SpaceX Starship Could Make Orbital AI Clusters Economically Viable

Oh and ref for that 1.4M figure:

I could have just called him a subhuman and be done with it, but since some people criticize my… didactic methods, I decided to try to be patient. Alas… it seems my detector is nearly flawless.

predictable, really Western commies cannot into space subhumans generally cannot into space

@algern0n79 This is the level of disconnect from reality a lot of people are operating at, sadly. In many areas of life😬

@BellikOzan And SpaceX will make their own chips. At something like 20% the cost of everyone else’s.

@BellikOzan 145x145mm tiles, 32 elements butler matrix (when the ITASAT flown) Amazonas-1 used a 3 by 3 arrangement of those tiles. T/R module for each of 4 zones for each tile. Amazonas-1 used a diffrent way to handle the beamforming network and DSP(and got lower sWap-c, huh)

@BellikOzan Economics doesn't really matter. The problem has always been that earth orbit is an incredibly unsafe location to put extremely valuable infrastructure that could work just as well on the ground. It's like building the datacenters in the Korean DMZ to save some money on power

@BellikOzan @marcrygamerbr Don’t you understand? SpaceX had it wrong this whole time. They could have been using super cheap PESA’s of arbitrarily small size and just told them to pretend to be a big, high throughput AESA.

If only they’d known antenna boy.

@BellikOzan Even ignoring Tesla which is run by the same guy they don't trust... at least Intel, Google and Anthropic are all gonna play some role. Maybe more allies incoming.

The idea that nobody in those companies can check the numbers or model future scenarios is surreal, almost funny 🙂

@BellikOzan Burning 8 GWh worth of methane to put the sats up there makes the ROI longer than the useful life of the hardware. This is a vanity/marketing project to pump the IPO. Like the Mars colony, there will be delays, then eventually the project will be scrapped, or delayed indefinitely

@BellikOzan A lot lower risk to throw a few debris in the datacenters orbit, just as a demonstration to kill a few sats than targeting the continental US. It's not even close

@Marc_Topaz I think power expenditure on compute will tend to grow indefinitely (w/ the possibility of short term dips), not shrink or remain flat. Jevons paradox...

@marcrygamerbr Cool, thank you, so under 2% of the area of a Starlink v2 mini antenna.

I hope you can see where I'm going with this...

@BellikOzan @marcrygamerbr They didn't say, but it would be logical to assume that what they supply for the satellite antenna is a mirror image of the terminal on the ground (with larger size and higher power), which I think includes the phase shifters for beam steering, maybe @olegkutkov can weigh in?

@marcrygamerbr Afaik it's 8 antennas with a total area of >11m2 and a few thousand elements apiece, handling complex multi-channel high bandwidth I/O. The control hardware isn't dirt cheap and scales superlinearly with antenna size, as do stiffness requirements.

The math is solid. In fact getting rid of the majority of thr infrastructure surrounding the processor nodes is highly beneficial.

However Elon Jaz already hedged something I've been saying. Eventually more dense power and cooling may outperform the individual sats. Elon Jaz said they will look into more dense nodes, Starcloud like, later.

The ability to select you 2GW datacentre, gpu mix, net and simply have it launched and self assemble into Starlink is hard to overstate the benefits.

@EmperorBag Not at all. The methane cost is less than 1% of the figures discussed here.

@BellikOzan @marcrygamerbr @olegkutkov SpaceX is doing that, they buy the components from ST which are maybe 2% of the total V3 satellite cost

@marcrygamerbr I can't tell if you're joking or have no sense of the implicit scale and precision.

@marcrygamerbr Sounds like the latter. Tell me, how many elements in that antenna and how big is it?