[throwaway13337]

Are there any reasonable analyses of the practicality of data centers in space?

I know Dwarkesh Patel was interviewing Elon and brought up the fact that power cost for data centers is only 20%. The number I could find is 7-18%? GPUs are the majority of the cost. I don't think Elon responded directly to that.

There's the argument that licensing to build these things is cheaper in space. But earth has a lot of space in the middle of nowhere that no one would object to. That seems cheaper than space.

And the heat dissipation argument against it seems like a good one but I don't know if it's actually just a small engineering problem that can be solved cheaply or more fundamental.

On the plus side, you could say there is better connectivity in orbit. But if you're running inference, you'd probably want to talk to the same server that has your context cached. As it whips around earth, your latency would vary a lot, right?

I'd love it if someone could point me to a better analysis. It's an interesting question in general. Not just because one of the highest valued companies in the world is based entirely on its feasibility.

[notahacker]

This has been done on HN a few times. The heat dissipation argument is a solved problem with known physics, it's just that pointing out the inefficiency of radiative cooling is the correct response to the daft claim that "space is cold" (the solution is "launch lots of additional mass into space every 5 years" which isn't exactly cheap). Though that claim is still better than Elon trying to tell idiots that SpaceX's ODCs are "much simpler than a Starlink satellite"...

Good coverage of the relevant problems are here https://peraspera.us/realities-of-space-based-compute/

Perhaps an understated one is chip obsolence. With a terrestrial datacentre you replace chips when they're uneconomical due to how much faster alternatives are or when they reach end of life; with an orbital datacentre you replace them on fixed cycles depending on how much propellant you launched with.

But nobody doubts you can build them, it's just hard to imagine a scenario in which a terrestrial equivalent isn't cheaper, more flexible and more reliable. Actual good reasons for adding compute capacity in space are, ironically, the latency: for some edge cases like autonomous control systems that matters more than the attractive unit economics of sticking computers in a building.

Still, the economic case for ODCs [eventually] is more compelling than the case for the value of that revenue stream to SpaceX exceeding current US GDP in the not too distant future...

[Freedom2]

Elon has mentioned before that he intends to place AI-powered factories in space that will manufacture and install new chips, while recycling the older ones. He says it's going to be fully self-sufficient.

[mingus88]

Haha sure.

He also said that cybertrucks would be deployed to Mars, it would be amphibious and bulletproof, people buying Teslas would make money turning them into to self driving taxis, and charged people for FSD in 2016.

If you believe anything he says, maybe you should get in on this spaceX stock deal!

[notahacker]

That's the least practical solution of all (I say that as one of the few people in the world actually working on an orbital recycling project), especially if your technical advantage is launch, not inference chip manufacture. Apart from anything else, these giant satellites in LEO still need refuelling every five years, otherwise they'll just burn up in the atmosphere.

Replacing chips with newly launched chips on a refuelling cycle (and keeping all the heavy structural elements) is well within the boundaries of near term technology, but still involves enough mission and design-for-serviceability complexity to make it not obviously better than the 5 year recycle cadence Starlink and other modern constellations operate on.

Manufacturing newer, better chips in a microgravity environment using exactly the same material and plant technology as older chips is another thing entirely, particularly if you don't happen to own Nvidia et al's closely guarded trade secrets on how to make inference chips that are actually competitive and still aren't fully self sufficient even after the magical technology has been introduced because you need to top your massive satellites up with gas on a periodic basis....

[stasomatic]

Good one. Are they going to synthesize silica in orbit as well, from dark matter?

[FloorEgg]

Two things about your line of thinking stand out to me:

The first is a subtle implication that the share price of SpaceX is dependent on the short term feasibility of data centers in space. I don't think it is. Its more dependent on public sentiment and hype, which is detached from the truth and can have its focus directed elsewhere as the company engaged in many activities. I would also argue that the company's culture matters more than any engineering specifics, especially with how diverse their assets, competencies and offerings are. Maybe you didn't mean to infer this though.

Second, if intelligence is something like electricity, in that it's fungible and translates fairly universally into value, then it's safe to assume civilization will pursue expanding it endlessly (and will never have that demand satiated) like a force of nature. If this is true, would we rather have 1000x or 1,000,000x the data centers we have today within earths atmosphere or out in space somewhere? Personally, at that scale I would prefer them far away. (Not even in LEO)

There is no core physics reason why AI data centers in space can't work. It's just really hard engineering. Kind of like what reusable rockets looked like 10 years ago... All the experts were nay-sayers.

Short term, who knows, SpaceX could struggle financially and be a terrible investment today. I have no idea. It seems overpriced to me now and it did at IPO. But on a 10 or 20 year investment horizon, it looks a lot more interesting. I don't own shares but maybe dollar cost averaging in at some point isn't a bad idea especially if the price comes down further and as part of a diversified portfolio.

[xorcist]

> Are there any reasonable analyses of the practicality of data centers in space?

The purpose of "datacenters -- in space!" isn't to actually lift server racks into orbit, one piece at a time.

It serves the same purpose as every Tesla serving extra as autonomous taxis, as people travelling by hyperloop instead of a functional train service, and as humanoid robots doing your dishes instead of the normal robots called dish washing machines.

None of these things are meant to be taken literally. Yet they all seems to have worked out just fine.

For these arguments to carry pathos they all need have a grain of truth to them. There will be local AI in satellites, for "special" applications. Think image processing for birds that have a very nice high resolution view of your back yard, and whose decisions must be low latency or when bandwith is an issue. The trick to the argument to to blur the issue by conflating it with the datacenter AI that anyone can use.

[beAbU]

No need for any technical analysis. Space-based DC's claim solve one problem: backlash and resistance against construction from local communities.

Yet, the DCs here on earth have other problems such as high energy consumption needing on-prem gas turbines and requiring large quantities of water for cooling requirements.

I'm looking forward to a terrestrial DC that solves all these issues: it's completely self contained and self sustaining requiring very little maintenance, no external power and self contained waterless cooling. Once we have that then we can ask how we can package up this building and put it in the sea. And then when we've had enough of that, we can put them in space.

[xnx]

> Are there any reasonable analyses of the practicality of data centers in space?

This video points out that even if you take the most unrealistically optimistic value of every variable, it still doesn't make sense:

"The Space Data Centers Situation is Insane" https://www.youtube.com/watch?v=_qpdUNMt2yg

[cyanydeez]

when the billionaires trigger a nuke fest, I assume they want their AI overload safe.