Space data centers are physically possible but don't make financial sense. The total cost of an orbital datacenter over five years is at least 2–3x that of a terrestrial one.
But those economics don't matter to SpaceX, because the main purpose of its orbital data centers is to create a use case for Starship. Starship has to fly frequently to iron out the kinks, encounter and fix rare (1/1000) failure situations, and optimize the launch cadence which pushes launch costs down. Plus Starship needs to fly a lot before it's ready for crewed flight. The long-term goal is a Starship optimized for crewed interplanetary travel. Orbital data centers are a payload that bring in some revenue, and provide a reason to launch constantly.
It's the same thing they did with Starlink to make Falcon 9 as reliable and rapidly reusable as it is.
I still don’t see what the advantage is. Of course it’s physically possible to build a datacenter in space, but I can’t imagine land prices being that high that the same data center on earth wouldn’t be cheaper. Even just due to launch costs and the more sophisticated equipment needed for space.
This isn't terribly practical. Yes, we can deal with heat. The trouble is cost, and dealing with high energy radiation both flipping bits and corrupting the silicon.
Allow me to propose a modest alternative to space data centers, namely mountaintop data centers. This would consist of a container full of servers and GPUs and what else goes into a data center, a wind turbine for power and a communication module (say laser or microwave) for communicating with a base station with a fiber connection. This would be lifted on top of a mountain by a helicopter and bolted in place. Cooling would be provided by heat sinks exposed to the outside air. Some of the nodes could relay traffic from other nodes on remote mountain tops out of sight of the base station.
This scheme has many advantages over space data centers including launch costs, cooling, connection latency, servicability and ease of recycling.
I think it's a vary valid option to launch swarms of datacenters into space. I think a few decades to a hundred years from now, it will be the norm. Until then, we can find plany of land to do it. Instead a launch, you just need a battery. Much cheaper. All the rest stays the same.
Whyyyy are we not building distributed data centers under driveways? I want one under my driveway to melt the snow. It can use my power and water hookups if it pays for them.
What I don't understand about building a space data center is that you need radiators to release heat. Otherwise, it will become a space thermos. What's even more incomprehensible is that you would need specialized equipment for space radiation, and GPUs are consumables. To make that profitable, you would need pricing that is many times higher than the cost of a regular data center. I don't understand why there are people who actually fall for this. If I say this, people will call me someone who mocks others' challenges, but it seems like they're saying physical problems can be overcome too easily.
Seems reasonable that the area needed would be less than the solar panels. Since it sould be more efficient to dump heat than collect energy from light.
Only stating the solar panel - radiator area ratio and not mentioning the actual area that a data center should have, is a sin of omission - and a great one. Also, hand-waving numerous engineering problems just by saying “we’ve solve harder problems before / you can come up with a few designs” is the last thing I want to see in anything that comes up in Hacker News.
It’s an interesting era to be alive in, where the rich and powerful (Trump, Musk) just assert things evidence free and then a whole community rolls in to come up with a post-facto rationalisation.
Musk hasn’t thought about any of this. He just says stuff. He agreed with a statement that “space cooling is free”[1], as a sign of how deeply considered it all is.
If I had a dollar for everytime someone told me they have an advanced physics degree and can unequivocally gaurantee me that it will never be physically feasible to cool GPU's in space.... How soon those people have disappeared!
Don't we also have to worry about heating of the solar panels themselves? 150W/m^2 isn't the incident power, it's the output power. Incident is something like ten times that. Some of that's going to be reflected, but not all of it.
All the comments are negative so I'll play the devil's advocate. Here's the steelman case for SpaceX orbital datacenters.
# EARTH IS FULL
It sounds ridiculous but the ability to build AI datacenters on Earth is nearly exhausted. The options are:
• USA, Australia. The electricity infrastructure has already bottlenecked and some datacenters like Colossus are being forced to build their own power plants, but that's also bottlenecked on gas turbine capacity. Hacks like recycling jet turbines only squeeze a bit more out. Terrestrial solar can't be used to escape this problem because you need the clusters to run at night too.
• Europe. Deindustrialized, EU Commission is anti AI, very expensive power, grid also bottlenecked. Forget it.
• Middle East. Had some datacenters until they got blown up by Iran.
• China. Got power but bottlenecked by trade sanctions. Might well do a big buildout when Ascend starts to be competitive, but Chinese demand is likely to absorb it.
• Latin America, Africa, south-east Asia, etc: bottlenecked by political stability, not pro-business enough, etc.
In space you don't need gas turbines because solar can be 24/7, political risks aren't there, you aren't bottlenecked by grid capacity. Even if it costs more to put stuff in space that doesn't matter if space is the only place you can put stuff.
