You can play around with the cost sliders to estimate the economics, but even being quite optimistic, space data centers cost ~2-3x of their terrestrial counterparts.
To remove heat by radiation there's a big benefit to running the GPUs hotter as the radiation will be proportional to the fourth power of temperature. This resource is using 85 deg C versus 60 deg C for OP which will improve cooling performance.
If Nvidia/SpaceX can make the chips run at a higher temperature that would help a lot although I assume it would have already been done if it were possible. Another option is to add a heat pump to raise radiator temperature if the smaller radiator mass can pay for the heat pump mass.
They'll pay the price if it means their data centers are immune to the anger of the commoners who are being surveilled, analyzed, and manipulated by systems running on these data centers.
That particular concern is currently covered as far as SpaceX is concerned since it launches under DoD coverage. See the California Coastal Commission and the fact that they had to apologize etc. for attempted overreach.
You are altogether externalizing the cost of carbon dioxide emissions that humanity has to pay for. The resulting global warming is not a joke. The heat problem cumulatively is bigger on Earth than it would be in space.
Remember space manufacturing? When the first modules of ISS were launched 27 years ago, that was supposed to be the next big thing. Space manufacturing, asteroid mining, and space data centers are pipedreams. The capital equipment is too heavy, and 100% forever automated operations are not realistic.
It's still full of delicious rocks with completely insane market value. They sit there waiting for our greed and capabilities to reach escape velocity. Some day there will be a TEMU space station drop shipping custom fidget spinners to wherever we desire them. Probably not any time soon but eventually...
All very interesting, but I'd wait and see what people with real-world experience running Spacecraft can come up with before dismissing the entire endeavor.
Only a single order of magnitude of cost? So a $4b data center on Earth delivers the same value as a $40b data center in Space? So if you have greater than a 95% operating margin and capacity constraints (either due to nation environmental law, power limits, or other such blockers) then this is a no-brainer - a 95% operating margin becomes a 50% operating margin. The return on capital is much much lower so you have to be anticipating cheaper capital, even cheaper costs than this in practice, or something else.
Inference margins are some 70% right now? So it needs quite some work on both sides but this seems to make it seem more achievable than I previously thought.
It assumes $44/kg to launch into orbit. Idk where that came from if it's a musk claim or what but it's ludicrous against current costs. Currently still at thousands/kg and dropping but nowhere near $44 and I haven't heard anyone claim that's likely who isn't selling a rocket company.
Musk is deliberately creating a business that needs high volume low cost launch services because in order to get low cost launch high volume is needed, and in order to get high volume a customer is needed. In the same way that Starlink launches were used to create the Falcon 9 economies of scale, the orbital AI will be used to create the Starship economies of scale. He also wants to bet the whole company on this which creates pressure to deliver and not just coast on what they've already accomplished. There are a bunch of related high risk bets but they ultimately connect back to bringing Mars within reach with frequent low cost launch capability.
The plan is to launch from Earth in the beginning stages and then switch to launching from the Moon which has no atmosphere so you don't need rockets and can use an electromagnetic gun instead.
That requires satellite manufacturing on the Moon and I guess the minerals required will come from asteroid mining.
That's my understanding of what Elon said about how it's going to work.
>So a $4b data center on Earth delivers the same value as a $40b data center in Space?
It may do so, initially, compared to a new-build DC.
But, critically, you can upgrade the racks in an already-built DC without demolishing the entire building. You can't do that in orbit, so the full lifecycle ROI is lower.
Orbital DCs are constrained by cooling and by bandwidth. Even a space-to-ground laser (which does not currently exist in a sufficiently-mature form) has a fraction of the bandwidth of a proper terrestrial fiber line. So you're paying at least 10x for essentially a disposable datacenter that can't move as much data in or out, and likely will not be as powerful as a terrestrial DC because of the cooling constraint, just to have it in space because reasons.
I don't see the business case, at all.
This is so transparently another hyperloop-esque pipe dream invented for the sole purpose of inflating SpaceX's valuation.
