Elon Musk’s new Space AI ambitions

[Ludus]

https://eu.36kr.com/en/p/3567865887390593

This is a brief summary of the idea. Elon Musk has been offering comments and posts for a few weeks on the topic.

I post the topic in with Starship because it’s clearly based on the projected capacity of Starship and the Starship based expansion of Starlink. Maybe it should be elsewhere.

This seems like a new landscape of possibilities he’s considering, a bit like the first exploration of Starlink. It also involves a convergence between X.ai, Tesla and SpaceX. Before this Musk rejected any important role for Space based solar power. He explicitly did not see that as a driver for profitably using the launch capacity of Starship. The enormous investments in scaling AI and the power requirements changed this.

He seems to have decided on a first principles basis that Space based solar powered AI clusters make sense and that this ought to be a priority.

This is an existing topic about the idea generally: https://forum.nasaspaceflight.com/index.php?topic=63177.0

This is about Google Moonshot with the same focus. https://forum.nasaspaceflight.com/index.php?topic=63819.0

[Twark_Main]

The idea, clearly, is that AI training tasks let you beam the value back to Earth (ie the finished trained model) without beaming the BTUs back to Earth (and vaporizing the planet, in the limiting growth case).

I admit this something I contemplated, but I never thought someone would be crazy enough to actually attempt it!

It's so funny seeing people attribute this to some sort of marketing hype. As usual, when you see too far people think you're mad... 

[DigitalMan]

Elon has spoken about inference tasks, with the reason being that a short delay is acceptable

[leovinus]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

[DanClemmensen]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

Yep. Remember the huge render farms that were needed in the year 2000 to create CGI that we would now consider barely adequate?

[Greg Hullender]

Elon has spoken about inference tasks, with the reason being that a short delay is acceptable

Other way around. Inference requires low latency. Training (which can take months) does not.

[Greg Hullender]

The idea, clearly, is that AI training tasks let you beam the value back to Earth (ie the finished trained model) without beaming the BTUs back to Earth (and vaporizing the planet, in the limiting growth case).

And that's not a bad idea, in theory. But just looking at the OP's link, there are all manner of practical problems--not the least of which is "how does anyone make money off of this?"

[meekGee]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

Yep. Remember the huge render farms that were needed in the year 2000 to create CGI that we would now consider barely adequate?

The flip side is that miniaturization and diminishing power draws of compute did not make the semiconductor industry shrink, but))⁹ instead enabled more applications.

I have no doubt that the cost of a single inference will go down with time.  The question is just how pervasive AI will get.

[DigitalMan]

Elon has spoken about inference tasks, with the reason being that a short delay is acceptable

Other way around. Inference requires low latency. Training (which can take months) does not.

He clearly said exactly the opposite. Feel free to argue with him whether he is right.

[launchwatcher]

Remember the huge render farms that were needed in the year 2000 to create CGI that we would now consider barely adequate?

Thanks to Moore's law and Denard scaling, those massive render farms now more or less fit in the GPU of a typical smartphone.   Unfortunately Denard scaling broke down in about 2006 and Moore's law has also been slowing down.

While hardware will continue to improve incrementally, I strongly suspect that major improvements in efficiency of both the "training" and "inference" operations will come from algorithmic cleverness rather than better hardware.

[Vultur]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

Yep. Remember the huge render farms that were needed in the year 2000 to create CGI that we would now consider barely adequate?

The flip side is that miniaturization and diminishing power draws of compute did not make the semiconductor industry shrink, but))⁹ instead enabled more applications.

I have no doubt that the cost of a single.inference will go down with time.  The question is just how pervasive AI will get.

I personally believe there is a fairly hard upper limit on demand. There are only so many people in the world, expected to peak somewhere in the general neighborhood of 10B sometime in the second half of this century, and no immediate prospect of getting everyone on Earth fully connected into the information economy either.

I tend to think that 20-30 years from now the "wave of the future" will be in stuff that has very little to do with computers or IT. Nearly every technology has a sort of A curve - slow start, rapid growth, plateau.

[meekGee]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

Yep. Remember the huge render farms that were needed in the year 2000 to create CGI that we would now consider barely adequate?

The flip side is that miniaturization and diminishing power draws of compute did not make the semiconductor industry shrink, but))⁹ instead enabled more applications.

I have no doubt that the cost of a single.inference will go down with time.  The question is just how pervasive AI will get.

I personally believe there is a fairly hard upper limit on demand. There are only so many people in the world, expected to peak somewhere in the general neighborhood of 10B sometime in the second half of this century, and no immediate prospect of getting everyone on Earth fully connected into the information economy either.

I tend to think that 20-30 years from now the "wave of the future" will be in stuff that has very little to do with computers or IT. Nearly every technology has a sort of A curve - slow start, rapid growth, plateau.

Yeah but how many toasters?

Musk has some insight into where AI is headed 10 years from now.

I discounted AI before LLMs came out, but obviously people on the inside knew.

[Ludus]

Elon has spoken about inference tasks, with the reason being that a short delay is acceptable

A full scale implementation of Starlink over the next 5 years with tens of thousands of V3 platform and later 2000kg+ satellites with robust laser links will handle a large fraction of long distance data for the earth and a lot of direct 5G+ mobile connection. A lightspeed lag to a AI cluster in geosynchronous orbit or similar is not so bad considering that Starlink is already faster than fiber by having lasers in vacuum and more direct routing. AI will likely be pretty good at parsing work with low latency tasks handled on the Edge and inference requests to high orbit reserved for appropriate deep thinking.

[steveleach]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

Yep. Remember the huge render farms that were needed in the year 2000 to create CGI that we would now consider barely adequate?

The flip side is that miniaturization and diminishing power draws of compute did not make the semiconductor industry shrink, but))⁹ instead enabled more applications.

I have no doubt that the cost of a single.inference will go down with time.  The question is just how pervasive AI will get.

I personally believe there is a fairly hard upper limit on demand. There are only so many people in the world, expected to peak somewhere in the general neighborhood of 10B sometime in the second half of this century, and no immediate prospect of getting everyone on Earth fully connected into the information economy either.

I tend to think that 20-30 years from now the "wave of the future" will be in stuff that has very little to do with computers or IT. Nearly every technology has a sort of A curve - slow start, rapid growth, plateau.

While talented people are using AI to do amazing things, the overwhelming majority of current AI output is low quality garbage. I can't believe that the post-hype future of AI is going to be just a scaled-up version of what we have now.

[meekGee]

While talented people are using AI to do amazing things, the overwhelming majority of current AI output is low quality garbage. I can't believe that the post-hype future of AI is going to be just a scaled-up version of what we have now.

Exactly.
AI is not just LLMs and video monitoring.

Just a few years ago what you see today was strictly Sci Fi, and already people think they can see all the way to the far wall...

The world market can accommodate at least 5 computers, was it?

[Giovanni DS]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

That will be fun to watch... Billions in investment under the assumption that inference cannot be performed more efficiently. At some point somebody will breakthrough and "poof".

[TrevorMonty]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

That will be fun to watch... Billions in investment under the assumption that inference cannot be performed more efficiently. At some point somebody will breakthrough and "poof".

I don't know much about AI chips but assume principle is same as math coprocessors and graphics processors where HW is optimised for particular job. Math coprocessors were easy to spec as math functions they perform have been known for decades.  AI functions are still in stage of flux so its gamble designing HW for particular function to find its obsolete next month.

[Vultur]

While talented people are using AI to do amazing things, the overwhelming majority of current AI output is low quality garbage. I can't believe that the post-hype future of AI is going to be just a scaled-up version of what we have now.

Exactly.
AI is not just LLMs and video monitoring.

Just a few years ago what you see today was strictly Sci Fi, and already people think they can see all the way to the far wall...

The world market can accommodate at least 5 computers, was it?

I don't think it is analogous to that at all. At the time that quote was said, computers were a super extreme niche thing. AI is already all over the Internet (eg every Google search pulls up an AI Overview).

Also, even if it were analogous... There's an upper limit on use of computers too. You could maybe get x10 the number of computerized devices we have now (if every person on Earth had a smartphone, smart watch, smart home etc) but not 100x (there just aren't going to be enough people to use that many devices)..

I am not claiming that current AI use is near peak (though it could be, if much of the current use is "artificial" demand - like AI Overviews on every Google search - which wouldn't be used once the true costs are passed on). I am only claiming that the peak is probably well below terawatt level power use.

[meekGee]

While talented people are using AI to do amazing things, the overwhelming majority of current AI output is low quality garbage. I can't believe that the post-hype future of AI is going to be just a scaled-up version of what we have now.

Exactly.
AI is not just LLMs and video monitoring.

Just a few years ago what you see today was strictly Sci Fi, and already people think they can see all the way to the far wall...

The world market can accommodate at least 5 computers, was it?

I don't think it is analogous to that at all. At the time that quote was said, computers were a super extreme niche thing. AI is already all over the Internet (eg every Google search pulls up an AI Overview).

Also, even if it were analogous... There's an upper limit on use of computers too. You could maybe get x10 the number of computerized devices we have now (if every person on Earth had a smartphone, smart watch, smart home etc) but not 100x (there just aren't going to be enough people to use that many devices)..

I am not claiming that current AI use is near peak (though it could be, if much of the current use is "artificial" demand - like AI Overviews on every Google search - which wouldn't be used once the true costs are passed on). I am only claiming that the peak is probably well below terawatt level power use.

Actually I think AI today is not 1% of being all over the place.

You could have made your statement when computers became personal ("all over the place"), but then came laptops, and phones, and scam currency, and now AI...

So fast forward to AI. You think the type of tasks it does is exhausted, and I think the types of activities haven't even started to manifest.  Right now AI made a few functional steps forward - advanced image recognition, video analysis, syntax synthesis like LLMs, and the "reasoning" it employs.

I have zero doubt many more capabilities and applications are being worked at in labs, and will define the upcoming decade. I'm also sure the likes of SpaceX and Google have insight into those processes, and are not basing the future of their companies on simple extrapolations.

I'm pretty sure "TettaWatts" was not thrown around without regard.

The world uses 3 TWatt-avg of electricity today, which is shy of 0.5 kWatt-avg per person.

I don't see a problem assuming that a third of the population will have a half-kWatt continuous "AI draw" footprint.

[Eric Hedman]

In my business we are using AI on a steadily growing basis.  We started by using it to develop marketing material with an AI generated avatar.  I've been absolutely amazed at what it is able to do.  Our next step is to start using Copilot in Visual Studio for software development.  The potential just seems boundless so I think the growth will be sustained for quite awhile.

Another thing I've been noticing is the rapidly growing opposition to AI data centers in my home state of Wisconsin.  There was just an $8 billion data center partly owned by Oracle approved by the city council in Port Washington.  Meta is trying to get a Billion dollar data center approved in Beaver Damn.  Black Rock is trying to get a $12 billion data center approved in DeForest.  I've been following the opposition groups to see what they are doing.  They are starting to coordinate with other opposition groups around the country.  Amazon just pulled out of a plan for a data center in Tucson due to fierce local opposition.  Nobody it seems wants them in their back yards.

A lot of companies are trying to build data centers all around the Great Lakes region because of all the water available for cooling and a somewhat stable power grid.  The locals don't like that where they've been built before that they drive up utility costs and usually want somewhat pristine land while providing few permanent jobs.  This is why I think there will be a strong push to build these centers in orbit even if they cost somewhat more for at least then next decade or two..

[Vultur]

So fast forward to AI. You think the type of tasks it does is exhausted,

No, that's not what I'm saying at all, here.

I'm saying that even if tons of new applications are invented there's still a fairly hard upper limit which is ultimately more or less set by the number of "technologically connected" people. And that hard limit, at least in the next century or so, is probably below terawatt level use.

Especially with efficiency improvements. And I think fairly dramatic efficiency improvements are ultimately necessary to make widespread use of AI cost effective if/when something closer to the true cost is passed on to the user, which currently often isn't the case.

I'm pretty sure "TettaWatts" was not thrown around without regard.

Oh, I think there's thought behind it. But I think it's a hypothetical of what can be done assuming unbounded demand (as well as not hitting any limits in scaling up manufacture, etc.) It's perfectly valid in that context, I just don't think that's the most likely scenario.

[RedLineTrain]

Elon has spoken about inference tasks, with the reason being that a short delay is acceptable

Other way around. Inference requires low latency. Training (which can take months) does not.

You're talking about different latencies.  Many or most inference jobs accept some latency to and from the user.  On the other hand, training jobs require extremely low latencies among the training CPUs in a coherent cluster.

Inference jobs calling agents/tools and moving data to/from the training cluster are separate discussions.

[meekGee]

So fast forward to AI. You think the type of tasks it does is exhausted,

No, that's not what I'm saying at all, here.

I'm saying that even if tons of new applications are invented there's still a fairly hard upper limit which is ultimately more or less set by the number of "technologically connected" people. And that hard limit, at least in the next century or so, is probably below terawatt level use.

Especially with efficiency improvements. And I think fairly dramatic efficiency improvements are ultimately necessary to make widespread use of AI cost effective if/when something closer to the true cost is passed on to the user, which currently often isn't the case.

I'm pretty sure "TettaWatts" was not thrown around without regard.

Oh, I think there's thought behind it. But I think it's a hypothetical of what can be done assuming unbounded demand (as well as not hitting any limits in scaling up manufacture, etc.) It's perfectly valid in that context, I just don't think that's the most likely scenario.

But doesn't that limit mean you're imagining only a "person interacting with an AI" kind of application?

What if the AI is driving robotics?  Or vehicles? Or corporate AI workers?

So many thoughts to be thought....

And even with people connectivity, what is the hard limit? A kWatt per person?

It adds up very quickly, I think.

Ugh. I forgot about the military. A lot of the people that don't have connectivity might actually consume AI compute power passively.

[DanielW]

I think Musk is talking inference not training for the initial build out.

Training data-centers need to be colocated with extremely high-speed interconnects between nodes. This requires On-Orbit assembly.

Inference also requires a great deal of power and compute because of the vast numbers of users, but it can be distributed. There is some baseline requirements driven by the size of the model, but generally a single query can be handled by a single rack or even a single GPU.

This means that you can off-load electricity consumption to space as fast as you can launch, without On-Orbit assembly, by just putting GPUs on a starlink satellite bus. You can do this until you saturate the need for inference. This frees up resources and goodwill on earth to be used for training.

Once inference has bootstrapped the process, you can worry about assembling training centers.

So I think Musk is thinking about what can we do easily right now rather than what makes the most sense from a pure physics standpoint.

Edit to add this from the MIT Technology Review. https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/

"It’s now estimated that 80–90% of computing power for AI is used for inference."

[launchwatcher]

I think Musk is talking inference not training for the initial build out.

Training data-centers need to be colocated with extremely high-speed interconnects between nodes. This requires On-Orbit assembly.

Google is investigating whether they can avoid on-orbit assembly by instead using free-space optical data links between satellites flying in close (under 1km) formation:

https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/

At the altitude of our planned constellation, the non-sphericity of Earth's gravitational field, and potentially atmospheric drag, are the dominant non-Keplerian effects impacting satellite orbital dynamics. In the figure below, we show trajectories (over one full orbit) for an illustrative 81-satellite constellation configuration in the orbital plane, at a mean cluster altitude of 650 km. The cluster radius is R=1 km, with the distance between next-nearest-neighbor satellites oscillating between ~100–200m, under the influence of Earth’s gravity.

