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

[thespacecow]

The idea that there is any advantage in putting AI in orbit or on the moon is so stupid, I'm wondering if Elon is just trying to make up an economic justification for having a moon colony. I'd be concerned if I thought he really believed this.

Greg, since I share(d) some of this: What are the reasons it is so stupid?  I want to compare notes.

1) The cost of lifting equipment into space is exorbitant.
2) You have to put up not just your hardware but also power supply and cooling.

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

3) Assembly. A big data center is big. You'd have to make dozens or hundreds of launches to send up the pieces, and then how do all the pieces get put together?

You don't have to assemble it if just used for inference, which is the majority of the AI compute.

But auto assembly is not that hard, we already did it with ISS.

4) Bandwidth from space is pitiful. Latency is bad too.

Starlink is already aiming for 1Gbps for consumers, and in LEO the latency is comparable to terrestrial network.

5) Radiation in space is going to mess with your hardware, unless you add lots and lots of shielding. MMOD shielding is another requirement.

I'm not seeing lots and lots of shielding on Starlink or Starcloud-1.

6) If you put something on the moon, you add in super-high speed-of-light latency.

The plan is not to put the compute on the moon, the plan is to use moon resources to build AI compute satellite and launch them using mass driver.

7) Even if you get it set up, how will you maintain it? You can't easily send someone up to pull out bad components and plug in new ones.

You don't do that on Earth either. But if you have to do it, use robots.

Now, what are the legitimate advantages of a data center in orbit? I cannot think of a single one.

Power and permit.

Robotbeat also already gave a reason in previous comment, you're just not thinking at the scale Elon Musk is thinking.

[thespacecow]

Wow, I didn't know SpaceX has a lunar business development group, it'd be fascinating to see what business ideas they have. I think some intermediate profitable business is needed to bridge the gap between today and the lunar AI satellite factory/launcher future:

https://x.com/DrPhiltill/status/1985347532575096908

I’m truly happy to hear this, since international policy for the Moon will be decided in just 10-20 years. There’s no time to wait. (BTW, SpaceX has a lunar business development group that’s been saying it for a few years.)

https://x.com/DrPhiltill/status/1985377930554274123

They have a lunar business development manager who I have met with a few times. He told me in 2023 that SpaceX plans to stay on the Moon. I was discussing with him some possible business cases that a university might develop, taking advantage of low-cost Starship flights.

[thespacecow]

Some more posts from Dr. Phil Metzger:

https://x.com/DrPhiltill/status/1985518725181665422

I will show research that supports Elon’s statement (below) at the Economist’s Space Economy Summit on Thursday. I’ve been modeling growth of in-space AI, including space-based manufacturing and space resource mining, putting in as much realism as possible, and it suggests that we are only O(50 years) from AI and it’s supporting industries exceeding the entire terrestrial economy. (This assumes only ordinary economic growth rates — not Singularity magic.) This is the path to Kardashev II, and although that destination is still distant, it is not far along the path before AI has to go into space and has to use space resources (Moon and asteroids). Going off planet has to occur in just decades, not centuries, and really starting yesterday since it is a geopolitical race.

https://x.com/DrPhiltill/status/1985539096878407777

Over the years Elon has evolved the concept on how to pay for Mars settlement, and I think his posts yesterday and today should be interpreted in that context.

The evolution:

1. (2016 Guadalajara) Settlers will sell their homes on Earth to buy tickets to Mars.

2. (2017 Adelaide)  Starship (BFR) will raise funds by doing point-to-point transportation on Earth.

3. (October 2018 Twitter) Elon stated that his personal wealth (Tesla, etc.) is earmarked to pay for Mars as bridge funding.

4. (May 2019 Starlink press conference) Revenue from Starlink is tied to paying for Mars settlement.

5. (A few days ago on X) Starlink will not be just an internet relay, but a distributed AI compute system.

6. (Yesterday and Today) Elon leans further into space-based compute and lunar resources to build it.

I think this formula (all of the above) will easily pay for Mars. I mean, building it won’t be easy at all, but the upside of space-based AI is unbounded, so Mars funding should be covered easily by this plan. I have long believed that economics will be the hardest part of Mars settlement — not radiation or gravity or technology — and I honestly think the economics can now be seen as a solved problem.

[Vultur]

This honestly reduces my confidence in it happening, because it means the funding disappears if there is a major AI bubble burst.

