Orbital data centers, part 1: There’s no way this is economically viable, right?

"This is not physically impossible; it’s only a question of whether this is a rational thing."

See full article...

 

[Asbestos Muffins]

i think we passed investments being rational a long time ago. oil and gas investments are barely rational in a world thats baking under their heat, AI is hemorrhaging money yet its utterly imperitive we spend on it, and financers are unmoored fron reality because they have been derisked to the point of being delusional

 

[monkeycid]

[T]he bare-bones cost of deploying 1 million satellites is more than a trillion dollars. SpaceX’s two biggest previous projects to date, the hyper-ambitious Starlink and Starship programs, each required on the order of $10 billion up front. So in terms of scope and cost, orbital data centers are two orders of magnitude larger.

Maybe time to pack it up and go home, then.

 

[AlbatrossMoss]

Yes, we can clearly conceive a data center in space. We can discuss the challenges in a rational manner. It can be done. Yes, it will be a tough compromise, but the advantages are worth it in the long run. And if you think of all the thi... huh? What? You mean it's the grifters trying to sell engineering-fiction? Again? Wow. Investors are very naive if they dump money on this. Especially coming from (more-or-less) the same grifters as the AI ones.

Wait. Investors are going to invest in this?? But of course they will, it's... money changing hands, in the form of a pyramid. They'll be extremely happy. What's not to like?

 

[kesawi]

I can't help but think of this previous Ars article.

https://meincmagazine.com/science/201...-do-you-send-more-computing-power-into-space/

How do orbital data centres address the impacts of radiation on processors? If you need to slow down your processors to mitigate radiation impacts, then that would have to change the economics.

The heat dissipation issues discussed in that previous article also don't really appear to be addressed.

I hope the next two parts of this article will go someway to answering these issues as I'm keen to learn how tackling them has progressed since 2019.

 

[truejim]

I can't help but think of this previous Ars article.

https://meincmagazine.com/science/201...-do-you-send-more-computing-power-into-space/

How do orbital data centres address the impacts of radiation on processors? If you need to slow down your processors to mitigate radiation impacts…

And when a generation of processors invariably grows obsolete or outright fails, do we deorbit all those satellites and launch replacement satellites? So like, it’s actually a million satellites being launched every several years? And we’re simultaneously discarding bus, comms, power and cooling along with each piece of deorbited silicon?

 

[Tactical Finesse]

Basically...the current population of speculators have a built in exploit. All you have to do is make some Star Trek style technobabble--and they hand over infinite money.

 

[NoSkill]

This seems neither economically nor ecologically feasable. Meanwhile, the consideration to allow astronomers to use ground-based telescopes will take a back seat to this scheme, and we already know that NASA won't be given a budget to put new telescopes in space.

 

[Jimmy the Geek]

I'm looking at those AWS data centers, and their massive roof space, and I'm wondering why the hell we don't require them to install solar. For that matter, cover the parking lots and put solar there too.

 

[schmod]

Physicist Matt Buckley has an excellent writeup of the idea on his blog.

If you do some optimistic handwaving, the launch costs aren't quite as prohibitive as I'd initially have guessed.

That is, until you start thinking about how you're going to handle the very difficult problem of keeping your orbital datacenter cool. Once you start considering even basic cooling, shielding, and maintenance requirements, the entire thing spirals out of control.

...and to what end? There's nothing an orbital datacenter can do that you couldn't do better and vastly more efficiently with a solar-powered data center here on Earth.

 

[LG11]

This is a balanced article, much appreciated. Perhaps missing somewhat from the economic and financial perspective: who benefits? While data centers (terrestrial or elsewhere) have a huge impact on the physical and social environment (directly and indirectly), the profits only end in in a very small number of pockets. Sure, some folks will find employment in that sector, but many more will mostly experience the negative externalities. Even space-based data centers have many negative effects that will be felt by more people than the possible financial returns.

