPUs.
For someone who is "increasingly bothered by the lack of rigor in the current discourse", the author sure has no qualms about using LLM outputs as primary sources. This is an immediate red flag that discredits the entire article.
This is AI slop in a pretty dress. It's fascinating that space-based datacenters are such a catastrophically bad investment that even limp apologia like this can, at best, argue that maybe it's not quite as bad as you think, as long as you still manage to ignore half the costs because you're a loser who delegated your thinking to a chatbot.
The only benefit as I perceive it re: orbital data center hardware is regulatory avoidance. Think...DDOS machines that can't be shut off; or financial hosting services for unsavory individuals. However, it's very expensive by all metrics (including those talked about in the article), and frankly, these satellites are sitting ducks for the hunter killer satellites the various space powers have, if they actually wanted to do something about these hypothtical data centers and the problems they would cause.
I guess I'd assume that the premise driving this would be that there will eventually be enough business in space that it's necessary for space-centric use, and that terrestrial use is just a fringe benefit or loss leader or something.
But oddly this doesn't seem to be how the concept is typically framed.
My second level curiosity is how much cheaper/competitive it'd be if we had space elevators.
space-elevators require various types of unobtanium and have their own logistics challenges not to mention failure modes that involve spattering fast moving debris round the entire equator
I sometimes wonder if there are people out there who just read too much Neuromancer, and they think they can construct their own Tessier-Ashpool orbital dynasty.
I suppose there are several other Oligarchs In Space stories and movies since then, but if the point of the space station is to host AI, that narrows it down a bit.
Or perhaps it's performative, designed to spook gullible politicians into changing laws to "keep" businesses that were never actually going to go somewhere else anyway.
- High risk of losing the entire equipment to a rocket failure (not infrequent even for modern launch vehicles)
- Supplying enough electrical power would be extremely difficult
- Cooling would be extremely difficult
- Geosynchronous orbits have at least 200ms of communication latency
- Lower orbits means the data center would not stay in place and require complicated tracking antennae and/or a communication mesh a la Starlink, again increasing latency and complexity
One of the main reasons for putting "compute" and "storage" in space is that it is out of reach for the general public and would allow for more stability in exceptionally intense tyranny.
It is much easier to blow things up on land than in space, and the 'negative externalities' simpler to make assumptions about.
The value of this to the people who would be in charge of this "compute" and "storage" is likely much larger than the difference in energy cost.
But if cost of the space GPUs is higher then the land GPUs, why would demand matter? Is there limited land? Are space GPUs better for some reason, like perhaps they can't be regulated as easily or because our AGI overlords will be less vulnerable to mobs with pitchforks?
I realize terrestrial data centers have environmental risks, but are the risks greater for an orbital data center? I would think space debris, solar flares, or a bad actor satellite with a laser could do a lot of damage. Good luck repairing the orbital data center.
the cost of having very fast up/down syncrounous data/compute,comunications to any random spot on the planet, at any time, in a very hard to interupt or detect manner, is, what it is
I'm not really interested in the problems that can come with orbital compute. We've seen them listed ad nauseam.
Have we seen any benefits to orbital computing by launching a cluster of raspberry pis to LEO? Surely this isn't an impossible task to test out on a smaller scale?
There isn’t really much benefit to having compute on orbit unless you’re working on VERY specific applications that have such tight latency requirements where you need to process the data immediately as it comes out of the sensor. In which case you just implement the algorithms in ASICs or FPGAs anyways.
There have been NVIDIA Jetsons or better on orbit since at least 2021 and that had no meaningful impact on any actual meaningful compute workloads beyond proof of concept demos.
> ...we should be actively goading more billionaires into spending on irrational, high-variance projects that might actually advance civilization. I feel genuine secondhand embarrassment watching people torch their fortunes on yachts and status cosplay. No one cares about your Loro Piana.
I 100% agree with this. There are ~2,600 billionaires in the world and we should encourage all of them to spend their money. Even buying a superyacht is a benefit to the economy. But the best billionaires, like Bill Gates and Elon Musk, are actually trying to advance the tech tree.
We are honestly lucky that Musk is wired funny. Any normal human being would retire and hang out on the beach with supermodels after all the abuse he has taken. But he takes it all as a personal challenge and doubles down. That is both his worst quality and his best.
I will agree that if he hadn't aligned himself with Trump/MAGA he would have a much lower profile.
But I think we'd be better off if taking a political position did not automatically piss off half the country. I think a lot of competent but normal people refuse the get involved in politics because of how toxic it is.
I wish Musk had stayed out of politics, but I'm glad he hasn't given up on Tesla/SpaceX just because of the enemies he's made. I think any normal person would have.
You should read F.A Hayek's essay on The Paradox of Savings [1]. Creation of capital like factories, education creating new specialists, or new processes lowers the cost of production and leads to real economic progress. Excessive spending without capital creation does nothing except keeping factors of production wastefully employed when they could be put to other uses and always ends in inflation.
>”we should encourage all of them to spend their money. Even buying a superyacht is a benefit to the economy.”
You’re falling victim to the ‘broken windows fallacy’ here; money which is invested is actually more productive in improving medium and long term economic productivity than ‘consumption’ goods. Even ‘retained’ money (under one’s mattress) is not net-negative, as it increases the value of its circulating counterparts.
Scenario A: Someone breaks a window and the homeowner buys a replacement.
Scenario B: A homeowner adds a new window to their home.
Scenario C: A homeowner buys an online-course to learn how to make windows and then adds one to their home.
Scenario A has approximately no benefit to the economy. The homeowner is no better off (same number of windows) but had to spend money. The window maker might be better off, but only to the same extent that the homeowner is worse off.
I totally agree that Scenario A is not a benefit to the economy. That's the "broken window fallacy".
But Scenario B is definitely better for the economy. The homeowner has decided that having a new window is better than having the money. So the homeowner is better off. The window maker is also better off because they get the money. This is what happens when a billionaire buys a yacht.
Scenario C is the best. The homeowner has a new skill, which they can use to add more windows to their house or maybe their neighbors' houses. Over time, the amount of money spent on window-making will decrease, but the number of windows will stay constant or increase. That's a net benefit. And the online-course creator still made money.
This is what Musk is doing. He is developing new technologies that enable new capabilities and/or make existing things cheaper (e.g., electric cars, access to space, rural internet connectivity).
There is also Scenario D: The homeowner doesn't buy a new window but just keeps his money under his mattress. This is clearly the worst for the economy. Hording money like that means that there is less money circulating and lowered economic activity. The window maker is worse off, and even the homeowner is worse off if they would like to have a new window.
Billionaires who don't spend their money are the real danger, not the ones who tweet too much.
Investing their money is slightly better in that it makes the price of borrowing cheaper. But that only helps up to a point. Someone has to spend money or else there's no point in being able to borrow some. So I wish more billionaires were following Scenario C.
Your description of Elmo applies to several other billionaires who have somehow avoided quixotic hyper-destructive rampages through American politics. I’m all for wired funny, but not when it comes with this much carnage.
I'm not sure who the whole 'Elmo' thing is for -- more than anything I cringe at the person saying it, rather than thinking less of Elon or whatever the hope is. Like 'drumpf', or the whole small hands thing, it just comes across like a redditism that escaped confinement.
Is the hope that Elon or fans of his read it and get offended? I doubt they care much, and I fail to see the point of it.
Elmo already cancelled out any progress made by buying a social media platform and getting the most anti-science anti-NASA admin in history elected. He's done a net negative on the world at this point, even if the scale is vastly larger than most people.
do it on mars then... that would have the added benefit of heating the planet so we could live on it. It seems so obvious if you think about it. Someone transfer eleventy trillion dollars to Elon Musk so we can get started.
this does not makes sense from dollars pov to me, I ran a back of napkin session with claude & gemini on this and the short of it is you need a magical weightless radiator for cooling and even then it wont work because the launch costs need to be sub $100 before this can be feasible. this does not even factors the 5y amortization and LEO orbit drag correction.