I just flatly do not believe that there is a long term stable market for AI power use exceeding the total rest of humanity's energy use - especially if that "AI" is LLM based and therefore unable to do very much useful.

So far LLM AI has mostly made the Internet worse.

[thespacecow]

Obviously if you don't think AI will work out, none of these makes any sense.

It's pretty clear that Elon is a firm believer that AI will work, massively. He has another trillion dollar company and a 3rd hundred billion company betting on this.

But I'll say this: Pinning the hope of lunar development on AI makes much more sense than anything we've heard before, such as: Building in-space habitats, building space based solar power, moving heavy industry to space, mining propellant on the Moon for travel to the rest of the solar system, etc.

[meekGee]

This honestly reduces my confidence in it happening, because it means the funding disappears if there is a major AI bubble burst.

I just flatly do not believe that there is a long term stable market for AI power use exceeding the total rest of humanity's energy use - especially if that "AI" is LLM based and therefore unable to do very much useful.

So far LLM AI has mostly made the Internet worse.

That part I have no problem seeing.

LLMs will evolve more, but I already find them (chatGPT mostly) very useful in both personal and professional life, and honestly quite fun...

I also recognize not everyone's on the same page, but this is a one way train.

This one is no bubble, it is as robust as can be.

[volker2020]

Finally something I can really have an professional opinion on.

While I firmly believe, that one day we will create an AI that surpass us intellectually, I have to say, that the current generation of generative AI has about peaked. Technically we can clearly improve further but further advances are flattening out. The problem is with the base logic, that at no point tries to set up a world model, and therefore only rely on the statistical correlation between words in certain context. So regardless how much more context we press into this model, it will never stop hallucinations. That means that we can never stop to check the results of AI very carefully, which ultimately limits the productivity gains, you can achieve.
Which brings us to the problem of the productivity gap. AI only makes sense, if the effort we put into it does surpass the costs connected. That is not sustainable at this point, and the 11 Billion open AI lost in the last quarter are just a reminder of this. Rather soon, this companies have to earn there money and that spells dot-com style disaster for most companies. Looking at the financial base data, we are already beyond the point where the dot-com bubble did burst.
I am sure that GenAI will stay as a technology with a number of real live applications, but taking the real costs a lot of the gimmick applications, currently on vogue, will die the moment the real costs have to be payed.

All that said, I am almost certain that we will never see large GenAI  satellites in space. So in my point of view, Musk is just collecting money here.
   

[Vultur]

This honestly reduces my confidence in it happening, because it means the funding disappears if there is a major AI bubble burst.

I just flatly do not believe that there is a long term stable market for AI power use exceeding the total rest of humanity's energy use - especially if that "AI" is LLM based and therefore unable to do very much useful.

So far LLM AI has mostly made the Internet worse.

That part I have no problem seeing.

LLMs will evolve more, but I already find them (chatGPT mostly) very useful in both personal and professional life, and honestly quite fun...

I also recognize not everyone's on the same page, but this is a one way train.

This one is no bubble, it is as robust as can be.

Well there are two separate issues here.

I do think there is likely a bubble (which does not mean the technology is useless - there was a huge Internet bubble burst about 25 years ago, but the Internet was certainly very useful & survived the burst) and that this is a potential threat not just to Mars funding, but to R&D/tech funding in general. (I also think a lot of current LLMs are pretty harmful,  but that's yet another issue.)

But letting that pass. Let's say investment doesn't decline and LLMs turn out to dramatically increase everybody's productivity... How do you go from that to 100TW of AI power use? The economy is ultimately based on goods and services. There are only so many people (world population is not growing that fast anymore) and they can only consume so much actual goods and services.

The idea of uncapped demand for *anything*, AI queries or not, with a capped world population, doesn't make sense to me.

(And Mars is not going to be a meaningful part of the total human population on this time scale).

OTOH even something on a much smaller scale (say 100GW rather than 100TW) would still be ridiculously game changing in terms of space development.

[Twark_Main]

Some more posts from Dr. Phil Metzger:

...

https://x.com/DrPhiltill/status/1985539096878407777

Over the years Elon has evolved the concept on how to pay for Mars settlement, and I think his posts yesterday and today should be interpreted in that context.

The evolution:

...

4. (May 2019 Starlink press conference) Revenue from Starlink is tied to paying for Mars settlement.

Nitpick, but this information was known long before 2019. It was first publicly revealed at the original Starlink announcement back in January 2015.

