Elon Musk quote: SpaceX will lean in big on the Moon

[InterestedEngineer]

FYI there is already a "built the data center in space" thread

https://forum.nasaspaceflight.com/index.php?topic=63177.msg2731431#new

[Greg Hullender]

Starcloud wrote a white paper about this, they think the business case closes if launch cost is low enough (i.e. what Starship can achieve): https://starcloudinc.github.io/wp.pdf

Okay, I read through the paper, plus some of the sources it cited, particularly AI Datacenter Energy Dilemma - Race for AI Datacenter Space, which I really rewards a close read.

The most salient single data point is that AI data center costs are not dominated by power requirements. They're dominated by the capital cost of the compute hardware. GPU Cloud Economics Explained – The Hidden Truth. That kind of takes the wind out of the sails of any project whose main goal is to reduce the cost of power. They are on the right track in focusing on training, since apparently 90% of current power use is on training--not inference.

But back to the original paper, it talks about a system with 16 square kilometers of solar panels, requiring 200 Starship launches with costs (to the customer) of about $10/kg. That's pretty ambitious! And the system they describe includes lots of moving parts, like heat pumps, and two-phase (i.e. liquid and gas) cooling systems. Assembly and maintenance issues are waved away with "robots will do it." Also, the system they describe is a little too large; components inside a data center need to be no more than 50m apart, but the Starcloud system is 100 to 200m across.

Anyway, Starcloud is a company that's at least trying to make this work, but I think their assumptions about the problem they're trying to solve are too pessimistic and their assumptions about the problems they have to solve are too optimistic. At least for the next decade or two.

[xvel]

Only reason for high cost of the compute hardware is nvidia and TSMC monopoly

[DanClemmensen]

Great paper!  This image captures the bandwidth problem.

Current state of the art is about 1.6Tbps east-west traffic (between racks).  If you consider a satellite a rack (say 100kw of compute) each satellite would need 1.6Tbps of optical links between satellitles.  Starlink is currently a mesh of 3ea 100Gpbs links.  So that's off by an order of magnitude.

However the paper shows how to solve this problem.

It may not be completely obvious, but those numbers say that three laser links consuming 5 Watts can provide more than enough bandwidth to support 100KW of computing. That is, the comms consume less than 0.1% of the total power.  This is all using itty-bitty 10 cm telescopes, one per beam, so they also use less than 0.1% of the mass.

Starlink is "off" because its satellites are a lot further apart than the paper's reference numbers, and because the current Starlink satellites do not need more than that amount of bandwidth. The paper identifies three separate ways to increase the bandwidth to easily achieve higher bandwidth through the same 10 cm telescopes, and they don't even mention the brute force approach of dramatically increasing the number of DWDM channels at the same power per channel.

[KF3K]

I think a big part of the push for moving data centers to orbit is extreme pessimism about the regulatory prospects of building datacenters and power sources for them on-planet.  If the NIMBYs convince regulators to say "you can't build that on Earth," it almost doesn't matter how much it costs to build in space, if that's your only remaining option.

I don't think we're realistically at the point of banning datacenter construction (at least in the U.S.—Europe is another story), but if I had billions of dollars at stake, I'd probably want to be working on a "plan B" for if that happens.

Datacenters are relatively benign compared to other industry, but they're an easy target for people worried about the clankers taking their jobs, or those preaching economic suicide as the only solution to climate change.  And producing 1s and 0s in space is a great early adoption case for moving industry off-planet, since the output can be returned down for use on Earth relatively cheaply versus manufactured goods.

[meekGee]

Starcloud wrote a white paper about this, they think the business case closes if launch cost is low enough (i.e. what Starship can achieve): https://starcloudinc.github.io/wp.pdf

Okay, I read through the paper, plus some of the sources it cited, particularly AI Datacenter Energy Dilemma - Race for AI Datacenter Space, which I really rewards a close read.

The most salient single data point is that AI data center costs are not dominated by power requirements. They're dominated by the capital cost of the compute hardware. GPU Cloud Economics Explained – The Hidden Truth. That kind of takes the wind out of the sails of any project whose main goal is to reduce the cost of power. They are on the right track in focusing on training, since apparently 90% of current power use is on training--not inference.

But back to the original paper, it talks about a system with 16 square kilometers of solar panels, requiring 200 Starship launches with costs (to the customer) of about $10/kg. That's pretty ambitious! And the system they describe includes lots of moving parts, like heat pumps, and two-phase (i.e. liquid and gas) cooling systems. Assembly and maintenance issues are waved away with "robots will do it." Also, the system they describe is a little too large; components inside a data center need to be no more than 50m apart, but the Starcloud system is 100 to 200m across.

