Moving The Cloud to orbit

[Vultur]

They do have a website
https://www.orbitai.global/

It claims a "Genesis" mission November 29 (with a countdown clock). Farther down it says "Launch partner: Galactic Energy". And "First satellite with onboard blockchain wallet performing the world's first in-orbit blockchain transaction signature."

This feels like a pile of techy buzzwords.

The DeStarAI section also has some bizarre claims like "vacuum provides natural cooling" and "Compute Capacity: ∞ TFLOPS".

The countdown clock is now to December 10.

Nextspaceflight.com doesn't show a Galactic Energy launch on December 10... Though there is a CAS Space Kinetica 1 with unknown payload. *Shrug*

[Twark_Main]

Using AI just for the sake of using AI isn't valuable. It only makes sense if it's cheaper than doing the same thing without AI, or gives real benefits that aren't otherwise attainable.

"Otherwise attainable" at what price?

Current AI is cripplingly flawed, a technology in its infancy. But I suspect it won't be that way for too much longer.

"At what price" is exactly the question.

Right now, given the amount of investment money going toward AI, end users are generally not paying the 'real' cost of it.

If vastly more energy/hardware efficient means are found, of course that changes. But vastly less energy use means the purpose of moving it to orbit goes away.

I think that depends on whether "vastly less" means 4x or 100x.  I suspect it's closer to the former.

Steam engines were around 5-10% before we developed a theory of heat engines, whereas the absolute limit is closer to 40-60%. That's not 100x.

I think "railgun" is being used here as a catch-all for any electric linear accelerator tech, as opposed to rocket launches, which when you pencil it out clearly wouldn't work.

Coil guns have major problems too... Is there any linear accelerator that would clearly be practical at multiple km/s for long enough lifetimes to compete with (say) $10-20/kg Earth launch costs?

If only we had a company with demonstrated success in blue-sky engineering space technologies that have never existed before...   

[Vultur]

Using AI just for the sake of using AI isn't valuable. It only makes sense if it's cheaper than doing the same thing without AI, or gives real benefits that aren't otherwise attainable.

"Otherwise attainable" at what price?

Current AI is cripplingly flawed, a technology in its infancy. But I suspect it won't be that way for too much longer.

"At what price" is exactly the question.

Right now, given the amount of investment money going toward AI, end users are generally not paying the 'real' cost of it.

If vastly more energy/hardware efficient means are found, of course that changes. But vastly less energy use means the purpose of moving it to orbit goes away.

I think that depends on whether "vastly less" means 4x or 100x.  I suspect it's closer to the former.

Steam engines were around 5-10% before we developed a theory of heat engines, whereas the absolute limit is closer to 40-60%. That's not 100x.

I don't think this is like engines at all, though.

A human brain is something like 20W (~20-25% of basal metabolic rate consumption of 80-100W).

So "intelligence" doesn't have a physical requirement to be crazy power hungry.

Now what current "AI" systems do is nothing at all like what a human brain does. But if the hope is a true general AI... OTOH maybe that's just not possible (brains are just not electronic computers, and the two radically different systems aren't interchangeable) and a lot of the more inflated hopes for AI are just never going to happen at all.

If only we had a company with demonstrated success in blue-sky engineering space technologies that have never existed before...   

There's a difference though. SpaceX has very efficiently implemented and scaled things that were impractical before (reusability, subcooled propellant, large scale low latency satellite internet) to achieve the world's highest launch rate. People questioned whether these things were economical, but I don't think anyone seriously thought in say 2002 that they weren't physically possible, just that the market wouldn't support the cost. Some of these things had been attempted in the 90s but failed, partly due to the dotcom bubble.

The issues with coilguns may be qualitatively worse, in the sense that "the materials you need just may not exist".

[Twark_Main]

Using AI just for the sake of using AI isn't valuable. It only makes sense if it's cheaper than doing the same thing without AI, or gives real benefits that aren't otherwise attainable.

"Otherwise attainable" at what price?

Current AI is cripplingly flawed, a technology in its infancy. But I suspect it won't be that way for too much longer.

"At what price" is exactly the question.