It then occurred to me that they (all major AI companies) know all of these facts but still pushing for it so there must be another reason. Then I recalled the offhand statement from the openAI lady about govt backstop for infra, which was strongly opposed by public and AI czar. this might be be a backdoor way of injecting that backstop capital in terms of subsidies now for results in 5 years or so. and needless to say after pilot programs those will fail spectacularly.
I suspect that this orbital data centers isn't entirely about dollars (No doubt dollars are important).
I suspect it is about the regulatory environment. The regulatory environment on data centers is moving quickly. Data centers used to be considered a small portion of the economy and thus benign and not worth extorting/controlling. This seems to be changing, rapidly.
Given that data centers only exchange information with their consumers they are a natural candidate for using orbit as a way to escape regulators.
Further, people are likely betting that regulators will take considerable time to adjust since space is multinational.
True, but businesses don't care about regulations except where it costs them money. Also, remember that time is money, so any regulatory delays cost real money to a business.
My point is that you can actually reduce it all to dollars. And I believe that the cost of orbital data centers will come down due to technological advances, while the cost of regulation will only go up, because of local and global opposition.
"My point is that you can actually reduce it all to dollars."
I'm not sure. A couple of points:
1) The regulatory landscape is enormous. It is unknown from which angle regulators will "slow you down."
2) As I mentioned the regulatory frameworks in this area are evolving very quickly. It is unknown what the regulations will be in 1, 2, 5 years and how that will impact your business.
I think it is also about security. It is impossible for ordinary people to break into such a data center.
It’s a bit like the cyberpunk future when the ultra riches live in moon bases or undersea bases and ordinary people fight for resources in a ruined earth.
Well the argument some of these companies are making is that it would be cheaper over 10 years (some things like power can be cheaper in space, and you can get it from solar nearly 24h a day). It seems likely to me (as it does many other people) that it won't be cheaper, but if it's the same price or mildly more expensive there might be a regulatory incentive to train a ML model in space instead of a place like the EU
I have no idea what this Grok assisted article is trying to say. But the data-center-in-space hype is irrational. It hand-waves cooling and bit flip errors. It does not explain why we need chat bots in space (we don't need them on earth either).
It is a nice talking point for the U.S. Saudi Investment Forum. The Saudis apparently buy anything:
For someone engaging in a lot of fun, sci fi utopian thinking, he still falls prey to libertarian thought:
"I'll go one step further and say the quiet part out loud: we should be actively goading more billionaires into spending on irrational, high-variance projects that might actually advance civilization. I feel genuine secondhand embarrassment watching people torch their fortunes on yachts and status cosplay. No one cares about your Loro Piana. If you've built an empire, the best possible use of it is to burn its capital like a torch and light up a corner of the future. Fund the ugly middle. Pay for the iteration loops. Build the cathedrals. This is how we advance civilization."
That can be done easily (and has been done many times in the past! And in the present, elsewhere in the world outside the US!) by TAXING the billionaires and using that money for government funded research programs such as DARPA, NSF, national space programs that are actually ambitious and risk taking and held to timelines.
Americans need to get over this idea that billionaires are gods that we must pray to and instead see them as just normal citizens who need to be taxed way more.
If I were to guess, my first bet would be grand PR damage control for all the Mexicans, Irish, and what have you as in “don’t worry, we’ll soon be in space and out of your backyard” (no, they won’t).
What really worries me is that I keep hearing "cooling is cheap and easy in space!" in a lot of these conversations, and it couldn't be farther from the truth. Cooling is _really_ hard and can't use efficient (i.e. advection-based air or water cooling) approaches and are limited to dramatically less efficient radiative cooling. It doesn't matter that space is cold because cooling is damned hard in a vacuum.
The article makes this point, but it's relatively far in and I felt it was worth making again.
With that said, my employer now appears to be in this business, so I guess if there's money there, we can build the satellites. (Note: opinions my own) I just don't see how it makes sense from a practical technical perspective.
Space is a much harder place to run datacenters.
Yeah, I don't see a way to get around the fact that space is a fabulous insulator. That's precisely how expensive insulated drink containers work so well.
If it was just about cooling and power availability, you'd think people would be running giant solar+compute barges in international waters, but nobody is doing that. Even the "seasteading" guys from last decade.
These proposals, if serious, are just to avoid planning permission and land ownership difficulties. If unserious, it's simply to get attention. And we're talking about it, aren't we?
You should read the linked article, they talk about it there. You radiate the heat into space which takes less surface area than the solar panels and you can just have them back to back.
In general I don't understand this line of thinking. This would be such a basic problem to miss, so my first instinct would be to just look up what solution other people propose. It is very easy to find this online.
12 replies →
But heat = energy, right? So maybe we don’t really want to radiate it, but redirect it back into the system in a usable way and reduce how much we need to take in? (From the sun etc)
1 reply →
"space is cold"
I've always enjoyed thinking about this. Temperature is a characteristic of matter. There is vanishingly little matter in space. Due to that, one could perhaps say that space, in a way of looking at it, has no temperature. This helps give some insight into what you mention of the difficulties in dealing with heat in space - radiative cooling is all you get.
I once read that, while the image we have in our mind of being ejected out of an airlock from a space station in orbit around Earth results in instant ice-cube, the reality is that, due to our distance from the sun, that situation - ignoring the lack of oxygen etc that would kill you - is such that we would in fact die from heat exhaustion: our bodies would be unable to radiate enough heat vs what we would receive from the sun.
In contrast, were one to experience the same unceremonious orbital defenestration around Mars, the distance from the sun is sufficient that we would die from hypothermia (ceteris paribus, of course).
Assuming merely attitude control, sure only radiative cooling is available, but its very easy to design for arbitrary cooling rates assuming any given operating temperature:
Budget the solar panel area as a function of the maximum computational load.
The rest of the satellite must be within the shade of the solar panel, so it basically only sees cold space, so we need a convex body shape, to insure that every surface of the satellite (ignoring the solar panels) is radiatively cooling over its full hemisphere.
So pretend the sun is "below", the solar panels are facing down, then select an extra point above the solar panel base to form a pyramid. The area of the slanted top sides of the pyramid are the cooling surfaces, no matter how close or far above the solar panels we place this apex point, the sides will never see the sun because they are shielded by the solar panel base. Given a target operating temperature, each unit surface area (emissivity 1) will radiate at a specific rate, and we can choose the total cooling rate by making the pyramid arbitrarily long and sharp, thus increasing the cooling area. We can set the satellite temperature to be arbitrarily low.
Forget the armchair "autodidact" computer nerds for a minute
3 replies →
Temperature is a property of systems in thermal equilibrium. One such system is blackbody radiation, basically a gas of photons that is in thermal equilibrium.
The universe is filled with such a bath of radiation, so it makes sense to say the temperature of space is the temperature of this bath. Of course, in galaxies, or even more so near stars, there's additional radiation that is not in thermal equilibrium.
> Temperature is a characteristic of matter. There is vanishingly little matter in space. Due to that, one could perhaps say that space, in a way of looking at it, has no temperature.
Temperature: NaN °C
A perfect vacuum might have no temperature, but space is not a perfect vacuum, and has a well-defined temperature. More insight would be found in thinking about what temperature precisely means, and the difference between it and heat capacity.
3 replies →
Related: what color is space?