When I was peripherally working on some HPC stuff, there was a comment by one of the hardware guys that it mattered which national lab you were building the supercomputer for, because the guys at high altitude like Los Alamos get a lot more bitflips than someone closer to sea-level like Argonne. Although that said, for an exascale supercomputer, the mean time between uncorrected bit flip somewhere in the machine is on the order of a few hours, which means that large supercomputer-scale workloads should actually expect to hit a bit flip in their computation.
Cosmic rays can probably be resolved with some parity bits and redundancy. Especially since these don't have to be located way up in geosynch, lower orbits get you more magnetosphere protection.
Not that it isn't a problem, but I think heat dissipation will have the edge.
I worked at the architecture level on designs to mitigate signal corruption inside the asic. I don't remember the exact numbers, and obviously it depends on the design. but you need to add error detection and correction on every path (busses, mixes, registers, function units, etc). the number I seem to recall was a 25% area overhead and a nominal decrease in clock rate. this was for an earth-bound very large machine, so idk if that would be sufficient for space. usually those designs have much larger nodes and much slower clock rates. primarily because of damage caused by ionizing radiation that accumulates over time.
so sure, the heat issue seems fatal, but rad-hard designs will certainly have a bottom line impact
I'm excited I think of the Cowboy Bebop episode (Jamming With Edward) where they hunt down a rogue satellite to get data off it.
Also think it would be crazy to have a worm spread across the servers/starlink, would there be an antivirus system onboard or maybe not applicable, says RTOS can get a virus
I'm not usually an "AI is going to kill us all" kind of person but on the off chance we do get Skynet/etc why would we give the machines the literal high ground?
I think it's less Skynet and more Wintermute: Some people saw the movie Elysium or read a summary of the book Neuromancer, and decided those were aspirational visions of how to be Untouchably Rich In Space.
Or perhaps some guy with lots of stock in Rocket Launches and Datacenters has a strong financial incentive to push a story where doing both things simultaneously is the big new thing.
Honestly, data centers in the ocean make so much more sense than data centers in space it's kind of silly, but Mr Moneybags owns a rocket company he's trying to pump and dump, not a submarine company, so here we are.
It's like the idea of Mars as a backup to human civilization. The technology required to make Mars livable and independent from Earth is so advanced it would allow you to survive basically anything here on Earth already.
I would suggest data centres near the arctic - it's close to most of the Northern hemisphere's users, presents an easy option for cooling (the place is a heat sink), and, with collapsing glaciers, there will be abundant hydro power to be used.
It has the killer feature of allowing a human to walk up to a rack and replace a component.
A problem with the arctic, is that infrastructure becomes a problem the longer north you go. Even in western countries like Norway, the power grid is simply too small to handle any data center in the northern-most counties. Not to mention the political issue, where those living there would face higher utility bills.
It is assumed that if you want to build lots of data centers up north, you also need to invest in infrastructure. I've seen discussion about smaller modular nuclear power plants, but those things take years. Another thing could be other renewable energy sources.
One huge advantage of ocean data centers is that you can do geothermal down there more cheaply and easily, and it's a very consistent source of power, so you don't even need hydro. Additionally, deep ocean water is consistently very near freezing, you could get below freezing out of the water but then you lose conduction efficiency.
The replacability is nice, but even in terrestrial data centers there are situations where something fails and it never gets replaced, just routed around, until the pod gets ripped out and replaced in its entirety.
>I would suggest data centres near the arctic - it's close to most of the Northern hemisphere's users, presents an easy option for cooling (the place is a heat sink), and, with collapsing glaciers, there will be abundant hydro power to be used.
Advocating for exacerbating the melting of the polar ice cap, which will endanger dozens of millions of people, just to have more convenient data centers for manufacturing AI slop, is peak HN.
I know Microsoft tried it. They had a press release saying it was successful but never did it again. Does anyone know more about that? I agree with you that data centers in the ocean seem a better bet.
What’s the regulatory path to something like this? Oceans seem to have all the problems of land except manifold: permitting is nigh impossible, and power is hard to line up.
To make it worse, underwater tech is notoriously hard to make operationally visible. Sabotage is trivial and undifferentiable from failure and honest error. When we used to work in trading subsea cable cuts in Asia would constantly ruin our best networks. Everyone had point to point microwave expressly because it wasn’t breakable in this way. Exposing compute to this rather than just networking would have doomed the entire enterprise.