The models show that, with satellites positioned just hundreds of meters apart, we will likely only require modest station-keeping maneuvers to maintain stable constellations within our desired sun-synchronous orbit.

[Vultur]

So fast forward to AI. You think the type of tasks it does is exhausted,

No, that's not what I'm saying at all, here.

I'm saying that even if tons of new applications are invented there's still a fairly hard upper limit which is ultimately more or less set by the number of "technologically connected" people. And that hard limit, at least in the next century or so, is probably below terawatt level use.

Especially with efficiency improvements. And I think fairly dramatic efficiency improvements are ultimately necessary to make widespread use of AI cost effective if/when something closer to the true cost is passed on to the user, which currently often isn't the case.

I'm pretty sure "TettaWatts" was not thrown around without regard.

Oh, I think there's thought behind it. But I think it's a hypothetical of what can be done assuming unbounded demand (as well as not hitting any limits in scaling up manufacture, etc.) It's perfectly valid in that context, I just don't think that's the most likely scenario.

But doesn't that limit mean you're imagining only a "person interacting with an AI" kind of application?

What if the AI is driving robotics?  Or vehicles? Or corporate AI workers?

That raises the limit but does not change the fundamentals. Those robots building stuff or vehicles delivering stuff are still ultimately providing goods and services to people. The total demand in the economy is still based on people.

"Technologically connected" people in this context doesn't necessarily just mean an internet connection, it means people who have decent access to/participation in the technological world economy (which isn't everyone).

There's no point in having automated factories building things there is no demand for. It's physically possible, but doesn't make sense.

Also I think the cost issues still hit. I don't think AI doing a lot of those things will end up being cost effective at current total costs, once (or if) those costs are passed on to the end user.

So the terawatt+ scenario isn't impossible, no. But I think it requires a fairly narrow needle to be threaded. If AI gets much more efficient, you don't get terawatt power use; if AI doesn't get more efficient, its use probably becomes more limited due to cost. There's a very narrow range in there (assuming those two scenarios don't overlap entirely, leaving no range) where AI is still energy hungry enough to demand terawatt+ power but cheap enough to be used widely once real costs catch up.

(And that's assuming other factors not directly related to the technology itself don't get in the way. Which strikes me as likely.)

[DanielW]

I think Musk is talking inference not training for the initial build out.

Training data-centers need to be colocated with extremely high-speed interconnects between nodes. This requires On-Orbit assembly.

Google is investigating whether they can avoid on-orbit assembly by instead using free-space optical data links between satellites flying in close (under 1km) formation:

https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/

At the altitude of our planned constellation, the non-sphericity of Earth's gravitational field, and potentially atmospheric drag, are the dominant non-Keplerian effects impacting satellite orbital dynamics. In the figure below, we show trajectories (over one full orbit) for an illustrative 81-satellite constellation configuration in the orbital plane, at a mean cluster altitude of 650 km. The cluster radius is R=1 km, with the distance between next-nearest-neighbor satellites oscillating between ~100–200m, under the influence of Earth’s gravity.

The models show that, with satellites positioned just hundreds of meters apart, we will likely only require modest station-keeping maneuvers to maintain stable constellations within our desired sun-synchronous orbit.

I think Musk is a step ahead. Since most power consumption is inference he can launch that without the added complexity of tight interconnects. If built into larger Starlink satellites, you increase the addressable market for Starlink while having built-in data connects for the AI workloads.

It will also be hard to beat the latency since the inference will most likely happen right over your head and beamed straight back down. (not as good as living close to a datacenter)

[JayWee]

It will also be hard to beat the latency since the inference will most likely happen right over your head and beamed straight back down. (not as good as living close to a datacenter)

This. Especially for robots like Optimus or drones.

That raises the limit but does not change the fundamentals. Those robots building stuff or vehicles delivering stuff are still ultimately providing goods and services to people. The total demand in the economy is still based on people.

"Technologically connected" people in this context doesn't necessarily just mean an internet connection, it means people who have decent access to/participation in the technological world economy (which isn't everyone).

There's no point in having automated factories building things there is no demand for. It's physically possible, but doesn't make sense.

Ultimately people benefit, yes. But the customers might not be nearby. There are people remote driving mine excavators from 1500km away in Australia right now.
With global low-latency inference, you can easily answer the call to "We want to deploy thousands of robots in the middle of Congo" by simply flying them in, instead of having to plan all the necessary support infrastructure locally taking months/years.

[Vultur]

That raises the limit but does not change the fundamentals. Those robots building stuff or vehicles delivering stuff are still ultimately providing goods and services to people. The total demand in the economy is still based on people.

"Technologically connected" people in this context doesn't necessarily just mean an internet connection, it means people who have decent access to/participation in the technological world economy (which isn't everyone).

There's no point in having automated factories building things there is no demand for. It's physically possible, but doesn't make sense.

Ultimately people benefit, yes. But the customers might not be nearby. There are people remote driving mine excavators from 1500km away in Australia right now.
With global low-latency inference, you can easily answer the call to "We want to deploy thousands of robots in the middle of Congo" by simply flying them in, instead of having to plan all the necessary support infrastructure locally taking months/years.

Of course.

But I am talking about an upper limit set by total world demand, not local demand.

1TW (about 1/3 of current total world electricity use) is admittedly imaginable ... IF AI tech threads the narrow needle between "much greater efficiency means less energy need" and "too expensive to be used in everything once full costs are passed on to end users". And IF practicalities of building it all (and making the factories to build it all, etc) don't get in the way, and IF external factors (political backlash against AI affecting job markets, etc) don't get in the way.

I don't know where that upper limit is. But I think there has to be one.

[DigitalMan]

I think Musk is talking inference not training for the initial build out.

Training data-centers need to be colocated with extremely high-speed interconnects between nodes. This requires On-Orbit assembly.

Google is investigating whether they can avoid on-orbit assembly by instead using free-space optical data links between satellites flying in close (under 1km) formation:

https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/

At the altitude of our planned constellation, the non-sphericity of Earth's gravitational field, and potentially atmospheric drag, are the dominant non-Keplerian effects impacting satellite orbital dynamics. In the figure below, we show trajectories (over one full orbit) for an illustrative 81-satellite constellation configuration in the orbital plane, at a mean cluster altitude of 650 km. The cluster radius is R=1 km, with the distance between next-nearest-neighbor satellites oscillating between ~100–200m, under the influence of Earth’s gravity.

The models show that, with satellites positioned just hundreds of meters apart, we will likely only require modest station-keeping maneuvers to maintain stable constellations within our desired sun-synchronous orbit.

I think Musk is a step ahead. Since most power consumption is inference he can launch that without the added complexity of tight interconnects. If built into larger Starlink satellites, you increase the addressable market for Starlink while having built-in data connects for the AI workloads.

It will also be hard to beat the latency since the inference will most likely happen right over your head and beamed straight back down. (not as good as living close to a datacenter)

I suspect the primary benefit he will see is reduced workloads running in 3rd party datacenters.

[meekGee]

So fast forward to AI. You think the type of tasks it does is exhausted,

No, that's not what I'm saying at all, here.

I'm saying that even if tons of new applications are invented there's still a fairly hard upper limit which is ultimately more or less set by the number of "technologically connected" people. And that hard limit, at least in the next century or so, is probably below terawatt level use.

Especially with efficiency improvements. And I think fairly dramatic efficiency improvements are ultimately necessary to make widespread use of AI cost effective if/when something closer to the true cost is passed on to the user, which currently often isn't the case.

I'm pretty sure "TettaWatts" was not thrown around without regard.

Oh, I think there's thought behind it. But I think it's a hypothetical of what can be done assuming unbounded demand (as well as not hitting any limits in scaling up manufacture, etc.) It's perfectly valid in that context, I just don't think that's the most likely scenario.

But doesn't that limit mean you're imagining only a "person interacting with an AI" kind of application?

What if the AI is driving robotics?  Or vehicles? Or corporate AI workers?

That raises the limit but does not change the fundamentals. Those robots building stuff or vehicles delivering stuff are still ultimately providing goods and services to people. The total demand in the economy is still based on people.

"Technologically connected" people in this context doesn't necessarily just mean an internet connection, it means people who have decent access to/participation in the technological world economy (which isn't everyone).

There's no point in having automated factories building things there is no demand for. It's physically possible, but doesn't make sense.

Also I think the cost issues still hit. I don't think AI doing a lot of those things will end up being cost effective at current total costs, once (or if) those costs are passed on to the end user.

So the terawatt+ scenario isn't impossible, no. But I think it requires a fairly narrow needle to be threaded. If AI gets much more efficient, you don't get terawatt power use; if AI doesn't get more efficient, its use probably becomes more limited due to cost. There's a very narrow range in there (assuming those two scenarios don't overlap entirely, leaving no range) where AI is still energy hungry enough to demand terawatt+ power but cheap enough to be used widely once real costs catch up.

(And that's assuming other factors not directly related to the technology itself don't get in the way. Which strikes me as likely.)

But if I'm understanding it correctly, you're saying that the breadth of AI activity is proportional to the population (e.g. times "participation fraction" times "utilization factor" etc..) so it can't grow exponentially.

Which is true.

But the utilization factor - AI footprint per participating person - that factor can still grow by 1000x or a 1,000,000x from what it is today.

There will be a ceiling, we just don't have any idea where that ceiling is.

I mean - look at semiconductor technology.  It's capped, but we're not close to scratching the limit, even 70 years in.

[volker2020]

Since we are all number persons here. Let's lay out some numbers and let us think:

One traditional GWh data center for AI on earth has initial setup costs of around 80.000.000.000$ and has self live of around 5 years, before the hardware has to be replace.

I am sure that a clever man with a lot of money can reduce that cost by using custom self produced AI chips and mass production, by some margin but assuming that will be more than 50% does not sound very realistic, especially thinking of the need to space hardening the whole setup, putting it into a rocket, make it save to 2g of acceleration .... Than there will be the extra costs of moving that hardware.

At the end, I guess we can agree that even achieving 100 Billion $ per data center is a doubtful task. So basically each of this data centers in space need to produce  around 20 Billion of revenue per year. So there must be a productivity gain, greater or at least equal to this cost.

But the productivity gain is severely limited by what some call hallucinations,  or people like me call unavoidable consequences of the underlying algorithm of all current generative AIs. As long as this (let's call it flaw of the underlying statistical model), is not solved (and that would need one really genius idea, that is not to predict so far) the usability of AI has it's limit. 3 years after introduction, there is still no company making serious money with this. Open AI currently make a minus of 15 Billion per quarter.  So who will pay for this. School children, cheating with there homework clearly is not the answer. In my line of work, I can realize a productivity increase of around 5% using AI, which is nothing to sneeze at, but is far off the promises given by the AI executives.

So no, I don't believe that much will happen in this round. That might change in the future, but that future has not presented itself so far. 

[launchwatcher]

But wait, there's more:

Sam Altman Has Explored Deal to Build Competitor to Elon Musk’s SpaceX

The OpenAI CEO has publicly talked about the possibility of building ‘a rocket company’ and the potential for developing data centers in space.

OpenAI Chief Executive Sam Altman has explored putting together funds to either acquire or partner with a rocket company, a move that would position him to compete against Elon Musk’s SpaceX.

Altman reached out to at least one rocket maker, Stoke Space, in the summer, and the discussions picked up in the fall, according to people familiar with the talks. Among the proposals was for OpenAI to make a series of equity investments in the company and end up with a controlling stake. Such an investment would total billions of dollars over time.

If you subscribe to the WSJ, see: https://www.wsj.com/tech/ai/sam-altman-has-explored-deal-to-build-competitor-to-elon-musks-spacex-01574ff7

[RedLineTrain]

Since we are all number persons here. Let's lay out some numbers and let us think:

One traditional GWh data center for AI on earth has initial setup costs of around 80.000.000.000$ and has self live of around 5 years, before the hardware has to be replace.

I am sure that a clever man with a lot of money can reduce that cost by using custom self produced AI chips and mass production, by some margin but assuming that will be more than 50% does not sound very realistic, especially thinking of the need to space hardening the whole setup, putting it into a rocket, make it save to 2g of acceleration .... Than there will be the extra costs of moving that hardware.

It's GW (gigawatt), not GWh (gigawatt-hours).  Regardless, the cost reduction for using self produced AI chips can be quite dramatic, and much more than 50%.  This is because NVidia has 73% gross margins.  Of course, this is true whether we are talking terrestrial or orbital.

[RedLineTrain]

But wait, there's more:

Sam Altman Has Explored Deal to Build Competitor to Elon Musk’s SpaceX

Musk and Altman are battling each other and Google/Microsoft/Meta/NVidia/Broadcom by threatening to verticalize the necessary supply chains.  If threats of vertical supply chains are the game, then Musk is the master by being able to extend it from the fab all the way to space, with integrated power generation and power management.

As an observer, I take all of these threats with a grain of salt and make allowances for posturing.  But I keep an open mind.  The space industry might benefit a bit through add'l investment because of all of the posturing, even if ultimately space is not part of the game.  Unsure.

[wes_wilson]

Will be a wild future if AI becomes the "thing" in space that closes the business case for becoming a spacefaring civilization. 

[Vultur]

So fast forward to AI. You think the type of tasks it does is exhausted,

No, that's not what I'm saying at all, here.

I'm saying that even if tons of new applications are invented there's still a fairly hard upper limit which is ultimately more or less set by the number of "technologically connected" people. And that hard limit, at least in the next century or so, is probably below terawatt level use.

Especially with efficiency improvements. And I think fairly dramatic efficiency improvements are ultimately necessary to make widespread use of AI cost effective if/when something closer to the true cost is passed on to the user, which currently often isn't the case.

I'm pretty sure "TettaWatts" was not thrown around without regard.

Oh, I think there's thought behind it. But I think it's a hypothetical of what can be done assuming unbounded demand (as well as not hitting any limits in scaling up manufacture, etc.) It's perfectly valid in that context, I just don't think that's the most likely scenario.

But doesn't that limit mean you're imagining only a "person interacting with an AI" kind of application?

What if the AI is driving robotics?  Or vehicles? Or corporate AI workers?

That raises the limit but does not change the fundamentals. Those robots building stuff or vehicles delivering stuff are still ultimately providing goods and services to people. The total demand in the economy is still based on people.

"Technologically connected" people in this context doesn't necessarily just mean an internet connection, it means people who have decent access to/participation in the technological world economy (which isn't everyone).

There's no point in having automated factories building things there is no demand for. It's physically possible, but doesn't make sense.

Also I think the cost issues still hit. I don't think AI doing a lot of those things will end up being cost effective at current total costs, once (or if) those costs are passed on to the end user.

So the terawatt+ scenario isn't impossible, no. But I think it requires a fairly narrow needle to be threaded. If AI gets much more efficient, you don't get terawatt power use; if AI doesn't get more efficient, its use probably becomes more limited due to cost. There's a very narrow range in there (assuming those two scenarios don't overlap entirely, leaving no range) where AI is still energy hungry enough to demand terawatt+ power but cheap enough to be used widely once real costs catch up.

