Comment by geertj
3 hours ago
The plan is to launch a constellation of smaller AI sats not a monolithic large data center. The calculations I have seen actually have a smaller radiator area than solar panels. Scott Manley’s has a video on this where he goes into some numbers.
Scott Manley's video uses 20kw as a reference number which isn't even half of the power usage of a modern GB200 rack. I.e. not even close to the power usage of an actual datacenter. In fact, not even 1% of the power usage of a datacenter...
Also, how is a constellation of satellites any easier in this case? They all need extremely large radiators, they all need maintenance, they all need high bandwidth communication.
If you calculate the actual cooling requirements for megawatts of server, you end up with needing many, many football fields of cooling.
It's nonsensical. Sure you can make the numbers sort of work for a single server, but a single server on earth costs MUCH MUCH less to launch, maintain, etc. So why bother doing it in space? We just end up with loads of unusable space servers as they gradually breakdown and cannot be repaired.
In fact, not even 1% of the power usage of a datacenter...
Right, but SpaceX has already filed plans with the FCC to launch a million of them, which is to say, 10K of your datacenter units. Tying back to the article, this plan is definitely going to require Starship and airline-like operations.
Yeah. I don't have any doubts that this is something that can be done. But doing it cheaply enough to be worth while is the difficult bit. Elon does have reputation for delivering impressive things, but not for finishing them on the deadlines he sets.
Ok, how many Starship launches have there been so far? Ok, 11 tests, of which only one has sent a dummy payload. Of the 11 tests, you can say that the latest one was closest to a final (sort of) working version. So we're still very early in the Starship launch program.
Let's do the math on "millions of datacenters" worth of launches.
In fact, let's try and do it for a single 50k GPU datacenter:
50,000 GPUs at GB200 density = 695 NVL72 racks at 1360 kg (1.36 tonnes) each, so the racks are roughly 950 tonnes.
GPUS = 950 tons
85MW of power needed for the GPUs. Latest solar power panels give roughly 120-150 watts per kg. Let's be generous and say 150 W/kg. So 85 MW / 150 W/kg = 570 tonnes of solar arrays
Power = 570 tons
Thermal management (radiators). Real space radiators are around 12 kg/m² for a heavy deployable radiator and its support structure, though ISS radiators are 8 kg per square meter, or 2.75 kg/m² if we only consider the exposed panels. (Using 8 kg/m² for an estimate). 200,000 m² × 8 kg/m² = 1,600 tonnes of radiators
Plus working fluid (ammonia or similar), pumps, manifolds, redundant loops: 150 tonnes.
Radiators Total: 1750 tons.
Structure, Propulsion, Comms, Avionics, Attitude Control Systems, plus Margin. Hard to estimate but conservatively several hundred tons extra. (Actual spacecraft programs always add roughly 20-30% mass margin).
Extras: 750 tons. (being very conservative).
Total = 950+570+1750+750 = 4020 tons.
And note, this is for a single 50k GPU datacenter with all the numbers being skewed to most optimistic.
That would be 40 (!!!) Starship launches. So far we've had 11 launches total with none being successful (100% successful I mean). Each of those launches currently costs 90M dollars. And note, we are assuming a fully working 100 ton payload for Starship of which none of the launches so far have been close to at all.
So our full datacenter to space would cost 3.6B dollars (at current SpaceX prices)... (just to launch it, not to actually buy the equipment). And realistically would cost far more than that...
Note, this is for (by today's standards) a small datacenter with only 50k GPUs and I haven't included any testing, R&D costs, costs of "maintenance", station keeping, replacements, etc etc.
Let alone the question of huge amounts more satellites in orbit, risks of space junk, Kessler syndrome, etc etc.
Later in the video he runs through the changes needed for 100kw per rack.
Wow, one 100kw rack!
Dude, you realize that right now there are 100+s of data centers in construction around the world often in the 500MW to 1PW range? I.e. there are many, many datacenters (100s) in construction, right now, with 100s of MW up to multiple PW?
Scott's analysis is out by several orders of magnitude!
Everyone knows its "theoretically" possible to have a single server or a single rack in space.
The big and most obvious glaring miss is, how would it be economically viable and operationally viable to have datacenters in space which compete with datacenters on the ground.
We are experts are creating economically profitable (very profitable) datacenters on the ground, which work really well for inference and training, which are cooled really well, can be maintained easily, etc. The idea that we are going to have 100s of MW clusters or 1PW clusters in space, and they are going to be competitive economically, and we are going to be able to maintain them, and they are going to actually WORK (i.e. how will the networking be competitive with datacenters on earth), is frankly laughable.
Scott is totally talking rubbish on this.