> I think the specific attraction to space is the copious massive amounts of free solar energy, isn't it?
Yes you get more energy harvest from solar above the atmosphere and can orient them to always be pointed towards the sun. But it is still so much more expensive than building out conventional solar in the Sun Belt, which is so much more expensive than just building a massive natural gas plant right next to people's homes.
No, space is desirable because there is no local permitting authority that can push back. People living near data centers have gotten wise to the fact that local people pay most of the externalities of data centers and AI, but the benefits mostly go elsewhere, and the jobs created during construction are temporary. In space, you don't have to lobby/bribe local politicians and astroturf a YIMBY movement.
> People living near data centers have gotten wise to the fact that local people pay most of the externalities of data centers and AI, but the benefits mostly go elsewhere.
That's true of just about every industrial, commercial, civic, or residential site. It's the fundamental premise behind every NIMBY protest ever. The benefit of each individual site always runs disproportionately to people further away. It's only in the aggregate, i.e. each individual enjoying the cumulative externalized benefits from far-off, that the equation could ever balance.
That is a problem, but you already have to contend with GPU refresh every ~6 years, it's likely you'd pre-build pods that you could sink and link, then you'd float them and refurbish at refresh time. With that cycle time it's a manageable engineering challenge.
Tyranny of the rocket equation[0] makes it hard for me to conceive how an orbital data centre could ever make sense.
If you just want to escape permitting laws, why not float it in international waters? That's a much more accessible and hospitable environment than LEO.
Space gets you 24/7 solar power. To run a data center for 12 hours without sunlight, you'd need massive batteries. Also oceans have weather and corrosion.
The issue with putting stuff in space isn't the kinetic energy required. In LEO that's about 30 megajoules per kilogram or $5 worth of propellant. The issue is that orbital launch vehicles are not reusable, so you must destroy an expensive rocket to get your payload to orbit. All of these space datacenter efforts are betting that Starship will be fully reusable.
> At ABI Research, where I work as an aerospace analyst, we did a rough total-cost-of-ownership comparison between a data center on Earth and one in space. It showed that the cost to launch and run a GPU in space for a year is at least an order of magnitude higher than the same feat in a terrestrial data center.
"However, there are niche applications where the much higher costs of computing in space could be justified. Examples include ... active collision avoidance in the increasingly crowded low Earth orbit."
Just so that nobody jumps in with what I first wanted to jump in with, this estimate was done with the assumption that Elon Musk's Starship is built, works as advertised, and launch costs are at the lower end of the projected range.
It wasn't an order of magnitude more because of how expensive rocket launches currently are.
(I'm glad that I read the article before arguing this one...)
The best thing for space GPUs would to run at a much high temperature, thereby allowing the radiator to be fairly small. This would require a special high-temperature GPU, with robust radiation protection and error correction, ideally in solar orbit. I have a video https://youtu.be/s7Mv_OcBXI8 covering the approach for heat rejection.
Economics calculator for Orbital vs Terrestrial Data Centers
https://andrewmccalip.com/space-datacenters
You can play around with the cost sliders to estimate the economics, but even being quite optimistic, space data centers cost ~2-3x of their terrestrial counterparts.
To remove heat by radiation there's a big benefit to running the GPUs hotter as the radiation will be proportional to the fourth power of temperature. This resource is using 85 deg C versus 60 deg C for OP which will improve cooling performance.
If Nvidia/SpaceX can make the chips run at a higher temperature that would help a lot although I assume it would have already been done if it were possible. Another option is to add a heat pump to raise radiator temperature if the smaller radiator mass can pay for the heat pump mass.
They'll pay the price if it means their data centers are immune to the anger of the commoners who are being surveilled, analyzed, and manipulated by systems running on these data centers.
I'm still expecting a crypto satellite network to phone app now that normal phones have became satellite connectible. Something outside oversight.
But, datacenter development in space does not deal with local NIMBYs, just centralized decision-makers that are easier to "influence".
The NIMBYs are coming for space launch sites, too.
