Jeff Bezos throws his hat in the ring for an orbital data center megaconstellation, too
- Thread starter JournalBot
- Start date
thelongdarkteatimeof
Ars Scholae Palatinae
So.. You're going to create an AI, that's in space, that we can't get to. So, when it becomes sentient, and tries to kill us all, we can't do anything because they are in space, Kessler syndrome has happened, and we can't launch anything. Great.
Upvote
0
(1
/
-1)
Stop being such a negative Kelvin! Everyone knows space is like super cold. Didn’t you see Gravity? George Clooney froze ASAP. If anything, I don’t know how they’ll keep the computers warm enough without burning any coal in space. /s
Upvote
7
(9
/
-2)
sigh. No you haven't left national laws behind. If you're a corporation you're registered in a nation, you launch from a nation, you need licenses to send signals and data to and from a nation, etc. Under the Outer Space Treaty, if you are a US company, the US government is responsible for your fuckups.
Upvote
4
(5
/
-1)
How about no to both of them? Thousands of heavy satellites raining down into the atmosphere and burning up, or not and hitting the ground, every year, doesn't sound great. It's already not great with starlink, and this will add several orders of magnitude more with heavier satellites.
Anyway, it doesn't make any sense except as a way to keep the bubble going by distracting people.
Anyway, it doesn't make any sense except as a way to keep the bubble going by distracting people.
Upvote
7
(7
/
0)
What do you mean, "we'll never know the details"?
Off the top of my head: Both Starcloud and Google recently published white papers purported to be in favor of orbital data centers.
I say, "purported" because neither are all that good. Both have serious issues with their citations not supporting what they claim they so, and both rely on extremely simplistic characterizations of what it is they hope to launch in the first place with unrealistically low cost estimates to boot. The Starcloud white paper reads like a high school assignment mostly composed the very day it was due
Given the long history of talking heads over-promising, under delivering and outright abandoning fantastic claims, is it so hard to accept the distinct possibility that orbital data centers may just be ideas that fall apart with sustained thinking? It wasn't too long ago that people were funnelling billions into NFTs and Metaverse.
Upvote
13
(13
/
0)
The only thing I've thought of is it hard to tamper with hardware in space once it has been put up there.
If this is why it is interesting to billionaires I've no idea
Upvote
0
(0
/
0)
I agree with your basic idea, but look at the millions of shell companies that are based in tax havens that compete to be the least restrictive financially. They will be even more willing to compete to have no protective regulations, since it doesn't even involve cutting their tax slice to the smallest possible fraction.
Upvote
0
(0
/
0)
No one really “tampers with hardware”. Hacking is the way to go for speed, deniability, safety, scalability and cost effectiveness.
Orbital data centers will actually have many more attacks possible than terrestrial because they will have much more extensive control systems that must be 100% remotely controlled. Hacking the thrusters would wipe out the entire data center in a single attack.
Upvote
7
(7
/
0)
Google? When did they enter the chat?
At any rate, you can assume that a public whitepaper represents their full internal analysis, if you like. It wouldn't be true at any place I've ever worked, but okay.
I think there's a distinct possibility that the ideas fall apart after engineering efforts fail to overcome the challenges involved (both known and unknown). A bit different that what you said, but close.
I'm not one to claim that the concept will be easy to implement, or be a sure-fire success either operationally or financially. My point was just that a Scott Manley analysis -- or any other third party analysis -- aren't the sorts of thing I'd weigh heavily.
Your point that large companies make bad decisions (i.e. Metaverse) is totally valid. Whether it's the case here or not, I can't say. I just know that SpaceX has a better-than-average track record when it comes to implementing concepts which others had either dismissed as unworkable or previously failed at.
Upvote
-4
(1
/
-5)
They are extremely determined to achieve a teraflop, just not in the way they imagined.
Upvote
6
(6
/
0)
3 MW? That's a really really small datacenter by today's standards. 300 MW would be closer.
Upvote
2
(2
/
0)
Upvote
1
(2
/
-1)
Ok. I didn't know how much a datacenter draws. But I guess 3MW is still a lot for a single installation consisting of just two single CPU computers plus supporting equipment.
