You seem to ignore some of the technical challenges of putting data centres into space right now, which makes your article a bit misleading. But relax, you are not the only one because you all think Musk will solve these challenges in the next few years. A year ago this idea was thrown out as some kind of science fiction ambition that seems to have obtained its own momentum.
For example, solar power needs to be collected on panels, very large and heavy ones . Like gigantic ones that are heavy to get into space and that would be insanely expensive. On the cooling side, yes space is a cold vacuum. But think about it, low earth orbits give you higher latency but you are not going to get 24 hours of sunlight. Also, the temperatures can vary enormously; from -160 degrees centigrade to + 120, ie not great for your electronics. So lets say you shoot for higher orbits and 24/7 sunlight, well then you have to deal with higher radiation levels and much lower latency. And how do you service these things in space? Processors can and do disfunction and are more likely to do so in space using current tech. And how do you cool stuff in a vacuum? Well you need radiative cooling which means more huge and heavy panels. There are no molecules in space to "air cool" your heat sinks.
There is also the little matter of the risk of some of these "Dyson Swarm" satellites experiencing Kessler Syndrome. Even at current density levels Spacex is having to do an insane number of flight path variations in order to avoid collisions. But apparently AI is going to take this over so everything is going to be ok. The list goes on.
Bottom line is that we are probably 10-15 years away from doing this commercially in space, at best. By this time who knows, we might have quantum computing, ASI , fusion, more underwater data centres OR whatever.
Hey, firstly, thank you for such a beautifully put and detailed response. Among many things this means that you read this and understand the subject deeply.
Here's my response to each of these points:
0. All of your counters are valid, so much so that I'll edit this article to include them and respond to them.
1. In space, the solar panels' output increases 2x-10x in space. So a 2-square-metre panel can run an H100 24x7. SpaceX's Starlinks are already powered by a 105 square-meter panels. So, a Starlink-sized solar panel could run an array of 50 H100s.
They're planning to launch a million such satellites. That's 50 million H100 clusters.
The Colossus 2 data centre for comparison has 300K H100s -- so that's about 160x bigger than what they already have on ground.
Please note that this is just napkin maths; with their current setup that they already know how to operate -- they could be doing 160x compute than the biggest datacenter that's live right now.
2. LEO gives you the lowest latency – yes. But I'm sure they'd divide these clusters into inference-only and training-only clusters.
Also, just like we have a CDN network, even between the inference-only clusters there'd be reasoning-only and non-reasoning clusters.
So, you could have non-reasoning clusters in LEO for lower latency. Reasoning clusters in higher orbits for reasoning tasks.
And both of these talking to each other? I mean, it is hard but it's not completely outside the realm of unsolvable math and computation.
3. Radiation levels -- this isn't an unsolved problem. Satellites already use radiation shielding, error correction and bunch of other means. But, again, these challenges could push us to devising new ways of doing compute in space.
What's the point of capital if not to push what we know and what's possible?
4. How do you service them in space? You don't? Even the current Starlink clusters are pulled back into the atmosphere. Maybe someday with humanoid robots we'll find a way to get them to do this -- making Tesla's Optimus' use cases more concrete.
5. Heat sinks = radiators. That's already done today on the ISS.
6. Kessler Syndrome -- very true, but I'm assuming they'll think it true. Doesn't make the whole endeavour moot.
All of your points are valid, but they need to be addressed and resolved anyway. We can't have SpaceX pioneer launches and stop at that. You need to see where we can actually take it.
The biggest challenge in all of this isn't even how it will work. We have a far bigger challenge: that's copper and other rare-earth metals. How do we get enough of them.
If space mining is the answer, we might as well start with eating the shit sandwich and getting our hands dirty in the process.
We have to go out anyway, might as well make money doing it.
The response is good and it helps my understanding - thank you. There would seem to be an opportunity to explore 'space-borne collateral' security. These types of 'races' tend to be competitive and unyielding.
How do you quantify the security risk of these satellites being targeted, taken down and / or compromised? Are we at risk of becoming too reliant on space technology and is a renewed Star Wars defence strategy developing alongside this proliferation of spaceborne technology and compute power?
The security risk remains as it is today, with the bunch of other satellites we already have in space.
We should be too reliant on space tech. Not because we can't solve the "building on land" problems, but because if we continue to solve what we already know, we'll never push through what we don't.
It's much like the Age of Exploration.
Finding new land to conquer and enslave wasn't a very moral idea -- and probably won't have gotten a green signal today –– but it pushed the idea and technology needed for transatlantic voyages. We understood more about the world; discoveries were made, and the world became an overall better place.
We have to solve space mining, humanity needs to become space-faring, and we need to traverse the universe -- like we do the international waters.
What better way for it than to solve an abstract problem of doing "computation in space".