That particular concern is currently covered as far as SpaceX is concerned since it launches under DoD coverage. See the California Coastal Commission and the fact that they had to apologize etc. for attempted overreach.
You are altogether externalizing the cost of carbon dioxide emissions that humanity has to pay for. The resulting global warming is not a joke. The heat problem cumulatively is bigger on Earth than it would be in space.
Cost = emmissions
No, cost is not directly proportional to emissions.
Heat emission in space is not problematic for humans
Remember space manufacturing? When the first modules of ISS were launched 27 years ago, that was supposed to be the next big thing. Space manufacturing, asteroid mining, and space data centers are pipedreams. The capital equipment is too heavy, and 100% forever automated operations are not realistic.
It's still full of delicious rocks with completely insane market value. They sit there waiting for our greed and capabilities to reach escape velocity. Some day there will be a TEMU space station drop shipping custom fidget spinners to wherever we desire them. Probably not any time soon but eventually...
All very interesting, but I'd wait and see what people with real-world experience running Spacecraft can come up with before dismissing the entire endeavor.
Only a single order of magnitude of cost? So a $4b data center on Earth delivers the same value as a $40b data center in Space? So if you have greater than a 95% operating margin and capacity constraints (either due to nation environmental law, power limits, or other such blockers) then this is a no-brainer - a 95% operating margin becomes a 50% operating margin. The return on capital is much much lower so you have to be anticipating cheaper capital, even cheaper costs than this in practice, or something else.
Inference margins are some 70% right now? So it needs quite some work on both sides but this seems to make it seem more achievable than I previously thought.
It assumes $44/kg to launch into orbit. Idk where that came from if it's a musk claim or what but it's ludicrous against current costs. Currently still at thousands/kg and dropping but nowhere near $44 and I haven't heard anyone claim that's likely who isn't selling a rocket company.
That seems pretty cheap, like, can I launch my cat's ashes into space or something for $50?
Musk is deliberately creating a business that needs high volume low cost launch services because in order to get low cost launch high volume is needed, and in order to get high volume a customer is needed. In the same way that Starlink launches were used to create the Falcon 9 economies of scale, the orbital AI will be used to create the Starship economies of scale. He also wants to bet the whole company on this which creates pressure to deliver and not just coast on what they've already accomplished. There are a bunch of related high risk bets but they ultimately connect back to bringing Mars within reach with frequent low cost launch capability.
The plan is to launch from Earth in the beginning stages and then switch to launching from the Moon which has no atmosphere so you don't need rockets and can use an electromagnetic gun instead. That requires satellite manufacturing on the Moon and I guess the minerals required will come from asteroid mining.
That's my understanding of what Elon said about how it's going to work.
So the majority of the hardware should stay on the moon where latency isn't so critical. There the cooling won't be the problem.
>So a $4b data center on Earth delivers the same value as a $40b data center in Space?
It may do so, initially, compared to a new-build DC.
But, critically, you can upgrade the racks in an already-built DC without demolishing the entire building. You can't do that in orbit, so the full lifecycle ROI is lower.
Orbital DCs are constrained by cooling and by bandwidth. Even a space-to-ground laser (which does not currently exist in a sufficiently-mature form) has a fraction of the bandwidth of a proper terrestrial fiber line. So you're paying at least 10x for essentially a disposable datacenter that can't move as much data in or out, and likely will not be as powerful as a terrestrial DC because of the cooling constraint, just to have it in space because reasons.
I don't see the business case, at all.
This is so transparently another hyperloop-esque pipe dream invented for the sole purpose of inflating SpaceX's valuation.
Anyone know why everyone focuses on the radiators, when the solar panels are significantly larger?
Why is everyone so focused on heat when cosmic rays randomly scrambling memory is going to be a much bigger problem?
When I was peripherally working on some HPC stuff, there was a comment by one of the hardware guys that it mattered which national lab you were building the supercomputer for, because the guys at high altitude like Los Alamos get a lot more bitflips than someone closer to sea-level like Argonne. Although that said, for an exascale supercomputer, the mean time between uncorrected bit flip somewhere in the machine is on the order of a few hours, which means that large supercomputer-scale workloads should actually expect to hit a bit flip in their computation.