Upvote
0
(0
/
0)
Actually you do, until you establish a self-sustaining colony and declare independence from Earth. Hmm, I wonder if any billionaires have had that thought?
Upvote
-1
(0
/
-1)
That's some of it. Another part of this later round of pitches is that once one of them has gone for a landgrab, they all have to. Including the smaller companies, because of course one thing they can do is sell themselves to bigger companies after the fact.
Upvote
2
(2
/
0)
So the alleged data center in space uses a really long USB cable to communicate with a single point on the ground? Or do smoke signals work in space?
Last edited:
Upvote
3
(3
/
0)
And they apparently think that denial also includes thermodynamics. Perpetual motion machines even work better in space!*
*But never seem match their name.
Upvote
1
(1
/
0)
But there is currently a “we ignore any laws that we don't like” regime in the US.
Upvote
6
(6
/
0)
Surface area of radiator needed to dissipate heat:
A = 1.3 * Q_dissipation / ( sigma x emissivity x T_rad^4)
For a 100W dissipation with a 30% Maximum Expected Value (MEV) for margin, a radiator temperature of T_rad=313K, assuming 0.87 emissivity , that requires 0.35 m^2 of surface area. Now, scale all that to data center rates, which I leave as an exercise to the reader.
A = 1.3 * Q_dissipation / ( sigma x emissivity x T_rad^4)
For a 100W dissipation with a 30% Maximum Expected Value (MEV) for margin, a radiator temperature of T_rad=313K, assuming 0.87 emissivity , that requires 0.35 m^2 of surface area. Now, scale all that to data center rates, which I leave as an exercise to the reader.
Upvote
0
(0
/
0)
I mean Ars published like 3 or 4 of these within a two-week span back when the Epstein class first started talking about data centers in space, and there was zero critique then as well.
Upvote
5
(6
/
-1)
So in addition to polar orbits to be always in sunlight, they also have the magical ability to refract their shadows to hit the earth? Are these satellites also being designed with LLM AI as well?
Upvote
0
(0
/
0)
Statistical
Ars Legatus Legionis
Sure but then terrestrial datacenters are still no different than orbital datacenters.
Upvote
3
(3
/
0)
Upvote
0
(0
/
0)
Guess Lex Luthor was feeling left out from all the grift gravy that's been flowing lately
Upvote
2
(2
/
0)
Now I see why Elon wants to move to space. That "pivot to (child) porn" Elon is doing.
Upvote
-3
(1
/
-4)
They only cross most of the other satellite orbits, especially the orthogonal equatorial ones. So obviously no problem, excepting for the for continual possibility of collisions, right? I mean there are only going to be 100K to a 1M or so, so they can afford to lose a hundred a day or so.
Upvote
-3
(0
/
-3)
The Nvidea GPU'S are not RAD Hard, they are using small geometry transistors. Their lifetime in space will be FAR shorter than being on the ground and inside the Van Allen Belts. Space rated semiconductors use decades old device geometries for good reasons. It is impractical to provide sufficient shielding to stop cosmic rays in space. For reasonable material thicknesses, one cosmic ray can send multiple impact particles into the silicon, causing even greater area damage. This is pie in the sky fantasy, all to support a stochastic, perverted parrot LLM AI. When the grifting isn't enough, add additional grifts to add additional delays and to get even more VC or governmental money.
Upvote
8
(9
/
-1)
They will be non-functional well before that time. They are not RAD Hard devices, and they cannot achieve that designation, without severely compromising their transistor count by moving to much larger devices.
Upvote
4
(4
/
0)
Perpetual motion machine validations should only be done by the perpetual motionists who build them. Sadly, intellectual property and company trade secrets prevent anyone else knowing how these devices work. Perhaps you have heard of Tesla FSD? /S (Just in case.)
Upvote
4
(5
/
-1)
Can we please, please, please, retire concerns about cooling in space? For all the problems the idea of datacenters on-orbit have, cooling is not one of them. Or, at least, the math actually closes. This is at least the third time I've walked through this exercise, but perhaps this time it will stick.