2 replies →
Jusssst had this conversation two nights ago with a smart drunk friend. To his credit when I asked "what's heat?" and he said "molecules moving fast" and I said "how many molecules are there in space to bump against?" He immediately got it. I'm always curious what ideas someone that isn't familiar with a problem space comes up with for solutions, so I canvased him for thoughts -- nothing novel, unfortunately, but if we get another 100 million people thinking about it, who knows what we'll come up with?
I got really annoyed when I first realized that heat and sound (and kinetic energy) are both "molecules moving," because they behave so dramatically differently on a human scale.
And yes, obviously they aren't moving in the same way, but it's still kind of weird to think about.
This article assumes that no extra mass is needed for cooling, i.e. that cooling is free. The list of model assumptions includes:
• No additional mass for liquid cooling loop infrastructure; likely needed but not included
• Thermal: only solar array area used as radiator; no dedicated radiator mass assumed
Author also forgot batteries for the solar shade transition period and then additional solar panels to charge these batteries during the solar "day" period. then insulation for batteries. Then power converters and pumps for radiators and additional radiators to cool the cooling infrastructure.
Overall not a great model. But on the other hand, even an amateur can use this model and imagine that additional parts and costs are missing, so if it's showing a bad outlook even in the favorable/cheating conditions for space DCs, then they are even dumber idea if all costs would be factored in fully. Unfortunately many serious journalists can't even do that mental assumption. :(
2 replies →
Yeah that's just flat out wrong then: you can't use the solar array as a radiator.
4 replies →
Cooling isn't anymore difficult than power generation. For example, on the ISS solar panels generate up to 75 W/m², while the EATCS radiators can dissipate about 150 W/m².
Solar panels have improved more than cooling technology since ISS was deployed, but the two are still on the same order of magnitude.
So just 13.3 million sq. meters of solar panels, and 6.67 million sq. meters of cooling panels for 1 GW.
Or a 3.651 km squared and 2.581 km squared butterfly sattelite.
I don't think your cooling area measures account for the complications introduced by scale.
Heat dissipation isn't going to efficiently work its way across surfaces at that scale passively. Dissipation will scale very sub-linearly, so we need much more area, and there will need to be active fluid exchangers operating at speed spanning kilometers of real estate, to get dissipation/area anywhere back near linear/area again.
Liquid cooling and pumps, unlike solar, are meaningfully talked about in terms of volume. The cascade of volume, mass, complexity and increased power up-scaling flows back to infernal launch volume logistics. Many more ships and launches.
Cooling is going to be orders of magnitude more trouble than power.
How are these ideas getting any respect?
I could see this at lunar poles. Solar panels in permanent sunlight, with compute in direct surface contact or cover, in permanent deep cold shadow. Cooling becomes an afterthought. Passive liquid filled cooling mats, with surface magnifying fins, embedded in icy regolith, angled for passive heat-gradient fluid cycling. Or drill two adjacent holes, for a simple deep cooling loop. Very little support structure. No orbital mechanics or right-of-way maneuvers to negotiate. Scales up with local proximity. A single expansion/upgrade/repair trip can service an entire growing operation at one time, in a comfortable stable g-field.
6 replies →
None of it is easy but neither is cooling impossible as many people are saying.
Doing like an 8xh200 server (https://docs.nvidia.com/dgx/dgxh100-user-guide/introduction-...) is 10.2kW.
Let’s say you need 50m^2 solar panels to run it, then just a ton of surface area to dissipate. I’d love to be proven wrong but space data centers just seem like large 2d impact targets.
8 replies →
There’s a big difference between “impossible” (it isn’t) and “practical” (it isn’t).
3 replies →
Space hardware needs to be fundamentally different from surface hardware. I don't mean it in the usual radiation hardenrining etc, but in using computing substrates that run over 1000c and never shut down. T^4 cooling means that you have a hell of a time keeping things cool, but keeping hot things from melting completely is much easier.
if you have a compute substrate at 1300K you don't have a cooling problem - you have an everything else problem
2 replies →
Who says that?
Every conversation I've seen is despite how many serious organizations with talented people, the "uhhh how do you cool it?" Is brought up immediately
Maybe hang out with different people?
Everyone I talked to (and everyone on this forums) knows cooling is hard in space.
It is always the number one comment on every news piece that is featured here talking about "AI in space".
I think the point is, yes, cooling is a significant engineering challenge in space; but having easy access to abundant energy (solar) and not needing to navigate difficult politically charged permitting processes makes it worthwhile. It's a big set of trade offs, and to only focus on "cooling being very hard in space" is kind of missing the point of why these companies want to do this.
Compute is severely power-constrained everywhere except China, and space based datacenters is a way to get around that.
Of course you can build these things if you really want to.
But there is no universe in which it's possible to build them economically.
Not even close. The numbers are simply ridiculous.
And that's not even accounting for the fact that getting even one of these things into orbit is an absolutely huge R&D project that will take years - by which time technology and requirements will have moved on.
2 replies →
I've done some reading on how they cool JWST. It's fascinating and was a massive engineering challenge. Some of thos einstruments need to be cooled to near absolute zero, so much so that it uses liquid helium as a coolant in parts.
Now JWST is at near L2 but it is still in sunlight. It's solar-powered. There are a series of radiating layer to keep heat away from sensitive instruments. Then there's the solar panels themselves.
Obviously an orbital data center wouldn't need some extreme cooling but the key takeaway from me is that the solar panels themselves would shield much of the satellite from direct sunlight, by design.
Absent any external heating, there's only heating from computer chips. Any body in space will radiate away heat. You can make some more effective than others by increasing surface area per unit mass (I assume). Someone else mentioned thermoses as evidence of insulation. There's some truth to that but interestingly most of the heat lost from a thermos is from the same IR radiation that would be emitted by a satellite.
The computer chips used for AI generate significantly more heat than the chips on the JWST. The JWST in total weighs 6.5 tons and uses a mere 2kw of power, which is the same as 3 H100 GPUs under load, each of which will weight what, 1kg?
So in terms of power density you're looking at about 3 orders of magnitude difference. Heating and cooling is going to be a significant part of the total weight.
For some decades now I’ve heard the debunk many times more than the bunk. The real urban myth appears to be any appreciable fraction of people believe the myth.
But space isn't actually cold, or at least not space near Earth. It's about 10 C. And that's only about a 10 C less than room temperature, so a human habitable structure in near earth space won't radiate very much heat. But heat radiated is O(Tobject^4 - Tbackground^4), and a computer can operate up to around 90C (I think) so that is actually a very big difference here. Back of the envelope, a data center at 90C will radiate about 10x the heat that a space station at 20C will. With the massive caveat that I don't know what the constant is here, it could actually be easy to keep a datacenter cool even though it is hard to keep a space station cool.
It's actually only about 3x.
As you intimated, the radiated heat Energy output of an object is described by the Stefan-Boltzmann Law, which is E = [Object Temp ]^4 * [Stefan-Boltzmann Constant]
However, Temp must be in units of an absolute temperature scale, typically Kelvin.
So the relative heat output of a 90C vs 20C objects will be (translating to K):
383^4 / 293^4 = 2.919x
Plugging in the constant (5.67 * 10^-8 W/(m^2*K^4)) The actual values for heat radiation energy output for objects at 90C and 20C objects is 1220 W/m^2 and 417 W/m^2
The incidence of solar flux must also be taken into account, and satellites at LEO and not in the shade will have one side bathing in 1361 W/m^2 of sunlight, which will be absorbed by the satellite with some fractional efficiency -- the article estimates 0.92 -- and that will also need to be dissipated.
The computer's waste heat needs to be shed, for reference[0] a G200 generates up to 700W, but the computer is presumably powered by the incident solar radiation hitting the satellite, so we don't need to add its energy separately, we can just model the satellite as needing to shed 1361 W/m^2 * 0.92 = 1252 W/m^2 for each square meter of its surface facing the sun.