(And that's assuming other factors not directly related to the technology itself don't get in the way. Which strikes me as likely.)

But if I'm understanding it correctly, you're saying that the breadth of AI activity is proportional to the population (e.g. times "participation fraction" times "utilization factor" etc..) so it can't grow exponentially.

Which is true.

That's part of what I'm saying, but not all of it.

The other half is that (once things settle out and it becomes clear what "AI" works best for what applications*) AI will presumably only be used when it's the more cost effective way of doing things.

So cost effectiveness sets a limit on that utilization factor. If the energy (and hardware) cost of the AI is greater than the cost of doing the same thing without AI, a rational business won't use AI for that task.

I think that if AI energy costs don't decrease a lot, it won't prove to be cost-effective for a lot of things once full costs are passed on.


*Which may well mean a move away from LLMs to more specialized, efficient, and reliable systems, using far less energy

But the utilization factor - AI footprint per participating person - that factor can still grow by 1000x or a 1,000,000x from what it is today.

This is what I disagree with, at least if you define footprint in terms of energy use. AI using that much energy won't be cost effective. Not even with space solar power.

Datacenters are already using several percent of total electricity use. Though some of that is non-AI stuff ... But AI is definitely well into the gigawatts. 1,000,000x would be well into petawatts, probably more than 1MW/person even at a hypothetical peak world population of maybe 10-11B. I don't see how it can be cost effective for 99.7% of civilization's energy use to be AI.

I mean - look at semiconductor technology.  It's capped, but we're not close to scratching the limit, even 70 years in.

We may not be close to the physical limit but we're into the diminishing returns phase. Moore's Law has effectively ended.

IMO we are (as a civilization) now investing far too much of our R&D effort into computer/IT technologies rather than technologies which are earlier on their S curve, where investment would be far more beneficial.

[meekGee]

That's part of what I'm saying, but not all of it.

The other half is that (once things settle out and it becomes clear what "AI" works best for what applications*) AI will presumably only be used when it's the more cost effective way of doing things.

So cost effectiveness sets a limit on that utilization factor. If the energy (and hardware) cost of the AI is greater than the cost of doing the same thing without AI, a rational business won't use AI for that task.

I think that if AI energy costs don't decrease a lot, it won't prove to be cost-effective for a lot of things once full costs are passed on.


*Which may well mean a move away from LLMs to more specialized, efficient, and reliable systems, using far less energy

But the utilization factor - AI footprint per participating person - that factor can still grow by 1000x or a 1,000,000x from what it is today.

This is what I disagree with, at least if you define footprint in terms of energy use. AI using that much energy won't be cost effective. Not even with space solar power.

Datacenters are already using several percent of total electricity use. Though some of that is non-AI stuff ... But AI is definitely well into the gigawatts. 1,000,000x would be well into petawatts, probably more than 1MW/person even at a hypothetical peak world population of maybe 10-11B. I don't see how it can be cost effective for 99.7% of civilization's energy use to be AI.

I mean - look at semiconductor technology.  It's capped, but we're not close to scratching the limit, even 70 years in.

We may not be close to the physical limit but we're into the diminishing returns phase. Moore's Law has effectively ended.

IMO we are (as a civilization) now investing far too much of our R&D effort into computer/IT technologies rather than technologies which are earlier on their S curve, where investment would be far more beneficial.

So lets split it into "doing the same things but cheaper", and "doing new things".  (Typed A and B for this discussion)

Obvious analogies: thermodynamic engines allowed us to do sailing ships and horse drawn carriages better, but also allowed us to do airplanes and rockets.

Semiconductor tech allowed us to do radios better (replacing vacuum tubes) but then went on to create a couple other things.

Right now AI does some of both A and B.

Your argument applies only to type A, though in many situations AI is much more power intensive, but still "does it better".

Type B meanwhile is entirely outside the scope of your argument.

As per the analogies, cars are better than horses even if they consume more energy per mile, because they're faster and more dependable.  So there are more cars than there ever were horse-drawn carriages.  Meanwhile airplanes and rockets enabled entire sectors of industry that were not imaginable in the late 19th and early 20th century.

[Vultur]

Eh. I don't really agree, but we'll have to see what happens ... And it might be a couple decades before the answer is clear.

What I'm more concerned about in the short term is an investment bubble burst bringing down space stuff along with everything else. This could happen *regardless* of the soundness of the underlying technology - the dot com bubble of the late 90s wrecked private space projects then, though both the Internet in general and satellite internet specifically ultimately did work out.

[meekGee]

Eh. I don't really agree, but we'll have to see what happens ... And it might be a couple decades before the answer is clear.

What I'm more concerned about in the short term is an investment bubble burst bringing down space stuff along with everything else. This could happen *regardless* of the soundness of the underlying technology - the dot com bubble of the late 90s wrecked private space projects then, though both the Internet in general and satellite internet specifically ultimately did work out.

I'm pretty sure that it'll go down a lot faster than it did with thermodynamic engines.

[Vultur]

Eh. I don't really agree, but we'll have to see what happens ... And it might be a couple decades before the answer is clear.

What I'm more concerned about in the short term is an investment bubble burst bringing down space stuff along with everything else. This could happen *regardless* of the soundness of the underlying technology - the dot com bubble of the late 90s wrecked private space projects then, though both the Internet in general and satellite internet specifically ultimately did work out.

I'm pretty sure that it'll go down a lot faster than it did with thermodynamic engines.

Probably*, but that doesn't really answer the underlying question. There was a very long time between the invention of steam engines and them really becoming widespread; 20-25 years or so (like the time gap between Teledesic and Starlink) would still be much faster.

Economic problems/investment bubble bursts/etc can happen on a much, much shorter timescale than that, and can happen almost regardless of the underlying value or not of the technology. (The dot com bubble didn't mean the Internet didn't ultimately get everywhere, the housing bubble didn't mean people stopped needing housing.)

I am *also* much less optimistic than you about the ultimate value of LLM based AI technology or "generative AI", but that's basically a *completely separate question*. The short term (next 4-5 years) concerns remain either way.

I just don't want SpaceX to get so economically tied to xAI or some other AI thing that an investment bubble burst pushes Moon and Mars exploration back 20-25 years. And I want them to keep focus on Mars for the next couple synods, at least as much as they can with Artemis obligations.

*Though I don't think the 19th-20th century industrial revolution analogy really holds. That was a period of very rapid demand growth driven by both overall population growth and much more of the world being drawn into the industrial economy.

[meekGee]

Eh. I don't really agree, but we'll have to see what happens ... And it might be a couple decades before the answer is clear.

What I'm more concerned about in the short term is an investment bubble burst bringing down space stuff along with everything else. This could happen *regardless* of the soundness of the underlying technology - the dot com bubble of the late 90s wrecked private space projects then, though both the Internet in general and satellite internet specifically ultimately did work out.

I'm pretty sure that it'll go down a lot faster than it did with thermodynamic engines.

Probably*, but that doesn't really answer the underlying question. There was a very long time between the invention of steam engines and them really becoming widespread; 20-25 years or so (like the time gap between Teledesic and Starlink) would still be much faster.

Economic problems/investment bubble bursts/etc can happen on a much, much shorter timescale than that, and can happen almost regardless of the underlying value or not of the technology. (The dot com bubble didn't mean the Internet didn't ultimately get everywhere, the housing bubble didn't mean people stopped needing housing.)

I am *also* much less optimistic than you about the ultimate value of LLM based AI technology or "generative AI", but that's basically a *completely separate question*. The short term (next 4-5 years) concerns remain either way.

I just don't want SpaceX to get so economically tied to xAI or some other AI thing that an investment bubble burst pushes Moon and Mars exploration back 20-25 years. And I want them to keep focus on Mars for the next couple synods, at least as much as they can with Artemis obligations.

*Though I don't think the 19th-20th century industrial revolution analogy really holds. That was a period of very rapid demand growth driven by both overall population growth and much more of the world being drawn into the industrial economy.

Agreed that there's much that is unknown, scope and timewise.

I don't know that SpaceX is betting the farm on AI though.  I think financially Starlink was already a good enough foundation.

Orbital AI is just one possible expansion, same as p2p. Maybe it will, maybe it won't.  I see potential and possibilities, not a sure thing.

But it's not just "if you don't try, you'll surely miss out".  It's that in Musk world, the act of trying is what potentially makes it happen. It's not about catching the wave, it's about making the wave. As in electric cars, LEO constellations, etc.

[Vultur]

I don't know that SpaceX is betting the farm on AI though.  I think financially Starlink was already a good enough foundation.

I think so, and hope so. It was the connection of the space datacenter stuff to Starlink v3 that started me worrying.

But it's not just "if you don't try, you'll surely miss out".  It's that in Musk world, the act of trying is what potentially makes it happen. It's not about catching the wave, it's about making the wave. As in electric cars, LEO constellations, etc.

That then raises the question of whether it's a good thing to make happen. I am far from convinced that AI that *is* cost-effective would actually be an overall net good for humanity, in a world with limited demand growth.

[seb21051]

That then raises the question of whether it's a good thing to make happen. I am far from convinced that AI that *is* cost-effective would actually be an overall net good for humanity, in a world with limited demand growth.

Some perspectives on what is good for humanity:

Seb's Theorem: Nature's Agenda is to enable as many viable species as possible to develop. The individuals of those species are useful only in their role to strive for the ongoing survival of their species. They have no other purpose. Certainly no such thing as Rights, Freedom, or the pursuit of happiness. That is a human thing, and in viewing as many other species as possible, purely a figment of our collective imagination. (I grew up in Africa, and had much time to observe other species in their daily fights for survival).

1. Humanity as a species is very successful, in a local stellar system sense, but individuals might not be.

2. To remain a successful species a sufficiency of survivors is required.

3. In order to survive, a species must be adaptable, or scratch out a unique niche to occupy.

4. So it may be that AI, in its most successful evolution, keeps some portion of humanity surviving, functional and evolving. Either as equals or pets.

5. I see AI as the next step in our species' exercise to be adaptable. It may lead to our extinction, eventually. But, as we develop Super AI, we play the roles of creators, sort of little demi-gods.

6. The biggest single problem I see in this mad rush of ours to evolve, is that it seems very unlikely that we will be able to keep control of our AI invention, as it becomes more intelligent than we are.

7. The next couple of hundred years are going to be fascinating to observe. I just wish I could be a fly on the wall. If only Neuralink would hurry up and make it possible to load my psyche up to the cloud. I estimate I have less than 10 years to see that happen.

To summarize: It is likely that the majority of humanity will suffer with the advent of more and more intelligent machines, if Nature's Agenda is as I see it. Pets or batteries. Take your pick.

[zubenelgenubi: Please use quote function when quoting.]

Fair enough.

[Twark_Main]

Since we are all number persons here. Let's lay out some numbers and let us think:

One traditional GWh data center for AI on earth has initial setup costs of around 80.000.000.000$ and has self live of around 5 years, before the hardware has to be replace.

I am sure that a clever man with a lot of money can reduce that cost by using custom self produced AI chips and mass production, by some margin but assuming that will be more than 50% does not sound very realistic, especially thinking of the need to space hardening the whole setup, putting it into a rocket, make it save to 2g of acceleration .... Than there will be the extra costs of moving that hardware.

Now add the cost of terrestrial solar + batteries to power it 24/7, vs ~1/5th the solar and no batteries if you're in space.

You also delete a huge mass of steel and concrete and glass needed for terrestrial PV, which is actually a large fraction of the finished cost (and ecological impact) nowadays.

[Twark_Main]

Eh. I don't really agree, but we'll have to see what happens ... And it might be a couple decades before the answer is clear.

What I'm more concerned about in the short term is an investment bubble burst bringing down space stuff along with everything else. This could happen *regardless* of the soundness of the underlying technology - the dot com bubble of the late 90s wrecked private space projects then, though both the Internet in general and satellite internet specifically ultimately did work out.

I'm pretty sure that it'll go down a lot faster than it did with thermodynamic engines.

Compared to thermodynamic engines, AI is in the early "play around and see what works" stage.

Eventually, with steam engines, we figured out the limiting laws that governed their operation (Carnot's efficiency limit and what would later become Odum's specific power limit), and this new theoretical understanding enabled rapid progress which quickly approached these limits.

I expect we'll see the same in AI, where we develop a "Carnot's limit" for the maximum algorithmic efficiency of computation and ML (ie a software equivalent of what Landauer's Principle is for compute hardware).

Eventually the early drama blows over as the field matures, all the clever tricks get incorporated into the status quo, and it all gets boiled to a boring flowchart sizing the "turbine" for a particular information processing task.

[meekGee]

Eh. I don't really agree, but we'll have to see what happens ... And it might be a couple decades before the answer is clear.

What I'm more concerned about in the short term is an investment bubble burst bringing down space stuff along with everything else. This could happen *regardless* of the soundness of the underlying technology - the dot com bubble of the late 90s wrecked private space projects then, though both the Internet in general and satellite internet specifically ultimately did work out.

I'm pretty sure that it'll go down a lot faster than it did with thermodynamic engines.

Compared to thermodynamic engines, AI is in the early "play around and see what works" stage.

Eventually, with steam engines, we figured out the limiting laws that governed their operation (Carnot's efficiency limit and what would later become Odum's specific power limit), and this new theoretical understanding enabled rapid progress which quickly approached that limit.

I expect we'll see the same in AI, where we develop a "Carnot's limit" for the maximum algorithmic efficiency of computation and ML (ie a software equivalent of what Landauer's Principle is for compute hardware).

Yup and cyber science in the 21st century moves a lot faster than the very primitive investigation of things like combustion dynamics that limited early 20th century development of thermodynamic engines (and are only robustly solved since maybe a couple decades ago)

[Vultur]

Eh. I don't really agree, but we'll have to see what happens ... And it might be a couple decades before the answer is clear.

What I'm more concerned about in the short term is an investment bubble burst bringing down space stuff along with everything else. This could happen *regardless* of the soundness of the underlying technology - the dot com bubble of the late 90s wrecked private space projects then, though both the Internet in general and satellite internet specifically ultimately did work out.

I'm pretty sure that it'll go down a lot faster than it did with thermodynamic engines.

Compared to thermodynamic engines, AI is in the early "play around and see what works" stage.

Eventually, with steam engines, we figured out the limiting laws that governed their operation (Carnot's efficiency limit and what would later become Odum's specific power limit), and this new theoretical understanding enabled rapid progress which quickly approached that limit.

I expect we'll see the same in AI, where we develop a "Carnot's limit" for the maximum algorithmic efficiency of computation and ML (ie a software equivalent of what Landauer's Principle is for compute hardware).

Yup and cyber science in the 21st century moves a lot faster than the very primitive investigation of things like combustion dynamics that limited early 20th century development of thermodynamic engines (and are only robustly solved since maybe a couple decades ago)

There is, however, the possible opposite effect of the difference between a period of very rapid demand growth (due *both* to rapid population growth *and* incorporation of more and more of the world population into the industrial economy) and a period where both of those drivers of demand growth are far lower.

