crdrost
5 days ago
So like I don't get to use my degree except to tutor kids and answer physics questions and the like, so indulge me for a second because this doesn't make much sense to me. So like the endgame is to create a new Pryor, OK facility in space, right? It doesn't make sense if we're not shipping something that's about the same as what you can build on Earth.
Epoch.AI says that this facility has 100,000 TPU chips, eats 370 megawatts of electricity, but designed to get rid of up to 480 megawatts of heat. Sun power is 1.36 kW/m^2, which is something like 7 megawatts per football field (assuming ~5000 m^2), but you have to put up something like 3-5x that because solar panels are only 20-30% efficient, right? Assume that you can arrange it so that the solar panels are one side, the radiators on the other, you can maybe get away with only 300-400 football fields maybe? So like 1.5 million m^2 or 1.2 km wide, right? At LEO, 650 km, you have (1.2km)/(650 km) * 180 degrees/pi is 0.11 degrees or 6.6 arcminutes of visual size. The moon is only like 30 arcminutes. And the paper talks about how they're going to not do one big monolithic construction but an oval of fridge-sized objects separated out 2-3 times this size -- so like I don't see how you get another Pryor, OK size data center in LEO without basically having it look like a second, smaller moon flying across the sky 10 times per day.
And if this were a wildly successful idea are we talking about having like 5-10 of them, a few for each big frontier lab? This just sounds like we're talking about the most profound shift to our night skies since we started having to deal with light pollution in our cities.
ben_w
5 days ago
There's a lot of different designs besides this one.
So far as I can work out, literally all of the plans are bad. The "why" varies, but they're all bad.
I'm too tired to double check your maths, so I will assume correct: one likely difference even for this plan is a terminator following sun-synchronous orbit, which means you'll only see it twice a day despite the orbital period being about 90-100 minutes, and when you see it will be specifically at sunrise and sunset.
Visibility is also a question of reflection, not just size. Terminator following orbits are worse than normal satellites, because one of the tricks for reducing e.g. Starlink visibility is to tilt them as they cross the terminator and you can't do that if they're always on the terminator.
The SpaceX plans (a million small ones) becomes a glitter band in some parts of the sky and will appear visually contiguous in other parts, though I need to double check my maths and assumptions about visibility given this happens during sunrise and sunset so the sky itself is pretty bright.
hedora
5 days ago
I think there are at least a few problems with the math’s assumptions:
1) Don’t the radiators need to have more surface area than the solar? (Unless the chips run very hot.)
2) The datacenter will have the pesky earth between it and the sun some fraction of the time, and have to either shut down or run off batteries. ~50%, assuming LEO, right? If you leave LEO, then the latency sucks, so they’re training-only clusters. At 50% the solar doubles and you need 370 megawatt hours per hour of darkness, or you run the machines 50% of the time, rebooting for each orbit. If you make the orbit shorter (so you can have smaller batteries), then they wear out faster. The batteries also emit heat. Plus, you need to double the solar so they charge while the workload is running.
3) How do they cope with cosmic rays? The standard approach is still to duplicate or triplicate all computation, or use larger/slower processes, right?
The obvious answer to each question makes the engineering design at least twice as dumb, and they stack. There are many other problems like these.
sobellian
5 days ago
The radiators are what need to be very hot. IIRC it works out to, order of magnitude, a similar area to the solar panels. These satellites would likely be placed in SSO, a polar orbit that precesses at the same rate that the Earth orbits the sun. So the fraction of time they see the sun would be very high. Cosmic rays are a little bit of an unknown, we don't know exactly how a commercial GPU running an LLM deals with cosmic ray radiation. I think the intuition is that it would be very little because a single bit flip in that vast array of matrix math and many times its result might not actually change any activations anyway.
I think the big problem with the idea is that GPUs have a failure rate and even if they didn't, they become obsolete. Most orbital DC plans are really more like flying server racks with no servicing in orbit. So when the GPUs die the satellite is a flying brick. All the power equipment, all the thermal equipment, all the comms equipment now depreciates at the same rate as GPUs. Very different economics from terrestrial DCs. And that's all assuming that you can launch everything up there quite efficiently. And the satellites take time to engineer but DCs are a more known quantity.
trhway
5 days ago
>1) Don’t the radiators need to have more surface area than the solar? (Unless the chips run very hot.)
to generate 1KW you need 5m2 of solar panels. And black body radiation of 1.5m2 at 70 C is 1KW.