 

@13:14  This is intended to generate significant revenue and help fund a city on Mars.

Looking into the long-term, what is needed to create a city on Mars? Well one thing's for sure, a lot of money. So we need things that will generate a lot of money.

[Greg Hullender]

Basically, the issue is that he sees AI as taking more power than you can build on Earth, and he actually thinks he can make a lot of money with it.

The idea of AI needing that much power is ridiculous, of course. But the idea that you can't build it on Earth is pretty questionable too. The Earth gets 173,000 TW from the sun, so the waste heat from 100 TW of power--ridiculous though that is--wouldn't affect the Earth's climate as long as it didn't come from fossil fuels. And 1 GW of solar power on Earth runs about $1 billion, so if you figure 10x the price for putting it in space, that's a quadrillion-dollar expense just for the power! And it would be rather strange if the actual AI hardware cost less than the power supply.

Finally, all that expense isn't going to get you the "magical AI" that people seem to be imagining. No more than if you tried to make a rocket ship by scaling up a truck to the size of a city.

So, yeah, it really is as dumb as it looks.

[Greg Hullender]

Starlink is already aiming for 1Gbps for consumers, and in LEO the latency is comparable to terrestrial network.

Just one quick comment, and then I'm going to read the document you linked to. I'm taking about the latency inside the data center, since that determines how long it takes parallel operations to synchronize. Even inference amounts to performing a monstrous matrix multiplication for each layer of the net. To do that efficiently in parallel, you want all the processors involved to be as close to each other as possible.

But that raises another question: If you're going to need 100 TW just for inference, does that mean you also need 1,000 TW to train the system? Or is the idea that AI would somehow be finished, so no one would need to train new networks?

[Roy_H]

Basically, the issue is that he sees AI as taking more power than you can build on Earth, and he actually thinks he can make a lot of money with it.

The idea of AI needing that much power is ridiculous, of course. But the idea that you can't build it on Earth is pretty questionable too.

I have a hard time with this too. Power requirements for AI are likely to come down dramatically as more efficient designs are discovered.
https://innatera.com
"Innatera’s ultra-efficient neuromorphic processors mimic the brain’s mechanisms for processing sensory data. Based on a proprietary analog-mixed signal computing architecture, Innatera’s processors leverage the computing capabilities of spiking neural networks to deliver ground-breaking cognition performance within a narrow power envelope. With an unprecedented combination of ultra-low power consumption and short response latency, these chips enable high-performance always-on pattern recognition capabilities for power-limited and latency-critical applications."

[meekGee]

Basically, the issue is that he sees AI as taking more power than you can build on Earth, and he actually thinks he can make a lot of money with it.

The idea of AI needing that much power is ridiculous, of course. But the idea that you can't build it on Earth is pretty questionable too. The Earth gets 173,000 TW from the sun, so the waste heat from 100 TW of power--ridiculous though that is--wouldn't affect the Earth's climate as long as it didn't come from fossil fuels. And 1 GW of solar power on Earth runs about $1 billion, so if you figure 10x the price for putting it in space, that's a quadrillion-dollar expense just for the power! And it would be rather strange if the actual AI hardware cost less than the power supply.

Finally, all that expense isn't going to get you the "magical AI" that people seem to be imagining. No more than if you tried to make a rocket ship by scaling up a truck to the size of a city.

So, yeah, it really is as dumb as it looks.

You say "of course", but that's hardly an argument.  I'll get back to that at the end.

---

I'm well aware of how easy it is to build more PV in Earth.  That was my main argument for a long time.  The upper value of orbital PV (assuming it is free) is the cost of PV on earth.  Can't benefit more than the terrestrial cost.

And then you need to subtract the actual cost of orbital PV, radiators, lack of maintenance ability, etc.

Been there.

But I've learned over he years that when Musk says stuff, even if counter-intuitive to me  I should listen.

He's not right 100% of the time, but he's up there above 90, which is not bad for a visionary.

So wrt orbital solar, I'm coming around.

I can see the cost equation coming a lot closer to balancing out, but also that cost is not where it's at.

The rub is in speed. Even with a pressing need, terrestrial PV still takes a lot of time (and mass) to build. It's hardly even keeping up with the current increase in terrestrial demand.

What if you wanted a faster increase?  Can you upscale terrestrial deployment?  How costly will that be?