We've already established that objects at 20C and 90C only radiate 1220 W/m^2 and 417 W/m^2, respectively, so to radiate 1252 W per square meter coming in from the sun facing side we'll need 1252/1220 = 1.026 times that area of shaded radiator maintained at a uniform 90C. If we wanted the radiator to run cooler, at 20C, we'd need 2.919x as much as at 90C, or 3.078 square meters of shaded radiator for every square meter of sun facing material.
[0] Nvidia G200 specifications: https://www.nvidia.com/en-us/data-center/h200/
4 replies →
The temperature that you raise to the fourth power is not Celsius, it's Kelvin. Otherwise things at -200 C would radiate more heat than things at 100 C. Also the temperature of space is ~3 K (cosmic microwave background), not 10 C.
4 replies →
Pressure matters
Will these space-based data centers run on rad-hard silicon (which is dog slow compared to anything on Earth) or just silently accept wrong results, hardware lockups and permanent failure due to the harsh space environment? Will they cool that hardware with special über-expensive high-temperature Peltiers that heat the radiators up to visible incandescence so that the heat can be shed with any efficiency? There's zillions of those issues. The whole idea is just bonkers.
Google did a study with their TPU v6
> For ML accelerators to be effective in space, they must withstand the environment of low-Earth orbit. We tested Trillium, Google’s v6e Cloud TPU, in a 67MeV proton beam to test for impact from total ionizing dose (TID) and single event effects (SEEs). > > The results were promising. While the High Bandwidth Memory (HBM) subsystems were the most sensitive component, they only began showing irregularities after a cumulative dose of 2 krad(Si) — nearly three times the expected (shielded) five year mission dose of 750 rad(Si). No hard failures were attributable to TID up to the maximum tested dose of 15 krad(Si) on a single chip, indicating that Trillium TPUs are surprisingly radiation-hard for space applications.
Do you have a link to this?
At Satellogic, we famously flew mostly just regular cellphone parts on orbit. We did have higher rates of various kinds of failures than is usual on Earth, but hardware failure can generally be masked by software redundancy.
RAM corruption is not cheap to protect against
2 replies →
It's very important in this case to specify which orbit the satellite is going to be in. If you're in LEO like the international space station you spend all day inside the Van Allen Belt protected from all those charged particles that the sun is pumping out. You're still lacking the atmosphere's protection from cosmic rays but that's not a huge dosage.
If you go out to MEO then suddenly you're outside that protective magnetic shield and you have to deal with charged particles smashing into you and you want a large mass of water or wax shielding if you don't have radiation tolerant electronics.
SSO, a low earth orbit whose plane is perpendicular to the direction of the sun so it gets constant sunlight, is harsher than normal LEO orbits because it passes over the poles where the protection from the Earth's magnetic field is weakest, but it's still a lot better than higher orbits. This is probably where you want a datacenter to get constant sunlight and as much protection as possible.
rad-hard silicon ... or just silently accept wrong results, hardware lockups and permanent failure
Somehow I don't think those are the only options. AFAIK Starlink is using a lot of non-rad-hard silicon already.
Starlink is however operating at ~500km where radiation is less of a concern, but where the lifetime of a satellite is only 2-3 years.
The unit economics of orbital GPUs suggest that we'll need to run them for much longer than that. This is actually one of the few good points of orbital data centers, normally older hardware is cycled out because it's not economic to run anymore due to power efficiency improvements, but if your power is "free" and you've already got sufficient solar power onboard for the compute, you can just keep running old compute as long as you can keep the satellite up there.
6 replies →
My understanding is non rad hardened method get around this by basically doubling or some multiple of repeating calculations and chexking data often.
Random errors will occur you just need to be checking fast enough to fix and update that bad bit flip.
I am sure there's all sorts of fun algorithms in this space but I am under the impression there is SOME tax to doing this. What is the tax? Is it 10% ir 60% I have no idea would love to know!
3 replies →
Your other options of fault tolerance typically achieved by doing everything at least twice and being willing to reboot (and accepting attrition from total ionizing radiation) or lots of shielding are fine for building functioning space hardware but suboptimal for building datacentre business models...
The radiation effects on the silicon solar cells is often underestimated, it's not just the GPUs!
they throw those satellites to a fiery doom on a regular cadence.
The LLMs they hope to have in those data centers already silently accept wrong results.
Say what you will about the data centers in space idea (I think it's transparently stupid), but ML is generally resistant to random undirected noise. It's almost a requirement by definition that a machine which takes pictures and accurately outputs the probability that they are pelicans has to be pretty robust to nigh-infinite amounts of minor variation. That's part of the reason all the super low precision stuff works. It's only in the control logic or maybe the absolute precise chokepoints of computation where flips are dangerous, so most of them are harmless.
At this scale could you do shielding?
Orbital data centers are impractical for a lot of reasons (to put it mildly) but radiation shielding isn’t one of them. Proportionally less shielding is needed as one scales up, due to lower surface/volume ratios.
There are ways in which shielding in space can do harm: really energetic particles get trapped and produce a shower of daughter particles and rays over a greater area. So you'd need even more shielding. Or you accept that such things will happen and use rad-hard parts, redundancy etc. When you have the whole atmosphere above, it's much less of a concern.
Besides, that's even more mass to be lofted. Pushing the economics further into the ludicrous end.
Sure. At the cost of lofting that shielding from the ground and taking the economics from 500x to 2000x crazy.
> or just silently accept wrong results
Silently wrong results are very fashionable these days, you know. Deterministic results are very 2010s.
> Will…
I think “won’t”. I could be wrong of course, but I imagine efforts to put servers into orbit will die before anything is launched. It’s just a bad idea. Maybe a few grifters will make bank taking suckers’ money before it becomes common knowledge that this is stupid, but I will be genuinely surprised if real servers with GPUs are launched.
I don’t mean to be facetious here. But saying “will” is treating it as inevitable that this will happen, which is how the grifters win.
> This is all to say that the current discourse is increasingly bothering me due to the lack of rigor; people are using back-of-the-envelope math, doing a terrible job of it, and only confirming whatever conclusion they already want. Calculating radiation and the cost of goods is not difficult. Run the numbers.
> References: Gemini, Gemini, ChatGPT, ChatGPT, Gemini, ChatGPT, Gemini, ChatGPT, Grok, Gemini (There are sub-references from these services in the GitHub.)
I think, if you're going to make statements like this - especially from a position of expertise, you should be personally verifying the numbers and citing their sources directly. What good is asking the reader to trust an AI on your behalf? They should trust you.
(To be clear, I suspect the conclusions drawn are still correct.)
>This isn't about talent. It's about integration... Vertical integration isn't a nice-to-have. It's the whole ballgame.
I'm going to assume there's tons of logical errors and oversights in the math, considering the author couldn't even be bothered to write the text of the post himself.
Scott Manley had a video about this last year.
https://www.youtube.com/watch?v=d-YcVLq98Ew
The short of it is that cooling is likely the biggest problem, given you will need to pump the heat to the backside and radiate it away, and the amount of mass you will need to dedicate to cooling works against deployments and increases the cost per unit significantly. Not to mention, the idea of these huge deployments runs into potential space debris issues.
Whenever one of these ventures actually manages to launch a proof of concept, I think we'll be able to quickly discern if there is actually a near-future here.
On the ground you don’t have to fling tons of sensitive electronics to orbital speeds, you can have someone pop by to check whether everything is ok if you get an alert, you have a thick atmosphere cutting down solar radiation, you can cool everything relatively cheaply, and you can run fiber and power cables directly into the building. In space ou have to design super robust and self maintaining machinery. It’s a cool marketing stunt, but I don’t get the economics behind doing it for real.