Anyway, Starcloud is a company that's at least trying to make this work, but I think their assumptions about the problem they're trying to solve are too pessimistic and their assumptions about the problems they have to solve are too optimistic. At least for the next decade or two.

You're reading, but you're not listening.

You say: "That kind of takes the wind out of the sails of any project whose main goal is to reduce the cost of power" - thus missing the entire point.

Exactly because power is not the main investment, even if its final cost (after improvement in PV performance but also the need for radiators and the cost of launching it) ends up being higher.  (Which it may or may not - it doesn't matter)

The reason for going to orbital PV is that nobody else has a path for installing 100 GWatt per yr. In the 10-year term.  Not terrestrial PV, not nuclear, not the US, not China.

[Greg Hullender]

The reason for going to orbital PV is that nobody else has a path for installing 100 GWatt per yr. In the 10-year term.  Not terrestrial PV, not nuclear, not the US, not China.

Nor in space either. Not soon, anyway. If a thing is impossible, then it won't happen.

However, the SemiAnalysis estimate sees 20 GW/year by 2028, not 100. And that's assuming there's no AI bust coming. To justify the type of investment you're talking about, AI would have to generate more value than that. At the moment, though, it's struggling to justify the money invested so far.

I understand that everyone's excited by the idea that there's some industry that really would justify industrializing space. I'd like to believe that too. I just don't believe this is it, and I don't think any sober look at the data supports it.

[meekGee]

The reason for going to orbital PV is that nobody else has a path for installing 100 GWatt per yr. In the 10-year term.  Not terrestrial PV, not nuclear, not the US, not China.

Nor in space either. Not soon, anyway. If a thing is impossible, then it won't happen.

However, the SemiAnalysis estimate sees 20 GW/year by 2028, not 100. And that's assuming there's no AI bust coming. To justify the type of investment you're talking about, AI would have to generate more value than that. At the moment, though, it's struggling to justify the money invested so far.

I understand that everyone's excited by the idea that there's some industry that really would justify industrializing space. I'd like to believe that too. I just don't believe this is it, and I don't think any sober look at the data supports it.

Impossible how?

I mean, I get it's ambitious, but how is it impossible?

(Not the relative pricing argument, the deployment rate.)

[sferrin]

Starship could deliver 100GW/year to high Earth orbit within 4 to 5 years if we can solve the other parts of the equation.

100TW/year is possible from a lunar base producing solar-powered AI satellites locally and accelerating them to escape velocity with a mass driver.

That was the plan for funding space colonies in SP-413 as I recall.  Only they would have been shooting ore from the moon to L5 and processing it there.

[Coastal Ron]

Starship could deliver 100GW/year to high Earth orbit within 4 to 5 years if we can solve the other parts of the equation.

100TW/year is possible from a lunar base producing solar-powered AI satellites locally and accelerating them to escape velocity with a mass driver.

That was the plan for funding space colonies in SP-413 as I recall.  Only they would have been shooting ore from the moon to L5 and processing it there.

I thought there was an NSF thread or conversation that was looking at mass drivers on the Moon, and they determined (or came to the conclusion) that in reality they really don't work as well as we thought.

For instance, whatever mass you launch is either going to be semi-orbital (which means you have to have a catcher in orbit that can snatch the payload before it falls back to the Moon) or it leaves the orbit of the Moon, in which case you have to go chase it somehow, but it is difficult to have a "catcher" prepositioned (something to do with unstable Moon orbits or??).

In any case, I think the burden of proof is on those that think we can "manufacture" anything on the Moon, much less something as high tech as solar cells in large volume (as well as everything it takes to install those panels so they can be useful), at any point in near future. I think it is more likely we could have a small Earth-supplied colony on Mars well before we could figure out solar cell manufacturing on the Moon.

My $0.02

[TrevorMonty]

Starship could deliver 100GW/year to high Earth orbit within 4 to 5 years if we can solve the other parts of the equation.

100TW/year is possible from a lunar base producing solar-powered AI satellites locally and accelerating them to escape velocity with a mass driver.

That was the plan for funding space colonies in SP-413 as I recall.  Only they would have been shooting ore from the moon to L5 and processing it there.

I thought there was an NSF thread or conversation that was looking at mass drivers on the Moon, and they determined (or came to the conclusion) that in reality they really don't work as well as we thought.

