Cowboy Space (I hate their name so much) published their consolidated response a week ago but we just noticed it.
-
@sundogplanets This is so colossally stupid. Like, I know AI + Space! is a perfect VC honeypot but has anyone even thought about the fact that GPUs go out of date inside like 2-3 years? Are they really planning to launch 10,000 satellites per year indefinitely?
Also, has anyone given ANY thought to radiation hardening? GPUs of the sort that people are using for AI are typically manufactured on a 3nm process or so. Most orbital hardware which we actually want to keep running is manufactured on like 50nm because radiation absolutely destroys tiny circuits.
But I'm sure it'll be fine...
@faithisleaping @sundogplanets GPUs also often wear out inside of months, or fail spontaneously, or are installed in systems that fail spontaneously, because compute hardware is just not built to last like this.
-
@sundogplanets Admittedly I don't know much about astronomy but I know a shit ton about data centers and server infrastructure, and this _has_ to be a grift of some kind.
Compute is nowhere near reliable enough to be thrown into orbit for 18+ months without something failing: a fan, a heat sink, a power supply, a DIMM, what have you—that kind of shit breaks DAILY once you get above a certain quantity of systems, and it's pure fiction to think you could just strand it all in space and expect nothing to go bad—even the very reliable enterprise-grade stuff is not reliable enough to be left unmaintained indefinitely.
And that doesn't even touch the stupidly complex problem of cooling high density, high power compute in a vacuum (which I assume their white paper just winks at) ... and the gobshites at the FCC won't push back on any of this, probably because they'll want to cash in on the scam, too.
@klausfiend Thanks for sharing this. And yeah, they said absolutely nothing at all in any of the documents about how this would actually WORK.
-
@faithisleaping @sundogplanets GPUs also often wear out inside of months, or fail spontaneously, or are installed in systems that fail spontaneously, because compute hardware is just not built to last like this.
@klausfiend Data center companies are already eating that cost. They don't replace individual GPUs. They just shut off the unit and leave it be.
But that failure rate will be WAY higher with zero gravity, the heat differentials of space, and radiation.
-
@klausfiend Data center companies are already eating that cost. They don't replace individual GPUs. They just shut off the unit and leave it be.
But that failure rate will be WAY higher with zero gravity, the heat differentials of space, and radiation.
@faithisleaping @sundogplanets No, they don't. 90% of my job is verifying that GPUs perform to spec and the rest is identifying which approach will satisfy the RMA gods. We don't host dead compute, because that still costs money.
-
Cowboy Space (I hate their name so much) published their consolidated response a week ago but we just noticed it. EDIT: It was due at 5pm EDT today. So... I guess I have to read this shit. Swearing a lot here will get me through it.
A reminder: this is the 4th company that wants to launch fucking enormous "AI data center" satellites into sun-synchronous terminator orbits, giving Earth a ring. This particular company wants 20,000 satellites in one ring.
@sundogplanets this is the company sucking up to their suckers, er, “investors,” so as to more easily separate them from their money. At least I hope that’s all it is.
-
This company also posted an ex parte filing (which means they had a super special extra FCC meeting), their satellites are too fucking huge to even be physically possible. They just can't fucking do this.
My estimate from looking at this ridiculous sketch is 8x60x10m solar panels... that's a reflecting area of 4800 m^2. The International Space Station, the biggest thing in orbit, has a solar panel area of about 3400 m^2.
Not to mention that the cooling is something I cannot imagine worth the effort. This project will fail, but first they will fill the skies with junk.
-
Cowboy Space (I hate their name so much) published their consolidated response a week ago but we just noticed it. EDIT: It was due at 5pm EDT today. So... I guess I have to read this shit. Swearing a lot here will get me through it.
A reminder: this is the 4th company that wants to launch fucking enormous "AI data center" satellites into sun-synchronous terminator orbits, giving Earth a ring. This particular company wants 20,000 satellites in one ring.