This is an interesting analysis, and I like the sliders that let you instantly show the impacts of system trades.
The one glaring hole that I see is the challenge of moving the data to/from the datacenter while it's on orbit. Bandwidth to/from space isn't free. FCC/ITU licenses are required, transmitters/receiviers/modems/DSP/antennas all add to SWAP (size, weight, and power). Ground-stations are needed to move the data up/down, but those have recently become a commodity too. Still, they're not free. (see: https://aws.amazon.com/ground-station)
There is also the added latency between earth-based users and space-based datacenters, which may be a deal breaker for some applications.
Another issue I don't see covered are the significant differences between space-based hardware and terrestrial hardware. The space stuff needs to be radiation tolerant, and that usually makes it a lot slower and a lot more expensive than the terrestrial stuff, all other things being equal.
In the end, space-based datacenters are highly impractical even if you assume that Starship can put anything into orbit very cheaply.
The most chased workload will inevitably be cryptographic research, proofs of mathematical statements are hard to find the proof for, but tend to be short and easy to verify once a putative proof is presented. Just send the proofs back to earth.
Here's some math on how affordable that abundant LEO solar energy is:
First you have to pay energy to get to LEO
A Starship Launch costs[0] 51.75 TJ of energy in terms of its methane fuel.
It will be able to take a payload of 150 tonnes or 331,000 pounds[1].
How many computers is that?
One online estimate says a computer weights 80 lbs or 35 kg.
So 150000 kg / 35 kg/computer = approximately 4285 computers that we can launch into orbit per Starship.
51.75TJ / 4285 computers = approximately 12.08 GJ per computer to place it in orbit.
Let's say each computer is a H200 and consumes 700 watts continuously. How long would it need to run in orbit before it used as much energy for computation as it took to launch it?
12.08 GJ / 700 W = 12,080,000,000 J / 700 J/s = approximately 17,257,143 seconds.
Or about 6.5 months to break even on energy.
That sounds pretty good, except my estimate for the weight of each compute unit and associated power system & cooling etc. are probably underestimates by one or two orders of magnitude. In which case you'd be looking at 5 to 50 years to break even on energy, by which time the chips are obsolete and need to be replaced anyway.
[0] https://space.stackexchange.com/questions/66480/how-much-ene... [1] https://en.wikipedia.org/wiki/SpaceX_Starship#Description
You are just launching computers, with no propulsion, no attitude control, no solar panels, no radio/laser systems, no radiators. So all of that will take mass away from the computing power. A starlink satellite already weighs about 1000kg, and that really is just the supporting infrastructure you need before you start adding computers...
So yes, 10-100x extra is probably reasonable.
Does anybody actually work with H100s and the like? Their failure rate is so high, I dont understand why anybody will even consider it feasible to put the machines in orbit or even the sea. By my ballpark estimate, if you have 800 H100s, after 6 months, about 100 would be overheating or throttling, and a few will disappear and one or two will crash the machine with load.
> Does anybody actually work with H100s and the like?
They don't. The expectation the cloud develops in people is that magic computers just appear. They're living at a virtualized layer where all the nitty gritty of real machines going down and needing to be serviced all the time is handled by unseen minions (sorry SREs and DC staff) and cluster management and provisioning software.
The reality is that datacenters in space is mind-boggling stupid, just from the infeasibility of maintenance alone.
The analyses here miss the economic realities of building datacenters. "Just use land", "just use nuclear", "just use water". All of this is contested. A system of lawsuits and regulations turns negative externalities (even ones you aren't convinced of!) into costs you can weigh against. So, like hydrogen vs. RP-1, it's not enough to pick a handful of physical performance metrics. It has to win holistically.
If you can produce any kind of economically productive compute node and add it to (for example) the Starlink network, and launch on a reusable vehicle, you carry on installing them as fast as you can build them.
So, the move is to turn the problem of contested land use into a manufacturing problem.
This is not so easy to pin down on a spreadsheet, and will be decided at the level of the business unit. If SpaceX can put a GPU/TPU on the grid more economically than the other guy, then it doesn't matter if they have ammonia in the pipes instead of water.
Grab your popcorn.
The article makes pretty much exactly this point somewhere in the middle.
> the list of organizations positioned to even try that is basically one.
Maaaaybe Blue Origin can join once they get a constellation going.
> That "why" is almost missing from the public conversation. People jump straight to hardware and hand-wave the business case, as if the economics are self-evident. They aren't.
But then he never answers that fundamental question, and jumps straight to the hardware and power and cost? What problems are orbital data centers trying to solve? What optimizations are they intended to deliver? Are these optimizations beneficial to everyone who uses a data centers, or just operators or users of orbiting satellite constellations?
> But the knock-on effects are why this keeps pulling at people. If you can industrialize power and operations in orbit at meaningful scale, you're not just running GPUs. You're building a new kind of infrastructure that makes it easier for humans to keep spreading out. Compute is just one of the first excuses to pay for the scaffolding.
This seems to be the closest we get to a “Why”, but it doesn’t make much sense. A constellation of 40,000 satellites with GPUs “infrastructure that makes it easier for humans to keep spreading out”? How?
> The target I care about is simple: can you make space-based, commodity compute cost-competitive with the cheapest terrestrial alternative? That's the whole claim. … Can you deliver useful watts and reject the waste heat at a price that beats a boring Crusoe-style tilt-wall datacenter tied into a 200–500 MW substation?
Isn’t the answer clearly “No”? The default settings of his model — which I assume he considers optimal — tell us that power for orbital data enters will cost 3.5X terrestrial ones. And that only SpaceX has the vertical integration to do even attempt to do this. So again, where is the competitive advantage?
Also, I don’t understand why he’s including satellite construction and launch costs for a 40,000 satellite constellations in his analysis, if he’s assuming SpaceX as he claims. Wouldn’t SpaceX simply implement these compute capabilities in the next gen of Starlink, so which would reduce costs significantly.
> It might not be rational. But it might be physically possible.
But isn’t that precisely what everyone has been saying? I don’t think the question has been whether orbital data centers are possible, it’s been whether they are rational. And that centers foremost h the unanswered question, Why is this a good idea?
> But then he never answers that fundamental question
The fundamental question is “is it economically viable”, and the answer from his model is “not really”
> A constellation of 40,000 satellites with GPUs “infrastructure that makes it easier for humans to keep spreading out”?
I think he’s claiming industrializing larger and more economical power generation in space, as well as the means to put it up there, would make it easier to transition to a theoretical space economy
> But isn’t that precisely what everyone has been saying?
From the article, he claims that people handwave the economics, so at least the people he has interacted with haven’t been saying that.
I love the sliders, but note that the numbers on this site literally came from ChatGPT, so there is plenty of room for disagreement.
Seems like according to this analysis it all hinges on launch cost and satellite cost. This site's default for Starship launch cost is $500/kg, but SpaceX is targeting much lower than that, more like $100/kg and eventually optimistically $10/kg (the slider doesn't even go that low). At $100/kg (and assuming all the other assumptions made on the site hold) then you break even on cost vs. terrestrial if you can make the satellites for $7/watt (excluding GPUs, as the whole analysis does).
Aerospace industry has a long history of missing lower cost/kg to orbit. I'm extremely suspicious of $500/kg, which is about a third of today's cost.
OTOH SpaceX has a pretty good history of undercutting the industry on cost. If Starship full reusability works I would be very surprised if it only lowered launch costs by a factor of three. Of course it's not guaranteed to work, but clearly SpaceX's orbital datacenter plans are predicated on Starship working.