 

[Tactical Finesse]

Physicist Matt Buckley has an excellent writeup of the idea on his blog.

If you do some optimistic handwaving, the launch costs aren't quite as prohibitive as I'd initially have guessed.

That is, until you start thinking about how you're going to handle the very difficult problem of keeping your orbital datacenter cool. Once you start considering even basic cooling, shielding, and maintenance requirements, the entire thing spirals out of control.

...and to what end? There's nothing an orbital datacenter can do that you couldn't do better and vastly more efficiently with a solar-powered data center here on Earth.

Cooling or powering.

Because a single GB300 rack needs 130kW. And data centers can have many hundreds or thousands of racks (depending on what you choose to use as your standard of a data center, since there isn't one). Sure one could not put an entire rack into a single satellite--but then you needs even more satellites. Also the depreciation schedule on computer hardware, because cap depreciation is a thing for tax purposes, in the corporate world is 100% in 4-5 years. Meaning all that money sunk into putting those satellites is literally worthless in 4-5 years. As bad as a sports car is for depreciation--it has nothing on computer hardware in the corporate realm.

The depreciation schedule alone makes the entire thing beyond stupid.

 

[fenris_uy]

His model estimates the cost of a Starlink V2 satellite, which has a mass of 1,250 kg, at about $22 per watt generated—which is highly efficient compared to, say, a NASA flagship mission that can cost hundreds of thousands of dollars per watt.

What a weird comparison. If you want to compare commercial sats, you should compare Starlink to Iridium or the like.

NASA sats costs aren't driven by their power needs, but by trying to get the most science out of a given mass budget. Webb didn't cost billions because they wanted to generate MW of power, but because it needed to operate at a really cold temperature and that drove costs up.

 

[Kurenai]

The very concept seems preposterous mainly due to the cooling problem which I assume will be covered by a future article on topic. The only counterpoint which gives me pause that I haven't seen in articles before is that AFAICT, running your datacenter in space radically changes the amortization schedule for the hardware. Once it's up there, there is is basically zero ongoing costs since they're not really maintainable anyway. Assuming the hardware itself is reliable enough (a huuuuge assumption given the environment it will be operating in), a server could be selling compute time until the second it dies, perhaps decades in the future. The reason why we don't do that in terrestrial data centers is because the power/maintenance costs favor replacing servers, but that simply doesn't apply because you can't perform maintenance anyway, and you've functionally frontloaded the entire power budget for the lifetime of the server into the launch cost. Today's high power training server could be a relatively low power (in terms of compute) inference server a decade later.

I also wonder if the nature of the environment might allow for changing how we think about the standard operating temperatures of servers. Perhaps the servers could be idling at 60k, and 150k at full power or something, or inversely idling at 400k/peak at 500k. Those temps would be difficult in a terrestrial datacenter, but maybe somehow viable in space, and could shift the performance characteristics in unexpected ways (though I suspect negatively).

Anyway, summing up, I'd be interested in looking at how the following incredibly optimistic factors might shift the seemingly poor economics:

• Long server lifetimes allowing for perpetually selling compute
• A general slowdown in the rate of compute density increases
• Different operating temperatures unlocking some non-obvious advantage over a terrestrial environment

 

[harlequin69]

Odd that the billionaires seem to be much more bullish about using solar power in space than they are on earth. Last time I checked there was plenty of unused acreage that could be used for solar farms. Pair that with cheap low efficiency batteries and you have a far more cost efficient data center. As a bonus you can upgrade the hardware when something better comes along instead of dumping it all in the ocean.

 

[JoHBE]

My modest proposal is to simply stop talking about this nonsense, instead of feeding the Monster.

 

[Phone Sterilizer]

Just the latency between satellites will make processer to processer communication stupidly slow, even given massive parallelism. Can’t figure out what their game is here.

 

[solomonrex]

I'm glad this will be addressed in a semi-public forum in a semi serious fashion.