A unit with 1 GPU, 2m2 radiator and 5m2 solar panels is say 20kg. At promised Starship price well under $100/kg, that is less than $2000 to put that unit into orbit. That is much cheaper than $15000 per 1KW of a ground-based datacenter, especially when additionally factoring in [expensive and climate change causing] ground-based electricity vs. free electricity in space once you launched the unit with its solar panels.
Add political opposition on Earth, Iranian drones hitting datacenters, various laws (i.e. costs) that your ground-based datacenter is subject too ... the space starts to look like a very cozy place for a datacenter :)
eutropia
4 days ago
at current prices it's above $1000/kg to orbit, and we're just conveniently ignoring all of the operational constraints of a completely unmanned datacenter upon which you can perform no maintenance.
What do you do when a micrometeorite causes pinprick leaks in your radiator loop?
is LEO sufficiently shielded from cosmic and solar radiation?
trhway
4 days ago
nobody is building real datacenters in space at $1000/kg - at that price the cost is comparable to ground-based and thus no big point. The explosion will start at $100/kg when as i mentioned the base cost would beat ground-based almost 10x, and thus would allow for the additional costs (with total cost still beating ground-based) of all the additional issues/concern you and the others usually mention in this context.
4 years ago we had the ChatGPT moment and notice how civilizational change has been accelerating since then. In a few years we're gong to have the seemingly profound Starship moment. Back at the time it was obvious, i may be said it is even here, that Starship low price will cause explosion of amount of launched payloads and a technological revolution as a result, yet it wasn't clear what kind of payloads it would really be. And now we have the AI - Starship and AI are basically ideally dove-tail each other. Two revolutionary technologies accelerating each other - Strarship lowering deployment cost and removing other obstacles for AI while AI providing guaranteed massive scale launch market for Starship - that will be a thing to watch (or participate if you're lucky :)
ben_w
2 days ago
As per another of my comments, Alphabet, applying learning curves arbitrarily far into the future, recon it will take SpaceX launching 370,000 tons to LEO to make the costs come down enough to be worth it: https://arxiv.org/pdf/2511.19468
That's not even $100/kg, that's $200/kg.
Even my bull case puts that ($200/kg) 10 years off, which is so far away it lacks relevance just because compute and AI models move so much faster than that timescale; my bear case says that's about 45 years off.
Always account for how over-optimistic Musk's public timelines are. Starship has only just a few days ago managed its first circular orbit, but when he first talked about it:
The basic game plan is like we're going to send a mission to Mars with every Mars opportunity from 2018 onwards. So and they occur approximately every 26 months. So you know, we're establishing cargo flights to Mars that people can count on for cargo.
trhway
2 days ago
>370,000 tons to LEO
SpaceX is building spaceport with 10 pads or something like this. At 1 Starship/pad per day that 1000 tons/day to LEO - i.e. 1 year.
The Starship is making great progress. Even if it takes 10 years to once-a-day reusability, the need will be even greater by then.
> compute and AI models move so much faster than that timescale
tech moves faster with each iteration, and datacenters and networks only grow with it. The ground-based datacenters has and will continue to grow in cost. While space-based costs will be cheaper and cheaper. Not the first time we have 2 such curves, and such curves dynamics usually leads to paradigm shift.
ben_w
a day ago
> SpaceX is building spaceport with 10 pads or something like this. At 1 Starship/pad per day that 1000 tons/day to LEO - i.e. 1 year.
When you making forecasts that assume success, they're generally over-optimistic.
So, here, if they launch 1 Starship per pad per day from 10 pads. And also if they double the current capacity per launch because the demonstrated launch capacity right now has "only" been 44 tons. (44 would be impressive! It's just that Musk's pronouncements are nowhere near what gets delivered).
> The Starship is making great progress. Even if it takes 10 years to once-a-day reusability, the need will be even greater by then.
Reality check: we're 14 launches in and development is so slow that we're more than 10 years past when he said he'd be sending things to Mars within two years and it's only just now managed to get a circular orbit.
Getting there from here requires funding at the intermediate cost, not the final cost. That means they have to spend order-of 300-400 billion dollars, which means they have to raise that many billion dollars in capital, which is around 4x times the 86 billion they actually raised in their IPO (not trillions: market cap is what you get if you could sell all the shares, and they didn't do that).
This is, inflation-adjusted, around 1.5 times the Apollo mission. Since this whole thing began, the development process has been slower and more expensive than the actual Saturn V launch program.
Right now, SpaceX needs more financing rounds to cover their existing spending commitments. Some of this has nothing to do with going to space, thanks to them having merged with xAI and the AI spending spree.
