Space data centers still make absolutely no sense.

Whether they make economic sense or not depends on launch costs. With current operational rockets (such as Falcon 9), the launch costs are too high for space compute to be profitable. All these projects depend on Starship succeeding. If you read Starcloud's white paper[1], they make the case that space data centers can be cheaper than terrestrial ones because of access to 24/7 solar power. I think failure is more likely than success for Starcloud and similar startups, but conditional on Starship lowering launch costs, it does look like the idea can make money.

1. https://starcloudinc.github.io/wp.pdf

▲

jubilanti 5 hours ago | parent | next [-]

And whether you can radiate all that heat in the vacuum of space

▲

ggreer 4 hours ago | parent [-]

If orbital compute fails, it won't be due to heat issues. Current radiators are around 8kg per square meter, and each square meter can radiate close to 1kW. So a 1 megawatt satellite would need less than 40 square meters of radiators weighing around 300kg. And that assumes there won't be any improvements in radiator efficiency.

The outcome is almost entirely determined by launch costs.

▲

UltraSane 3 hours ago | parent [-]

You need both enormous solar panels and radiator panels. 1MW needs about 7000 square meters of solar panels. This would require 1,500 square meters of radiator panels along with a lot of coolant pipes that have to be kept edge on to the sun at all times. And all of this will require thousands of kilograms of structural support.

▲

ggreer 3 hours ago | parent [-]

What sort of solar panel gets only 140 watts per square meter in space? Even with efficiency as low as 20%, you'd get 1 megawatt from 3,700 square meters. And you'd design the satellite to minimize heat transfer between the panels and main body, because panels can run at higher temperatures than the computers. Since heat radiation goes up with the 4th power of temperature, the panels would not need large radiators. In most satellites today, a coating on the back of the panel is all that is required.

Edit: Oh wait, you're the guy who makes bold claims and then refuses to bet on them because you know you'll lose money. 10 months ago you said, "no Tesla has EVER driven without a driver with Tesla assuming liability for crashes. And I don't think they ever will."[1] When I tried to get you to bet money on your prediction (which is now false), you ignored me. A few months later, you said, "Tesla doesn't even have the permits to operate Robotaxis in California and probably never will because they are actually stricter vs Texas."[2] and again I tried to bet you that Tesla would have Robotaxis in California before the end of 2028, and again you refused.

I really wish HN had a way to mark or mute users, because if I knew who I was replying to, I'd collapse the thread and not waste my time engaging with you.

1. https://news.ycombinator.com/item?id=45661368

2. https://news.ycombinator.com/item?id=47816921

▲

UltraSane 2 hours ago | parent [-]

That includes overprovisioning for times the satellite is behind the earth.

And the Tesla robotaxi is still a embarrassing failure.

▲

ggreer 2 hours ago | parent [-]

I'm sorry, but you do not know what you are talking about. If you'd read the Starcloud white paper I linked to, you'd see that the plan is to put satellites in a dawn-dusk sun-synchronous orbit, meaning they would be in constant sunlight and the backs of the solar panels would radiate to deep space. No orbital compute project is planning to put satellites in an orbit where Earth blocks the sun, as that would require large battery banks to power the computers while in shadow.

	▲	piloto_ciego an hour ago \| parent [-]
		This is the correct take here. SSO is where they'll park these things.

▲

UltraSane 3 hours ago | parent | prev [-]

Launch costs need to decrease by a factor of 20 to 30 for orbital data centers to make sense. And even then there are enormous issues with irreversible radiation damage.

▲

lesuorac 8 hours ago | parent | prev [-]

When nobody lets you build a land data center they're going to make a lot of sense.

▲

runako 6 hours ago | parent | next [-]

A hypothetical operational space data center will cost a multiple over a similar terrestrial facility. It would be cheaper/more efficient to offer a municipality or its citizens a fraction of that multiple to secure approvals, now that the bullying approach is running out of steam.