There are two major factors missing: the AI boom will probably bust before new chip fabs come online. The whole economy is wonky of course. But the AI bubble is more predictable. They're already laying off people because of AI or in anticipation of AI, junior software engineers are in a mini recession already.

Chip fabs take a long time and lots of money, and the current AI LLM model looks unsustainable on both ends (AI is fed by free forums paid for with search advertising and frequented by white collar workers, both of which are going to be undercut by AI). And I wouldn't bet on this company, of all businesses, getting it done promptly. 10 years would be quick for building a fab and learning to package whole data centers into space, even if the world economy were perfectly stable.

Second, and surprisingly absent, is the lag time to get cutting edge hardware into space. By the time a new chip is fabbed and then launched, integrated and (hopefully) doesn't fail up there, will it still be relevant for inference workloads? Vs terrestrial NVIDIA?

Now, I would add a third consideration, which I don't expect this author to address. Elon is still very active in politics (distracted) and his reactionary politics might be just as self-defeating for AI as it is for EVs. Trump might promote AI but he also set it up as the fall guy for energy prices.

It's hard to look at Tesla's history and assume that this is anything other than a big plan that came out precisely for the upcoming IPO.

 

[Mentil]

If it's an economical, rational thing depends on the economics and rationality of investors. Looking at the stock market today (TSLA says 'hi!') I'm gonna say definitely a maybe.

 

[Eldorito]

Depending on who you ask, the environmental costs estimates of earth-based data centers vary. In terms of water use alone, we see estimates of 560 billion liters annually, and other estimates are much higher. This is especially problematic for arid regions, such as Tucson, Arizona, which successfully pushed back on a large Amazon data center project for this very reason. Ground-based data centers also produce a lot of greenhouse gases from energy consumption.

By contrast, once operational, data centers in space have zero impact on emissions and use no water for cooling. Andrew Dessler, a professor of climate science at Texas A&M University who also writes at The Climate Brink, said there are clear climate benefits from moving this energy generation into space. He considered the potential benefits from a SpaceX constellation generating 100 GW of energy in orbit. The equivalent amount of power from natural gas on Earth would generate around 2 gigatons of carbon dioxide over five years. The Starship launches to put such a constellation into space might produce the equivalent of 100 megatons of carbon dioxide into the atmosphere.

This feels misleading. Hey, let's compare the environmental cost of solar in space to natural gas on earth! Let's look at the excessive water usage on earth compared to no water usage in space! That's not a technology or cost problem, that's a regulatory one. One that is easily fixable, and suddenly Grok in Space makes zero sense.

The reason natural gas and high water usage are a problem is that they're an incredibly cheap, convenient and legal thing to do. Applying technologies that fix those problems on earth will always be cheaper than doing the same thing in space. As the article mentions there's plenty of scope for solar on the planet, there's also that solar panels rated and deployed for space definitely aren't cheaper and likely never will be. We can build huge amounts of batteries (if only Musk owned a company that built batteries he could use...) that give 24/7 power.

The water usage we do now is not because it's efficient or even a good idea, but because it's cheap. Rather than extract the heat from the water it's easier to just pump in more, cooler water. Something that is not an option in space. The real problem is that water has been discounted out the wazoo to corporations and farms, rather than making them pay fair value for it. Increase those costs and efficiency will suddenly shoot up without a whole lot of cost attached. Then suddenly a whole lot of industries will need to face a stark reality.

I assume it's leading up to part 3 being "this makes no financial sense, and Musk is an idiot in every way possible". Because nothing about putting it in space is cost efficient, there's all the downsides of not being able to send a human in to fix a problem, all the radiation hardening, etc. Starlink might make sense because it replaces millions of km of fibre optics, but data centres? Makes no sense at all.

 

[Wandering Monk]

I was just complaining yesterday that Ars hadn’t done an in-depth look at this. So I’m glad that it’s being addressed (Eric was probably mostly finished with this when I was complaining). Hopefully this series becomes something that gets referenced a lot, even beyond the normal Ars community.