@sundogplanets Let's use them for target practice! If it goes ahead.
What's the link? I might send it to my Astonomy and Astrophysics department, and other government sites and organisations which have a stake in it.
-
Cowboy Space (I hate their name so much) published their consolidated response a week ago but we just noticed it. EDIT: It was due at 5pm EDT today. So... I guess I have to read this shit. Swearing a lot here will get me through it.
A reminder: this is the 4th company that wants to launch fucking enormous "AI data center" satellites into sun-synchronous terminator orbits, giving Earth a ring. This particular company wants 20,000 satellites in one ring.
@sundogplanets What's the lifespan of that orbit? They are going to be using a lot of fuel to keep their path. Maybe 2 or 3 years? Then all that compute will deorbit and be destroyed?
-
@sundogplanets Admittedly I don't know much about astronomy but I know a shit ton about data centers and server infrastructure, and this _has_ to be a grift of some kind.
Compute is nowhere near reliable enough to be thrown into orbit for 18+ months without something failing: a fan, a heat sink, a power supply, a DIMM, what have you—that kind of shit breaks DAILY once you get above a certain quantity of systems, and it's pure fiction to think you could just strand it all in space and expect nothing to go bad—even the very reliable enterprise-grade stuff is not reliable enough to be left unmaintained indefinitely.
And that doesn't even touch the stupidly complex problem of cooling high density, high power compute in a vacuum (which I assume their white paper just winks at) ... and the gobshites at the FCC won't push back on any of this, probably because they'll want to cash in on the scam, too.
Cosign 100%. I have managed enterprises with massive server load. There is never a day when something doesn’t need to be touched.
Even ignoring physics, which on face seem to make this idea absurd, the practical implementation is pure fantasy. You cannot put server farms in space unless those servers are connected to a space station with full time personnel.
We have neither the technical capacity, nor the will to create massive space server farms serviced by terrestrial operators. The whole idea is science fiction fwapping.
-
@klausfiend Thanks for sharing this. And yeah, they said absolutely nothing at all in any of the documents about how this would actually WORK.
I'm sure there are abundant water sources in Low Earth Orbit they can use for cooling, just like they do here on Earth. *checks note* Oh.
-
@Red_Shirt_no2 The pictures in the marketing material show radiators. Are they enough? Probably not. Do they know that? Of course not!
They also seem to think that radiators don't emit IR...
-
@faithisleaping @sundogplanets No, they don't. 90% of my job is verifying that GPUs perform to spec and the rest is identifying which approach will satisfy the RMA gods. We don't host dead compute, because that still costs money.
@klausfiend @sundogplanets Okay, I might be out of date in how these things are managed. I was thinking back to when Microsoft (I think) was bragging about their modular data centers and how they shipped everything in sealed containers and never had to have a tech on site. But it’s been a long time and I haven’t paid much attention to that stuff.
-
And then there's a whole section where they say that the FCC isn't allowed to consider environmental regulations, and they point to Reflect Orbital getting approved where the FCC specifically said they're not allowed to consider anything except radio broadcasts. This section might actually make me cry.
Next section is just... complaining, I think?
And then they end with saying that clearly they need to be approved.
Fuck you, FCC. Kessler Syndrome is on you.
@sundogplanets And THAT is where using RF spectrum licensing as a proxy for space resource allocation has come to bite humanity in the arse. A cooperative consensus based system fails as soon as one party doesn't respect the commons. It says a lot about the LEO bros that they broke a system that functioned throughout the cold war.
-
@sundogplanets Admittedly I don't know much about astronomy but I know a shit ton about data centers and server infrastructure, and this _has_ to be a grift of some kind.
Compute is nowhere near reliable enough to be thrown into orbit for 18+ months without something failing: a fan, a heat sink, a power supply, a DIMM, what have you—that kind of shit breaks DAILY once you get above a certain quantity of systems, and it's pure fiction to think you could just strand it all in space and expect nothing to go bad—even the very reliable enterprise-grade stuff is not reliable enough to be left unmaintained indefinitely.
