M&A: xAI, A SpaceX Company

[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.