 

[VoterFrog]

One thing to keep in mind when discussing the economics of these things: the price that consumers of data center capacity are willing to pay is not linear. They pay a premium for a full, working node of GPUs but if even one of those is broken the going rate drops like a rock.

So you might think you can just calculate the return on your fleet by just assuming that if X% of the GPUs are still working, you still make X% revenue. The reality will be that about a third of your nodes will only command like half the rate of the others in the first year.

About a hundred things have to hit the most optimistic protections (and not all of those will be in Elon's control) in order to bring the costs down that you can pay off these satellites before all their GPUs are dead.

 

[dennis.kext]

The heat dissipation issues discussed in that previous article also don't really appear to be addressed.

Scott Manley just did a really great and in-depth vid on cooling servers in space. It is doable.


View: https://www.youtube.com/watch?v=FlQYU3m1e80

 

[Pieter2]

Physicist Matt Buckley has an excellent writeup of the idea on his blog.

If you do some optimistic handwaving, the launch costs aren't quite as prohibitive as I'd initially have guessed.

That is, until you start thinking about how you're going to handle the very difficult problem of keeping your orbital datacenter cool. Once you start considering even basic cooling, shielding, and maintenance requirements, the entire thing spirals out of control.

...and to what end? There's nothing an orbital datacenter can do that you couldn't do better and vastly more efficiently with a solar-powered data center here on Earth.

The thing that jumps out to me from that article:
"the useful lifetime for a GPU in datacenter on Earth is a year or so. So we’re spending 60 billion dollars to put this datacenter in orbit, and if it works, it only works for a year"
So why are we still talking about this?
Or are there AI capable GPU's with a much larger lifespan?

 

[jason8957]

I think an alternate history sci-fi book based on if all Musk's hype had turned out to be true would be interesting.

 

[Tactical Finesse]

The thing that jumps out to me from that article:
"the useful lifetime for a GPU in datacenter on Earth is a year or so. So we’re spending 60 billion dollars to put this datacenter in orbit, and if it works, it only works for a year"
So why are we still talking about this?
Or are there AI capable GPU's with a much larger lifespan?

Because people selling rocket launch services (Elon) and people selling AI pickaxes (Jenson Huang) during a goldrush...need someone to keep paying for their services that statistically no consumer is really wanting to pay for.

 

[daddyboomalati]

I foresee a day when people will look to the night sky and say, "is that a star??". As a backyard astrophotographer (and human being), I am so tired of this dystopian, straight from Hollywood script, nightmare that we are living.

 

[thelongdarkteatimeof]

Microsoft has already proved that putting a data centre in an unlikely place, the bottom of the ocean, is neither economically viable, nor serviceable.

This is literally the stupidest idea ever given a second thought.

 

[SuperOuss]

Space is a vacuum. Heat leaves servers through convection and conduction — neither of which works in a vacuum. The ISS needs massive radiator panels just to manage a few kilowatts. A real data center would need radiator surface areas that are physically absurd.

If someone is seriously pitching this, they're either skipping the physics or hoping you will. Fools and money parting is apparently still a reality even in supposedly knowledgeable circles.

The xAI/Twitter valuation situation is its own chapter — using one inflated asset to prop up another while calling it a merger is hilariously bold, I'll give it that.

I hope the market reads it the same way because I have prepared a steady supply of pop corn!

edit:typos.

 

[mathguru]

Why not build them on the moon? Or how about on an aircraft that’s constantly in flight and refueled aerially? Oh I know, how about under water in the ocean? The mind boggles at the investment opportunities!!!

 

[remyporter]

I can't help but think of this previous Ars article.

https://meincmagazine.com/science/201...-do-you-send-more-computing-power-into-space/

How do orbital data centres address the impacts of radiation on processors? If you need to slow down your processors to mitigate radiation impacts, then that would have to change the economics.