# INFERENCE NOT TRAINING
Trying to do backpropagation in space would be a bad idea. You need extreme locality in a single physical location for networking reasons. But a lot of modern AI load including training load is just inference, which only requires small pods not entire clusters, and bandwidth needs in/out aren't that high. Inferencing can fit on a satellite.
Space radiation isn't necessarily a problem. Bit flips can be tolerate to quite a high degree for inferencing because models can recover from corruptions in the activation stream or even some bad tokens.
# COOLING
As the article lays out this isn't necessarily the problem people are assuming. Also there are candidate designs from decades ago for ferrofluid droplet radiators. These might be overkill but can in theory radiate huge amounts of heat without needing to launch big radiators.
# COST
Unlike terrestrial data centers which are always bespoke projects, inferencing satellites can be mass manufactured. SpaceX and Elon in general are good at setting up mass production lines, and it seems apparent that SpaceX has no intention of throwing very high margin Nvidia hardware into orbit. The plan is to use Tesla's AI chips i.e. SpaceX could acquire accelerators at cost. This changes the cost calculations quite a bit. Although these accelerators might not be useful for training or research, most training workloads would stay on Earth so that doesn't matter (the inferencing loads moved into space would free up terrestrial hardware for training anyway).
The real wild card is if there's enough demand for a 'good enough' model that it's predicted to last the lifetime of the satellite. In that case the weights could be fabbed directly into the chips like Taalas does, and so the energy consumption would be far lower.
# BUSINESS CASE
It's possible that datacenter construction goes the same way as nuclear and becomes impossibly expensive here on Earth. If so then SpaceX can end up with a near monopoly on new inferencing capacity, making them the gateway to AI and the new Nvidia.
What's especially confounding is that the mere existence of orbital inferencing might actually create that outcome, because politicians would find it much easier to squash datacenter / power projects to please activists if there is a genuine alternative!
This article made me think of a strange claim by Elon Musk at 07:08 in this [1] interview:
"Cooling is actually much easier in space than it is on earth. You can just radiate to the vacuum."
I don't think that follows. The radiator is only the final heat sink. You still need to move heat from very dense chips into a deployable, space-rated radiator, and handle pumps, loops, leaks, redundancy, radiation damage, replacement, eclipses, Earth IR/albedo, and launch mass.
The incoming power is not the electrical power generated by the solar panels, but the entire power of the light that is absorbed by the solar panels and by the body of the satellite.
Even with a perfectly reflecting body and with SOTA solar panels, the amount of incoming power is at least double in comparison with the electrical power consumed by the datacenter.
Also, the heat radiated is smaller than in TFA, because no radiator is perfectly black at the radio waves in the frequency range corresponding to the ambient temperature.
I am too lazy to make the correct computation, but there was another article linked on HN some days ago where a more plausible computation was done and the conclusion was that the minimum area of the radiators is slightly larger than the area required for solar panels.
This would still be feasible, but in reality the area would have to be even larger, because the radiator cannot have a uniform temperature, the parts where the cooling fluid is incoming will be hotter than the parts where the cooling fluid is outgoing. Moreover, the pumping of the cooling fluid requires extra power that must be added to the power budget.
There is no doubt that it is possible to build a space datacenter, if much more GPUs are installed in it than necessary, to enable to correct the more severe transient errors and to preserve enough capacity after many GPUs become permanently defective, but the cost will not be competitive with terrestrial datacenters any time soon.
The make the whole calculation more detailed, e.g. include the orbit, there are specialized programs.
The one ESA used to use is called
ESATAN
I used to share office with a colleague who mostly worked with it. This was at a time when the fad of naming products with an initial "e" had just faded. The surviving victim is "ebay" I guess, but at a time there were many like it.
So my natural reading of the huge ASCII art rending on ESATAN's title screen has always been E-SATAN. Sorry ESA.
21 comments
[ 3.3 ms ] story [ 54.6 ms ] threadBut those economics don't matter to SpaceX, because the main purpose of its orbital data centers is to create a use case for Starship. Starship has to fly frequently to iron out the kinks, encounter and fix rare (1/1000) failure situations, and optimize the launch cadence which pushes launch costs down. Plus Starship needs to fly a lot before it's ready for crewed flight. The long-term goal is a Starship optimized for crewed interplanetary travel. Orbital data centers are a payload that bring in some revenue, and provide a reason to launch constantly.
It's the same thing they did with Starlink to make Falcon 9 as reliable and rapidly reusable as it is.
https://www.youtube.com/watch?v=FlQYU3m1e80
This scheme has many advantages over space data centers including launch costs, cooling, connection latency, servicability and ease of recycling.
[1] https://taranis.ie/datacenters-in-space-are-a-terrible-horri...
Musk hasn’t thought about any of this. He just says stuff. He agreed with a statement that “space cooling is free”[1], as a sign of how deeply considered it all is.
[1] https://nitter.net/elonmusk/status/1998483552669937682?lang=...