Right now, given the amount of investment money going toward AI, end users are generally not paying the 'real' cost of it.

If vastly more energy/hardware efficient means are found, of course that changes. But vastly less energy use means the purpose of moving it to orbit goes away.

I think that depends on whether "vastly less" means 4x or 100x.  I suspect it's closer to the former.

Steam engines were around 5-10% before we developed a theory of heat engines, whereas the absolute limit is closer to 40-60%. That's not 100x.

I don't think this is like engines at all, though.

A human brain is something like 20W (~20-25% of basal metabolic rate consumption of 80-100W).

So "intelligence" doesn't have a physical requirement to be crazy power hungry.

I expect this is mostly because of hardware power inefficiencies, not software inefficiencies. But this is an open question AFAIK.

If only we had a company with demonstrated success in blue-sky engineering space technologies that have never existed before...   

There's a difference though. SpaceX has very efficiently implemented and scaled things that were impractical before (reusability, subcooled propellant, large scale low latency satellite internet) to achieve the world's highest launch rate. People questioned whether these things were economical, but I don't think anyone seriously thought in say 2002 that they weren't physically possible, just that the market wouldn't support the cost. Some of these things had been attempted in the 90s but failed, partly due to the dotcom bubble.

They did. I was there.

The issues with coilguns may be qualitatively worse, in the sense that "the materials you need just may not exist".

If you have a reason to believe that (as Elon says) "success is not one of the possible outcomes," then perhaps you could articulate it?

[Tywin]

Interesting tweet:

https://twitter.com/CJHandmer/status/1997906033168330816

[Vultur]

If you have a reason to believe that (as Elon says) "success is not one of the possible outcomes," then perhaps you could articulate it?

Depends on the definition of success.

I don't doubt that with sufficient money and development time, a railgun or similar system able to launch payloads from the Moon to lunar escape velocity could be built and would work if built.

What I do doubt is that in an environment with sufficiently cheap launch to create a space industry large enough to justify it, that a railgun / mass driver system could compete on cost with that cheap Earth launch, once full system costs are taken into account.

In this kind of scenario (Starship launching a megaton per year) Earth launch costs will necessarily be very very very cheap (if they're not, there won't be money to launch that much). Maybe $20/kg in LEO, $50/kg in cislunar space?

Let's say that the Lunar mass driver operates with zero human oversight or maintenance, so cost once built is effectively zero (hardware amortization only). Let's also say that all the industry needed to make the payloads the mass driver is launching is already paid for by a separately existing, non SpaceX moonbase.

Given these incredibly favorable (I'd say completely unrealistic) assumptions, the lunar mass driver only makes sense if (its cost divided by its lifetime payload) is less than the cost to launch payloads from Earth.

If the lunar mass driver has a lifetime of 10,000 uses and launches 1000kg payloads at a time... It has to cost less than $500M to compete with $50/kg launch costs. Less than $5B if it has a 100,000 launch lifetime or launches 10 ton payloads. Etc.

To have any chance of making this work economically you need incredibly long rail lifetimes (or whatever components you use). That's what I'm saying may not be within the performance of existing materials, not making the railgun work in the first place.

[StraumliBlight]

Aetherflux Announces Orbital Data Center; Targets Q1 2027 [Dec 9]

Today, Aetherflux announced a Q1 2027 target for its first orbital data center satellite, which leverages solar power in space to address the massive energy needs for artificial intelligence. The project, dubbed "Galactic Brain", offers a bypass to the current five-to-eight year time horizon for data centers to be built on Earth.

Access to energy is one of the primary bottlenecks for scaling artificial intelligence. This problem is driven by infrastructure timelines: securing real estate, establishing utility connections and constructing new data centers can take more than half a decade.

"The race for artificial general intelligence is fundamentally a race for compute capacity, and by extension, energy. The elephant in the room is that our current energy plans simply won't get us there fast enough," said Baiju Bhatt, founder and CEO of Aetherflux, and co-founder of Robinhood. "Galactic Brain puts the sunlight next to the silicon and skips the power grid entirely."

Aetherflux's first data center node for commercial use is targeted for Q1 2027; subsequent satellite launches will build a constellation of nodes to scale capacity.

