M&A: xAI, A SpaceX Company

[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.

[meekGee]

Related: https://x.com/StockSavvyShay/status/2001391832241181103

Elon Musk told xAI staff that winning the AI race comes down to surviving the next 2–3 years & scaling compute, power & capital faster than anyone else.

Key points from the meeting:
• xAI could achieve AGI as early as 2026
• xAI has access to ~$30B per year in funding
• Grok 5 carries ~10% probability of achieving AGI
• Optimus humanoid robots could eventually operate data centers
• Plans to scale from ~200k GPUs to ~1M GPUs across its Colossus data centers
• Long-term concepts include data centers in space & Mars-related infrastructure

This is when I find myself a little bit on Vultur's side of the argument...

Musk is probably the greatest industrialist that ever lived, but when it comes to having a handle on future computer tech, sometimes I just don't know ...

Like some 10 years ago, when he said his engineers are using motion capture to do CAD like Tony Stark does, or Tesla's ever imminent self-driving capabilities.

I mean LLMs caught me off guard and I'm super impressed, but it's also clear there's no I in that AI.  So...  GAI around the corner?  I dunno.

[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.

[Coastal Ron]

...
I mean LLMs caught me off guard and I'm super impressed, but it's also clear there's no I in that AI.  So...  GAI around the corner?  I dunno.

There is an NSF thread called "What Billionaire Tech CEOs Get Wrong About The Future, with Adam Becker", and in a post I made I said:

...So for me, I don't blindly accept what any CEO says, because they have vested interests in specific outcomes, and they will use the power of their position to influence their potential customers and their competitors.

Elon Musk has a vested interest in everyone being excited about xAI, Optimus, Tesla FSD, and now SpaceX data centers in space. With SpaceX in the past, there was no direct financial connection between SpaceX success with reusability and me, but with AI there is because Elon Musk wants to take SpaceX public, and he is relying on the public to help not only fund the company, but increase his financial wealth.

So when someone has vested interests in specific outcomes, I don't tend to 100% believe them. And since nobody has been able to come up with a definition of AGI (Artificial General Intelligence) that can be measured, Musk, Altman, and everyone else can promise whatever they want.

[meekGee]

...
I mean LLMs caught me off guard and I'm super impressed, but it's also clear there's no I in that AI.  So...  GAI around the corner?  I dunno.

There is an NSF thread called "What Billionaire Tech CEOs Get Wrong About The Future, with Adam Becker", and in a post I made I said:

...So for me, I don't blindly accept what any CEO says, because they have vested interests in specific outcomes, and they will use the power of their position to influence their potential customers and their competitors.

Elon Musk has a vested interest in everyone being excited about xAI, Optimus, Tesla FSD, and now SpaceX data centers in space. With SpaceX in the past, there was no direct financial connection between SpaceX success with reusability and me, but with AI there is because Elon Musk wants to take SpaceX public, and he is relying on the public to help not only fund the company, but increase his financial wealth.

So when someone has vested interests in specific outcomes, I don't tend to 100% believe them. And since nobody has been able to come up with a definition of AGI (Artificial General Intelligence) that can be measured, Musk, Altman, and everyone else can promise whatever they want.

I'm very conflicted on this.  And it bothers me, simce it's a very impressive topic.  This is not social networking or krypto, it's actually stuff that matters.

I spend a lot of time with chatGPT.  I'm amazed at the level of conversation it can hold, and yet at the same time still be an automaton.

Just saying, Musk's superpowers are a bit shaky when it comes to computing, and I think it's a genuine limitation, not a cynical ploy.

OTOH he does have a lot of insight into things that make him look... delusional... but then actually hold water.

We'll see, I guess.

[thespacecow]

Related: https://x.com/StockSavvyShay/status/2001391832241181103

Elon Musk told xAI staff that winning the AI race comes down to surviving the next 2–3 years & scaling compute, power & capital faster than anyone else.

Key points from the meeting:
• xAI could achieve AGI as early as 2026
• xAI has access to ~$30B per year in funding
• Grok 5 carries ~10% probability of achieving AGI
• Optimus humanoid robots could eventually operate data centers
• Plans to scale from ~200k GPUs to ~1M GPUs across its Colossus data centers
• Long-term concepts include data centers in space & Mars-related infrastructure

This is when I find myself a little bit on Vultur's side of the argument...

Musk is probably the greatest industrialist that ever lived, but when it comes to having a handle on future computer tech, sometimes I just don't know ...

Like some 10 years ago, when he said his engineers are using motion capture to do CAD like Tony Stark does, or Tesla's ever imminent self-driving capabilities.

I mean LLMs caught me off guard and I'm super impressed, but it's also clear there's no I in that AI.  So...  GAI around the corner?  I dunno.

I don't see how he has trouble with computer tech, he got his first $200M from internet after all. The Tesla FSD delays are just Elon Time, they're making good progress with latest releases. He's hardly the only person who under-estimated the time needed for advancing AI, Geoffrey Hinton - one of the godfathers of deep learning - famously said in 2016 that we should stop training radiologists since AI will replace them in 5 years, which of course turns out to be completely wrong.

Elon foresaw the potential of AI very early, he's an early investor in DeepMind and tried to buy it before Google.

So AGI probably won't be as fast as he predicted, but neither is Starship, so this is par for the course.

[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.