This is however a longer term issue than a potential investment bubble burst, which is a "next few years" question.

EDIT: the recent discussion of an IPO doesn't make me any more comfortable that this isn't a diversion from original core goals.

[Ludus]

<blockquote class="twitter-tweet"><p lang="en" dir="ltr">SpaceX has way more satellites in orbit than the rest of the world combined, so maybe we know a thing or two about the subject 🤣

Starlink V3 will be 20kW and launched at scale around Q4 next year. No problem to scale that to &gt;100kW if the satellite mass is shifted towards solar…</p>&mdash; Elon Musk (@elonmusk) December 10, 2025 <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>

The shift to Space based AI seems to be the main reason for the 2026 SpaceX IPO plans. It’s widely projected at $1.5 T market cap raising $30B+.

The near term project seems to sun synchronous orbit with the V3 Starlink platform minus antennas and related electronics which is the majority of its mass then scaling power and thermal dissipation from 20 kW to 100+kW. Laser comm connecting it to Starlink net.

Sats are mass produced and launched just like Starlink V3 on Starship.

The hyperscalers are already reaching the terrestrial limits. Every new data center becomes a problem competing for limited locations that have power and local regulatory tolerance.Expanding AI compute in space can be done systematically effectively just by scaling up mass production, of Starships, Starship infrastructure, Starlink platform satellites, etc.

One obvious prediction from this is that the SpaceForce will come out with approval of Starship launch and landing facility at SLC6 Vandenberg very soon - because this requires LOTS of Starship launches into Sun Synchronous orbit.

[JaimeZX]

....as we develop Super AI, we play the roles of creators, sort of little demi-gods.

6. The biggest single problem I see in this mad rush of ours to evolve, is that it seems very unlikely that we will be able to keep control of our AI invention, as it becomes more intelligent than we are.

The biggest problem is that the demi-gods are motivated by quarterly financial reporting and not What Is Best For The Species.

Pets or batteries. Take your pick.

Iain Banks' "Minds" are best-case scenario.
Pets is almost second-best, and almost certainly preferable to our extinction.

[seb21051]

....as we develop Super AI, we play the roles of creators, sort of little demi-gods.

6. The biggest single problem I see in this mad rush of ours to evolve, is that it seems very unlikely that we will be able to keep control of our AI invention, as it becomes more intelligent than we are.

The biggest problem is that the demi-gods are motivated by quarterly financial reporting and not What Is Best For The Species.

Pets or batteries. Take your pick.

Iain Banks' "Minds" are best-case scenario.
Pets is almost second-best, and almost certainly preferable to our extinction.

We are doomed to live in what the Chinese call "Interesting Times".

What I would give to be able to be a fly on the wall for the next 200 years. Neuralink needs to get its Psyche-Upload-To-The-Cloud service launched.

[Vultur]

What I would give to be able to be a fly on the wall for the next 200 years. Neuralink needs to get its Psyche-Upload-To-The-Cloud service launched.

I am very skeptical that uploading minds is even theoretically possible, even with arbitrarily advanced technology. The brain doesn't store information in the same way an electronic computer does; how would you get all the information out without destroying the 3D structure that is key to storing it? I think you'd need something like Star Trek scanners, which probably aren't physically possible.

(Anyway even if possible it would be the Star Trek transporter problem .. it's a copy of you not *you*. It's not immortality just a nonbiological form of reproduction, a sort of mental cloning.)

[seb21051]

What I would give to be able to be a fly on the wall for the next 200 years. Neuralink needs to get its Psyche-Upload-To-The-Cloud service launched.

I am very skeptical that uploading minds is even theoretically possible, even with arbitrarily advanced technology. The brain doesn't store information in the same way an electronic computer does; how would you get all the information out without destroying the 3D structure that is key to storing it? I think you'd need something like Star Trek scanners, which probably aren't physically possible.

(Anyway even if possible it would be the Star Trek transporter problem .. it's a copy of you not *you*. It's not immortality just a nonbiological form of reproduction, a sort of mental cloning.)

All I can do is live in hope, lol. I really wouldn't be worried what version of me lives on if it became a possibility.

On the other hand, look at the technological advances we have made in the last 100 years! I sometimes think of how I would try to explain the present to a Rip Van Winkeled Leonardo Da Vinci. That is after getting him to bathe, get a haircut and dressing him in cargo pants, nikes and a hawaiian shirt. Silly, I know.

[Vultur]

What I would give to be able to be a fly on the wall for the next 200 years. Neuralink needs to get its Psyche-Upload-To-The-Cloud service launched.

I am very skeptical that uploading minds is even theoretically possible, even with arbitrarily advanced technology. The brain doesn't store information in the same way an electronic computer does; how would you get all the information out without destroying the 3D structure that is key to storing it? I think you'd need something like Star Trek scanners, which probably aren't physically possible.

(Anyway even if possible it would be the Star Trek transporter problem .. it's a copy of you not *you*. It's not immortality just a nonbiological form of reproduction, a sort of mental cloning.)

All I can do is live in hope, lol. I really wouldn't be worried what version of me lives on if it became a possibility.

Eh, I guess this is a philosophical divide. From my perspective a copy of me wouldn't be "me" in any meaningful sense, there would be no continuity of identity. (Ship of Theseus, etc )

But this is off topic.

Back on topic ... I actually don't think this is a "wait 200 years" thing. With the end of Moore's law, I think these sorts of computer based things will either work out in the next 20 years, or never. After that the primary direction of advancement probably won't be computer/IT at all.

[meekGee]

What I would give to be able to be a fly on the wall for the next 200 years. Neuralink needs to get its Psyche-Upload-To-The-Cloud service launched.

I am very skeptical that uploading minds is even theoretically possible, even with arbitrarily advanced technology. The brain doesn't store information in the same way an electronic computer does; how would you get all the information out without destroying the 3D structure that is key to storing it? I think you'd need something like Star Trek scanners, which probably aren't physically possible.

(Anyway even if possible it would be the Star Trek transporter problem .. it's a copy of you not *you*. It's not immortality just a nonbiological form of reproduction, a sort of mental cloning.)

Just like you are just a copy of yesterday's you.  The real yesterday's you died overnight, as happens every night. Yup, that conversation from early childhood - how to define continuity of consciousness... 

Good thing current plans are not dependent on such pseudo-technology.

[Twark_Main]

Summarizing, there seem to be two main arguments in favor of space-based AI compute:

    1. Fewer PV and batteries needed vs terrestrial solar. Because AI requires so much power 24/7, this is actually a big deal.

    2. It doesn't deplete the (large but still finite) anthropogenic waste heat rejection capacity of the surface of the Earth. This is the "if you tried to beam the Sun's power to Earth the planet would melt" reason.


#1 is the short-term reason, and it's the reason why Musk says space-based AI will be there cheapest option in 2-3 years.


#2 is the long-term reason, and it doesn't effect current economics because we don't have a Joule Tax yet. This would be like a coal company planning around a Carbon Tax in 1890. It's too early.


And yes, before someone says it, I'm aware that PV-powered chips don't change the Earth's radiant power balance, but it does still "leach" exergy (aka "useful work") from the biosphere and agriculture and other industrial processes. There is a finite (but large) limit on the amount of available exergy that can be siphoned off from the Earth's total exergy budget before those other systems start being starved of useful work, due to unavoidable physics (thermodynamic) constraints.



_{Hopefully we all understand thermodynamics enough that we can skip the back-and-forth about entropy vs exergy vs energy.}

[meekGee]

Summarizing, there seem to be two main arguments in favor of space-based AI compute:

    1. Fewer PV and batteries needed vs terrestrial solar. Because AI requires so much power 24/7, this is actually a big deal.

    2. It doesn't deplete the (large but still finite) anthropogenic waste heat rejection capacity of the surface of the Earth. This is the "if you tried to beam the Sun's power to Earth the planet would melt" reason.


#1 is the short-term reason, and it's the reason why Musk says space-based AI will be there cheapest option in 2-3 years.


#2 is the long-term reason, and it doesn't effect current economics because we don't have a Joule Tax yet. This would be like a coal company planning around a Carbon Tax in 1890. It's too early.


And yes, before someone says it, I'm aware that PV-powered chips don't change the Earth's radiant power balance, but it does still "leach" exergy (aka "useful work") from the biosphere and agriculture and other industrial processes. There is a finite (but large) limit on the amount of available exergy that can be siphoned off from the Earth's total exergy budget before those other systems start being starved of useful work, due to unavoidable physics (thermodynamic) constraints.



_{Hopefully we all understand thermodynamics enough that we can skip the back-and-forth about entropy vs exergy vs energy.}

#2 is true in theory but is sooooo far away that other mechanisms will kick in first.

You're talking about directly influencing the energy balance of the planet, not just messing with greenhouse gasses.

Starship: 1-100 GWatt/yr (reasonable estimate)
Lunar: 0.1-10 TWatt/yr (very handwavy)
Earth energy uptake 120,000 TWatt

So...  If we install 10 TWatt/yr over 10 millennia, we'll have ourselves a competition.

Until then, it's mostly reason #1 and some other related ones.

[Twark_Main]

Summarizing, there seem to be two main arguments in favor of space-based AI compute:

    1. Fewer PV and batteries needed vs terrestrial solar. Because AI requires so much power 24/7, this is actually a big deal.

    2. It doesn't deplete the (large but still finite) anthropogenic waste heat rejection capacity of the surface of the Earth. This is the "if you tried to beam the Sun's power to Earth the planet would melt" reason.


#1 is the short-term reason, and it's the reason why Musk says space-based AI will be there cheapest option in 2-3 years.


#2 is the long-term reason, and it doesn't effect current economics because we don't have a Joule Tax yet. This would be like a coal company planning around a Carbon Tax in 1890. It's too early.


And yes, before someone says it, I'm aware that PV-powered chips don't change the Earth's radiant power balance, but it does still "leach" exergy (aka "useful work") from the biosphere and agriculture and other industrial processes. There is a finite (but large) limit on the amount of available exergy that can be siphoned off from the Earth's total exergy budget before those other systems start being starved of useful work, due to unavoidable physics (thermodynamic) constraints.



_{Hopefully we all understand thermodynamics enough that we can skip the back-and-forth about entropy vs exergy vs energy.}

#2 is true in theory but is sooooo far away that other mechanisms will kick in first.

You're talking about directly influencing the energy balance of the planet, not just messing with greenhouse gasses.

Starship: 1-100 GWatt/yr (reasonable estimate)
Lunar: 0.1-10 TWatt/yr (very handwavy)
Earth energy uptake 120,000 TWatt

...

Until then, it's mostly reason #1 and some other related ones.

Indeed! I thought I made that clear (in fact, pointing out this distinction was my main reason for posting), but it's always good to re-iterate and re-phrase the point.


"#2... would be like a coal company planning around a Carbon Tax in 1890. It's too early."

"(large but still finite) anthropogenic... heat capacity"


We may wish that fossil fuel companies had the foresight to address global warming from the start. Well Tesla is doing exactly what we might wish, but the popular reaction is instead to ridicule the fix as "uneconomical."

Fortunately, reason #1 is enough to make space-based AI economical even today. But you gotta give points for forward thinking... 

So...  If we install 10 TWatt/yr over 10 millennia, we'll have ourselves a competition.

Growth is exponential, more like 2% CAGR.


Thanks. A rare opportunity to use "exponential" in its real math meaning instead of "very big." 

[meekGee]

Summarizing, there seem to be two main arguments in favor of space-based AI compute:

    1. Fewer PV and batteries needed vs terrestrial solar. Because AI requires so much power 24/7, this is actually a big deal.

    2. It doesn't deplete the (large but still finite) anthropogenic waste heat rejection capacity of the surface of the Earth. This is the "if you tried to beam the Sun's power to Earth the planet would melt" reason.


#1 is the short-term reason, and it's the reason why Musk says space-based AI will be there cheapest option in 2-3 years.


#2 is the long-term reason, and it doesn't effect current economics because we don't have a Joule Tax yet. This would be like a coal company planning around a Carbon Tax in 1890. It's too early.


And yes, before someone says it, I'm aware that PV-powered chips don't change the Earth's radiant power balance, but it does still "leach" exergy (aka "useful work") from the biosphere and agriculture and other industrial processes. There is a finite (but large) limit on the amount of available exergy that can be siphoned off from the Earth's total exergy budget before those other systems start being starved of useful work, due to unavoidable physics (thermodynamic) constraints.



_{Hopefully we all understand thermodynamics enough that we can skip the back-and-forth about entropy vs exergy vs energy.}

#2 is true in theory but is sooooo far away that other mechanisms will kick in first.

You're talking about directly influencing the energy balance of the planet, not just messing with greenhouse gasses.

Starship: 1-100 GWatt/yr (reasonable estimate)
Lunar: 0.1-10 TWatt/yr (very handwavy)
Earth energy uptake 120,000 TWatt

...

Until then, it's mostly reason #1 and some other related ones.

Indeed! I thought I made that clear (in fact, pointing out this distinction was my main reason for posting), but it's always good to re-iterate and re-phrase the point.


"#2... would be like a coal company planning around a Carbon Tax in 1890. It's too early."

"(large but still finite) anthropogenic... heat capacity"


We may wish that fossil fuel companies had the foresight to address global warming from the start. Well Tesla is doing exactly what we might wish, but the popular reaction is instead to ridicule the fix as "uneconomical."

Fortunately, reason #1 is enough to make space-based AI economical even today. But you gotta give points for forward thinking... 

So...  If we install 10 TWatt/yr over 10 millennia, we'll have ourselves a competition.

Growth is exponential, more like 2% CAGR.


Thanks. A rare opportunity to use "exponential" in its real math meaning instead of "very big." 

Awright!

---

Meanwhile double checking:

Start with 1 GWatt/yr and grow by 10% every year, getting to 100,000 TWatt is a 170 years.

So yup, about the same time as from the industrial revolution to now.

Though, chat argues that worldwide energy use, since 1800, only grew by a factor of 30, not 100,000,000...

(I pushed back a bit but he's defending that 30x pretty well.)

[mikegi]

Your brain and mine require something like 20 to 25W of power. I don't think that the current models of AI capture what's really going on (by many orders-of-magnitude, even correcting for the magnitude of analog-to-digital conversion). I don't know the correct algorithm, but I do know that this isn't it. The current algorithms *might* be equivalent to the retina, but that's about it.

[Twark_Main]

Summarizing, there seem to be two main arguments in favor of space-based AI compute:

    1. Fewer PV and batteries needed vs terrestrial solar. Because AI requires so much power 24/7, this is actually a big deal.

    2. It doesn't deplete the (large but still finite) anthropogenic waste heat rejection capacity of the surface of the Earth. This is the "if you tried to beam the Sun's power to Earth the planet would melt" reason.


#1 is the short-term reason, and it's the reason why Musk says space-based AI will be there cheapest option in 2-3 years.


#2 is the long-term reason, and it doesn't effect current economics because we don't have a Joule Tax yet. This would be like a coal company planning around a Carbon Tax in 1890. It's too early.