And that doesn't even touch the stupidly complex problem of cooling high density, high power compute in a vacuum (which I assume their white paper just winks at) ... and the gobshites at the FCC won't push back on any of this, probably because they'll want to cash in on the scam, too.
@klausfiend @sundogplanets Yeah, if they thought the price of GPUs and RAM was high on earth wait until they see the cost of radiation hardened equivalents. And then the compliance costs when they find EAR dual use applies to many rad-hard things are not trivial.
-
This company also posted an ex parte filing (which means they had a super special extra FCC meeting), their satellites are too fucking huge to even be physically possible. They just can't fucking do this.
My estimate from looking at this ridiculous sketch is 8x60x10m solar panels... that's a reflecting area of 4800 m^2. The International Space Station, the biggest thing in orbit, has a solar panel area of about 3400 m^2.
@sundogplanets This sketch of the radiators is not encouraging. All the power that comes into the solar arrays needs to go out through the radiators, and my guess is they will need substantially more radiator area than solar array area. Doesn't help that the radiators must be shielded from both the Sun and the Earth -- the Earth being almost an entire hemisphere.
-
@klausfiend Thanks for sharing this. And yeah, they said absolutely nothing at all in any of the documents about how this would actually WORK.
@sundogplanets @klausfiend Spaaaaaaace viiiiibes!!
-
I'm sure there are abundant water sources in Low Earth Orbit they can use for cooling, just like they do here on Earth. *checks note* Oh.
@davidtheeviloverlord @sundogplanets You _can_ do it, but the costs are (without irony, word choice intended) astronomical. You could liquid cool the inside of an orbital DC, but you'd still need a TON of surface area on external heat exchangers to passively radiate that heat away, since you can't rely on either convection or conduction—which are much much faster for heat transfer—but also aren't possible in a vacuum. A DC in space is basically the inside of a giant Yeti thermos, and that inside has an operational temperature range up to about 80C, but without cooling will easily get up to 140C or hotter—at which point stuff fails catastrophically; you have to move a lot of Joules to keep things operating in range.
Which is all to say that it's not _impossible_, but it IS nonetheless wildly impractical. But it is also calculable: DC math isn't hard, and manufacturers include power inputs and thermal outputs in their spec sheets—so omitting that seems very deliberate.
-
Cowboy Space (I hate their name so much) published their consolidated response a week ago but we just noticed it. EDIT: It was due at 5pm EDT today. So... I guess I have to read this shit. Swearing a lot here will get me through it.
A reminder: this is the 4th company that wants to launch fucking enormous "AI data center" satellites into sun-synchronous terminator orbits, giving Earth a ring. This particular company wants 20,000 satellites in one ring.
-
And then there's a whole section where they say that the FCC isn't allowed to consider environmental regulations, and they point to Reflect Orbital getting approved where the FCC specifically said they're not allowed to consider anything except radio broadcasts. This section might actually make me cry.
Next section is just... complaining, I think?
And then they end with saying that clearly they need to be approved.
Fuck you, FCC. Kessler Syndrome is on you.
@sundogplanets they've been perfecting their craft for a long time
-
@faithisleaping @sundogplanets
Yep. That's what I told a student of mine, who brought up the idea of these AI satellites.
I remember the time when NASA was paying silly money for original old Intel chips because they needed them for space applications that ran longer than anticipated.
BTW, you don't even need to go to space for that kind of problem: I once built a "supercomputer" with a load of RAM and you could measure solar activity from the ECC error logs
@tschenkel @faithisleaping @sundogplanets I've worked in automotive safety and we have a spreadsheet for likely error rate depending on height at the Earth's surface and the amount of memory on the MCU. I guess since these LLM processes spit out random garbage anyways the error rate doesn't matter?