Meanwhile, he's sitting on a fully reusable ship that can go multiple times per day at a projected cost of $1M a pop, for say 100 tons.

Add to that the unbearable lightness of space structures, the 4x continuous insolation, and it starts stacking up.

A "data center" won't look anything like terrestrial "giant compute centers". If built from scratch as orbital infrastructure, I can see it being a superior solution for large-scale deployment, maybe even cheaper.

---

So back to scale.

You can't see AI requiring more power than is currently consumed by non-AI?  Already AI's power footprint is moving the needle, and AI has hardly started to be deployed.

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.

Not just browser embedded chatbots, but persistent video monitoring ones, and "thinking" ones that are not limited to query-response cycles.

An engaged AI circuit that runs full time is in the 100s of Watts.  Won't take more than a few Billions of them to hit that benchmark you're so skeptical about.

---

So I'm getting on board with orbital.  Not quite seeing lunar though, except possibly in the very long run as material supply to high orbit.  But that's much further away than just orbital AI.

[Action]

And 1 GW of solar power on Earth runs about $1 billion, so if you figure 10x the price for putting it in space, that's a quadrillion-dollar expense just for the power!

I'm somewhat skeptical of the whole idea too, but this part seems wrong.

Part of the point of the whole thing is that, shortly, electric power is likely to be quite a bit cheaper in space than on the ground.  That part, at least, seems realistic.  Solar really, really wants to be in space and launch costs are plummeting.  If you assume it's much more expensive, you're effectively assuming away the biggest motivation for doing this.

Cooling and debris mitigation are much more serious concerns.  Cooling might be okay if you're really clever with geometry.  I don't see how you manage debris mitigation from very large objects in sun-synchronous orbits without large changes in the management of LEO.  Changes that require cooperation from people who are not inclined to be cooperative.

[hplan]

Starlink is already aiming for 1Gbps for consumers, and in LEO the latency is comparable to terrestrial network.

Just one quick comment, and then I'm going to read the document you linked to. I'm taking about the latency inside the data center, since that determines how long it takes parallel operations to synchronize. Even inference amounts to performing a monstrous matrix multiplication for each layer of the net. To do that efficiently in parallel, you want all the processors involved to be as close to each other as possible.

But that raises another question: If you're going to need 100 TW just for inference, does that mean you also need 1,000 TW to train the system? Or is the idea that AI would somehow be finished, so no one would need to train new networks?

Inference requires far less computation than training--a handful of large processors (H100s, say) is sufficient, a small enough number that you could fit many such clusters on a single satellite. For training, synchronization may be more relevant if the cluster in use is too big to fit on a single satellite. But it's done on earth, with training taking place on clusters distributed over multiple data centers. Here's a paper on minimizing the need to synchronize:

https://arxiv.org/pdf/2502.11058

But training is so much rarer than inference that it might make sense to keep it on earth.

The power (TW) figure is really being used as a measure of compute power, I think. OpenAI has started talking that way -- installing 10 new data centers with 1 GW each of compute power. The estimated cost of the compute hardware for each data center is ~$50 billion, not counting the nuclear reactor needed to power it. So Musk is talking about compute power that today would cost trillions.

All of this is highly speculative on Musk's part--who knows what the land will look like 10 years from now? Inference seems likely to move to the edge of the network (your cellphone); hardware architectures are going to get far more power efficient at the same time that they get faster.

[xvel]

Starlink is already aiming for 1Gbps for consumers, and in LEO the latency is comparable to terrestrial network.

Just one quick comment, and then I'm going to read the document you linked to. I'm taking about the latency inside the data center, since that determines how long it takes parallel operations to synchronize. Even inference amounts to performing a monstrous matrix multiplication for each layer of the net. To do that efficiently in parallel, you want all the processors involved to be as close to each other as possible.

But that raises another question: If you're going to need 100 TW just for inference, does that mean you also need 1,000 TW to train the system? Or is the idea that AI would somehow be finished, so no one would need to train new networks?

Inference requires far less computation than training--a handful of large processors (H100s, say) is sufficient, a small enough number that you could fit many such clusters on a single satellite. For training, synchronization may be more relevant if the cluster in use is too big to fit on a single satellite. But it's done on earth, with training taking place on clusters distributed over multiple data centers. Here's a paper on minimizing the need to synchronize:

https://arxiv.org/pdf/2502.11058

But training is so much rarer than inference that it might make sense to keep it on earth.