You seem to ignore some of the technical challenges of putting data centres into space right now, which makes your article a bit misleading. But relax, you are not the only one because you all think Musk will solve these challenges in the next few years. A year ago this idea was thrown out as some kind of science fiction ambition that seems to have obtained its own momentum.
For example, solar power needs to be collected on panels, very large and heavy ones . Like gigantic ones that are heavy to get into space and that would be insanely expensive. On the cooling side, yes space is a cold vacuum. But think about it, low earth orbits give you higher latency but you are not going to get 24 hours of sunlight. Also, the temperatures can vary enormously; from -160 degrees centigrade to + 120, ie not great for your electronics. So lets say you shoot for higher orbits and 24/7 sunlight, well then you have to deal with higher radiation levels and much lower latency. And how do you service these things in space? Processors can and do disfunction and are more likely to do so in space using current tech. And how do you cool stuff in a vacuum? Well you need radiative cooling which means more huge and heavy panels. There are no molecules in space to "air cool" your heat sinks.
There is also the little matter of the risk of some of these "Dyson Swarm" satellites experiencing Kessler Syndrome. Even at current density levels Spacex is having to do an insane number of flight path variations in order to avoid collisions. But apparently AI is going to take this over so everything is going to be ok. The list goes on.
Bottom line is that we are probably 10-15 years away from doing this commercially in space, at best. By this time who knows, we might have quantum computing, ASI , fusion, more underwater data centres OR whatever.
Hey, firstly, thank you for such a beautifully put and detailed response. Among many things this means that you read this and understand the subject deeply.
Here's my response to each of these points:
0. All of your counters are valid, so much so that I'll edit this article to include them and respond to them.
1. In space, the solar panels' output increases 2x-10x in space. So a 2-square-metre panel can run an H100 24x7. SpaceX's Starlinks are already powered by a 105 square-meter panels. So, a Starlink-sized solar panel could run an array of 50 H100s.
They're planning to launch a million such satellites. That's 50 million H100 clusters.
The Colossus 2 data centre for comparison has 300K H100s -- so that's about 160x bigger than what they already have on ground.
Please note that this is just napkin maths; with their current setup that they already know how to operate -- they could be doing 160x compute than the biggest datacenter that's live right now.
2. LEO gives you the lowest latency – yes. But I'm sure they'd divide these clusters into inference-only and training-only clusters.
Also, just like we have a CDN network, even between the inference-only clusters there'd be reasoning-only and non-reasoning clusters.
So, you could have non-reasoning clusters in LEO for lower latency. Reasoning clusters in higher orbits for reasoning tasks.
And both of these talking to each other? I mean, it is hard but it's not completely outside the realm of unsolvable math and computation.
3. Radiation levels -- this isn't an unsolved problem. Satellites already use radiation shielding, error correction and bunch of other means. But, again, these challenges could push us to devising new ways of doing compute in space.
What's the point of capital if not to push what we know and what's possible?
4. How do you service them in space? You don't? Even the current Starlink clusters are pulled back into the atmosphere. Maybe someday with humanoid robots we'll find a way to get them to do this -- making Tesla's Optimus' use cases more concrete.
5. Heat sinks = radiators. That's already done today on the ISS.
6. Kessler Syndrome -- very true, but I'm assuming they'll think it true. Doesn't make the whole endeavour moot.
All of your points are valid, but they need to be addressed and resolved anyway. We can't have SpaceX pioneer launches and stop at that. You need to see where we can actually take it.
The biggest challenge in all of this isn't even how it will work. We have a far bigger challenge: that's copper and other rare-earth metals. How do we get enough of them.
If space mining is the answer, we might as well start with eating the shit sandwich and getting our hands dirty in the process.
We have to go out anyway, might as well make money doing it.
It makes sense. ✅
The response is good and it helps my understanding - thank you. There would seem to be an opportunity to explore 'space-borne collateral' security. These types of 'races' tend to be competitive and unyielding.
True, the immediate gains or losses are modest compared to what could be at stake if it eventually happens.
How do you quantify the security risk of these satellites being targeted, taken down and / or compromised? Are we at risk of becoming too reliant on space technology and is a renewed Star Wars defence strategy developing alongside this proliferation of spaceborne technology and compute power?
The security risk remains as it is today, with the bunch of other satellites we already have in space.
We should be too reliant on space tech. Not because we can't solve the "building on land" problems, but because if we continue to solve what we already know, we'll never push through what we don't.
It's much like the Age of Exploration.
Finding new land to conquer and enslave wasn't a very moral idea -- and probably won't have gotten a green signal today –– but it pushed the idea and technology needed for transatlantic voyages. We understood more about the world; discoveries were made, and the world became an overall better place.
We have to solve space mining, humanity needs to become space-faring, and we need to traverse the universe -- like we do the international waters.
What better way for it than to solve an abstract problem of doing "computation in space".