Cosmic rays can probably be resolved with some parity bits and redundancy. Especially since these don't have to be located way up in geosynch, lower orbits get you more magnetosphere protection.
Not that it isn't a problem, but I think heat dissipation will have the edge.
You can also place the most sensitive electronics inside a module within the propellant tank. That also should help a lot.
I worked at the architecture level on designs to mitigate signal corruption inside the asic. I don't remember the exact numbers, and obviously it depends on the design. but you need to add error detection and correction on every path (busses, mixes, registers, function units, etc). the number I seem to recall was a 25% area overhead and a nominal decrease in clock rate. this was for an earth-bound very large machine, so idk if that would be sufficient for space. usually those designs have much larger nodes and much slower clock rates. primarily because of damage caused by ionizing radiation that accumulates over time.
so sure, the heat issue seems fatal, but rad-hard designs will certainly have a bottom line impact
I mean, LLMs are already a pile of stochastic output .. maybe that just adds to the fun? /s
Finally, the classic BOFH excuse of sunspots causing issues will be true
“Solved problem”. https://en.wikipedia.org/wiki/Matrioshka_brain
It merely remains to build one.
Silly human, won't be us building that. Just hope that you get one of the nice gilded cages!
I'm excited I think of the Cowboy Bebop episode (Jamming With Edward) where they hunt down a rogue satellite to get data off it.
Also think it would be crazy to have a worm spread across the servers/starlink, would there be an antivirus system onboard or maybe not applicable, says RTOS can get a virus
I'm not usually an "AI is going to kill us all" kind of person but on the off chance we do get Skynet/etc why would we give the machines the literal high ground?
I think it's less Skynet and more Wintermute: Some people saw the movie Elysium or read a summary of the book Neuromancer, and decided those were aspirational visions of how to be Untouchably Rich In Space.
Or perhaps some guy with lots of stock in Rocket Launches and Datacenters has a strong financial incentive to push a story where doing both things simultaneously is the big new thing.
Because that's the next generation's problem, and we can do more to mortgage their future for our own luxury :)
Honestly, data centers in the ocean make so much more sense than data centers in space it's kind of silly, but Mr Moneybags owns a rocket company he's trying to pump and dump, not a submarine company, so here we are.
It's like the idea of Mars as a backup to human civilization. The technology required to make Mars livable and independent from Earth is so advanced it would allow you to survive basically anything here on Earth already.
I would suggest data centres near the arctic - it's close to most of the Northern hemisphere's users, presents an easy option for cooling (the place is a heat sink), and, with collapsing glaciers, there will be abundant hydro power to be used.
It has the killer feature of allowing a human to walk up to a rack and replace a component.
A problem with the arctic, is that infrastructure becomes a problem the longer north you go. Even in western countries like Norway, the power grid is simply too small to handle any data center in the northern-most counties. Not to mention the political issue, where those living there would face higher utility bills.
It is assumed that if you want to build lots of data centers up north, you also need to invest in infrastructure. I've seen discussion about smaller modular nuclear power plants, but those things take years. Another thing could be other renewable energy sources.
One huge advantage of ocean data centers is that you can do geothermal down there more cheaply and easily, and it's a very consistent source of power, so you don't even need hydro. Additionally, deep ocean water is consistently very near freezing, you could get below freezing out of the water but then you lose conduction efficiency.
The replacability is nice, but even in terrestrial data centers there are situations where something fails and it never gets replaced, just routed around, until the pod gets ripped out and replaced in its entirety.
>I would suggest data centres near the arctic - it's close to most of the Northern hemisphere's users, presents an easy option for cooling (the place is a heat sink), and, with collapsing glaciers, there will be abundant hydro power to be used.
Advocating for exacerbating the melting of the polar ice cap, which will endanger dozens of millions of people, just to have more convenient data centers for manufacturing AI slop, is peak HN.
I know Microsoft tried it. They had a press release saying it was successful but never did it again. Does anyone know more about that? I agree with you that data centers in the ocean seem a better bet.
What’s the regulatory path to something like this? Oceans seem to have all the problems of land except manifold: permitting is nigh impossible, and power is hard to line up.