> tech moves faster with each iteration, and datacenters and networks only grow with it. The ground-based datacenters has and will continue to grow in cost. While space-based costs will be cheaper and cheaper. Not the first time we have 2 such curves, and such curves dynamics usually leads to paradigm shift.
False. Any argument that "datacenters and networks only grow with it" applies regardless of where those datacenters are placed, ground or space.
They get more expensive in proportion to how much compute you have, and less in proportion to the tech of the processors themselves; but place is irrelevant to the compute tech, price of location only shifts with regard to the transport.
Why this matters:
Applying Koomey's law for 10 years makes any give unit of compute 14 times more energy efficient, so what was once a 1kW data centre GPU becomes a 70 watt laptop integrated graphics processor, and what was once limited to the 70 watt power envelope of a laptop becomes the default performance of a bargain basement phone.
Also, the models seem to still be getting more efficient at fixed parameter count, so model performance which can be had on your phone today would have needed hundreds of gigabytes and a beefy GPU just a few years back.
Either by themselves, let alone both, means it's quite plausible that all the data centre investment, both on the ground and in space, is worthless within a decade.
jeezfrk
4 days ago
Upgrading old and obsolete GPUs?
This is all wildly short term mania.
trhway
4 days ago
sysadmin/hardware jobs would get s3xy again :)
ben_w
4 days ago
1) Not even if the PV was 100% efficient, which is impossibly good.
Imagine this side on, as a T-shape: the top bar is the PV, sunlight coming downwards, and because of that only one side is illuminated. The radiator is the vertical bar, in shadow, but crucially it's two-sided, so 1m^2 of material is 2m^2 of surface. If they're the same area of material, thanks to that 2:1 advantage the radiator has naturally, this would only need to run at 58°C: https://www.wolframalpha.com/input?i=%281361+watt+%2F+%281*σ...
(Adjust as you prefer for power, area, emissivity is 1 here which isn't possible either but realistic radiators are more like 0.8-0.95 and even 0.8 only raises the above to 77°C)
One of the bigger problems for radiator size is putting them in LEO. Earth is warm and a big fraction of the sky at that altitude. Less of a problem as you get hotter, because radiated power is proportional to the fourth power of temperature in Kelvin, but the closer you operate to Earth's temperature the worse it is.
However, one of the things I am trying to sort out for the blog post is what happens exactly, as a function of temperature, if you just put this all on the ground, because down here you have convection as well as radiation; I'm sure I've seen someone do this and their conclusion was that any radiator good enough to work in space will actually work better on the ground for realistic operating temperatures.
If that rings a bell for anyone else, and you can remember a link to who already worked that out, please let me know :)
2) Not necessarily, but different people pursuing this have different plans: that's why I said the thing about terminator-following sun synchronous orbit, it's a special class of orbits that are always on the edge between day and night at all times.
If you don't do use that class of orbit, then yes, you lose half the supposed benefit on the power front.
3) The current plans seem to be "wing it". Given how resilient LLMs are to noise, this might even work.
lumost
4 days ago
Starship’s economics will make such projects inevitable, if it’s not data centers it will be something else. Ideally, the solar panels will be efficient enough to minimize albedo.
A facility in space will likely have better uptime then a terrestrial facility with lower costs for disaster mitigation/insurance. The economics are surprisingly close when you do the back of the napkin kg math, but you can mass-produce orbital data centers in a way that’s not possible for terrestrial facilities.
Granted, the same pitch could have been made for orbital telecommunications vs terrestrial wires… that bet hasn’t paid off yet.
sjs382
4 days ago
> lower costs for disaster mitigation/insurance
Say more about that
Keyframe
5 days ago
To take it to more absurd levels, further out and spread out we could create a mesh like structure and throw a bit (a tiny bit) of a shade on the planet and fight global warming.
trhway
5 days ago
>further out and spread out we could create
a formation in the shape of the client's logo.
justinclift
4 days ago
Re-arrangeable every week for the new weekly sponsor!
Maybe even a new "Shade as a Service" business model. ;)
devindotcom
5 days ago
you're correct. this video was a pretty good visual exploration of some of it: https://www.youtube.com/watch?v=_qpdUNMt2yg
my personal understanding is that in-orbit compute is perfectly practical up to some obvious limits like the ones you describe. a few reasonably sized clusters up there (tens of kilowatts) doing high priority processing jobs paid by the flop is a great idea. localized compute on existing satellites already does some of this but some earth observation company being able to rapidly scale up image processing for an hour is a great option to have. the really big stuff is just a fantasy.
user
5 days ago
MisterKent
5 days ago
Is the point to just capture sun energy off planet and solve global warming? /conspiracy
aaron695
5 days ago
[dead]