• USA, Australia. The electricity infrastructure has already bottlenecked and some datacenters like Colossus are being forced to build their own power plants, but that's also bottlenecked on gas turbine capacity. Hacks like recycling jet turbines only squeeze a bit more out. Terrestrial solar can't be used to escape this problem because you need the clusters to run at night too.
• Europe. Deindustrialized, EU Commission is anti AI, very expensive power, grid also bottlenecked. Forget it.
• Middle East. Had some datacenters until they got blown up by Iran.
• China. Got power but bottlenecked by trade sanctions. Might well do a big buildout when Ascend starts to be competitive, but Chinese demand is likely to absorb it.
• Latin America, Africa, south-east Asia, etc: bottlenecked by political stability, not pro-business enough, etc.
In space you don't need gas turbines because solar can be 24/7, political risks aren't there, you aren't bottlenecked by grid capacity. Even if it costs more to put stuff in space that doesn't matter if space is the only place you can put stuff.
Trying to do backpropagation in space would be a bad idea. You need extreme locality in a single physical location for networking reasons. But a lot of modern AI load including training load is just inference, which only requires small pods not entire clusters, and bandwidth needs in/out aren't that high. Inferencing can fit on a satellite.Space radiation isn't necessarily a problem. Bit flips can be tolerate to quite a high degree for inferencing because models can recover from corruptions in the activation stream or even some bad tokens.
As the article lays out this isn't necessarily the problem people are assuming. Also there are candidate designs from decades ago for ferrofluid droplet radiators. These might be overkill but can in theory radiate huge amounts of heat without needing to launch big radiators. Unlike terrestrial data centers which are always bespoke projects, inferencing satellites can be mass manufactured. SpaceX and Elon in general are good at setting up mass production lines, and it seems apparent that SpaceX has no intention of throwing very high margin Nvidia hardware into orbit. The plan is to use Tesla's AI chips i.e. SpaceX could acquire accelerators at cost. This changes the cost calculations quite a bit. Although these accelerators might not be useful for training or research, most training workloads would stay on Earth so that doesn't matter (the inferencing loads moved into space would free up terrestrial hardware for training anyway).The real wild card is if there's enough demand for a 'good enough' model that it's predicted to last the lifetime of the satellite. In that case the weights could be fabbed directly into the chips like Taalas does, and so the energy consumption would be far lower.
It's possible that datacenter construction goes the same way as nuclear and becomes impossibly expensive here on Earth. If so then SpaceX can end up with a near monopoly on new inferencing capacity, making them the gateway to AI and the new Nvidia.What's especially confounding is that the mere existence of orbital inferencing might actually create that outcome, because politicians would find it much easier to squash datacenter / power projects to please activists if there is a genuine alternative!
Note: I'm not invested in SpaceX.
"Cooling is actually much easier in space than it is on earth. You can just radiate to the vacuum."
I don't think that follows. The radiator is only the final heat sink. You still need to move heat from very dense chips into a deployable, space-rated radiator, and handle pumps, loops, leaks, redundancy, radiation damage, replacement, eclipses, Earth IR/albedo, and launch mass.
[1] https://youtu.be/D_1j5dVWNYI?si=R77VeVKlRXRhaBk5&t=428
The incoming power is not the electrical power generated by the solar panels, but the entire power of the light that is absorbed by the solar panels and by the body of the satellite.
Even with a perfectly reflecting body and with SOTA solar panels, the amount of incoming power is at least double in comparison with the electrical power consumed by the datacenter.
Also, the heat radiated is smaller than in TFA, because no radiator is perfectly black at the radio waves in the frequency range corresponding to the ambient temperature.
I am too lazy to make the correct computation, but there was another article linked on HN some days ago where a more plausible computation was done and the conclusion was that the minimum area of the radiators is slightly larger than the area required for solar panels.
This would still be feasible, but in reality the area would have to be even larger, because the radiator cannot have a uniform temperature, the parts where the cooling fluid is incoming will be hotter than the parts where the cooling fluid is outgoing. Moreover, the pumping of the cooling fluid requires extra power that must be added to the power budget.
There is no doubt that it is possible to build a space datacenter, if much more GPUs are installed in it than necessary, to enable to correct the more severe transient errors and to preserve enough capacity after many GPUs become permanently defective, but the cost will not be competitive with terrestrial datacenters any time soon.
The one ESA used to use is called
ESATAN
I used to share office with a colleague who mostly worked with it. This was at a time when the fad of naming products with an initial "e" had just faded. The surviving victim is "ebay" I guess, but at a time there were many like it.
So my natural reading of the huge ASCII art rending on ESATAN's title screen has always been E-SATAN. Sorry ESA.
Joking aside, you can download ESATAN here:
https://www.esatan-tms.com
It's not obvious what the download allows but expect restrictions. I remember the version we used had a HW dongle with heavy price tag every month.