And yes, before someone says it, I'm aware that PV-powered chips don't change the Earth's radiant power balance, but it does still "leach" exergy (aka "useful work") from the biosphere and agriculture and other industrial processes. There is a finite (but large) limit on the amount of available exergy that can be siphoned off from the Earth's total exergy budget before those other systems start being starved of useful work, due to unavoidable physics (thermodynamic) constraints.



_{Hopefully we all understand thermodynamics enough that we can skip the back-and-forth about entropy vs exergy vs energy.}

#2 is true in theory but is sooooo far away that other mechanisms will kick in first.

You're talking about directly influencing the energy balance of the planet, not just messing with greenhouse gasses.

Starship: 1-100 GWatt/yr (reasonable estimate)
Lunar: 0.1-10 TWatt/yr (very handwavy)
Earth energy uptake 120,000 TWatt

...

Until then, it's mostly reason #1 and some other related ones.

Indeed! I thought I made that clear (in fact, pointing out this distinction was my main reason for posting), but it's always good to re-iterate and re-phrase the point.


"#2... would be like a coal company planning around a Carbon Tax in 1890. It's too early."

"(large but still finite) anthropogenic... heat capacity"


We may wish that fossil fuel companies had the foresight to address global warming from the start. Well Tesla is doing exactly what we might wish, but the popular reaction is instead to ridicule the fix as "uneconomical."

Fortunately, reason #1 is enough to make space-based AI economical even today. But you gotta give points for forward thinking... 

So...  If we install 10 TWatt/yr over 10 millennia, we'll have ourselves a competition.

Growth is exponential, more like 2% CAGR.


Thanks. A rare opportunity to use "exponential" in its real math meaning instead of "very big." 

Awright!

---

Meanwhile double checking:
 
Start with 1 GWatt/yr and grow by 10% every year, getting to 100,000 TWatt is a 170 years.

So yup, about the same time as from the industrial revolution to now.

Exactly, it's hardly some unimaginable span of time! Even at a more reasonable (and more historical) 2-3% CAGR, we're down from "ten millennia" to less than one. 

Though, chat argues that worldwide energy use, since 1800, only grew by a factor of 30, not 100,000,000...

(I pushed back a bit but he's defending that 30x pretty well.)

Sure, that's exactly in line with what we should be expecting. That's a CAGR of 1.5%.

[Twark_Main]

Your brain and mine require something like 20 to 25W of power. I don't think that the current models of AI capture what's really going on (by many orders-of-magnitude, even correcting for the magnitude of analog-to-digital conversion). I don't know the correct algorithm, but I do know that this isn't it. The current algorithms *might* be equivalent to the retina, but that's about it.

Landauer's Limit is a long way down, so a lot of those orders-of-magnitude can still come from the hardware side.

Still, this is pretty uncontroversial. I think you'd be hard-pressed to find anyone who thinks today's algorithms are the end. Everyone expects future algorithms will improve efficiency, so it's already "baked into the cake" as it were.

[Twark_Main]

New article on the economics of space-based AI.

https://andrewmccalip.com/space-datacenters

Discussed here https://news.ycombinator.com/item?id=46281288


Notably their numbers assume fossil fuels are used to power the terrestrial datacenter, whereas I presume Musk is running numbers for solar + batteries because it's the only scalable solution. It also uses rather OldSpace-like assumptions for satellite and launch costs.

It does have a cool calculator function though, so with how these things go I assume this will become the "go-to" de facto source of misinformation on the subject... 

[meekGee]

Your brain and mine require something like 20 to 25W of power. I don't think that the current models of AI capture what's really going on (by many orders-of-magnitude, even correcting for the magnitude of analog-to-digital conversion). I don't know the correct algorithm, but I do know that this isn't it. The current algorithms *might* be equivalent to the retina, but that's about it.

Nobody is claiming it is...

But does it have to work the same way?

[xvel]

Your brain and mine require something like 20 to 25W of power. I don't think that the current models of AI capture what's really going on (by many orders-of-magnitude, even correcting for the magnitude of analog-to-digital conversion). I don't know the correct algorithm, but I do know that this isn't it. The current algorithms *might* be equivalent to the retina, but that's about it.

A small LLM speaks dozens of languages fluently and remembers the entire Wikipedia. It has superhuman abilities in some areas and is lacking in others. It is not a brain, it is something else, but it is not a simple system, as you are trying to imly.
Your digression on analog-to-digital conversion also suggests that you have not much idea about either of the two things as deep learning models are fundamentally continuous analog systems that are approximated by digital computers and spiking neurons of biological brains are actually very digital in nature.

[Vultur]

I see no reason to expect exponential growth in energy use over the next century or so. That's been driven so far by exponential growth in world population + more and more of the world population being brought into one connected economy. The first is slowing drastically and the second is self-limiting (there are no more markets as big as China and India).

Growth may effectively end. It isn't a fundamental law that it will/must continue.

[meekGee]

I see no reason to expect exponential growth in energy use over the next century or so. That's been driven so far by exponential growth in world population + more and more of the world population being brought into one connected economy. The first is slowing drastically and the second is self-limiting (there are no more markets as big as China and India).

Growth may effectively end. It isn't a fundamental law that it will/must continue.

If you're talking about the conversation between TM and myself, I don't think either of us took it literally or seriously.  Just juggling nomenclature and OOMs.

Nothing remains exponential across that many OOMs, some other limit always kicks in.

[InterestedEngineer]

I see no reason to expect exponential growth in energy use over the next century or so. That's been driven so far by exponential growth in world population + more and more of the world population being brought into one connected economy. The first is slowing drastically and the second is self-limiting (there are no more markets as big as China and India).

Growth may effectively end. It isn't a fundamental law that it will/must continue.

If you're talking about the conversation between TM and myself, I don't think either of us took it literally or seriously.  Just juggling nomenclature and OOMs.

Nothing remains exponential across that many OOMs, some other limit always kicks in.

exponential curves are an interesting topic.  Yes, they all eventually turn into logistics (S) curves, due to some limitations.  See for example DRAM prices, basically no change for 10 years.  We have hit a physical cost/bit limit.

OTOH, you CAN turn something that has matured (top of the S curve), into a new exponential growth.  And there's no better example than "new space" (SpaceX), regarding launch costs per kilogram.

Did physics change?  No, the same chemical limitations in general apply, we haven't gone with fission or fusion.

So what changed?

1.  Material Science (high volume 3D printing, new alloys)
2.  CFD being able to simulate an entire rocket engine - which was due to exponential reduction in compute price/performance in the last 30 years.  The Methalox engine, it can be argued, isn't truly possible to make reliable without the ability to simulate it.  This is happening in many fields.
3. The ability to reuse rocket hardware, which is a combination of compute power, better engines, sensors, and maturing engineering, and the willingness to try and fail a lot (itself a societal product of  the 1990s software boom)
4.  40+ years of reusable heat shield improvements (slow but steady), reaching critical mass.
5.  Probably stuff I'm missing.

Falcon-9 is about 3x cost reduction, and Starship will be another 10-40x on top of that.  That's definitely in the exponential growth area of the S-curve

So, getting around to the topic of energy, do we have such an exponential growth?  Kinda.  Not in amount of watts produced for sure - we are nearing the top of that s-curve.   But in the ability to produce those watts cheaply in space from solar power?  Absolutely.  It's now less than $1/watt of capacity to install solar power now on Earth, it was a lot more 2 decades ago.  Mix that with the  exponential drop in launch costs, and plenty of sunshine in space, and you have a new opportunity to grow the power produced by the human race.

And rather than beaming it down (which Elon says will never be efficient enough), we can just beam down the small few hundred megabytes of results from compute (AI, whatever), which is micropennies of energy.

So I think if one counts the net power produced by the human race, we will escape our current "top of the S curve is readily apparent" limits and go on another exponential binge.   Doing this in space for compute is good, we aren't taking away electricity from starving babies.

[meekGee]

I see no reason to expect exponential growth in energy use over the next century or so. That's been driven so far by exponential growth in world population + more and more of the world population being brought into one connected economy. The first is slowing drastically and the second is self-limiting (there are no more markets as big as China and India).

Growth may effectively end. It isn't a fundamental law that it will/must continue.

If you're talking about the conversation between TM and myself, I don't think either of us took it literally or seriously.  Just juggling nomenclature and OOMs.

Nothing remains exponential across that many OOMs, some other limit always kicks in.

exponential curves are an interesting topic.  Yes, they all eventually turn into logistics (S) curves, due to some limitations.  See for example DRAM prices, basically no change for 10 years.  We have hit a physical cost/bit limit.

OTOH, you CAN turn something that has matured (top of the S curve), into a new exponential growth.  And there's no better example than "new space" (SpaceX), regarding launch costs per kilogram.

Did physics change?  No, the same chemical limitations in general apply, we haven't gone with fission or fusion.

So what changed?

1.  Material Science (high volume 3D printing, new alloys)
2.  CFD being able to simulate an entire rocket engine - which was due to exponential reduction in compute price/performance in the last 30 years.  The Methalox engine, it can be argued, isn't truly possible to make reliable without the ability to simulate it.  This is happening in many fields.
3. The ability to reuse rocket hardware, which is a combination of compute power, better engines, sensors, and maturing engineering, and the willingness to try and fail a lot (itself a societal product of  the 1990s software boom)
4.  40+ years of reusable heat shield improvements (slow but steady), reaching critical mass.
5.  Probably stuff I'm missing.

Falcon-9 is about 3x cost reduction, and Starship will be another 10-40x on top of that.  That's definitely in the exponential growth area of the S-curve

So, getting around to the topic of energy, do we have such an exponential growth?  Kinda.  Not in amount of watts produced for sure - we are nearing the top of that s-curve.   But in the ability to produce those watts cheaply in space from solar power?  Absolutely.  It's now less than $1/watt of capacity to install solar power now on Earth, it was a lot more 2 decades ago.  Mix that with the  exponential drop in launch costs, and plenty of sunshine in space, and you have a new opportunity to grow the power produced by the human race.

And rather than beaming it down (which Elon says will never be efficient enough), we can just beam down the small few hundred megabytes of results from compute (AI, whatever), which is micropennies of energy.

So I think if one counts the net power produced by the human race, we will escape our current "top of the S curve is readily apparent" limits and go on another exponential binge.   Doing this in space for compute is good, we aren't taking away electricity from starving babies.

That's all well understood and agreed.

The question was whether the use of solar in space could grow to be comparable with the total solar insolation the Earth receives.

Because exponents will get you there, what's 100,000,000x between friends.

My point was, no.  It'll stop long before that, simply because Earth itself as a consumer is finite.

Maybe when we're all living in space it'll become relevant, but then the whole comparison to Earth's energy balance is irrelevant.

[Twark_Main]

The question was whether the use of solar in space could grow to be comparable with the total solar insolation the Earth receives.

Because exponents will get you there, what's 100,000,000x between friends.

My point was, no.  It'll stop long before that, simply because Earth itself as a consumer is finite.

Coal owners in the 1800s comforted themselves with similar hand-waving stories why the effects of CO2 could never change the total energy balance of Earth. Earth was equally finite then too.

When we first invented plastics, we never imagined that microplastic pollution could ever get so massive that it becomes an issue. Yet here we are.

I, for one, welcome our strange new "driving by looking out the front windshield" overlords.   


Also, if you think humans can't effect the total exergy (heat engine) of Earth, do a quick check on the current impact of a little thing called "agriculture" on the total primary energy (*exergy) available for the biosphere.

This isn't some theoretical / unimaginable scope and scale. We're applying major stress to this huge system already, and this would be on top of that.


Note that impact can greatly exceed the watts of power you generate as a human. Cutting down a rain forest and replacing it with solar panels comes with no guarantee that the solar panels use energy as effectively as the previous rain forest.   

[Vultur]

The question was whether the use of solar in space could grow to be comparable with the total solar insolation the Earth receives.

Because exponents will get you there, what's 100,000,000x between friends.

My point was, no.  It'll stop long before that, simply because Earth itself as a consumer is finite.

Coal owners in the 1800s comforted themselves with similar hand-waving stories why the effects of CO2 could never change the total energy balance of Earth. Earth was equally finite then too.

Not really. I don't think the idea was considered at all (even to be dismissed) in the 1800s. Svante Arrhenius write something brief about it on the early 1900s though. (But not as a problem. Not only was the amount of fuel burned far smaller then, the *second order effects* that cause the real problems weren't understood. A 1 or 2 degree rise would have sounded more beneficial than harmful in northern/western Europe.)

But that aside, there's a huge difference. Not only was world population way lower then, only a tiny proportion of it (pretty much Western Europe and the Northeast/Midwest US) was industrialized at all. So the accessible demand, and thus energy use, grew enormously over the course of the 20th century, as world population grew AND a much larger proportion of the world population was brought into the world economy.

That's not really happening any more. Sure not all the world population is participating in the global economy ... But with rural electrification in China and India, there aren't any *equivalently large* untapped markets.

This situation hasn't existed in at least 200 years; there's no real history to go by as to how that will interact with a technological/industrial economy, as the entire industrial age has been one of very rapid demand growth.

[Oersted]

Relevant:

https://twitter.com/elonmusk/status/2000331928109809757

[Twark_Main]

The question was whether the use of solar in space could grow to be comparable with the total solar insolation the Earth receives.

Because exponents will get you there, what's 100,000,000x between friends.

My point was, no.  It'll stop long before that, simply because Earth itself as a consumer is finite.

Coal owners in the 1800s comforted themselves with similar hand-waving stories why the effects of CO2 could never change the total energy balance of Earth. Earth was equally finite then too.

Not really. I don't think the idea was considered at all (even to be dismissed) in the 1800s. Svante Arrhenius write something brief about it on the early 1900s though.

(But not as a problem. Not only was the amount of fuel burned far smaller then, the *second order effects* that cause the real problems weren't understood. A 1 or 2 degree rise would have sounded more beneficial than harmful in northern/western Europe.)

Arrhenius's paper was in 1896.  I chose my dates carefully. 

Not that it would make a lick of difference, of course. "No no, it's totally different today. The difference is that we can see the problem coming, which is why we shouldn't do anything...."   

But that aside, there's a huge difference. Not only was world population way lower then, only a tiny proportion of it (pretty much Western Europe and the Northeast/Midwest US) was industrialized at all. So the accessible demand, and thus energy use, grew enormously over the course of the 20th century, as world population grew AND a much larger proportion of the world population was brought into the world economy.

That's not really happening any more. Sure not all the world population is participating in the global economy ... But with rural electrification in China and India, there aren't any *equivalently large* untapped markets.

I imagine that in 1896 you would've argued that everyone only "needs" an electric clothes iron and a toaster, so there's no way this industrial revolution thing has much steam left in it anyway.   

There doesn't seem to be a ceiling on quality-of-life. Those rural Indian and Chinese people aren't "done." They want all the same creature comforts, including those from AI, and they should have access to it. People usually get (rightfully) upset if you suggest otherwise.

This situation hasn't existed in at least 200 years; there's no real history to go by as to how that will interact with a technological/industrial economy, as the entire industrial age has been one of very rapid demand growth.

"And that's why I'm so confident in my prediction." 

But seriously, I agree this is an uncertain moment.  What we set into motion now will shape things for decades and even centuries to come. That's why I'm not unhappy to see a little extra "tread carefully" mindset among the people shaping that future.