7 replies →
I am struggling with a why for this (other than “huh cool, that will get investors”). All the jurisdiction and regulation arguments and the “we could get the costs down” seem to meet the objection of “for the same investment we could do just as well or better on the ground”.
The one that does not is the physics of the whole thing. I struggle to work out how exactly but being slightly time dilated compared to the ground does not seem like a win, but being able to gather data from opposite sides of the planet slightly faster than cables does seem like a potential win. Most stock exchanges make a significant chunk of their revenues renting out data space, so it seems a possibility.
But either way it seems very niche.
I’m not one for conspiracy theories, but since SpaceX is the only launch services provider that could actually put one of these in orbit, this smells a lot like hyperloop to me — an unserious proposal that serves as a distraction and furthers Musk’s aims, and benefits anyone who can get close enough to the piles of cash that VCs will drop on this.
You know what’s easier and cheaper than putting a data center in space? Putting one literally anywhere else other than space.
OP here. Happy to answer any questions, appreciate the boost.
François Chollet https://x.com/fchollet/status/1999982683708150014 :
"Datacenters in space" make for a catchy narrative and an interesting demo, but the math simply doesn't work.
When considering factors like launch cost, maintenance complexity, and the cost of high-bandwidth communications (latency included), there is no realistic set of economic and engineering assumptions under which orbiting datacenters become cost-competitive with simply building conventional nuclear-powered (or renewable energy-powered) datacenters on the ground.
In fact we're off by 50-100x. Dramatic launch cost reductions still won't make it work. And of course if you invest a lot in specific lines of tech to make it work you then have to consider that the same can also be invested in better ground-based nuclear, bringing the cost of power down for everyone.
So the guy who shows his work says 3.4x and the guy who doesn't says 50-100x.
OP here.
man who quotes big number - often made a fool
Did a similar back-of-the-napkin and got 5x $ / MW of orbital vs. terrestrial. This article's analysis is ~3.4x.
I do wonder, at what factor of orbital to terrestrial cost factor it becomes worthwhile.
The greater the terrestrial lead time, red tape, permitting, regulations on Earth, the higher the orbital-to-terrestrial factor that's acceptable.
A lights-out automated production line pumping out GPU satellites into a daily Starship launch feels "cleaner" from an end-to-end automation perspective vs years long land acquisition, planning and environment approvals, construction.
More expensive, for sure, but feels way more copy-paste the factory, "linearly scalable" than physical construction.
It becomes worthwhile if its actually cheaper (probably significantly cheaper given R&D and risk), or if you're processing data which originates in space and the data transfer or latency is an issue
You can set up plant manufacturing chips in shipping containers and sending them to wherever energy/land is cheapest and regulation most suitable, without having to seek the FCCs approval to get launch approved and your data back...
people use aws despite it being 2x-10x the cost of self hosting. cost isnt everything.
Hilarious how they could easily validate this by PoC. Even putting a small RasPi cluster in orbit. Do that and make a profit. Then ask for money.
Aside from the economics, the question is why do it in orbit vs on land (or sea)?
What are the regulatory/legal gains? Lack of jurisdiction means open slather?
What are the national security gains? Redundancy and resiliency by each satellite being a "micro-compute" connected by high speed laser links? So more resilient to attack?
Why do it at all?
I think the only reason is for legal purpose.
If data is downloaded illegally from space, stored in space and model trained on it... it will be a mess juridically if someone complain.
Same for model inference, it will be hard for a government to put controls on the model output.
I think the main draw is its elegance. You have very efficient power from the sun, put that directly into your compute, radiate it out. Energy is ~free, no heavy infrastructure required, just a closed circuit for computing.
Elegance compared to a PV/Storage facility built next door to a data centre?
It doesn't make sense right now, and won't for at least 5-10 years.
By which time, this current round of hype will have burned up ~$1T if it doesn't fall apart from the current internal contradictions and lack of market/customers/uses.
We're still on the uphill ride of the Gartner hype cycle, not even at the "Peak of Inflated Expectations" yet.
We also have very efficient power from the sun here on earth. I don't get how that is an argument.
Very doable with Photonic compute in orbit: Way less waste heat (huge when you can only radiate it away), lower power draw per FLOP (smaller solar arrays), and photons don't give a damn about cosmic radiation causing bit flips.
Economics for who? The builder of the data center and plethora of all contractors and sub-contractors would see great economics though. Even the sponsor/owner of the data center might see economics work out, if you consider the reputational gain (why did we land on Moon? what's the economics?), experience gained and considering the burn of someone else's money. The money of mega companies that go into this kind of "monuments" is not exactly theirs.
Who's money is it?
Tax payers
2 replies →
When Starcloud put together that whitepaper the first thing I looked at was the launch costs[1]. It references a $5M cost to launch, which right away made absolutely no sense to me. Just a cursory search shows launch costs are around $50M per launch, if not more.
It's great that this site drills down even further to demonstrate that there is absolutely no point at which the launch costs ever make this economical or viable, so I really don't understand what people are doing.
Especially because this site was harping for years about the cost of launches and putting things in to orbit, the whole reason why SpaceX got started and has grown as it has. As soon as that became an inconvenient number, we now just make things up (Just pretend that launch costs are 10% of what they actually are to get people to invest?).
[1]: https://starcloudinc.github.io/wp.pdf
Don't confuse launch _price_ with launch _cost_. It's been estimated the internal F9 launch costs are around $15M-$20M.
The $5M is a marginal cost-target for fully reusable Starship.
> Just a cursory search shows launch costs are around $50M per launch
I think datacentres in space are predicated on Starship bringing launch costs down. Way down.
Spinlaunch is also promising drastically reduced cost per launch. The payload size for their first launcher is pretty small and they appear to be struggling to get the kinetic launcher online.
1 reply →
The 'Satellite Size' slider does not seem to impact the cost, also not in relation to changing the other settings?
Terrestrial may be cheaper but it can be burned down by peasant mobs. To become an immortal god you must remove all meat-based legacy threats.
tbh this feels lot like people throwing Drake equation around. You put in whatever random numbers together and you can get any result you want.
We know the upper bound for most of those numbers. SpaceX already achieves internal marginal launch costs of ~$1000/kg, for instance. We know their rough costs per satellite. In contrast, we know little to nothing about the inputs to the Drake equation.
The numbers don't quite work out in favor of orbital datacenters at the current values. But we can tell from analyses like this what has to change to get there.
You use of the word "quite" is highly "innovative".
OP here. Thanks SyzygyRhythm
The cooling and density challenges of datacenters here on earth are not trivial , but in space this is multiple magnitudes more difficult.
These numbers are just random bullsh*t numbers.
And what problems do orbital datacenters solve? They still need uplink, so not libertarian we can do what we want, you have no jurisdiction here thing.
This is just a sci-fi idea that is theoretically possible and is riding the ai bubble for users and investors that don’t know better.
I’d love to see these variables fitted to learning curves. That would give you a forecast for when, if learning continues as predicted, the economics could be competitive. (If it doesn’t, you need a new paradigm first.)
I am reminded of how space exploration has come largely before deep ocean exploration, seems like a human bias.
Putting data centers under water makes way way more sense than into space.
https://en.wikipedia.org/wiki/Project_Natick shut down in 2024 (though apparently it hadn't been in the water since 2020.) It seems like it basically worked, but it wasn't clear that the cooling advantage was all that big relative to the hassle of having them in a difficult-to-maintain environment.
> Putting data centers under water makes way way more sense than into space
You need permits underwater. You don’t in space.
The FCC regulates satellites launched from or communicating with the US, including stuff which extends beyond spectrum licensing like mandatory 5 year deorbiting capability for newly launched LEO satellites. Europe, China and India are not regulation-free utopias either.