The power (TW) figure is really being used as a measure of compute power, I think. OpenAI has started talking that way -- installing 10 new data centers with 1 GW each of compute power. The estimated cost of the compute hardware for each data center is ~$50 billion, not counting the nuclear reactor needed to power it. So Musk is talking about compute power that today would cost trillions.

All of this is highly speculative on Musk's part--who knows what the land will look like 10 years from now? Inference seems likely to move to the edge of the network (your cellphone); hardware architectures are going to get far more power efficient at the same time that they get faster.

there shouldn't be any assumptions made, you are talking about very specific supervised learning of transformer like architectures, it's very unlikely this will be that important in the future, for example reinforcement learning is based mostly on "inference like" computations and there will be many other new algorithms, we are actually right now close to the wall of what can be done with just that stupid supervised training and that wall will be the bubble burst probably, and what will happen next is difficult to say.

[TheRadicalModerate]

Obviously there's a lot SpaceX can do with the Moon once they get Starship working as intended, but the question always comes down to funding: Where would the money come from?

A few possibilities:
1. Self-fund: They could use Starlink revenue, sort of invest in the future. Similar to how they used profit from launch to invest in Starlink, although I dare say this investment would be more risky than Starlink.
2. VC investment: They could raise more money based on their valuation. Elon will need to convince investors that build AI compute on the Moon would be actually profitable, not an easy task.
3. Debt financing: Very common for building infrastructure. This is how a lot of AI data centers are being funded right now, use GPUs as collateral. For the Moon, maybe they can use lunar real estate and infrastructure as collateral?
4. NASA funding: Not a lot but a steady stream. However they'll need to fight Congress to get it, also not an easy task.

It would be interesting to see how this plays out.

I would add xAI funding to the list.

[TheRadicalModerate]

But I do think there’s no reason to use the permanent craters. You can just make your own with a sunshade. And it doesn’t need to be on the Moon.

One nice property that has nothing to do with the shadowing:  If you're gonna have to make a zillion dilution fridges, it's nice to have the He-3 supply for them close by.

[meekGee]

I'm still not clear on the moon case, and lost track of exactly what Musk said about it.

Disclaimer: When Musk says stuff, I listen, but sometimes he misses the mark.  The reusable Mars ships were one such example.  One day we'll have symmetrical transport, but not until Mars is close to being on equal footing with Earth.

There are two lunar-related things you can do on the moon.

1: Build AI centers in dark craters, and collect PV at the rim.   That doesn't sound brilliant at all.

2: Mine and fabricate the satellites (or major subassemblies) on the moon so you don't have to lift so much from Earth.  Suppose you start just with radiators.  That may be plausible but the scales are crazy large even at Musk level.

When the excitement dies down a little, I think we'll find out that orbital AI is an actionable goal right now, with mass deployment say 3 years in the future, at truly mind boggling launch rates, while Lunar anything will remain aspirational for a long time.

That doesn't mean SpaceX won't do more on the surface, but not anywhere near as intense as the orbital campaigns.

[RedLineTrain]

Yes, it is a little hard to follow.  But he was pretty clear in his tweets today.  #2 only.  Solar cells and radiators should be relatively easy to do with only slightly modified machines from Earth.  These are not among the more difficult manufacturing challenges.

I would have thought that this was at least 10-15 years out because terrestrial solar seems like it has a ways to run before community rejection kicks in.  But AI scaling is becoming eye-watering.  Maybe $1 trillion in spending in each of the next half dozen years?

Peter Hague @peterrhague·8h
At the time I posted this, Elon viewed the Moon and its resources as a distraction from the Mars mission. But now he is talking about 100TW/year solar power using lunar resources as a long term goal.

This would, going on my beermat calculations (cheap 12% efficient cells, 350g/square metre), require about 210 million tonnes of silicon per year. Getting this from lunar silica, you would yield about 240 million tonnes of oxygen per year as a byproduct. You could vent >99% of this into space and still have more than you could possible use for the Mars fleet.

Now it would obviously be foolish to put this on the critical path for Mars - that will have to use Earth launched oxygen until lunar industry ramps up, and should not wait - but long term, it would mean a Starship to Mars would only require ~1 tanker instead of ~6. Moon, Mars, asteroids etc. are complimentary goals, not opposed ones. Developing one helps the others.

Elon Musk @elonmusk·5h
Scaling AI is what changes the equation

https://x.com/elonmusk/status/1985743650064908694