The heat dissipation issues discussed in that previous article also don't really appear to be addressed.

I hope the next two parts of this article will go someway to answering these issues as I'm keen to learn how tackling them has progressed since 2019.

In a terrestrial orbit, radiation isn't that big a deal. Like, it's a concern, but one that can generally be addressed with some basic error correction. The Earth's magnetic field protects against most charged radiation.

 

[nthdegree1]

Odd that the billionaires seem to be much more bullish about using solar power in space than they are on earth. Last time I checked there was plenty of unused acreage that could be used for solar farms. Pair that with cheap low efficiency batteries and you have a far more cost efficient data center. As a bonus you can upgrade the hardware when something better comes along instead of dumping it all in the ocean.

That's why part 4 of this series will explore beautiful, clean coal in space

 

[numerobis]

data centers are notoriously avaricious consumers of electricity

Voracious? Avaricious means obsessively stingy, which is kind of the opposite.

 

[Elektronika719]

Will there eventually be edge compute nodes, including those that perform inference, on orbit in the next decade? Almost a certainty.

Is a single H100 on orbit even the beginnings of a data center? Not even close - it’s nothing more than a stunt to dupe people out of their money. Not one technical advancement was made by doing this. Pure MLM-style nonsense to “build the business,” I mean pyramid.

Will we have general purpose orbital data center scale infrastructure on orbit within the next few decades? Almost a guaranteed no, unless it is forced by a limited set of users (technocrat wannabes) attempting to bypass earthbound legal restrictions for some nefarious purposes:

There is otherwise no remotely economical benefit to launching, operating, maintaining and continuously upgrading large scale data centers in space. Those claiming otherwise are the equivalents of Elizabeth Holmes and SBF.
All of the tech bros babbling on about this nonsense quickly shutdown discussion on the realities of the thermal, radiation and debris environment of space. As if all of the engineers who have been working around these problems for multiple decades are idiots because they’re not tech bros.

 

[Diabolical]

I'll be very curious to see what the Ars folks make of the cooling conundrum and the space weather questions.

Radiation cooling is.... relatively crap.
Single event upsets, geomagnetic storming... hell, stray micrometeoroid impacts.

Very curious indeed.

 

[DrewW]

I can't help but think of this previous Ars article.

https://meincmagazine.com/science/201...-do-you-send-more-computing-power-into-space/

How do orbital data centres address the impacts of radiation on processors? If you need to slow down your processors to mitigate radiation impacts, then that would have to change the economics.

If you think AI hallucinates now, just wait until they start getting bombarded with high energy particles. Shielding could also increase weight and potentially complicate cooling.

Instead, I’m sure Elon will dictate they call the space based constellation Hulk.

Edit from “Shielding would” to “Shielding could” because I’m sure there are options

 

[Promathion]

Question 1: You need to solve the heat dissipation problem in space without resorting to evaporative cooling. If you have the solution to that, could you port it back to earth and then not need the millions of gallons of water consumption for your terrestrial datacenters?

Question 2: Regarding the accumulation of lithium, copper, aluminum in the upper atmosphere, do we even know which WAY this affects climate change? Do they act as a greenhouse layer even way up there, or is their effect to block sunlight from entering the atmosphere? If we don't know, that sounds like an important thing to learn before proceeding with this plan (or really any plan involving 1000s of space launches).

 

[jonah]

Disappointed to not see any real discussion of cooling in the article. How, exactly, do they plan to cool these things? A single modern GPU emits as much heat as entire spacecraft do. They want to what, pack X number of such things into a satellite? How on earth would it keep them cool? Radiative cooling requires enormous amounts of mass, so you're getting bigger satellites (or a bunch of small shitty ones). If each satellite can only cool one or two GPUs, then why bother?

The answer is that the economics (and physics) will never work, but Elon is betting the market can stay irrational longer (since it already has with everything related to him) than it takes for those economic and physics concerns to bite.