[benl]

The bet is that demand for AI tokens will continue to grow exponentially. And that SpaceX will be able to deploy and rent out GPUs to serve those tokens faster than anyone else.

The wrinkle is that they are planning to deploy those GPUs in space. That’s what people are most skeptical about, I think!

[Alive-in-2025]

Space data centers need years of time to design, build, and deploy, 5-10 at least, and that's after they solve their multiple very difficult or impossible problems. How will they cool them? There are just simple ideas like giant structures to radiate the heat away, but you say you need to put lots of mass in orbit?

Like fsd, will take decades to figure things out.

[benl]

Well yes it will be hard, and hence maybe not economical, and that’s why many people are skeptical of the business case (myself included btw).

But satellite cooling already exists (Starlink v2 satellites dissipate heat at over a kilowatt I believe), so that’s why other people find it plausible.

[XorNot]

They also need Starship at minimum, which is now a 10+ year old project still exploding regularly.

Starship is at minimum a 2030 project at this point.

And even producing the volume of chips needed for the type of growth space data centers would need to have to justify this would be another decade if construction started now on those fabs.

[doxeddaily]

A minimum of 2030? That seems excessively pessimistic.

[XorNot]

I don't see how: Starship is a very long running project at this point, and progress has been incremental. Productionizing the basic logistics systems like turning around re-usable launch vehicle took years for Falcon 9, and Starship's haven't even done that yet.

By the end of the year if it's not landing intact yet, now you're 2027.

I'd say 2030 is optimistic (the 2028 moon landing with Artemis straight up isn't going to happen IMO).

[SJC_Hacker]

Cooling in space does not seem like a hard problem to me. You absorb a certain amount of energy in a given time in the form of solar energy, you should be able to emit that. On top of that, in LEO you are only in solar orbit roughly 50% of the time

[jgoldshlag]

It is in fact very hard, and LEO is not "solar orbit". You want your datacenters in sunlight 100% of the time, to not need heavy batteries, which is possible, but cooling is in fact very hard

[doxeddaily]

SpaceX already has 10,000 satellites on orbit that are basically preview versions of space data centers. They've already paid 5 years of that 5-10 year timeline you outlined.

[adgjlsfhk1]

the math doesn't work. a starlink satellite has ~10kw power consumption. A single ai optimized server rack (GB300) is 140kw. Starlink works because you get a massive benefit from putting networking in space for rural users. no one has made a convincing case as to why putting a data center in space is a benefit that can come anywhere near the drawbacks (inability to service, launch cost, cooling etc)

[JumpCrisscross]

> no one has made a convincing case as to why putting a data center in space is a benefit that can come anywhere near the drawbacks

Permitting. And the main drawback is cooling. If you want to sell a company to SpaceX, build a better radiator.

I’m not saying the math maths. But it isn’t fundamentally fucked in the way a lot of armchair commentary has been making it out to be.

[adgjlsfhk1]

Even permitting isn't a clear win. You are changing from land permitting (where you can pick the location to be wherever you want) to launch permitting (where you have to coordinate with the federal government for airspace and water closures). Not to mention that with the current regulatory status, a rocket explosion can easily lead to a multi-month mandatory safety review that blocks all new launches.

[JumpCrisscross]

> changing from land permitting (where you can pick the location to be wherever you want) to launch permitting (where you have to coordinate with the federal government for airspace and water closures)

One of these is orders of magnitudes longer and more complicated than the other. Land permitting always involves multiple layers of government. And most of them are causing months- to yearslong delays. (Power hook-up is another source of delay.) Launch permits are predictably issued by, essentially, a single regulator.

> a rocket explosion can easily lead to a multi-month mandatory safety review that blocks all new launches

Which is equivalent to a regular permiting delay.

The tradeoff is between the cost to launch radiator mass and the delays local and state governments cause in permiting. The first is mediated through launch costs. The latter through interest rates. And right now, the former is going down and the latter going up.