 I hope we can eventually circle back to Reason #1. I drew the distinction because all the "real" economics is in there.

[InterestedEngineer]

I"m confused.  How does using googlewatts of power in space negatively affect the Earth at all?

The fact that we don't have to generate the power on Earth is a positive effect.  What's the negative?  (leaving aside AI armageddon predictions, let's stick with energy)

[Vultur]

Arrhenius's paper was in 1896.  I chose my dates carefully. 

Hmm, ok. Last time I looked i thought that was just the greenhouse effect in general, the connection to coal burning wasn't until a later publication. But I could well be wrong.

I imagine that in 1896 you would've argued that everyone only "needs" an electric clothes iron and a toaster, so there's no way this industrial revolution thing has much steam left in it anyway.   

There doesn't seem to be a ceiling on quality-of-life. Those rural Indian and Chinese people aren't "done." They want all the same creature comforts, including those from AI, and they should have access to it. People usually get (rightfully) upset if you suggest otherwise.

But the connection between energy use and quality of life isn't linear. In many cases, new technologies are more energy efficient - an electric car's motor might be 90%+ efficient at converting stored energy to motion whereas a gas car's engine is more like 30%. LED bulbs are more energy efficient than fluorescents which are more energy efficient than incandescents. Etc.

And I don't think current AI improves quality of life. IMO it is a net *detriment*. Google, and the Internet in general, has distinctly become less useful and reliable in the last four years.

And that's why demand growth or the lack thereof matters. So far, jobs displaced by technological advancement have been replaced by more jobs elsewhere in the economy. But in a low demand growth world, that may no longer be true - and Ai/automation may no longer increase quality of life, rather decrease it.

"And that's why I'm so confident in my prediction." 

I'm *not*. What I'm questioning is the apparent assumption that rapid growth *will* continue and the question is who wins out/makes money from it.

What I don't see a lot of talk about is the scenario where that ... Just doesn't happen. There's no "next industrial revolution", but also no catastrophic collapse.

[Vultur]

I"m confused.  How does using googlewatts of power in space negatively affect the Earth at all?

The fact that we don't have to generate the power on Earth is a positive effect.  What's the negative?  (leaving aside AI armageddon predictions, let's stick with energy)

It doesn't, other than jobs being displaced (which is possibly a larger concern in an overall low growth world than it would have been a couple decades ago).

My concern is about the investment in space AI being wasted if it turns out that the current huge AI demand is basically because it's free to most of the end users (or bundled as Copilot with Microsoft Office, etc) and once the real costs start being charged there turns out to be actually relatively little demand for AI.

[meekGee]

The question was whether the use of solar in space could grow to be comparable with the total solar insolation the Earth receives.

Because exponents will get you there, what's 100,000,000x between friends.

My point was, no.  It'll stop long before that, simply because Earth itself as a consumer is finite.

Coal owners in the 1800s comforted themselves with similar hand-waving stories why the effects of CO2 could never change the total energy balance of Earth. Earth was equally finite then too.

When we first invented plastics, we never imagined that microplastic pollution could ever get so massive that it becomes an issue. Yet here we are.

I, for one, welcome our strange new "driving by looking out the front windshield" overlords.   


Also, if you think humans can't effect the total exergy (heat engine) of Earth, do a quick check on the current impact of a little thing called "agriculture" on the total primary energy (*exergy) available for the biosphere.

This isn't some theoretical / unimaginable scope and scale. We're applying major stress to this huge system already, and this would be on top of that.


Note that impact can greatly exceed the watts of power you generate as a human. Cutting down a rain forest and replacing it with solar panels comes with no guarantee that the solar panels use energy as effectively as the previous rain forest.   

Yup but those are qualitative hand-wavy analogies.  They don't hold up to the 100,000,000x mismatch.

I thought world power consumption since the 1700s would be 100,000x.  Turns out it's "only" less than 100x.

Broadly, even though our great great grandfathers didn't think emissions would be the problem that we later figured it is, it does not follow that solar panels on Earth risk directly affecting Earth's energy balance, even if we don't think so today...  It's a word analogy, but not a logical argument.

Not even in 200 years.

[Twark_Main]

The question was whether the use of solar in space could grow to be comparable with the total solar insolation the Earth receives.

Because exponents will get you there, what's 100,000,000x between friends.

My point was, no.  It'll stop long before that, simply because Earth itself as a consumer is finite.

Coal owners in the 1800s comforted themselves with similar hand-waving stories why the effects of CO2 could never change the total energy balance of Earth. Earth was equally finite then too.

When we first invented plastics, we never imagined that microplastic pollution could ever get so massive that it becomes an issue. Yet here we are.

I, for one, welcome our strange new "driving by looking out the front windshield" overlords.   


Also, if you think humans can't effect the total exergy (heat engine) of Earth, do a quick check on the current impact of a little thing called "agriculture" on the total primary energy (*exergy) available for the biosphere.

This isn't some theoretical / unimaginable scope and scale. We're applying major stress to this huge system already, and this would be on top of that.


Note that impact can greatly exceed the watts of power you generate as a human. Cutting down a rain forest and replacing it with solar panels comes with no guarantee that the solar panels use energy as effectively as the previous rain forest.   

Yup but those are qualitative hand-wavy analogies.  They don't hold up to the 100,000,000x mismatch.

I thought world power consumption since the 1700s would be 100,000x.  Turns out it's "only" less than 100x.

Broadly, even though our great great grandfathers didn't think emissions would be the problem that we later figured it is, it does not follow that solar panels on Earth risk directly affecting Earth's energy balance, even if we don't think so today...  It's a word analogy, but not a logical argument.

Not even in 200 years.

The bold says you're still thinking linearly, not exponentially.

100x in 200 years? That's quite fast! That means 100,000x only takes 500 years. That's a moderately recent pub in certain parts of the world. 

[meekGee]

The question was whether the use of solar in space could grow to be comparable with the total solar insolation the Earth receives.

Because exponents will get you there, what's 100,000,000x between friends.

My point was, no.  It'll stop long before that, simply because Earth itself as a consumer is finite.

Coal owners in the 1800s comforted themselves with similar hand-waving stories why the effects of CO2 could never change the total energy balance of Earth. Earth was equally finite then too.

When we first invented plastics, we never imagined that microplastic pollution could ever get so massive that it becomes an issue. Yet here we are.

I, for one, welcome our strange new "driving by looking out the front windshield" overlords.   


Also, if you think humans can't effect the total exergy (heat engine) of Earth, do a quick check on the current impact of a little thing called "agriculture" on the total primary energy (*exergy) available for the biosphere.

This isn't some theoretical / unimaginable scope and scale. We're applying major stress to this huge system already, and this would be on top of that.


Note that impact can greatly exceed the watts of power you generate as a human. Cutting down a rain forest and replacing it with solar panels comes with no guarantee that the solar panels use energy as effectively as the previous rain forest.   

Yup but those are qualitative hand-wavy analogies.  They don't hold up to the 100,000,000x mismatch.

I thought world power consumption since the 1700s would be 100,000x.  Turns out it's "only" less than 100x.

Broadly, even though our great great grandfathers didn't think emissions would be the problem that we later figured it is, it does not follow that solar panels on Earth risk directly affecting Earth's energy balance, even if we don't think so today...  It's a word analogy, but not a logical argument.

Not even in 200 years.

The bold says you're still thinking linearly, not exponentially.

100x in 200 years? That's quite fast! That means 100,000x only takes 500 years. That's a moderately recent pub in certain parts of the world. 

No it does not, since exponential growth in reality doesn't persist like that. Something always kicks in.  Otherwise, each species of bacteria would have outweighed the planet already.

In our case, power consumption on earth can't increase again and again by the factor of 100x (30x actually but rounding up) that it did during the industrial revolution.

It might do so as humans go into space in the next centuries, but that wasn't your premise - you were talking about affection the energy balance of the earth.

It's just that not every extrapolation is true, you neednto always look wt underlying limitations, resource constraints, etc.

[Twark_Main]

The question was whether the use of solar in space could grow to be comparable with the total solar insolation the Earth receives.

Because exponents will get you there, what's 100,000,000x between friends.

My point was, no.  It'll stop long before that, simply because Earth itself as a consumer is finite.

Coal owners in the 1800s comforted themselves with similar hand-waving stories why the effects of CO2 could never change the total energy balance of Earth. Earth was equally finite then too.

When we first invented plastics, we never imagined that microplastic pollution could ever get so massive that it becomes an issue. Yet here we are.

I, for one, welcome our strange new "driving by looking out the front windshield" overlords.   


Also, if you think humans can't effect the total exergy (heat engine) of Earth, do a quick check on the current impact of a little thing called "agriculture" on the total primary energy (*exergy) available for the biosphere.

This isn't some theoretical / unimaginable scope and scale. We're applying major stress to this huge system already, and this would be on top of that.


Note that impact can greatly exceed the watts of power you generate as a human. Cutting down a rain forest and replacing it with solar panels comes with no guarantee that the solar panels use energy as effectively as the previous rain forest.   

Yup but those are qualitative hand-wavy analogies.  They don't hold up to the 100,000,000x mismatch.

I thought world power consumption since the 1700s would be 100,000x.  Turns out it's "only" less than 100x.

Broadly, even though our great great grandfathers didn't think emissions would be the problem that we later figured it is, it does not follow that solar panels on Earth risk directly affecting Earth's energy balance, even if we don't think so today...  It's a word analogy, but not a logical argument.

Not even in 200 years.

The bold says you're still thinking linearly, not exponentially.

100x in 200 years? That's quite fast! That means 100,000x only takes 500 years. That's a moderately recent pub in certain parts of the world. 

No it does not, since exponential growth in reality doesn't persist like that. Something always kicks in.  Otherwise, each species of bacteria would have outweighed the planet already.

In our case, power consumption on earth can't increase again and again by the factor of 100x (30x actually but rounding up) that it did during the industrial revolution.

It might do so as humans go into space in the next centuries, but that wasn't your premise - you were talking about affection the energy balance of the earth.

It's just that not every extrapolation is true, you neednto always look wt underlying limitations, resource constraints, etc.

Again, in earlier times you could've easily argued that "energy use can't grow further" because there isn't enough wood that grows in all the world's forests. We all know why that prediction was wrong (coal).

Similarly, saying "power consumption on earth can't increase" ignores the fact that we can use power in space to serve needs on Earth (which, reminder, is the subject of this thread).


How many times have people predicted the "end of growth?"  How many times have they been wrong, so far?   


The favorable economics of computing in space are independent of any particular AI or compute technology. This isn't like predicting "everyone will own a DeSoto automobile," this is like predicting "wheeled vehicles will be the dominant form of medium-distance transport."

[meekGee]

The question was whether the use of solar in space could grow to be comparable with the total solar insolation the Earth receives.

Because exponents will get you there, what's 100,000,000x between friends.

My point was, no.  It'll stop long before that, simply because Earth itself as a consumer is finite.

Coal owners in the 1800s comforted themselves with similar hand-waving stories why the effects of CO2 could never change the total energy balance of Earth. Earth was equally finite then too.

When we first invented plastics, we never imagined that microplastic pollution could ever get so massive that it becomes an issue. Yet here we are.

I, for one, welcome our strange new "driving by looking out the front windshield" overlords.   


Also, if you think humans can't effect the total exergy (heat engine) of Earth, do a quick check on the current impact of a little thing called "agriculture" on the total primary energy (*exergy) available for the biosphere.

This isn't some theoretical / unimaginable scope and scale. We're applying major stress to this huge system already, and this would be on top of that.


Note that impact can greatly exceed the watts of power you generate as a human. Cutting down a rain forest and replacing it with solar panels comes with no guarantee that the solar panels use energy as effectively as the previous rain forest.   

Yup but those are qualitative hand-wavy analogies.  They don't hold up to the 100,000,000x mismatch.

I thought world power consumption since the 1700s would be 100,000x.  Turns out it's "only" less than 100x.

Broadly, even though our great great grandfathers didn't think emissions would be the problem that we later figured it is, it does not follow that solar panels on Earth risk directly affecting Earth's energy balance, even if we don't think so today...  It's a word analogy, but not a logical argument.

Not even in 200 years.

The bold says you're still thinking linearly, not exponentially.

100x in 200 years? That's quite fast! That means 100,000x only takes 500 years. That's a moderately recent pub in certain parts of the world. 

No it does not, since exponential growth in reality doesn't persist like that. Something always kicks in.  Otherwise, each species of bacteria would have outweighed the planet already.

In our case, power consumption on earth can't increase again and again by the factor of 100x (30x actually but rounding up) that it did during the industrial revolution.

It might do so as humans go into space in the next centuries, but that wasn't your premise - you were talking about affection the energy balance of the earth.

It's just that not every extrapolation is true, you neednto always look wt underlying limitations, resource constraints, etc.

Again, in earlier times you could've easily argued that "energy use can't grow further" because there isn't enough wood that grows in all the world's forests. We all know why that prediction was wrong (coal).

Similarly, saying "power consumption on earth can't increase" ignores the fact that we can use power in space to serve needs on Earth (which, reminder, is the subject of this thread).


How many times have people predicted the "end of growth?"  How many times have they been wrong, so far?   


The favorable economics of computing in space are independent of any particular AI or compute technology. This isn't like predicting "everyone will own a DeSoto automobile," this is like predicting "wheeled vehicles will be the dominant form of medium-distance transport."

You're strawmaning.

I didn't say energy use can't grow.  I said it can't keep growing exponentially, to a factor of 100,000,000.

So the rest of the post is just poking holes in something I never said.

And my comments are all about the statement that we need to take PV to orbit because sooner or later PV use on Earth will affect the planet's energy balance.

Which btw you made with some qualifiers that made me think you're not really invested in that argument, so imagine my surprise...

[Vultur]

Again, in earlier times you could've easily argued that "energy use can't grow further" because there isn't enough wood that grows in all the world's forests. We all know why that prediction was wrong (coal).

Similarly, saying "power consumption on earth can't increase" ignores the fact that we can use power in space to serve needs on Earth (which, reminder, is the subject of this thread).


How many times have people predicted the "end of growth?"  How many times have they been wrong, so far?   

Supply vs demand though.

I agree it has historically been a very bad bet that technology won't increase supply to match demand, e.g. the mid 20th century fears of vast overpopulation famines.

Transition from wood to coal to oil to current energy mix increased supply, but demand was already increasing rapidly due to both rapid population growth and more and more of the world population being brought into the world economy.

Given current falling population growth rates, and the pace of industrialization in the highest population parts of the world such as China and India, and the fact that new technologies are often more energy efficient to serve the same needs (with current LLMs being a huge exception to that general trend, admittedly) it is entirely plausible that world growth in energy *demand* will end by the middle of this century or so.

I think datacenters can raise the peak but not extend the timing ... Either vastly more efficient ways to run "AI" in some sense (LLMs or not) will be found, or costs will increase sufficiently that  many current uses will be non viable. And I think one or the other will happen way before 2050.

If Space Datacenters do end up being cheaper ... Well they're still not free. Question is whether they can be cheap enough to be funded by advertising money and actual subscriptions. AI still needs to be supported by the rest of the economy.

