> permits? connectivity?
SpaceX may not looked upon favorably by the next US administration, with regards to permits to get to orbit.
back
1 comments
From a rationalist perspective, provided SpaceX can launch at a low cost AND the problem of radiation shielding, cooling, maintenance can be solved, it will be almost a no-brainer to launch data centers into space.
Either the next administration declares data-centers to be unimportant (which is unlikely, China etc.), or they would want to offload them as far as possible. If space is an option, it will be taken provided it's cost efficient.
From another rationalist perspective, it will always be cheaper to run a data center somewhere on earth under any conceivable scenario. It would be easier and cheaper to put a data center in the ocean. Just a few of the significant issues:
1. Selling an old compute unit is a significant cost savings when they are replaced every five years. In space you basically have to write them off.
2. Cooling in space is limited. Not impossible but limited. And these things throw off massive amounts of heat. You are talking about huge heat sinks which take valuable space and add massive weight that limits space on the rocket.
3. The size per data center will be relatively minuscule OR they somehow have to assemble in space adding additional complexity.
4. Any launch, deployment, connectivity, or hardware failure is a total loss. So you have to account for that.
It is super easy to overestimate the cooling problem, and to underestimate existing satellites. A Starlink v3 satellite that SpaceX already launches in large numbers has solar panels that produce 20kW of electricity, 20kW of compute and radios, >20kW of heat dissipation, and more internal volume than a standard 42u rack. The average rack in the average colocation facility uses less than 20kW of power, so in you could almost just package the whole system into a Starlink shell and launch it into space today. Cooling for the 20kW of heat generated by the computers, plus the extra heat absorbed from sunshine, earthshine, and moonshine, is provided purely by the surface area of the hull; no complicated cooling system needed. Simple passive heat pipes or direct contact with the hull is sufficient to cool the entire system. No bulky heat sinks are required either. Of course a less conventional system, such as a rack full of GPUs, may require more power than this. That’s merely a matter of engineering.
The Starlink deployment model works for a datacenter’s worth of racks too. Assembling them into a single structure in space is doable, but unnecessary and unproven. Latency and bandwidth between the satellites in a fleet may not be as good as that between the racks in a terrestrial data center, but the laser links that Starlink uses are excellent and will suffice for most users.
Launch and deployment failures are quite rare, and covered by insurance. Look at Starlink again for data.
Hardware failures can be a problem, but good system design mitigates it. If one system in your rack fails, do you throw out the whole rack? No, obviously not. The average engineer at a hyperscaler doesn't even notice that kind of failure because all of their jobs still get run in a timely fashion, distributed across all the working nodes. There might be a tech at the data center who unplugs broken nodes and plugs working ones in their place. Since only a small fraction of systems truly fail after installation, the inefficiency should be small enough for most operations to tolerate.
And deploying in a remote location on Earth may cost more than you realize. There will be no power company to purchase power from, so you have to build your own power plant. No water company, so you have to source your own water, clean and filter it, all at your own expense. You have to house the workers who are doing the construction. Worse, you'll probably have to build all the roads, power lines, and communication links yourself too. You could easily spend more on all of that than for a rocket launch to orbit, and that’s just to put a data center in the middle of a desert or a jungle. The ocean? Straight up impossible. Orbit is _easier_ than the ocean.
Let's do some math with two pretty well understood variables. The cost of a SpaceX rocket launch and the cost to build a terrestrial 100 MW data center.
An average rocket launch for SpaceX costs 74 million. A business they break even on. Let's assume they get a bit more efficient and they get that down to 70 million per launch. According to Elon Musk's own estimate, the space data center will be 150 KW, roughly one Nvidia AI rack. It probably has to be under volted a bit and run at 120 - 140 kw, but, whatever, roughly equivalent.
A 100 MW data center costs up to 5 billion and to power it costs roughly 150 million a year, so let's just say 1 billion more for five years of power, so 6 billion dollars for the first 5 years of the data center's operation. Why just 5 years? Because that is how long a Starlink low-earth orbit satellite stays up before it is scuttled and burns up in Earth's atmosphere, so after 5 years you can write that thing off.
Anyway, the benchmark is set. 100 MW data center at 6 billion dollars. But hey, there are roadblocks here on earth; land isn't cheap and costs keep going up, so let's worst case scenario this and say the cost of a terrestrial data center doubles to 10 billion dollars and power costs quadruple. So that's a worst case scenario baseline cost of an astonishing 14 billion dollars for a terrestrial 100 MW data center with power for five years.
70 million per rocket launch seems like pennies compared to 14 billion dollars and power is free in space! But wait, you would of course need 666 of these 150 KW mini data centers to be equivalent to one 100 MW data center. So, 666 SpaceX rocket launches, at 70 million dollars per, equals 46+ billion dollars. And that is before you factor in any other, you know, data center costs.
So no, there is no life time where it is cheaper to have data centers in space, but it may be a thing Elon does just to keep the stock price high.
Good job moving the goalposts :) I didn't say that the math would work out, only that you cannot use cooling or capacity as arguments against the idea.
That said, I will note that you calculated the launch costs using the Falcon Heavy, but Elon is assuming that the launches will be done on the Starshi. He believes that Starship will reduce those costs by quite a lot; possibly even an order of magnitude. If those costs go from 46 billion down to only 4.6 billion things may start to look different. Also, unlike the Starlink constellation, he proposes launching these server systems into higher sun–synchronius orbits with constant sunlight and a longer lifetime than just 5 years.
"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."
https://news.ycombinator.com/item?id=48491425
The idea is marketing for the stock price and enterprise value. If SpaceX does do it, it'll be analogous to Tesla's "Supervised Full Self Driving." "We run a Kubernetes cluster with some models on the satellites."
https://andrewmccalip.com/space-datacenters
If you see a datacenter in orbit, it means something went wrong on Earth - https://news.ycombinator.com/item?id=48935384 - July 2026 (0 comments)
The Space Data Centers Situation is Insane [video] - https://www.youtube.com/watch?v=_qpdUNMt2yg - June 27th, 2026
Thermodynamics rules future orbital data centers - https://news.ycombinator.com/item?id=48490094 - June 2026 (107 comments)
(reply to your deleted share link from Claude: please don't reply with AI slop)
> space data centers cost ~2-3x of their terrestrial counterparts
Which is very cheap if it means they can launch it quickly instead of being in permit hell for a year. You can also avoid sabotage (including via lawfare) and dishonest TV reporting à la "I'm standing here in front Big Bad Evil Datacenter that the local community says causes cancer and bad hair".
They also get lower latency to pretty much everybody except the neighbours of such a terrestrial data center.
The latter is an honest advantage. The former is a way of routing around damage. Both can be very valuable.