For instance, whatever mass you launch is either going to be semi-orbital (which means you have to have a catcher in orbit that can snatch the payload before it falls back to the Moon) or it leaves the orbit of the Moon, in which case you have to go chase it somehow, but it is difficult to have a "catcher" prepositioned (something to do with unstable Moon orbits or??).

In any case, I think the burden of proof is on those that think we can "manufacture" anything on the Moon, much less something as high tech as solar cells in large volume (as well as everything it takes to install those panels so they can be useful), at any point in near future. I think it is more likely we could have a small Earth-supplied colony on Mars well before we could figure out solar cell manufacturing on the Moon.

My $0.02

Blue have already demostrated they can manufacture a solar cell from regolith (artificial). Extracted silicon by electrolysis and then produced cell from it.

Spinlaunch has already produced operational mass driver capable of 5000mph. Lunar escape velocity is 5325mph.

[sferrin]

Spinlaunch has already produced operational mass driver capable of 5000mph. Lunar escape velocity is 5325mph.

You definitely don't want to have to deal with the constraints Spinlaunch imposes.    An EM linear accelerator would be the way to go.

[Vultur]

Spinlaunch has already produced operational mass driver capable of 5000mph. Lunar escape velocity is 5325mph.

You definitely don't want to have to deal with the constraints Spinlaunch imposes.    An EM linear accelerator would be the way to go.

IIRC, EM linear accelerators at this velocity are far harder than used to be thought back when O'Neill and others were writing; the Navy's railgun projects had huge troubles with rail erosion. This isn't my field, and it may be solvable, but I'm not at all sure a lunar escape velocity EM accelerator *with a long enough operational lifetime to make economic sense* would be within current technological capability - building it almost entirely with robotic labor sounds impossible.

[Vultur]

Starcloud wrote a white paper about this, they think the business case closes if launch cost is low enough (i.e. what Starship can achieve): https://starcloudinc.github.io/wp.pdf

Okay, I read through the paper, plus some of the sources it cited, particularly AI Datacenter Energy Dilemma - Race for AI Datacenter Space, which I really rewards a close read.

The most salient single data point is that AI data center costs are not dominated by power requirements. They're dominated by the capital cost of the compute hardware. GPU Cloud Economics Explained – The Hidden Truth. That kind of takes the wind out of the sails of any project whose main goal is to reduce the cost of power. They are on the right track in focusing on training, since apparently 90% of current power use is on training--not inference.

But back to the original paper, it talks about a system with 16 square kilometers of solar panels, requiring 200 Starship launches with costs (to the customer) of about $10/kg. That's pretty ambitious!

That does seem problematic. I think Starship will drastically reduce launch cost, but $10/kg is a stretch. That might be doable for a really mature Starship-type system (as mature as cars or airliners are today), where fuel cost is a significant part of the lifetime cost, but not anytime soon.

F9 fully loaded is maybe $3000/kg to low LEO at sale price, maybe $1000-$1500/kg for SpaceX internal?

$10/kg requires $2M/launch for a fully loaded 200,000kg Starship v4. That's 3-4x cheaper per launch than Electron (and Starcloud won't get SpaceX internal pricing, like Starlink does)..
.

I think a big part of the push for moving data centers to orbit is extreme pessimism about the regulatory prospects of building datacenters and power sources for them on-planet.  If the NIMBYs convince regulators to say "you can't build that on Earth," it almost doesn't matter how much it costs to build in space, if that's your only remaining option.

I agree.

Datacenters are relatively benign compared to other industry,

I don't think that's true anymore when you are talking about the truly ridiculous power levels some are now discussing. It may still be true "per megawatt" - but from 2005 to about today, US power use was essentially flat, with efficiency improvements keeping pace with population growth. Building lots of new power has significant environmental costs, as right now a lot of it is natural gas not solar, in the US. There's a lot of industries that are way messier "per megawatt" but none of them are explosively expanding the way datacenters are.

[StraumliBlight]

IIRC, EM lineR accelerators at this velocity are harder than used to be thought; the Navy's railgun projects had huge troubles with rail erosion. This isn't my field, and it may be solvable, but I'm not sure a lunar escape velocity EM accelerator would be an easy thing - building it almost entirely with robotic labor sounds impossible.

Auriga Space is currently building a Mach 6 rail gun. Also General Atomics has continued development, as well as countries like Japan, China, etc.

[Twark_Main]

There are different projections on the future of AI, but this whole idea is based on the premise that "thinking interfaces" will be everywhere.  Every happy door, every smart elevator, every empath coffee maker.

Who's asking for that? This is literally a satirical bit from Douglas Adams.

"Here I am, brain the size of a planet, and they ask me to take you up to the bridge. Call that job satisfaction? 'Cause I don't."