That doesn't mean it's forever, but it does mean future demand growth would not be driven by a continuation of current trends. It would have to be by either a reversal of current trends (e.g. if total human population starts increasing again either due to cultural or socioeconomic changes or off-earth settlements growing on their own) or something entirely new (e.g. if super projects like an interstellar probe launching laser are built).

[meekGee]

Again, in earlier times you could've easily argued that "energy use can't grow further" because there isn't enough wood that grows in all the world's forests. We all know why that prediction was wrong (coal).

Similarly, saying "power consumption on earth can't increase" ignores the fact that we can use power in space to serve needs on Earth (which, reminder, is the subject of this thread).


How many times have people predicted the "end of growth?"  How many times have they been wrong, so far?   

Supply vs demand though.

I agree it has historically been a very bad bet that technology won't increase supply to match demand, e.g. the mid 20th century fears of vast overpopulation famines.

Transition from wood to coal to oil to current energy mix increased supply, but demand was already increasing rapidly due to both rapid population growth and more and more of the world population being brought into the world economy.

Given current falling population growth rates, and the pace of industrialization in the highest population parts of the world such as China and India, and the fact that new technologies are often more energy efficient to serve the same needs (with current LLMs being a huge exception to that general trend, admittedly) it is entirely plausible that world growth in energy *demand* will end by the middle of this century or so.

I think datacenters can raise the peak but not extend the timing ... Either vastly more efficient ways to run "AI" in some sense (LLMs or not) will be found, or costs will increase sufficiently that  many current uses will be non viable. And I think one or the other will happen way before 2050.

If Space Datacenters do end up being cheaper ... Well they're still not free. Question is whether they can be cheap enough to be funded by advertising money and actual subscriptions. AI still needs to be supported by the rest of the economy.

That doesn't mean it's forever, but it does mean future demand growth would not be driven by a continuation of current trends. It would have to be by either a reversal of current trends (e.g. if total human population starts increasing again either due to cultural or socioeconomic changes or off-earth settlements growing on their own) or something entirely new (e.g. if super projects like an interstellar probe launching laser are built).

And even if demand is there, why would there be exponential growth over a factor of hundreds of millions?

Musk is talking about a 100 GWatts/yr growth with Starship. The world right now is at 10 TWatt total.

He's talking about TWatts/yr later on with lunar sourced materials.

None of that remotely touches the kind of growth we're discussing.  What Musk describes is taking the Earth power generation capacity and multiplying it by 10.  Maybe by 100 if you're thinking each person on Earth will consume 100 kWatt of compute 24/7.  But not by 100 million.  Nobody thinks that's a thing.

Here: 10 GPeople x 1 kWatt is 10 TWatt, which is the current electric generation capacity of Earth.  So if each person consumes that much, we will have doubled Earth's demand. Let's loosely equate 10 TWatt with 1 TWatt/yr launch rate. (And 10 year life span}

To create a demand for 100 TWatt/yr, it means each person is eating up 100 kWatt of compute.  Or maybe we have 100B people on the planet and they each eat up 10kWatt of compute, 24/7.

Already these numbers are not realistic. 

But 100,000,000 is six OOMs even beyond that. It just doesn't make any sense.

[Robotbeat]

Total electricity of the earth is about 2.8TW right now. The 21TW figure is “primary energy” and includes all the waste heat. The 100GW/year is added solar power and would be equivalent to adding about 300-400GW per year in primary energy (electricity is responsible for around 40% of the total primary energy consumption of the world). If we assume that the 100GWe/year lasts 10-20 years on average, that’s about 1-2TW steady state, about equal to half to all the electricity generated by humanity today.

It is kind of annoying when public figures muddle this. A watt of electricity is worth about 3-4 Watts of primary energy.

Using primary energy is kind of a weird thing, though. I mean if we’re talking bioenergy, do we count the input energy of sunlight? Plants are only around 1% efficiency at converting that to chemical energy, so it would have the effect of inflating humanity’s energy use figure to count things this way.

[Robotbeat]

World average per capita energy consumption is 2500W primary energy. US per capita primary energy consumption is 4 times that, 10kW.

If you spend an hour flying a Learjet per day, that’s about 25 times that, 250kW.

I think there’s still quite a lot of room for global energy use to increase without more population growth. Waste heat will become a constraint even if it’s all carbon neutral.

[meekGee]

World average per capita energy consumption is 2500W primary energy. US per capita primary energy consumption is 4 times that, 10kW.

If you spend an hour flying a Learjet per day, that’s about 25 times that, 250kW.

I think there’s still quite a lot of room for global energy use to increase without more population growth. Waste heat will become a constraint even if it’s all carbon neutral.

How much room? 10x? 100x?  I'm ok with all those.  The future is wild and unpredictable.

It's the 100,000,000x I took issue with.

And yes, everything you said about energy is true, I figured PV comes to 1:1 comparison with electrical generation, but all those differences don't make up even one order of magnitude so I let it be.

[spacenut]

Well, the supply side of electricity can still grow.  The World Bank gave the Congo enough money to build 3 dams on the Congo River that would have put out enough electricity to power not only the Congo but two adjacent countries.  The Congo government blew the money.  Also, more nuclear power plants can be built, especially safer molten salt reactors. 

On the use side, most lighting in America has changed from incandescent bulbs to LED lighting.  Also appliances have become more energy efficient.  50 years ago, gas furnaces were only about 50% efficient.  Today they can be up to 95-98% efficient and heat pumps have become more efficient.  Homes have better insulation. 

Now, we can still produce more power on earth, but at what cost.  Cost will be the driver for in space solar power for use with AI or even beamed power.  Off world manufacturing may become cheaper with solar in space than building the manufacturing on earth. 

There are so many variables to be weighed in that we may not really know.  However, Musk somehow looks at all the alternatives.  He predicted Starlink over 10 years ago.  He pushed reusable rockets, now it seems everyone is trying to get in the game.  He made electric cars better and more affordable.  Even Amazon made internet shopping commonplace.  What and IBM in the late 1940's only thought the world would only need a total of about 3 computers. 

[Robotbeat]

World average per capita energy consumption is 2500W primary energy. US per capita primary energy consumption is 4 times that, 10kW.

If you spend an hour flying a Learjet per day, that’s about 25 times that, 250kW.

I think there’s still quite a lot of room for global energy use to increase without more population growth. Waste heat will become a constraint even if it’s all carbon neutral.

How much room? 10x? 100x?  I'm ok with all those.  The future is wild and unpredictable.

It's the 100,000,000x I took issue with.

And yes, everything you said about energy is true, I figured PV comes to 1:1 comparison with electrical generation, but all those differences don't make up even one order of magnitude so I let it be.

After 100x, you need a bunch of the energy to be generated and used in-space. That can be done. It will be done.

[TheRadicalModerate]

The idea, clearly, is that AI training tasks let you beam the value back to Earth (ie the finished trained model) without beaming the BTUs back to Earth (and vaporizing the planet, in the limiting growth case).

The problem is that training is only 10-20% of the compute load.  Inference latencies can be non-trivial, but I'd guess that a half second RTT is about all you can handle if you want to stay competitive.  So you might be able to use some kind of subsync orbit, well away from both LEO and those pesky protons.

[Robotbeat]

Depends on the model. Training can be 40% or more of total energy use, particularly if models become more sophisticated.

And quite a bit of AI usage could tolerant seconds of latency, such as deep research or video/image generation. Short, low-latency prompts don’t use much energy. Really long prompts or extensive code generation do.

I think short prompts will continue to be served very near point of use. But long research, “thinking” models, ensembles of multiple models, very long input and output, not to mention video and image generation already can take minutes to produce results and use tremendous energy to do so, and those can be offloaded to beyond LEO.

[DanClemmensen]

The idea, clearly, is that AI training tasks let you beam the value back to Earth (ie the finished trained model) without beaming the BTUs back to Earth (and vaporizing the planet, in the limiting growth case).

The problem is that training is only 10-20% of the compute load.  Inference latencies can be non-trivial, but I'd guess that a half second RTT is about all you can handle if you want to stay competitive.  So you might be able to use some kind of subsync orbit, well away from both LEO and those pesky protons.

On average, the actual LEO resource you are communicating with is approximately half an orbit away from you , which is 10,000 Km or about 60 ms round trip. A slightly lower orbit will have little effect.

[Robotbeat]

That sort of depends. I think some may end up being served directly from “regular” orbit Starlinks not on the terminator. That could reduce latency to like sub 10 millisecond.

Note I was looking at SOTA models and they tend to have latencies from 0.3 to 4s to first token. It’s not necessarily a problem to put them in like MEO or GSO, or even out to lunar orbit.

[DanClemmensen]

That sort of depends. I think some may end up being served directly from “regular” orbit Starlinks not on the terminator. That could reduce latency to like sub 10 millisecond.

Note I was looking at SOTA models and they tend to have latencies from 0.3 to 4s to first token. It’s not necessarily a problem to put them in like MEO or GSO, or even out to lunar orbit.

How are you serving a customer from directly overhead? unless every satellite has all the data needed, this does not work.

[Robotbeat]

That sort of depends. I think some may end up being served directly from “regular” orbit Starlinks not on the terminator. That could reduce latency to like sub 10 millisecond.

Note I was looking at SOTA models and they tend to have latencies from 0.3 to 4s to first token. It’s not necessarily a problem to put them in like MEO or GSO, or even out to lunar orbit.

How are you serving a customer from directly overhead? unless every satellite has all the data needed, this does not work.

Yea, the models aren’t that big and can fit in a 2t satellite like v3.

[TheRadicalModerate]

Eh. I don't really agree, but we'll have to see what happens ... And it might be a couple decades before the answer is clear.

What I'm more concerned about in the short term is an investment bubble burst bringing down space stuff along with everything else. This could happen *regardless* of the soundness of the underlying technology - the dot com bubble of the late 90s wrecked private space projects then, though both the Internet in general and satellite internet specifically ultimately did work out.

I've attached a semi-log chart of historical internet traffic, in average petabytes per month, by year.  Look around 2000, the time of the dot-com crash.  There's not even a blip.  That's because people were still building out hardware and using it pretty much up to its capacity, even though bad business models were getting wiped out.

I expect the same thing to happen with any putative AI bubble:  there are plenty of good applications, and they'll simply take over the compute resources of the bad apps.  Meanwhile, the software will get better and the applications more useful.  The computronium market is going to remain solid for decades.  If you deploy it, the models will come.

[TheRadicalModerate]

Depends on the model. Training can be 40% or more of total energy use, particularly if models become more sophisticated.

And quite a bit of AI usage could tolerant seconds of latency, such as deep research or video/image generation. Short, low-latency prompts don’t use much energy. Really long prompts or extensive code generation do.

I think short prompts will continue to be served very near point of use. But long research, “thinking” models, ensembles of multiple models, very long input and output, not to mention video and image generation already can take minutes to produce results and use tremendous energy to do so, and those can be offloaded to beyond LEO.

I do think there are going to be applications that are incredibly latency sensitive:  driving, autonomous robotics, AI-assisted teleoperation.  But there are also going to be things that read batches of medical imaging and return results in a few minutes.  Same thing with research AIs that are looking for signal in reams of experimental noise, or surveillance systems that need sub-5-second response, but are fine up to that limit.  You can put stuff at L4/L5 and host those kinds of apps.

[DanielW]

That sort of depends. I think some may end up being served directly from “regular” orbit Starlinks not on the terminator. That could reduce latency to like sub 10 millisecond.

Note I was looking at SOTA models and they tend to have latencies from 0.3 to 4s to first token. It’s not necessarily a problem to put them in like MEO or GSO, or even out to lunar orbit.

Tool use is becoming more and more important to ground AI in reality. Reading and editing code and documentation from the user's computer on the fly. That isn't particularly latency tolerant. You could get clever with batching for text i/o but you'll want to be fairly real-time once AI starts driving your CAD interface for you.

[Vultur]

Eh. I don't really agree, but we'll have to see what happens ... And it might be a couple decades before the answer is clear.

What I'm more concerned about in the short term is an investment bubble burst bringing down space stuff along with everything else. This could happen *regardless* of the soundness of the underlying technology - the dot com bubble of the late 90s wrecked private space projects then, though both the Internet in general and satellite internet specifically ultimately did work out.

I've attached a semi-log chart of historical internet traffic, in average petabytes per month, by year.  Look around 2000, the time of the dot-com crash.  There's not even a blip.  That's because people were still building out hardware and using it pretty much up to its capacity, even though bad business models were getting wiped out.

I expect the same thing to happen with any putative AI bubble:  there are plenty of good applications, and they'll simply take over the compute resources of the bad apps.  Meanwhile, the software will get better and the applications more useful.  The computronium market is going to remain solid for decades.  If you deploy it, the models will come.

Internet use may not have dropped, but Teledesic and various reusable private launch attempts of the late 1990s *did* die. *That's* my concern regarding a bubble here, the economics of AI messing up the economics of space. (Especially if it affects the whole economy, which seems reasonably likely.)

Satellite internet did show up again (OneWeb, Starlink, Amazon LEO) but it took 20+ years.

"Bubble" issues can, and historically do, cause problems even when the underlying technology turns out to be perfectly sound.

World average per capita energy consumption is 2500W primary energy. US per capita primary energy consumption is 4 times that, 10kW.

If you spend an hour flying a Learjet per day, that’s about 25 times that, 250kW.

I think there’s still quite a lot of room for global energy use to increase without more population growth. Waste heat will become a constraint even if it’s all carbon neutral.

There's room for global energy use to increase, but not to the point where waste heat (as opposed to greenhouse gases) becomes relevant. There's probably a point of diminishing returns where more energy use doesn't really increase quality of life, it's just wasteful.

Total energy use for Americans is about 10kW/person yeah, but if combustion engines were replaced with electrics (and nothing else changed) that would drop a lot. The US is also higher energy use for same standard of living than a lot of the developed world because it's larger and less population dense = longer transportation distances etc.

But assume all that efficiency is replaced by increasd AI demand, so it's still 10 kW/person for a peak late 21st century world population of about 10B = 100TW. That's not a lot compared to solar energy input. And if most of it is from solar power (as seems likely in a post fossil fuel world) there might not be much "extra" waste heat at all.

[InterestedEngineer]

It's already happening

https://www.geekwire.com/2025/starcloud-power-training-ai-space/

[Robotbeat]

100TW extra waste heat is actually significant. It’s comparable to the heating from CO2 from the 1-2TWe average of fossil fuel power plants.

[meekGee]

It's already happening

https://www.geekwire.com/2025/starcloud-power-training-ai-space/

He does confirm what we knew, that the compute hardware itself is the most expensive component.

The "power" play is mis-named.  It's not about providing cheaper power, since not only is the hardware the most important part, it'll get more expensive when made space-worthy.

The true play is about siting and deployment.

For all the King's ransom, nobody can deploy 100 GWatt/yr on the planet. That's the long and short of it.

This (and similar) proposal is about a direct deployment pipeline from fab to site, fully under the deployer's control. Power's included, but it's not the cost of it, it's the availability of it.