[Vultur]

"The race for artificial general intelligence is fundamentally a race for compute capacity, and by extension, energy. The elephant in the room is that our current energy plans simply won't get us there fast enough," said Baiju Bhatt, founder and CEO of Aetherflux, and co-founder of Robinhood.

I don't see where the idea that current AI tech scaled up can give AGI comes from. They're fundamentally different kinds of things.

(Unless by AGI they mean something much more limited than true "strong AI").

[meekGee]

Everyone seams to be jumping on the bandwagon. 

This is certainly true today ... The question is cost. The real cost of LLM queries (energy, hardware depreciation, etc) is currently not really being charged to users - otherwise you couldn't get Google AI Overviews, free ChatGPT use, Gemini on Android phones, etc. I think the key question is how many uses would/will survive if/when real costs have to be charged to end users.

(Or whether dramatically more energy efficient [non-LLM?] AI technologies will be developed, which might remove the entire issue but also remove the desirability of putting AI in orbit.]

But we don't pay for the cost of today's "regular" data centers either, and they're not cheap.

The cost is borne by the entire business model of Google.

So once everyone is used to making conversational queries and getting complex results, there's no going back, and ads will get more expensive then they currently are, which is no big deal since he price point is arbitrary anyway.

Nothing has ever been free, and this won't be either, it'll just be part of the ecosystem.

[Vultur]

Everyone seams to be jumping on the bandwagon. 

This is certainly true today ... The question is cost. The real cost of LLM queries (energy, hardware depreciation, etc) is currently not really being charged to users - otherwise you couldn't get Google AI Overviews, free ChatGPT use, Gemini on Android phones, etc. I think the key question is how many uses would/will survive if/when real costs have to be charged to end users.

(Or whether dramatically more energy efficient [non-LLM?] AI technologies will be developed, which might remove the entire issue but also remove the desirability of putting AI in orbit.]

But we don't pay for the cost of today's "regular" data centers either, and they're not cheap.

The cost is borne by the entire business model of Google.

So once everyone is used to making conversational queries and getting complex results, there's no going back

But it's not paid for by Google's business model today, it's paid for by massive unsustainable infusions of investment capital in hopes for returns which are very unlikely to actually materialize.

(And Google results today are notably worse than a few years ago.)

It doesn't ultimately matter whether advertisers or end users are paying for it, if neither can afford to pay for it. It doesn't (IMO) make any sense for the majority of resources in an economy, or at least a large fraction, to be used for glorified search. Especially if that glorified search is questionably any better than a mid 2010s search that used vastly less resources.

But eh. I'm not sure this line of argument is worth pursuing much further. In a few years the answer should be clear.

[zubenelgenubi]

Moderator:
Made slight change to thread title--some members don't proofread their posts and the coding error for quotation marks slip in.

[meekGee]

Everyone seams to be jumping on the bandwagon. 

This is certainly true today ... The question is cost. The real cost of LLM queries (energy, hardware depreciation, etc) is currently not really being charged to users - otherwise you couldn't get Google AI Overviews, free ChatGPT use, Gemini on Android phones, etc. I think the key question is how many uses would/will survive if/when real costs have to be charged to end users.

(Or whether dramatically more energy efficient [non-LLM?] AI technologies will be developed, which might remove the entire issue but also remove the desirability of putting AI in orbit.]

But we don't pay for the cost of today's "regular" data centers either, and they're not cheap.

The cost is borne by the entire business model of Google.

So once everyone is used to making conversational queries and getting complex results, there's no going back

But it's not paid for by Google's business model today, it's paid for by massive unsustainable infusions of investment capital in hopes for returns which are very unlikely to actually materialize.

(And Google results today are notably worse than a few years ago.)

It doesn't ultimately matter whether advertisers or end users are paying for it, if neither can afford to pay for it. It doesn't (IMO) make any sense for the majority of resources in an economy, or at least a large fraction, to be used for glorified search. Especially if that glorified search is questionably any better than a mid 2010s search that used vastly less resources.

But eh. I'm not sure this line of argument is worth pursuing much further. In a few years the answer should be clear.