Or:

"All the doors in this spacecraft have a cheerful and sunny disposition. It is their pleasure to open for you and their satisfaction to close again with the knowledge of a job well done."

...

Honestly sounds like an awful future. "You know that smart device you didn't want to buy but were forced to anyway? Well now you need to maintain a complex emotional relationship with it, too." 

Yes, I didn't just stumble on the examples just randomly, what would be the odds of that....  I figured most people here would get the reference without explicitly calling it out.

(Yes, that was a more advanced reference.  I can do this all day, and it's only 8 am)

It's how you used it. It didn't occur to me that someone would know the reference and so thoroughly miss the message. 

If you realize it's a satirical criticism of useless and unwanted AI, maybe it doesn't make sense to use it as backbone of your argument in favor of scaling AI?


Relevant meme attached.

[TrevorMonty]

Spinlaunch has already produced operational mass driver capable of 5000mph. Lunar escape velocity is 5325mph.

You definitely don't want to have to deal with the constraints Spinlaunch imposes.    An EM linear accelerator would be the way to go.

What constraints?. Bulk water and metal shipments don't care about high Gs involved. Delicate items like humans will use landers.
Best to run launchers in pairs, energy recovered while slowing one down to reload after a launch can be used to accelerate another to launch RPMs.

[Twark_Main]

Spinlaunch has already produced operational mass driver capable of 5000mph. Lunar escape velocity is 5325mph.

You definitely don't want to have to deal with the constraints Spinlaunch imposes.    An EM linear accelerator would be the way to go.

What constraints?. Bulk water and metal shipments don't care about high Gs involved. Delicate items like humans will use landers.
Best to run launchers in pairs, energy recovered while slowing one down to reload after a launch can be used to accelerate another to launch RPMs.

Is Spinlaunch still dropping the counterweight? AFAIK most of the energy will be lost with the counterweight.

[Greg Hullender]

That does seem problematic. I think Starship will drastically reduce launch cost, but $10/kg is a stretch. That might be doable for a really mature Starship-type system (as mature as cars or airliners are today), where fuel cost is a significant part of the lifetime cost, but not anytime soon.

So much depends on this assumption that SpaceX will offer launch services at cost (and at the low-end of what even Elon has estimated for those costs) that it's worth harping on this a little. SpaceX will not offer customers prices that are much lower than what its competitors charge! This is elementary economics. If competitors are charging $1000/kg, then SpaceX does better to charge $999/kg even if it only costs them $10/kg. Then they pocket the huge margins and apply that money towards their Mars project. They'd be stupid not to, and although Elon may be a fool, he's not that kind of fool!

The only exception would be if lowering the prices increased demand enough to make up for it. E.g. if at $100/kg you got more than a factor of ten increase in business. Even if it did, a sudden price drop would not instantly increase business that much. That kind of increase would take years to happen. Never mind the factor of 100 to make it worth dropping the price to $10/kg even if the launches cost SpaceX nothing!

Anyway, the question, then, is when will SpaceX's competitors be charging under $100/kg for mass to orbit? Before that happens, this project is clearly economically nonviable. I think that might happen between 2030 and 2040. Maybe.

[StraumliBlight]

What constraints?. Bulk water and metal shipments don't care about high Gs involved. Delicate items like humans will use landers.
Best to run launchers in pairs, energy recovered while slowing one down to reload after a launch can be used to accelerate another to launch RPMs.

There was a cool concept in Kim Stanley Robinson's Red Moon, where a Lunar launch rail can also decelerate and catch spaceships.

IEEE Spectrum: You invented a completely new technology for landing on the moon. It seems to combine a maglev train, a railgun, and a hyperloop. Can you briefly describe how that works and how you came up with it?

Kim Stanley Robinson: I got the idea from a lunatic friend of mine. It's basically the reverse of the magnetic launch rails that have been postulated for getting off the moon ever since the 1930s: These take advantage of the moon's light gravity and its lack of atmosphere, which allow a spaceship to be accelerated to a very high speed while still on the surface, after which the ship could just zoom off the moon going sideways, because there is no atmosphere to burn up in on the way out. If you just reverse that process, apparently you can land a spaceship on the moon according to the same principle.

It blew my mind. I asked about the tolerance for error; how precise would you have to be for the system to work? My friend shrugged and said it would be a few centimeters. This while going about 8,000 miles an hour (12,900 kilometers per hour)! But without an atmosphere, a landing can be very precise; there won't be any winds or turbulence, no friction. It was so fantastic a notion that I knew I had to use it.