M&A: xAI, A SpaceX Company

[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.