Initial regular data centers were investment-funded too, but now they are priced in.

Why won't AI follow the same model?

Eventually the consumers pay for it, just like they paid for Radio and TV towers...

[Tywin]

Will this data centers move with SEP, electric propulsion?, if is yes, then I hope we see a revolution in the future in this tech.

[edzieba]

Everyone seams to be jumping on the bandwagon. 

This is certainly true today ... The question is cost. The real cost of LLM queries (energy, hardware depreciation, etc) is currently not really being charged to users - otherwise you couldn't get Google AI Overviews, free ChatGPT use, Gemini on Android phones, etc. I think the key question is how many uses would/will survive if/when real costs have to be charged to end users.

(Or whether dramatically more energy efficient [non-LLM?] AI technologies will be developed, which might remove the entire issue but also remove the desirability of putting AI in orbit.]

But we don't pay for the cost of today's "regular" data centers either, and they're not cheap.

The cost is borne by the entire business model of Google.

So once everyone is used to making conversational queries and getting complex results, there's no going back

But it's not paid for by Google's business model today, it's paid for by massive unsustainable infusions of investment capital in hopes for returns which are very unlikely to actually materialize.

(And Google results today are notably worse than a few years ago.)

It doesn't ultimately matter whether advertisers or end users are paying for it, if neither can afford to pay for it. It doesn't (IMO) make any sense for the majority of resources in an economy, or at least a large fraction, to be used for glorified search. Especially if that glorified search is questionably any better than a mid 2010s search that used vastly less resources.

But eh. I'm not sure this line of argument is worth pursuing much further. In a few years the answer should be clear.

Initial regular data centers were investment-funded too, but now they are priced in.

Where did this idea that datacentres were built speculatively come from? I can't think of any time in the past at which that has been true. They're built to meet demand, and basically from day 1 the revenue from them is far in excess of their operating costs. This is not the case for any 'AI' company thus far: operating costs exceed revenue, let alone if you then have to amortise the spending on the silicon packed into them (for perspective, just to meet the $1.6tn invested in 'AI' thus far, every single US citizen would ned to subscribe to some sort of AI service to the tune of ~$40 per month for the next decade).


LLMs are in the 'underpants gnome' phase: everyone assumes there will be some profit at some point, but not a single company has any idea what they revenue source will be, let alone has actually achieve it. Nvidia, TSMC et al are making out like gangbusters selling shovels, but everyone is digging in advance of the first gold nugget found under the assumption that if they just dig hard enough, there must be gold, because otherwise why would everyone be digging so hard, and if we don't dig ever faster than someone else will find the gold first?

[Vultur]

This is not the case for any 'AI' company thus far: operating costs exceed revenue, let alone if you then have to amortise the spending on the silicon packed into them (for perspective, just to meet the $1.6tn invested in 'AI' thus far, every single US citizen would ned to subscribe to some sort of AI service to the tune of ~$40 per month for the next decade).


LLMs are in the 'underpants gnome' phase: everyone assumes there will be some profit at some point, but not a single company has any idea what they revenue source will be, let alone has actually achieve it.

Yeah, exactly. There is no clear path to any revenue sources of the appropriate scale.

[Twark_Main]

Interesting tweet:

inference Starlink (Star Thought?) satellites

"StarThink" was right. There.     

[sanman]

Interesting tweet:

inference Starlink (Star Thought?) satellites

"StarThink" was right. There.     

Why not just go with SkyNet, for brand-name recognition and pop-culture's sake?

If you're also going to build a large army of humanoid robots, you want them marching under the right name/banner.

[Vultur]

They do have a website
https://www.orbitai.global/

It claims a "Genesis" mission November 29 (with a countdown clock). Farther down it says "Launch partner: Galactic Energy". And "First satellite with onboard blockchain wallet performing the world's first in-orbit blockchain transaction signature."

This feels like a pile of techy buzzwords.

The DeStarAI section also has some bizarre claims like "vacuum provides natural cooling" and "Compute Capacity: ∞ TFLOPS".

The countdown clock is now to December 10.