[TheRadicalModerate]

Internet use may not have dropped, but Teledesic and various reusable private launch attempts of the late 1990s *did* die. *That's* my concern regarding a bubble here, the economics of AI messing up the economics of space. (Especially if it affects the whole economy, which seems reasonably likely.)

You've mangled the analogy.  The internet bubble had nothing to do with space.  My point was that the people who invested in internet capacity (fiber, routers, switches, LAN equipment, etc.) did just fine.  What we're talking about here is exactly that capacity, not the applications that run on it.  Models and apps will come and go.  Computronium and its power supply are forever (where "forever" is defined as the point where the asset's useful life ends).

[Vultur]

Internet use may not have dropped, but Teledesic and various reusable private launch attempts of the late 1990s *did* die. *That's* my concern regarding a bubble here, the economics of AI messing up the economics of space. (Especially if it affects the whole economy, which seems reasonably likely.)

You've mangled the analogy.  The internet bubble had nothing to do with space. 

I'm pretty sure it did? Or are you saying the Iridium / Globalstar bankruptcies & failure of Teledesic were unrelated to the dotcom bubble despite happening around the same time?

Rotary Rocket, and i think other private launch efforts of that era, were attempting to serve that market. So those bankruptcies affected early private orbital launch attempts too.

[InterestedEngineer]

This (and similar) proposal is about a direct deployment pipeline from fab to site, fully under the deployer's control. Power's included, but it's not the cost of it, it's the availability of it.

I seem to recall something about supply/demand curves in Econ 101 class.

Are you saying the price won't go up if the demand is far higher than the availability?

[meekGee]

This (and similar) proposal is about a direct deployment pipeline from fab to site, fully under the deployer's control. Power's included, but it's not the cost of it, it's the availability of it.

I seem to recall something about supply/demand curves in Econ 101 class.

Are you saying the price won't go up if the demand is far higher than the availability?

Price and cost are not the same.  We're talking basic costs here.  The price may end up above or below cost.

Also the econ 101 model is just a simplified concept to illustrate a point. Don't take it literally.

[InterestedEngineer]

No, what I'm saying is this sentence makes no sense to me:

This (and similar) proposal is about a direct deployment pipeline from fab to site, fully under the deployer's control. Power's included, but it's not the cost of it, it's the availability of it.

If there's no availability of power, and there's demand, the cost goes up until either the demand drops or the availability increases because someone built more power plants (and natural gas wells, etc etc)

Statically analyzing the current cost of compute vs. power also makes no sense to me.  If power triples in costs, the ratio changes.  Someone will either further optimize the power usage of that compute (hello Apple and their power sipping ARM CPUs with neural net accelerators), build more power plants, or the ROI will be so amazing for the rich that babies will starve.  (the "elastic" in "elastic demand" can sometimes mean the innocent bear the cost).  Oh and a possibility of an AI market crash because the ROI is terrible if the power costs triple. Ask Germany about that with their $0.40/kwHr "green" energy.

When you build AI in space, you increase the supply of power (though at a higher cost of hardware), and therefore your power costs get capped both on earth and in space.   Somewhere in the fuzzy middle we hope the babies don't starve, I get 1 second responses on 200K size token contexts for my code agent (for example), and everyone is mostly happy.

In general in human history the more power we can utilize the less starvation we have and the better the civilization, let's hope the trend for more power continues.  If that means more power is used in space so much the better.  We get lots of beautiful land with no ugly windmills, solar panels, or steam belching nuke plants, or smoke belching coal plants.

[Twark_Main]

No, what I'm saying is this sentence makes no sense to me:

This (and similar) proposal is about a direct deployment pipeline from fab to site, fully under the deployer's control. Power's included, but it's not the cost of it, it's the availability of it.

If there's no availability of power, and there's demand, the cost goes up until either the demand drops or the availability increases because someone built more power plants (and natural gas wells, etc etc)

Statically analyzing the current cost of compute vs. power also makes no sense to me.  If power triples in costs, the ratio changes.  Someone will either further optimize the power usage of that compute (hello Apple and their power sipping ARM CPUs with neural net accelerators), build more power plants, or the ROI will be so amazing for the rich that babies will starve.  (the "elastic" in "elastic demand" can sometimes mean the innocent bear the cost).  Oh and a possibility of an AI market crash because the ROI is terrible if the power costs triple. Ask Germany about that with their $0.40/kwHr "green" energy.

When you build AI in space, you increase the supply of power (though at a higher cost of hardware), and therefore your power costs get capped both on earth and in space.   Somewhere in the fuzzy middle we hope the babies don't starve, I get 1 second responses on 200K size token contexts for my code agent (for example), and everyone is mostly happy.

In general in human history the more power we can utilize the less starvation we have and the better the civilization, let's hope the trend for more power continues.  If that means more power is used in space so much the better.  We get lots of beautiful land with no ugly windmills, solar panels, or steam belching nuke plants, or smoke belching coal plants.

They're both true, but the limitation here is "how much AI can be deployed with a factory making XYZ solar panels per year," not "how much money is available to buy solar panels."

The dog wags the tail, not the other way around. In this case, factory capacity is the dog and price is the tail.

[meekGee]

No, what I'm saying is this sentence makes no sense to me:

This (and similar) proposal is about a direct deployment pipeline from fab to site, fully under the deployer's control. Power's included, but it's not the cost of it, it's the availability of it.

If there's no availability of power, and there's demand, the cost goes up until either the demand drops or the availability increases because someone built more power plants (and natural gas wells, etc etc)

Statically analyzing the current cost of compute vs. power also makes no sense to me.  If power triples in costs, the ratio changes.  Someone will either further optimize the power usage of that compute (hello Apple and their power sipping ARM CPUs with neural net accelerators), build more power plants, or the ROI will be so amazing for the rich that babies will starve.  (the "elastic" in "elastic demand" can sometimes mean the innocent bear the cost).  Oh and a possibility of an AI market crash because the ROI is terrible if the power costs triple. Ask Germany about that with their $0.40/kwHr "green" energy.

When you build AI in space, you increase the supply of power (though at a higher cost of hardware), and therefore your power costs get capped both on earth and in space.   Somewhere in the fuzzy middle we hope the babies don't starve, I get 1 second responses on 200K size token contexts for my code agent (for example), and everyone is mostly happy.

In general in human history the more power we can utilize the less starvation we have and the better the civilization, let's hope the trend for more power continues.  If that means more power is used in space so much the better.  We get lots of beautiful land with no ugly windmills, solar panels, or steam belching nuke plants, or smoke belching coal plants.

Ok one note about the supply/demand model.

It assumes that both suppliers and consumers are ignorant of economics, that they're essentially dumb and interchangeable.  It also assumes all dependencies happen in real time, there's no planning or memory, psychology, or everything else that actually drives an economy.

We have here a single largest buyer who's also a potential (largest) supplier and a planning horizon of 20 years. So you can't boil everything down to Supply/Demand.

[InterestedEngineer]

We'll have cool AI satellites in space generating Gigawatts of power, and your average person will use it for this

[Twark_Main]

The "power" play is mis-named.  It's not about providing cheaper power, since

 ...

the [compute] hardware is the most [expensive] part

Bad reasoning. That would be like saying "the guy who drives a water powered car [sic] doesn't spend much on gas, so saving money on gas can't be the advantage." 

The real question is how much would a terrestrial data centers need to pay for that same power and cooling, over its lifespan?

The "power" play is mis-named.  It's not about providing cheaper power, since

...

it'll get more expensive when made space-worthy.

I wouldn't take it for granted that this will always be true.

Most of the cost of terrestrial solar isn't the PV cells. It's the wind-loading structure. This requires unsexy yet costly components like strong metal frames, cover glass, steel mounting hardware, and concrete footers. Plus the installation and/or assembly of all that (on-site being especially costly).

Space-based power can delete or greatly downside all those parts. The panels are folded during launch so you can carry that load without large structures,  You use the same modern high-efficiency silicon PV modules that are used in terrestrial solar (ala Starlink). Your biggest loads occur when it's still compact and folded, so carrying even those largest loads requires a lot less mass. There's zero on-site labor or custom engineering work, it "installs" (deploys) itself autonomously after launch.

In theory, given attainably cheap launch prices, there's nothing that says 1 watt of PV capacity in space can't be cheaper than 1 watt of terrestrial PV.


On top of that, of course, you need ~5x fewer panels and zero batteries to achieve 24 hour power. So space PV doesn't even require cost parity.

[meekGee]

The "power" play is mis-named.  It's not about providing cheaper power, since

 ...

the [compute] hardware is the most [expensive] part

Bad reasoning. That would be like saying "the inventor of the water powered car [sic] doesn't budget much money for gas, so saving money on gas can't be the advantage." 

The real question is how much would a terrestrial data centers need to pay for that same power and cooling, over its lifespan?

The "power" play is mis-named.  It's not about providing cheaper power, since

...

it'll get more expensive when made space-worthy.

I wouldn't take it for granted that this will always be true.

Most of the cost of terrestrial solar isn't the PV cells. It's the wind-loading structure. This requires unsexy yet costly components like strong metal frames, cover glass, steel mounting hardware, and concrete footers. Plus the installation and/or assembly of all that (on-site being especially costly).

Space-based power can delete or greatly downside all those parts. The panels are folded during launch so you can carry that load without large structures,  You use the same modern high-efficiency silicon PV modules that are used in terrestrial solar (ala Starlink). Your biggest loads occur when it's still compact and folded, so carrying even those largest loads requires a lot less mass. There's zero on-site labor or custom engineering work, it "installs" (deploys) itself autonomously after launch.

In theory, given attainably cheap launch prices, there's nothing that says 1 watt of PV capacity in space can't be cheaper than 1 watt of terrestrial PV.


On top of that, of course, you need ~5x fewer panels and zero batteries to achieve 24 hour power. So space PV doesn't even require cost parity.

I thought we did the math early on and using terrestrial power cost numbers, power didn't come close to the cost of the hardware. (My argument was that because of this, even if space power is free, it still didn't justify going to orbit)

And yeah ok maybe space hardware will beocme cheaper, but it's a race.  More radiators, cost of launch, cost of no-maintenance vs. mass reduction and maybe other redictions I can think of.

That's why for the longest time I thought this idea doesn't hold water financially.

Then came the understanding that it's not cost.  For all the King's ransom, nobody can give you 100 GWatt/yr. The ultrascalers don't have a solution for that currently.

[billh]

I thought we did the math early on and using terrestrial power cost numbers, power didn't come close to the cost of the hardware. (My argument was that because of this, even if space power is free, it still didn't justify going to orbit)

And yeah ok maybe space hardware will beocme cheaper, but it's a race.  More radiators, cost of launch, cost of no-maintenance vs. mass reduction and maybe other redictions I can think of.

That's why for the longest time I thought this idea doesn't hold water financially.

Then came the understanding that it's not cost.  For all the King's ransom, nobody can give you 100 GWatt/yr. The ultrascalers don't have a solution for that currently.

Permits and regulation, local resistance to land use, water consumption, transmission lines...

There's a whole list of things this avoids. But it's still amazing to think it might actually be economical.

[TheRadicalModerate]

Where are we on high-power infrared lasers?  Last I checked, the reason for using microwaves in space solar power was not only to penetrate the atmosphere but also because that was where the power transmission tech was mature.

In the long run, it's a lot more efficient to disaggregate the power generation from the power consumption.  But it requires using lasers to limit diffraction between the source and destination.

[alang]

This is probably a lot about Musk needing to create a market for space launch, as he did with StarLink, in order to justify more launches and bring down the cost to orbit per kilo.

[TheRadicalModerate]

This is probably a lot about Musk needing to create a market for space launch, as he did with StarLink, in order to justify more launches and bring down the cost to orbit per kilo.

You're implying that Starlink was created to create a market for F9 launch, which isn't true.  Starlink was created to leverage the F9 platform to create a cash cow.

If there's an orbital computronium business to be had, it's because it's highly profitable, using Starship as a mechanism to drive down its costs, not the other way around.

[TheRadicalModerate]

There's a whole list of things this avoids. But it's still amazing to think it might actually be economical.

I do wonder when (not if) people will start freaking out about the amount of NOx and water high launch cadences put into the upper stratosphere and mesosphere.  I also wonder whether their freakouts will be justified.

Space economies make a lot more sense when the bulk of materials they use come from off-planet.  If your economy requires tens of thousands of launches from Earth to support, you're doing something wrong.

[meekGee]

This is probably a lot about Musk needing to create a market for space launch, as he did with StarLink, in order to justify more launches and bring down the cost to orbit per kilo.

It's both.

Musk actually delivered in establishing a new industry (where others have failed before)

Of course he did that because he saw he was going to have a ton of launch capacity, but the point is that he built a real application, not some fake hype trend.

Same will happen with AI.  Obviously the impetus was "hey I have spare launch capacity 18 months out of every 24", but it's a real solution to a real problem.

[Ludus]

https://spacenews.com/space-force-offers-new-vandenberg-launch-site/

On cue, a Vandenberg RFI for a new SLC-14 with details that pretty much only match Starship.

A full Starship launch complex at Vandenberg is a necessity for Sun synchronous orbit AI sat launching at scale and would also support Starlink/Starshield V3.

Starbase TX and KSC/Canaveral Starship launch each have their own factory complex building Starship/Superheavy. How would it work at Vandenberg?

[catdlr]

https://spacenews.com/space-force-offers-new-vandenberg-launch-site/

On cue, a Vandenberg RFI for a new SLC-14 with details that pretty much only match Starship.

A full Starship launch complex at Vandenberg is a necessity for Sun synchronous orbit AI sat launching at scale and would also support Starlink/Starshield V3.

Starbase TX and KSC/Canaveral Starship launch each have their own factory complex building Starship/Superheavy. How would it work at Vandenberg?

On cue, Chris answers your question.

Starship Barge identified ahead of Starbase to KSC transports
written by Chris Bergin January 1, 2026

[DanClemmensen]

https://spacenews.com/space-force-offers-new-vandenberg-launch-site/

On cue, a Vandenberg RFI for a new SLC-14 with details that pretty much only match Starship.

A full Starship launch complex at Vandenberg is a necessity for Sun synchronous orbit AI sat launching at scale and would also support Starlink/Starshield V3.

Starbase TX and KSC/Canaveral Starship launch each have their own factory complex building Starship/Superheavy. How would it work at Vandenberg?

Speculation:

Send a booster and a backup booster to Vandenberg via barge. It's about a three-week trip through the Panama Canal. Send a replacement when/if a booster catch fails.

Send Ships by launching them from Starbase or KSC and catching them at Vandenberg. Only works for reusable Ships.

[spacenut]

There are several problems with terrestrial power.  You have location, transmission lines, private property, environmental concerns, type of power plants to generate the needs, property to buy for location.  Then you have federal, state, and local permitting to deal with.  All this is very time consuming and can be costly.   

Space locations have only a few; launch permits, location in space, and a reusable rocket system to install it.  No transmission lines, solar is the power being produced, no private property rights in space to deal with.  No state or local permitting involved. 

It makes sense to put AI servers in space. 

[catdlr]

Posting this here (too) as Hegseft's speech is heavy on AI

https://twitter.com/elonmusk/status/2010873739349823764