You've actually got more option to jurisdiction-shop with underwater data, but I'm not convinced that's the major issue with building datacentres anyway.
Ultimately there are latency and minimise data-transfer arguments for doing certain types of data processing on local machines in space, but the generalised compute and model-training argument only works if the unit economics stack up as sufficiently good to cover the risk and R&D, and they're not obviously favourable compared with cold place on earth with clear skies and access to cold water even assuming launch costs become minimal. (It's slightly amusing to see how much some advocates of that other controversial futurist vision of spaced-based solar power - whose chances of success equally depend on low launch costs - viscerally hate the latest wave of datacentres-in-space hype...)
2 replies →
Yeah try tell average eco joe you are planning to warm up oceans by 0.00000001% of what sun does already.
(I agree right now it probably makes sense, but decades and centuries away we probably don't want to warm up earth anymore. If anything space datacenters could provide shade for earth lol.)
AFAIK compute heavy datacenter in space don't work. But if you already have a vast fleet of laser connected LEO satellites throwing some efficient SSDs into them can make a lot of sense. A large portion of the traffic is fairly static, e.g. video based content or even model weights. Caching that will save you ground to space side of the transmission. This will let you put more user beams on satellites and use less ground stations.
Really interesting article. Curious to see the ChatGPT link attached
But when I click on it, I get this error.
Failed to load shared conversation. Request is not allowed. Please try again later. (403, 9aebe525df75165e-BLR)
This is very well done! I love including all the sliders so that we can play with the (reasonable looking) assumptions the author has made. Like the author, I share their surprise that the result did not come out even more in favor of terrestrial GPUs.
it would only make sense if land availability for DCs is 0 or truly expensive…
Someone might be foreseeing an scenario like that? Are satellite launchers behind this hype?
For someone who is "increasingly bothered by the lack of rigor in the current discourse", the author sure has no qualms about using LLM outputs as primary sources. This is an immediate red flag that discredits the entire article.
This is AI slop in a pretty dress. It's fascinating that space-based datacenters are such a catastrophically bad investment that even limp apologia like this can, at best, argue that maybe it's not quite as bad as you think, as long as you still manage to ignore half the costs because you're a loser who delegated your thinking to a chatbot.
The only benefit as I perceive it re: orbital data center hardware is regulatory avoidance. Think...DDOS machines that can't be shut off; or financial hosting services for unsavory individuals. However, it's very expensive by all metrics (including those talked about in the article), and frankly, these satellites are sitting ducks for the hunter killer satellites the various space powers have, if they actually wanted to do something about these hypothtical data centers and the problems they would cause.
I guess I'd assume that the premise driving this would be that there will eventually be enough business in space that it's necessary for space-centric use, and that terrestrial use is just a fringe benefit or loss leader or something.
But oddly this doesn't seem to be how the concept is typically framed.
My second level curiosity is how much cheaper/competitive it'd be if we had space elevators.
Space elevators are pretend, you may as well ask if it would be cheaper if we had dilithium crystals
space-elevators require various types of unobtanium and have their own logistics challenges not to mention failure modes that involve spattering fast moving debris round the entire equator
Obviously I don't expect one next Tuesday. I just think it'd be interesting to see how it alters the picture.
I sometimes wonder if there are people out there who just read too much Neuromancer, and they think they can construct their own Tessier-Ashpool orbital dynasty.
I suppose there are several other Oligarchs In Space stories and movies since then, but if the point of the space station is to host AI, that narrows it down a bit.
Or perhaps it's performative, designed to spook gullible politicians into changing laws to "keep" businesses that were never actually going to go somewhere else anyway.
Right now all my favourite pirate sites are based in Russia but moving them to extrateritorial space would be genius!
Americans couldn't shoot at it in fear of igniting a space war with China.
so the conclusion is orbital datacenter is a very bad idea ?
Cons of orbital data centers:
- Ludicrously expensive to setup
- Need radiation-hardened silicon
- Ludicrously expensive maintenance requiring highly specialized operators (a.k.a astronauts)
- High risk of losing the entire equipment to a rocket failure (not infrequent even for modern launch vehicles)
- Supplying enough electrical power would be extremely difficult
- Cooling would be extremely difficult
- Geosynchronous orbits have at least 200ms of communication latency
- Lower orbits means the data center would not stay in place and require complicated tracking antennae and/or a communication mesh a la Starlink, again increasing latency and complexity
Pros of orbital data centers:
- ??????
...why are we doing this again?
One of the main reasons for putting "compute" and "storage" in space is that it is out of reach for the general public and would allow for more stability in exceptionally intense tyranny.
It is much easier to blow things up on land than in space, and the 'negative externalities' simpler to make assumptions about.
The value of this to the people who would be in charge of this "compute" and "storage" is likely much larger than the difference in energy cost.
As an expert in this field, I can assure that cost is not the most important factor. Demand is.
But if cost of the space GPUs is higher then the land GPUs, why would demand matter? Is there limited land? Are space GPUs better for some reason, like perhaps they can't be regulated as easily or because our AGI overlords will be less vulnerable to mobs with pitchforks?
Demand for what? A sealab like head in the sand view of lack of data regulation requirements?
specific software applications you can run in space that you can't run on Earth or can't run cheaper on Earth.
16 replies →
I realize terrestrial data centers have environmental risks, but are the risks greater for an orbital data center? I would think space debris, solar flares, or a bad actor satellite with a laser could do a lot of damage. Good luck repairing the orbital data center.
Now imagine that the risks aren't being evaluated objectively. Imagine if the risks of terrestrial data centers have been exaggerated 100x.
the cost of having very fast up/down syncrounous data/compute,comunications to any random spot on the planet, at any time, in a very hard to interupt or detect manner, is, what it is
and can not be compared to anything else
I'm not really interested in the problems that can come with orbital compute. We've seen them listed ad nauseam.
Have we seen any benefits to orbital computing by launching a cluster of raspberry pis to LEO? Surely this isn't an impossible task to test out on a smaller scale?
There isn’t really much benefit to having compute on orbit unless you’re working on VERY specific applications that have such tight latency requirements where you need to process the data immediately as it comes out of the sensor. In which case you just implement the algorithms in ASICs or FPGAs anyways.
There have been NVIDIA Jetsons or better on orbit since at least 2021 and that had no meaningful impact on any actual meaningful compute workloads beyond proof of concept demos.
> ...we should be actively goading more billionaires into spending on irrational, high-variance projects that might actually advance civilization. I feel genuine secondhand embarrassment watching people torch their fortunes on yachts and status cosplay. No one cares about your Loro Piana.
I 100% agree with this. There are ~2,600 billionaires in the world and we should encourage all of them to spend their money. Even buying a superyacht is a benefit to the economy. But the best billionaires, like Bill Gates and Elon Musk, are actually trying to advance the tech tree.
We are honestly lucky that Musk is wired funny. Any normal human being would retire and hang out on the beach with supermodels after all the abuse he has taken. But he takes it all as a personal challenge and doubles down. That is both his worst quality and his best.
> Any normal human being would retire and hang out on the beach with supermodels after all the abuse he has taken.
First, he seeks and creates conflict. He isn't 'taking' abuse, except in the sense that he is reaching out and grasping at it.
Everyone in that position takes lots of abuse. If they built their own fortune, they generally don't retire to the beach or they would have long ago.
I will agree that if he hadn't aligned himself with Trump/MAGA he would have a much lower profile.
But I think we'd be better off if taking a political position did not automatically piss off half the country. I think a lot of competent but normal people refuse the get involved in politics because of how toxic it is.
I wish Musk had stayed out of politics, but I'm glad he hasn't given up on Tesla/SpaceX just because of the enemies he's made. I think any normal person would have.