Instead of installing MWs of illegal gas generation and also still causing local electricity rates to spike, why not arrange to buy down the electricity of local ratepayers and also comply with existing laws & regulations?

It's worth asking why operators aren't doing that yet. I suspect it's because neither OAI nor Anthropic have an extra $50B-$100B to put out at the moment. We'll see what happens post-IPO.

▲

mchusma 6 hours ago | parent | prev | next [-]

Yeah, the Data Center NIMBY crowd is making these look real good right now.

▲

wat10000 5 hours ago | parent | prev | next [-]

Never going to happen. Yeah, we might get to a point where lots of people object to them in lots of places. But nobody allowing them anywhere? Mongolia wouldn't let you build in a little corner of the vast steppe? Algeria wouldn't let you plop down in a nice patch of unused Sahara?

The worst that would happen is that it becomes difficult to build in convenient places. That certainly would be annoying. But the problem with space as the solution is that space is by far the least convenient place imaginable, with the possible exception of the bottom of the deep ocean.

▲

grim_io 7 hours ago | parent | prev | next [-]

That's an if, not a when. A very big, hugely improbable if. Nobody? really?

▲

piloto_ciego 7 hours ago | parent [-]

I know it sounds bananas, indeed I thought it sounded bananas for a long time, but in a world where we can't get high speed rail built anywhere because of permitting, and there's an artificial scarcity of available land to build, they make a lot of sense.

In space you can just start. Nobody is really going to stop you.

▲

jazzyjackson 6 hours ago | parent [-]

People are going to be even more pissed off when there are more satellites than stars

(I know nominally they’re not visible but they are shortly after launch while moving into position, and at large scale that would be happening all the time, not to mention the other side of burning up in the atmosphere after a few years)

▲

MarkusQ 4 hours ago | parent [-]

Do you have any idea how many stars there are?

▲

jazzyjackson 4 hours ago | parent [-]

Depends on what city you’re in :p

▲

MarkusQ 4 hours ago | parent [-]

Actually, no, they're still there even if you're indoors with your eyes closed.

	▲	piloto_ciego 3 hours ago \| parent [-]
		<DONALD HOFFMAN INTENSIFIES>

▲

arrosenberg 6 hours ago | parent | prev [-]

The cost of getting them there will be absurd and even if someone was willing to fund it, where do you put the heat they will generate?

▲

piloto_ciego 3 hours ago | parent [-]

If trying to do it terrestrially is because the city council will riot then doing it on orbit seems like the better play. Just saying.

“I can do it in this challenging environment or I can not do it at all” is kind of a false choice if you think AGI is imminent. Personally, I think we’re already there for KVM stuff, it’s just the goal posts are accelerating at an exponential rate the same with the capabilities.

▲

arrosenberg 2 hours ago | parent [-]

"If you think AGI is imminent" is doing a lot of heavy lifting. The rate of improvement has already leveled off significantly and there is little-to-no evidence that LLMs are close to or capable of achieving general intelligence.

If you assume we aren't going to create a singularity in the next few decades, then no, it doesn't make sense economically to do it. Even ignoring that we'd have to develop a design that works in space, and even if SpaceX is able to significantly reduce cost per kg to orbit, it would still cost 15-30x more to build the same data center, and you have to pay that premium every time a piece of equipment burns out.

It would be cheaper to build aircraft carrier sized ships to hold them or give every adult in the town where they are building a couple million dollars to fuck off. And again, that's only if SpaceX can produce an as-of-yet mythical level of scale. If they stop anywhere short of that, then we're talking about billions of dollars to build in space what you could build on earth for tens of millions.

	▲	piloto_ciego an hour ago \| parent [-]
		Everyone says this about the rate of improvement every month, and yet, I'm not just basically conjuring the code into existence that's paying my bills and building my company. Personally, I know it's just my anec-data, but I'm pretty much convinced that for kvm work, we're already there or extremely close to agi in any meaningful sense.