To make it worse, underwater tech is notoriously hard to make operationally visible. Sabotage is trivial and undifferentiable from failure and honest error. When we used to work in trading subsea cable cuts in Asia would constantly ruin our best networks. Everyone had point to point microwave expressly because it wasn’t breakable in this way. Exposing compute to this rather than just networking would have doomed the entire enterprise.
He must have that cave submarine lying around somewhere.
Corrosion is one hell of a problem in salt water.
I think the specific attraction to space is the copious massive amounts of free solar energy, isn't it?
(In reality, they want to build the torment nexus at the Lagrange points because that would just be edgy-as-fuck)
> I think the specific attraction to space is the copious massive amounts of free solar energy, isn't it?
Yes you get more energy harvest from solar above the atmosphere and can orient them to always be pointed towards the sun. But it is still so much more expensive than building out conventional solar in the Sun Belt, which is so much more expensive than just building a massive natural gas plant right next to people's homes.
No, space is desirable because there is no local permitting authority that can push back. People living near data centers have gotten wise to the fact that local people pay most of the externalities of data centers and AI, but the benefits mostly go elsewhere, and the jobs created during construction are temporary. In space, you don't have to lobby/bribe local politicians and astroturf a YIMBY movement.
> People living near data centers have gotten wise to the fact that local people pay most of the externalities of data centers and AI, but the benefits mostly go elsewhere.
That's true of just about every industrial, commercial, civic, or residential site. It's the fundamental premise behind every NIMBY protest ever. The benefit of each individual site always runs disproportionately to people further away. It's only in the aggregate, i.e. each individual enjoying the cumulative externalized benefits from far-off, that the equation could ever balance.
That is a problem, but you already have to contend with GPU refresh every ~6 years, it's likely you'd pre-build pods that you could sink and link, then you'd float them and refurbish at refresh time. With that cycle time it's a manageable engineering challenge.
Tyranny of the rocket equation[0] makes it hard for me to conceive how an orbital data centre could ever make sense.
If you just want to escape permitting laws, why not float it in international waters? That's a much more accessible and hospitable environment than LEO.
[0] https://archive.is/qp4gP
Space gets you 24/7 solar power. To run a data center for 12 hours without sunlight, you'd need massive batteries. Also oceans have weather and corrosion.
The issue with putting stuff in space isn't the kinetic energy required. In LEO that's about 30 megajoules per kilogram or $5 worth of propellant. The issue is that orbital launch vehicles are not reusable, so you must destroy an expensive rocket to get your payload to orbit. All of these space datacenter efforts are betting that Starship will be fully reusable.
Gravity. That's what rules orbital-anything.
> At ABI Research, where I work as an aerospace analyst, we did a rough total-cost-of-ownership comparison between a data center on Earth and one in space. It showed that the cost to launch and run a GPU in space for a year is at least an order of magnitude higher than the same feat in a terrestrial data center.
And in the next paragraph:
"However, there are niche applications where the much higher costs of computing in space could be justified. Examples include ... active collision avoidance in the increasingly crowded low Earth orbit."
A self-justifying purpose!
Just so that nobody jumps in with what I first wanted to jump in with, this estimate was done with the assumption that Elon Musk's Starship is built, works as advertised, and launch costs are at the lower end of the projected range.
It wasn't an order of magnitude more because of how expensive rocket launches currently are.
(I'm glad that I read the article before arguing this one...)
The best thing for space GPUs would to run at a much high temperature, thereby allowing the radiator to be fairly small. This would require a special high-temperature GPU, with robust radiation protection and error correction, ideally in solar orbit. I have a video https://youtu.be/s7Mv_OcBXI8 covering the approach for heat rejection.
I appreciate the sci-fi quality of this but when the IPO comes remember the "fi" stands for "fiction."
>Why Orbital Data Centers Are Harder Than Silicon Valley Thinks
Who thinks they are easy? Elon Musk? The guy that spews obvious (even to him) bullshit like we're going to Mars in the next 5 years?
I don't even want to read the article. It is obvious that Orbital Data Centers have MASSIVE engineering challenges. They may never be cost-effective.