1 reply →
You should read F.A Hayek's essay on The Paradox of Savings [1]. Creation of capital like factories, education creating new specialists, or new processes lowers the cost of production and leads to real economic progress. Excessive spending without capital creation does nothing except keeping factors of production wastefully employed when they could be put to other uses and always ends in inflation.
[1] https://mises.org/mises-daily/hayek-paradox-saving
>”we should encourage all of them to spend their money. Even buying a superyacht is a benefit to the economy.”
You’re falling victim to the ‘broken windows fallacy’ here; money which is invested is actually more productive in improving medium and long term economic productivity than ‘consumption’ goods. Even ‘retained’ money (under one’s mattress) is not net-negative, as it increases the value of its circulating counterparts.
Scenario A: Someone breaks a window and the homeowner buys a replacement.
Scenario B: A homeowner adds a new window to their home.
Scenario C: A homeowner buys an online-course to learn how to make windows and then adds one to their home.
Scenario A has approximately no benefit to the economy. The homeowner is no better off (same number of windows) but had to spend money. The window maker might be better off, but only to the same extent that the homeowner is worse off.
I totally agree that Scenario A is not a benefit to the economy. That's the "broken window fallacy".
But Scenario B is definitely better for the economy. The homeowner has decided that having a new window is better than having the money. So the homeowner is better off. The window maker is also better off because they get the money. This is what happens when a billionaire buys a yacht.
Scenario C is the best. The homeowner has a new skill, which they can use to add more windows to their house or maybe their neighbors' houses. Over time, the amount of money spent on window-making will decrease, but the number of windows will stay constant or increase. That's a net benefit. And the online-course creator still made money.
This is what Musk is doing. He is developing new technologies that enable new capabilities and/or make existing things cheaper (e.g., electric cars, access to space, rural internet connectivity).
There is also Scenario D: The homeowner doesn't buy a new window but just keeps his money under his mattress. This is clearly the worst for the economy. Hording money like that means that there is less money circulating and lowered economic activity. The window maker is worse off, and even the homeowner is worse off if they would like to have a new window.
Billionaires who don't spend their money are the real danger, not the ones who tweet too much.
Investing their money is slightly better in that it makes the price of borrowing cheaper. But that only helps up to a point. Someone has to spend money or else there's no point in being able to borrow some. So I wish more billionaires were following Scenario C.
1 reply →
Any normal human being may hang out with models for a while but I guess they will promote their own ideas very soon. $$$ brings power.
I am still waiting for someone to fund some open-source github projects instead of some new library or museum in their name.
Yes, absolutely! I think this will happen as more billionaires come from the tech world.
Your description of Elmo applies to several other billionaires who have somehow avoided quixotic hyper-destructive rampages through American politics. I’m all for wired funny, but not when it comes with this much carnage.
I'm not sure who the whole 'Elmo' thing is for -- more than anything I cringe at the person saying it, rather than thinking less of Elon or whatever the hope is. Like 'drumpf', or the whole small hands thing, it just comes across like a redditism that escaped confinement.
Is the hope that Elon or fans of his read it and get offended? I doubt they care much, and I fail to see the point of it.
1 reply →
He plans to IPO SpaceX and xAI, so this space data center fantasy benefits both:
https://finance.yahoo.com/news/musks-net-worth-hits-600-2022...
He is a walking billboard.
Elmo already cancelled out any progress made by buying a social media platform and getting the most anti-science anti-NASA admin in history elected. He's done a net negative on the world at this point, even if the scale is vastly larger than most people.
[dead]
do it on mars then... that would have the added benefit of heating the planet so we could live on it. It seems so obvious if you think about it. Someone transfer eleventy trillion dollars to Elon Musk so we can get started.
this does not makes sense from dollars pov to me, I ran a back of napkin session with claude & gemini on this and the short of it is you need a magical weightless radiator for cooling and even then it wont work because the launch costs need to be sub $100 before this can be feasible. this does not even factors the 5y amortization and LEO orbit drag correction.
It then occurred to me that they (all major AI companies) know all of these facts but still pushing for it so there must be another reason. Then I recalled the offhand statement from the openAI lady about govt backstop for infra, which was strongly opposed by public and AI czar. this might be be a backdoor way of injecting that backstop capital in terms of subsidies now for results in 5 years or so. and needless to say after pilot programs those will fail spectacularly.
I suspect that this orbital data centers isn't entirely about dollars (No doubt dollars are important).
I suspect it is about the regulatory environment. The regulatory environment on data centers is moving quickly. Data centers used to be considered a small portion of the economy and thus benign and not worth extorting/controlling. This seems to be changing, rapidly.
Given that data centers only exchange information with their consumers they are a natural candidate for using orbit as a way to escape regulators.
Further, people are likely betting that regulators will take considerable time to adjust since space is multinational.
True, but businesses don't care about regulations except where it costs them money. Also, remember that time is money, so any regulatory delays cost real money to a business.
My point is that you can actually reduce it all to dollars. And I believe that the cost of orbital data centers will come down due to technological advances, while the cost of regulation will only go up, because of local and global opposition.
"My point is that you can actually reduce it all to dollars."
I'm not sure. A couple of points:
1) The regulatory landscape is enormous. It is unknown from which angle regulators will "slow you down."
2) As I mentioned the regulatory frameworks in this area are evolving very quickly. It is unknown what the regulations will be in 1, 2, 5 years and how that will impact your business.
1 reply →
Interestingly, the humans running the "unregulated space datacenter" are still on Earth, subject to Earth's laws.
I think it is also about security. It is impossible for ordinary people to break into such a data center.
It’s a bit like the cyberpunk future when the ultra riches live in moon bases or undersea bases and ordinary people fight for resources in a ruined earth.
How on earth does that justify the astronomic expense difference?
Well the argument some of these companies are making is that it would be cheaper over 10 years (some things like power can be cheaper in space, and you can get it from solar nearly 24h a day). It seems likely to me (as it does many other people) that it won't be cheaper, but if it's the same price or mildly more expensive there might be a regulatory incentive to train a ML model in space instead of a place like the EU
[flagged]
I have no idea what this Grok assisted article is trying to say. But the data-center-in-space hype is irrational. It hand-waves cooling and bit flip errors. It does not explain why we need chat bots in space (we don't need them on earth either).
It is a nice talking point for the U.S. Saudi Investment Forum. The Saudis apparently buy anything:
https://xcancel.com/elonmusk/status/2000603814249079165#m
For someone engaging in a lot of fun, sci fi utopian thinking, he still falls prey to libertarian thought:
"I'll go one step further and say the quiet part out loud: we should be actively goading more billionaires into spending on irrational, high-variance projects that might actually advance civilization. I feel genuine secondhand embarrassment watching people torch their fortunes on yachts and status cosplay. No one cares about your Loro Piana. If you've built an empire, the best possible use of it is to burn its capital like a torch and light up a corner of the future. Fund the ugly middle. Pay for the iteration loops. Build the cathedrals. This is how we advance civilization."
That can be done easily (and has been done many times in the past! And in the present, elsewhere in the world outside the US!) by TAXING the billionaires and using that money for government funded research programs such as DARPA, NSF, national space programs that are actually ambitious and risk taking and held to timelines.
Americans need to get over this idea that billionaires are gods that we must pray to and instead see them as just normal citizens who need to be taxed way more.
There was already an article of how this thing is just ain’t gonna happen:
https://taranis.ie/datacenters-in-space-are-a-terrible-horri...
—
If I were to guess, my first bet would be grand PR damage control for all the Mexicans, Irish, and what have you as in “don’t worry, we’ll soon be in space and out of your backyard” (no, they won’t).
https://www.nytimes.com/2025/10/20/technology/ai-data-center...