Nextspaceflight.com doesn't show a Galactic Energy launch on December 10... Though there is a CAS Space Kinetica 1 with unknown payload. *Shrug*

Power Bank (a solar power company that seems to be involved somehow) has a press release claiming l this was launched?

https://powerbankcorp.com/powerbank-announces-launch-of-the-first-satellite-in-the-orbital-cloud-project-with-smartlink-ai/

It has a lot of the same buzzwords but doesn't say what rocket it was launched on. On Nextspaceflight.com, the only launches listed for December 10 (UTC) are a Starlink launch and Kinetica-1. So it pretty much has to be Kinetica-1,  but the thread on that launch https://forum.nasaspaceflight.com/index.php?topic=63701.20 has descriptions of the satellites, none of which seem to be this.

So I'm stumped. Did this really launch? If so, on what rocket?

Something is weird here

[Robotbeat]

This is not the case for any 'AI' company thus far: operating costs exceed revenue, let alone if you then have to amortise the spending on the silicon packed into them (for perspective, just to meet the $1.6tn invested in 'AI' thus far, every single US citizen would ned to subscribe to some sort of AI service to the tune of ~$40 per month for the next decade).


LLMs are in the 'underpants gnome' phase: everyone assumes there will be some profit at some point, but not a single company has any idea what they revenue source will be, let alone has actually achieve it.

Yeah, exactly. There is no clear path to any revenue sources of the appropriate scale.

I mean, people said the same thing about the Internet. And now pretty much everyone does everything online. "Oh, sure, we'll spend $280 per month on broadband, cable tv, netflix, Disneyplus, mobile data plans, etc (source: https://www.reviews.org/internet-service/cost-of-internet-streaming-and-cell-phone-bills/) , but $40 on self-driving cars, delivery robots, house cleaning, yard service, etc, is too much." Plus companies can benefit to the tune of potentially thousands of dollars per month in increased productivity per employee from doing the same, eventually.

The global e-commerce market is currently around $7 trillion in revenue per year. Americans alone drive their automobiles 3 trillion miles per year, and Americans are just 1/20th of the world population.

Globally, the Internet is worth, I dunno, maybe $30 trillion or so?

[Vultur]

This is not the case for any 'AI' company thus far: operating costs exceed revenue, let alone if you then have to amortise the spending on the silicon packed into them (for perspective, just to meet the $1.6tn invested in 'AI' thus far, every single US citizen would ned to subscribe to some sort of AI service to the tune of ~$40 per month for the next decade).


LLMs are in the 'underpants gnome' phase: everyone assumes there will be some profit at some point, but not a single company has any idea what they revenue source will be, let alone has actually achieve it.

Yeah, exactly. There is no clear path to any revenue sources of the appropriate scale.

I mean, people said the same thing about the Internet. And now pretty much everyone does everything online. "Oh, sure, we'll spend $280 per month on broadband, cable tv, netflix, Disneyplus, mobile data plans, etc (source: https://www.reviews.org/internet-service/cost-of-internet-streaming-and-cell-phone-bills/) , but $40 on self-driving cars, delivery robots, house cleaning, yard service, etc, is too much." Plus companies can benefit to the tune of potentially thousands of dollars per month in increased productivity per employee from doing the same, eventually.

The global e-commerce market is currently around $7 trillion in revenue per year. Americans alone drive their automobiles 3 trillion miles per year, and Americans are just 1/20th of the world population.

Globally, the Internet is worth, I dunno, maybe $30 trillion or so?

I do see this point, but I think there are major counterpoints:

- In the short term, *there was still a very major internet bubble*. So comparing AI to the early Internet isn't an argument that it's not a bubble; it's only an argument that it's a bubble like the internet bubble or the railroads bubble bursting in 1873 (where the technology *ultimately* produced great value) rather than like the tulip bubble (where no value was produced).

- The kind of AI that most of the investment is going into right now (LLMs) is ... Not a good fit for those kinds of physical tasks. "AI" is a very broad term, including a lot of different technologies. There is an entirely possible scenario where "AI" in some form is massively successful but all or nearly all the current efforts and investment is irrelevant to the technology that actually becomes profitable.

- this isn't the 1990s any more, and the overall economic situation is quite different