Let's stop and consider your 144 kW per rack of electrical power. If we're in sun-synchronous orbit at the terminus, we can expose our solar panels 100% of the time, but they're only going to be about 30% efficient at converting solar power into electricity.
For every square meter of solar panel pointed perfectly perpendicular to the sun, there's 1361 W/m^2 of solar power of which 400 W/m^2 will turn into useful electrical power. That means that for every 144 kW rack you describe above, you need to have 352 m^2 of solar panel to power it. That's a square 18.7 m on a side. Are you starting to see why the cooling issue closes?
The important thing to realize is that for all the massive real estate I'm going to have to allow for solar panels also provides a massive amount of real estate for radiators. And since the solar panels are always pointed at the sun, the back side is always pointed at empty space.
At this point, we have to make assumptions about surface temperatures, etc. but there are a few points to realize. If the radiators and the solar panels are the same temperature (given that the former is on the back side of the latter), then we have to consider that both sides of the solar panel / radiator are actually acting as radiators. Sure, in the direction of the sun the solar panels are certainly receiving a net amount of energy. However, for most of the hemisphere where a photon can try to emit, it's empty space too. If you're going to use heat pumps or the like to increase the radiator temperature, then you can mostly ignore the radiation from the solar panel side and have to consider just the area of the radiator (especially since you'll need more power to drive the heat pumps and therefore more solar panels).
So, for grins, let's consider a double-sided system (front and back) that has to reject (1361 W/m^2) / 2 . Use your favorite online blackbody calculator to determine the surface temperature of an ideal surface that can reject the amount of power that the sun puts on the panels. In the end, you'll find that you need 331 K or a 58 deg. C skin temperature. That's hot to touch, but certainly not too hot to run integrated chips or solar panels.
If we distribute your 144 kW of compute in that 352 m^2 area, there won't even need to be hot spots. If you're being smart and having a separable compute node that connects to cooling and power, you'll probably want a liquid loop of some sort to draw the heat away. You'll probably need 10 deg. C higher temperatures on your GPUs than the final 58 deg. C mark, but that's still very reasonable.
So please, feel free to crap on the idea of datacenters in space for any number of compute or financial reasons. But let's retire the idea that cooling in space is a technology bottleneck.
Last edited:
Upvote
0
(8
/
-8)
Yes, but their analysis includes factors like “if I can massage the assumptions to make this work, I can keep a well-paid job for as long as the boss’s money and enthusiasm last”, “the rocket I want to build will solve all your problems. Now what are your problems?”, and “my stock options have/are about to vest, so I win too if we juice the value”.
Upvote
4
(6
/
-2)
The pollution as all that e-waste burns in the upper atmosphere might be a problem.
Upvote
4
(4
/
0)
This gets thrown around a lot, but is misleading.
Depreciation is an accounting practice. 5 years is standard for capital equipment (although it depends on the type of equipment, 3 or 7 are also common). But this is based on tax rules and generally accepted accounting practice (GAAP) rules. Companies mostly want to depreciate as fast as allowed because it lets them defer taxes.
But that doesn't mean the equipment becomes useless after 5 years, it just means that it has reached it's minimum book value. Some equipment does indeed get scrapped after that period even if it's still working, but plenty keeps running for years after that generating essentially pure profit. And in fact right now companies are extending service lives of GPUs and AI servers even those they planned to decomission exactly because demand is so high and RAM shortages so acute.
Upvote
3
(4
/
-1)
You aren't missing anything. A data center in space is in all ways worse than one on the ground. And it can be pretty much anywhere on the ground.
Yep.
Upvote
4
(4
/
0)
This thorough, factual comment is very welcome. Contrast with people here complaining about the insane space data center gold rush while at the same time skim-reading the content and comments they're replying to in their own desperate rush to get their comments onto the first page.
I see your comment got more downvotes than upvotes, but no counterarguments. Therefore all the negative vote count indicates is the degree to which the downvoters lost the argument.
Upvote
-4
(5
/
-9)